Interpretable gene-program analysis offers a principled way to extract disease-relevant biology from single-cell RNA sequencing (scRNA-seq) while preserving cell-type specificity and reducing the complexity of gene-level results. Here, we present a methodological framework to quantify and compare cell-type-specific gene-program activation across conditions, and evaluate it in idiopathic pulmonary fibrosis (IPF), a progressive fibrotic lung disease with limited therapeutic options.

Our approach combines unsupervised and supervised matrix factorization-based methods to infer gene-program activity from scRNA-seq data, utilizing the Cytopus knowledge graph to define biologically grounded programs. Differential program activation is then tested with statistical models that account for patient-level dependence and cell-cycle effects.

To assess robustness and interpretability, we benchmark this workflow against conventional disease-analysis workflows, including single-cell and pseudo-bulk differential expression followed by over-representation and gene set enrichment analyses, by finely tuning a comparative evaluation of diverse computational methodologies.

Applied to scRNA-seq data from IPF and healthy lungs, the framework enabled a systematic evaluation of concordance across conventional and matrix factorization-based methods, leading to the identification of highly robust consensus disease-associated programs. Importantly, each method also detected biologically relevant dysregulation independently highlighting the importance of a comparative, multi-methodological approach.

Beyond prioritizing dysregulated processes, we reconstructed regulatory and interaction networks for relevant programs to identify core interactors and shared regulators with therapeutic potential for drug targeting and repurposing.

This study evaluates interpretable gene program dysregulation analysis in IPF, showing that it can serve as a valuable complement to gene-level workflows and a biologically relevant framework for target identification and drug repurposing.

Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful tool for analyzing complex mixtures due to its ability to manage matrix complexity, provide detailed molecular insights, and preserve sample integrity. Various NMR experiments, such as NOESY, CPMG, diffusion-edited, and J-resolved spectroscopy (JRES), offer complementary insights into biofluids like serum and plasma. For instance, CPMG selectively detects small molecules, diffusion-edited emphasizes macromolecules, NOESY captures both, and JRES aids signal assignment. However, acquiring multiple spectra can be resource-intensive and time-consuming, especially for high-throughput studies.
Here, we present a simple strategy to computationally derive CPMG, diffusion-edited, and projected JRES (pJRES) spectra from a single NOESY acquisition using Partial Least Squares (PLS) regression. Serum samples were used as a case study. We analyzed 1H-NMR spectra from 1842 individuals across 18 recruitment centers, acquired on a Bruker 600 MHz spectrometer. Data from 17 centers were split into training (80%) and validation (20%) sets, while 232 samples from an independent center served as the test set.
PLS models used NOESY spectra as predictors and the other experiments as targets. Experimental and predicted spectra were compared in regions with signal intensities at least three times above noise. Performance metrics included median relative error (MRE%), root mean square error (RMSE), coefficient of determination (R²), and ratio of performance to deviation (RPD). In the independent test set, MRE% values were 6%, 4%, and 13% for CPMG, diffusion-edited, and pJRES spectra, respectively[1]. A future direction is applying this approach to low-field NMR instruments, such as 100 MHz benchtop spectrometers.

Systems biology approaches increasingly rely on the integration of metabolomics data across multiple biological matrices to capture systemic metabolic variation. However, a major limitation remains the lack of standardized workflows enabling comparable metabolomic profiling across different human biofluids.
High-field Nuclear Magnetic Resonance (NMR) spectroscopy is a cornerstone of metabolomics, but its high cost and the need for specialized infrastructure limit its translation into clinical practice. Recently, benchtop low-field (LF) NMR spectrometers have emerged as a more accessible alternative, that could break down the cost barrier potentially enabling metabolomic analyses in virtually any setting.
In this study, we developed and validated a standardized benchtop NMR workflow for metabolomic profiling of multiple human biofluids, including serum and urine, to enable harmonized metabolic measurements across matrices.
The workflow includes sample preparation protocols, standardized benchtop NMR acquisition parameters, and automated spectral processing. Protocol robustness was validated through cross-platform comparison between low-field (100 MHz) and high-field (600 MHz) NMR spectroscopy.
As a proof-of-concept, the workflow was applied to a pilot cohort including serum and urine samples collected from patients before and after treatment. Multivariate analyses revealed treatment-associated metabolic shifts, demonstrating the potential of the platform for longitudinal metabolic profiling.

Accurate molecular representations are essential for AI-driven drug discovery. However, standard Graph Neural Network (GNN) pretraining predominantly relies on structural self-supervision. This often results in latent spaces that are insensitive to biological context, failing to capture activity cliffs or generalise to novel protein targets in drug-target interaction (DTI) tasks. To address this issue, we present a biologically supervised contrastive pretraining framework. Using the extensive chemogenomic LCIdb dataset (containing ~271,000 molecules and over 2,000 human proteins), we systematically trained topological (MolCLR), hierarchical (HiMol) and geometric (MolGDL) GNNs. Our modified supervised contrastive objective implicitly maps multi-target pharmacological relationships into the molecular embedding space by clustering ligands with shared bioactivity profiles. Comprehensive benchmarking reveals that this pre-training successfully enriches structural embeddings with actionable pharmacological signals. While this contrastive bottleneck acts as a restrictive filter for certain task-agnostic physicochemical properties, it decisively improves out-of-distribution generalisation in downstream DTI predictions. When evaluated on DrugBank, BindingDB and BIOSNAP under stringent structural splits, our pretrained GNNs achieved significant performance improvements in 'cold-drug', 'cold-target' and the highly challenging 'cold-both' scenarios. Furthermore, feature importance analysis of downstream nonlinear classifiers revealed a systematic increase in reliance on the molecular signature, confirming the embedding's enriched biological utility. Aligning the latent space with pharmacological outcomes ultimately provides a highly robust, target-aware foundation for navigating novel chemical entities and unseen proteins.

Human organoids emerged as pivotal models for studying organs and tissues, modeling diseases, or discovering novel therapeutics. This surge in organoid research has been accompanied by the production of large volumes of heterogeneous omics datasets. To enable cross-study integration and complex semantic queries (e.g., identifying all datasets associated with specific perturbagens, disease-related genes, or accessing datasets using specific sequencing techniques across diverse organoid types), there is a critical need for a structured, interconnected knowledge representation, i.e. a Knowledge Graph (KG) of organoid-related omics datasets.

We designed an automated workflow based on Large Language Models (LLM) to extract and standardize metadata from summary files associated with human organoid omics datasets archived in public repositories. We used Retrieval-Augmented Generation (RAG) to supply the LLM with relevant context. To ensure accuracy and standardized annotations, the pipeline integrates real-time access to the EBI Ontology Lookup Service (OLS), enabling retrieval of ontology-compliant values and reducing hallucination risks. Output fidelity is ensured through a triple-redundancy execution model, in which final annotations require majority consensus across independent runs. Any conflicting outputs are reviewed by a “Judge LLM” based on contextual evidence. The pipeline was validated against a gold-standard corpus of 50 manually annotated files, achieving 88% extraction accuracy. This high-precision approach will allow us to consolidate 993 organoid-related omics datasets into a KG.

This resource offers a scalable foundation for the organoid community, supporting not only efficient data reuse but also downstream meta-analyses and biomarker identification across diverse organoid systems and perturbagen conditions.

The circadian clock orchestrates physiological and behavioral processes over a 24-hour cycle
and plays a critical role in human health. Circadian disruption is associated with metabolic,
malignant, and psychological disorders. Developing an in silico representation of this complex
system could help identify key mechanisms underlying circadian desynchronization and support
preventive strategies.
Due to the intrinsic complexity of circadian regulation, involving multi-scale and multi-formalism
interactions, a systemic and hybrid modeling approach is required. In this work, we propose a
multi-agent framework integrating circadian regulation, metabolism, and hormonal dynamics.
Each group of cells (e.g., myocytes, hepatocytes, pancreatic beta cells) is modeled as an
autonomous agent interacting through both a shared environment representing the bloodstream
and explicit message exchanges.
Discrete formalisms based on the René Thomas framework are used to capture circadian
oscillations, while continuous formalisms with dynamic thresholds describe hormonal secretion
and metabolic fluxes. Such hybrid architecture enables the coupling of regulatory networks with
physiological dynamics across scales while avoiding excessive model complexity.
First simulation results of the integrated model reproduce key qualitative features of circadian
physiology, including oscillatory hormonal secretion patterns, synchronization between spatially
separated circadian clocks through indirect interactions, daily glucose fluctuations, and coherent
energy production cycles. These results are consistent with known biological behaviors and
support the relevance of the proposed approach.
This work constitutes a first step toward comprehensive systemic modeling of circadian
regulation and provides a foundation for studying pathological desynchronization, as well as for
the design of hybrid multi-agent models in systems biology.

Pseudomonas putida is an important microbial host in industrial biotechnology; however, its growth and production performance is highly sensitive to culture conditions such as medium composition and pH, which are often inadequately represented by static or empirical models. Here, we present a dynamic, mechanistic bioprocess model based on ordinary differential equations for P. putida cultivation that integrates microbial growth, metabolism, and pH dynamics across multiple medium formulations and experimental set‑ups.

The model explicitly describes the dynamic evolution of pH as a function of microbial metabolic activity, buffering capacity, and relevant acid–base equilibria, capturing the bidirectional coupling between cellular physiology and the extracellular environment. Importantly, the formulation allows the representation of both uncontrolled pH conditions, where pH changes are solely driven by microbial activity, and actively controlled strategies, such as titration with acid or base during bioreactor operation. This level of mechanistic detail enables the model to reproduce transient pH profiles and their impact on growth and substrate utilization under a wide range of cultivation scenarios.

Model parameterization and validation were performed using experimental cultivation data obtained under varying initial pH values, medium compositions, and bioreactor configurations, demonstrating the model's ability to accurately describe system dynamics across distinct operational conditions.

Overall, this work provides a versatile computational framework for P. putida bioprocess modeling that supports process understanding, model‑based optimization, and the rational design of microbial cultivation strategies under industrially relevant conditions.

Perturbation single-cell RNA-seq (scRNA-seq) data from drug-treated cancer-cell lines offer a powerful opportunity to dissect cellular responses to therapy at high resolution. However, these data are typically analyzed using gene-centric or purely data-driven approaches that do not explicitly capture pathway-level dynamics, limiting mechanistic interpretability.

To address this, we developed a computational framework coupling trajectory inference with footprint-based analysis, enabling the extraction of pseudo-time-resolved functional activities across signaling pathways and transcription factors. Applied to >32k single cells from 169 cancer cell lines treated with Trametinib (MEK inhibitor), this approach robustly distinguishes responsive from non-responsive tumor models. Moreover, trajectory-based cell ordering faithfully reconstructs MAPK pathway dynamics from single-time-point perturbation scRNA-seq, recapitulating patterns observed in complex and costly time-series experiments. Clustering of inferred activity trajectories reveals distinct dynamic response archetypes, disentangling heterogeneous adaptation strategies and reconstructing the functional rewiring of intracellular signaling networks following perturbation.

Importantly, these pseudo-time-resolved transcription factor activity profiles provide a basis for mechanistic modeling of tumor cell signaling responses to drug perturbations. Integrating them with curated prior knowledge signaling networks of intracellular interactions, we derive context-specific mechanistic signaling pathway models capturing causal relationships underlying drug-induced dynamics. These models enable the characterization of pathway deregulations in specific cellular contexts and support the prediction of responses to targeted therapies and their combinations. To this end, we extend our analysis to multiple targeted drugs acting on proteins within a curated signaling network regulating tumor-immune interactions, demonstrating how trajectory-informed functional activities can constrain and optimize mechanistic models even without explicit time-series measurements.

Reconstructing gene regulatory networks is a common method in the field of bioinformatics for uncovering the transcriptional regulation driving patterns of disease. In these networks, the relationship between transcription factor and target genes is commonly measured. However, these tools are commonly not user-friendly, requiring a fair amount of prior programming experience to be able to reconstruct the networks and perform the downstream network. Therefore, we have developed a new tool called SiSaNA (Single-Sample Network Analysis) that allows users to reconstruct and analyze single-sample gene regulatory networks all via the command line, eliminating the need for prior programming experience. SiSaNA seamlessly integrates into the NetZoo ecosystem, utilizing the PANDA and LIONESS algorithms from NetZooPy, but providing a more user-friendly experience for reconstructing, comparing and interpreting single-sample regulatory networks, all via the command line. SiSaNA requires little to no programming proficiency other than basic familiarity with the command line, while providing high-quality figures as output.

National data portals focused on infectious disease efforts have been shown to help to minimise the impact of outbreaks on society. They do this primarily by bringing together and promoting existing research resources (e.g. data and tools). This enables those working in research to quickly locate data and tools that they can reuse, rather than committing time and effort into needlessly recreating resources. National portals can also encourage greater collaboration between research groups, as it is typically easiest to identify potential collaborators on such portals.

Although national infectious disease data portals inherently focus on efforts within a country, their impact need not be localised to that country. In fact, such portals are arguably most effective when they are closely linked with international efforts. This could include, for example, helping to raise awareness of, and adherence to, international standards for data management. This is crucial in enabling data from different efforts to be brought together. National portals can also work together as part of a network to promote pandemic preparedness by ensuring the expedient transfer of information.

Here we intend to describe best practices in establishing and maintaining a national data portal focused on infectious disease research. We will draw on our experience with building and operating the Swedish Pathogens Portal over the last 6 years.

Cisplatin is a widely used chemotherapeutic agent for ovarian adenocarcinoma that ultimately leads to DNA damage and apoptosis. We previously demonstrated that dying tumor cells secrete extracellular vesicles (EVs) enriched with spliceosomal components. These EVs stimulate DNA repair and cell cycle regulation in chemo-naive recipient cells, enhancing their resistance to cisplatin. To identify the cellular processes driving cytoprotective secretome formation, we analyzed scRNA-seq from cisplatin-treated SKOV3 cells. Genes associated with cell division, DNA repair, and splicing show a slight increase in expression during the first 7-10 hours, followed by a decline through 24 hours.
To explain the expression trends of repair and splicing genes, we integrated cisplatin-induced DNA damage maps and excision repair sequencing data for immortalized GM12878 cells. Linear model indicated that cell division and stress response genes exhibit higher repair rates. The subsequent decline in expression could stem from either transcript degradation or export. mRNA integrity analysis for SKOV3 RNA-seq data showed that only one-third of the downregulated genes underwent degradation, suggesting that the decline after 7 hours may be primarily driven by the export of accumulated transcripts. SKOV3 EVs transcriptomic and proteomic analysis confirmed the selective secretion of DNA repair and spliceosomal transcripts and proteins.
To summarize, cisplatin initially triggers the preferential repair of survival-critical genes (DNA repair, cell cycle, mRNA splicing). Following critical DNA damage, transcription of these genes is inhibited, and the accumulated components are increasingly exported via EVs to mediate pro-survival signaling.
This research was supported by the Russian Science Foundation (project 25-75-10152).

Motivation: Protein--protein interactions (PPIs) underpin a wide range of biological functions in living organisms. Experimental identification of new PPIs is expensive and time-consuming. The experimental bottleneck has implied an imbalance in terms of data availability: while certain species have been screened exhaustively, other species have not been sufficiently examined. An AI driven protocol for PPI prediction that leverages the massive data accumulated for certain species means a decisive boost for so far understudied species.
Results: We present ANIMA (Artificial Neural Interaction Model for Animals), an AI supported cross-species PPI prediction model trained on popular species to predict PPIs in under-researched species. Our experiments demonstrate that our model, when trained on 200 diversely selected animal species, can successfully predict PPIs in other species: ANIMA achieves 95.3% accuracy on other animal species, 91.1% on other eukaryotes, and 82.5% on non-eukaryotes. For a more fine-grained evaluation of the model, we stratify performance rates by the evolutionary distance of test to training sets. We also stratify results by a novel, alignment based score ("representation score") which allows for fine-grained evaluation in terms of its capacity to generalize to unseen interactions. As expected, results demonstrate increasing performance on increasing evolutionary similarity and on increasing identity of interacting proteins, while still showing excellent performance on proteins entirely lacking counterparts in the training set. In comparison with the state of the art, ANIMA demonstrates substantial superiority in terms of performance rates.
Availability: https://github.com/0nurB/ANIMA

Combinatorial drug therapy holds great promise for tackling complex diseases, but the vast number of possible drug combinations makes exhaustive experimental testing infeasible. Computational models have been developed to guide experimental screens by assigning synergy scores to drug pair–cell line combinations, where they take input structural and chemical information on drugs and molecular features of cell lines. The premise of these models is that they leverage this biological and chemical information to predict synergy measurements.

In this study, we demonstrate that replacing drug and cell line representations with simple one-hot encodings results in comparable or even slightly improved performance across diverse published drug combination models. This unexpected finding suggests that current models use these representations primarily as identifiers and exploit covariation in the synergy labels. Our synthetic data experiments show that models can learn from the true features; however, when drugs and cell lines recur across drug–drug–cell triplets, this repeating structure impairs feature-based learning. While the current synergy prediction models can aid in prioritizing drug pairs within a panel of tested drugs and cell lines, our results highlight the need for better strategies to learn from intended features and to generalize to unseen drugs and cell lines.

Plant-based fermented foods often lack the nuanced, dairy-like flavor and texture consumers expect. We developed a community metabolic model-guided workflow to design microbial communities that enhance desired flavor compounds while minimizing undesired ones in plant-based kefir and cheese fermentations. For 90 bacterial and yeast strains, we reconstructed genome-scale metabolic models using CarveMe (bacteria) and CarveFungi (yeasts) and curated them with strain-specific growth evidence. Simulation environments we formulated to mimic five plant-based substrates based on their chemical characteristics.

We combined subsets of strain-specific models into ~300000 kefir and ~180000 cheese community metabolic models (CMM) across various sizes and substrates. To predict community behavior, we employed a two-stage CMM optimization workflow. First, we determined the maximum theoretical growth yield of each strain within the community in the substrate-specific environment. Second, the individual optima were used to impose relative strain-specific lower bounds on biomass yields to prevent competitive exclusion. The community objective was defined as the sum of individual growth rates, which was maximized to determine community performance. Then, communities were ranked using a composite score reflecting the predicted flavor enrichment and suppression of off-notes.

To refine community selections, we integrated model predictions with enzyme-level evidence by assessing key flavor pathways using Foldseek (structural homology), MMseqs (sequence similarity), and HMMER (Pfam domains). Top-ranked communities were prioritized for experimental implementation.

This workflow linked genome-scale metabolic modelling, substrate-aware media design for model simulations, and enzyme-level evidence for rationally engineering flavor-optimized plant-based fermentations with broad transferability to functional microbiome design in other fields.

Carbapenems are last-resort antibiotics, so it is crucial to prevent the spread of carbapenem-resistant Gram-negative bacteria (CRGNB) in hospitals. In 2011, the state of Hesse, Germany, has established a notification system for reporting detected CRGNB cases. To optimally take advantage of these data, we developed a computational tool that predicts outbreak events among the patients reported in a given time window, ranks them according to statistical significance and provides tailored visualizations for individual institutions. It uses a machine learning framework combining supervised and unsupervised techniques with patient-centric and hospital-centric approaches to systematically check for patient clusters with consistent resistance types and possible transmission events in the patient history of hospital stays. Leveraging patient history by custom algorithms is the key novelty of this work, becoming necessary because a colonization with CRGNB may remain undiscovered for a long period. Therefore, place or time of notifications do not hold information on the potential transmission source, and common epidemiological surveillance mechanisms are not applicable. The tool comprehensively recovered human-annotated clusters with different characteristics, such as single or multiple bacterial species, single or double resistances and specific or unspecific resistance types. The top significant additional predictions are investigated via sequencing. Furthermore, cluster recovery dropped drastically when including only the last hospital stay before the patient's CRGNB check, highlighting the importance of reporting earlier hospital stays. Adoption of automated data analysis and transparent presentation of evidence will increase awareness, leading to more complete data, precise predictions and timely countermeasures.

Rapid viral host identification is critical for public health, yet discovery often relies on high-throughput sequencing reads or partial contigs. While numerous machine learning (ML) prediction tools exist, performance is typically evaluated using random cross-validation splits of training and test data. These methods fail to address real world challenges, such as temporal viral evolution and potential host shifts, where models must predict outcomes for future sequences using only historical data. Furthermore, many existing methods lack transparency regarding database curation and feature engineering, impeding reproducibility and standardization. This study investigates the performance of host prediction models by comparing traditional random partitioning against a temporal split strategy that more accurately reflects realistic scenarios. We evaluated the performances of random versus temporal split models using viral genome sequences from family Coronaviridae. K-mer frequencies were extracted using simulated NGS reads and refined via distance-based dimensionality reduction. We assessed the effects of different normalization schemes and centroid vectors on feature robustness, subsequently, benchmarking multiple supervised ML classifiers to determine the optimal virus-host prediction framework. Preliminary analyses compare predictive accuracy of ML models across two strategies: a random, accession- and host-stratified split and a temporal split. We anticipate that the temporal split will provide a more rigorous assessment of the model generalizability on future unseen data, whereas the random split will identify patterns within known evolutionary clusters. These results will identify the optimal combination of normalization, centroid selection, and model architecture to determine the ideal framework for a hierarchical classification pipeline.

Microbial communities are abundant in nature, drive major biogeochemical cycles, and are integral to our health. Metabolic models promise mechanistic insights into community metabolism, including interactions such as competition and cross-feeding of nutrients.
Although community metabolic models are built from linear models of individual organisms, their solution space is non-linear, rendering standard single-organism methods insufficient.
Here, we present a new geometric perspective on community metabolic modelling. We extend metabolic pathway analysis to microbial communities by defining community-level elementary flux modes (cEFMs). These modes generalize single-organism EFMs and form minimal building blocks of community metabolism. Each cEFM explicitly captures nutrient uptake and metabolite exchange between all community members, which further admits a direct ecological interpretation, distinguishing interaction types such as mutualistic or commensal modes.
Crucially, this cEFM-based perspective also resolves a long-standing problem in metabolic modeling: thermodynamically infeasible cycles (TICs) across organisms. In community models, TICs commonly arise when different organisms catalyse opposing conversions, leading to inflated flux values and spurious cross-feeding predictions. We show that upon blocking external metabolite inflow, enumeration of cEFMs reveals all TICs present in the system. Incorporating simple activity-based constraints from these modes into an MILP yields cycle-free solutions, even at the genome scale.
The presented geometric view provides an important improvement in the analysis of microbial communities by removing technical artefacts in the prediction of interactions and by offering a fine-grained ecological interpretation of community metabolism.

Biomedical research relies on the generation and integration of an enormous diversity of omics data. Research communities are often underutilizing existing data and waste resources on re-analysing datasets, which can be avoided with the proper data infrastructure and standardisation practices. Major hurdles for effectively utilising data across research groups are finding datasets, missing metadata, difficulties comparing them, non-reproducible analysis, and the lack of knowledge about the best practices.
At the UK Dementia Research Institute, we provide a robust software infrastructure bundled into an omics framework that follows FAIR principles. We are developing reproducible and modular Nextflow pipelines for state-of-the-art processing, quality control, and downstream analysis that produce data in community standard formats.
We enable sharing data through a comprehensive data portal, Dementia Data Nexus (DDNexus), which is searchable by standardised metadata and shows the similarity between studies based on the overlap of characteristic gene sets.
The data pipelines and DDNexus seamlessly integrate with our interactive viewers for exploring omics data, Neuromics Explorer and DataMap. Neuromics Explorer allows in-depth investigation of QC metrics and custom differential expression analysis. DataMap exploits an extensive dementia-specific knowledge graph for highly flexible and feature-rich network analysis of gene sets` functional and disease context.
Our ecosystem covers all steps from raw omics data processing to insightful analysis while tracking details at each stage and preserving standardized metadata and data. By continuous benchmarking and development, we integrate the best scalable approaches for the ever-growing amount of omics data. Our tools make re-using, investigating, and comparing datasets simple.

Most regulatory network inference methods estimate population-average models and therefore overlook biologically important patient-level variation. Here, we performed a large-scale pan-cancer analysis of single-sample gene regulatory networks to characterize inter-patient heterogeneity across 33 cancer types from The Cancer Genome Atlas. We integrated transcription factor binding priors, protein-protein interactions, and transcriptomic profiles using PANDA, followed by LIONESS algorithms to reconstruct 9,778 patient-specific regulatory networks. Network-derived gene targeting signatures preserved tissue-specific structure while revealing regulatory variation not captured by expression profiles alone. Using pathway-level analyses to quantify coordinated regulatory heterogeneity, we identified recurrent dysregulation of immune, cell-cycle, and signaling pathways. PD-1 signaling emerged as the most frequently dysregulated pathway, showing significant heterogeneity across 23 cancer types and associations with patient outcomes in 12 cancers. These patterns were largely independent of established molecular subtypes and immune cell composition, suggesting alternative regulatory mechanisms.
Our findings demonstrate the power of single-sample gene regulatory networks to uncover hidden layers of tumor biology, improve patient stratification, and inform future research in immunotherapy and precision oncology.

As bioinformatics tools and databases scale, navigating extensive user documentation across disparate projects becomes a significant hurdle for researchers. To address this challenge, we present DocBot, an EMBL-wide AI assistant designed to provide streamlined, centralized access to documentation across 15 major services in EMBL.

Powered by Large Language Models (LLMs), DocBot employs a Retrieval-Augmented Generation (RAG) architecture. It leverages a vector database populated with embeddings from over 15,000 documents that are continuously updated and reindexed directly from official EMBL documentation sources. When users submit queries, DocBot retrieves the most relevant information to generate highly accurate, context-aware responses based strictly on its curated knowledge base.

DocBot offers several robust features designed to optimize the user experience. Most notably, as a centralized knowledge hub, rather than requiring researchers to identify the correct resource before searching, DocBot automatically routes queries to the most relevant service's documentation. Researchers can also flexibly narrow their search to specific projects if desired. The system retains conversational context within a session, enabling dynamic, multi-turn interactions. Thanks to this conversational awareness, it can gracefully handle complex queries that span multiple team domains, automatically prompting users for clarification when necessary to refine its answers.

Furthermore, DocBot ensures transparency and trust by appending direct source citations to every generated response. This allows researchers to easily verify information and navigate to the original documentation sections for deeper exploration. DocBot hence significantly enhances the researcher experience and serves as a scalable model for centralized, AI-driven assistance in large-scale computational biology institutes.

Identifying key regulators that drive cellular states and state transitions from transcriptomic data has been a longstanding objective in systems biology. However, this remains challenging due to the complexity of biological regulation and the difficulty to systematically evaluate inferred regulators from orthogonal data.

We previously developed interpretable neural networks that infer regulatory proteins, known as knowledge-primed neural networks (KPNNs). Similar to other visible or biology-inspired neural networks, KPNNs integrate prior biological knowledge directly into their architecture. In KPNNs, this enables attribution of model predictions to biologically meaningful entities such as transcription factors. However, KPNN interpretations, and those of similar models, have not been systematically assessed against ground truth or compared with other inference methods.

Here, we present a ground truth-based evaluation to assess the reliability of KPNN inferred regulators across multiple datasets with varying types of ground truth, and to compare KPNNs with the state-of-the-art regulator inference tool decoupleR. We first optimised KPNN implementation using simulated transcriptomic data generated from a known regulatory network with SERGIO. We then validated KPNNs against large experimental perturbation datasets, comparing inferred regulators with the experimental ground truth from genetic and chemical perturbations or pathway stimulations. We find that KPNNs, in most scenarios, prioritise biologically meaningful regulators and outperform baseline and alternative approaches.

Together, our study provides the first systematic ground truth evaluation and benchmarking of KPNNs and establishes practical guidelines for their application in data-driven biological discovery.

While traditional toxicology has historically relied on animal testing, ethical concerns and high costs are driving a paradigm shift toward Next-Generation Risk Assessment (NGRA). This transition requires computational tools capable of integrating diverse biological data to predict hazards in humans independently of in vivo testing. To address this, we have developed the Chemical Effect Predictor (CEP), an end-to-end Graph Neural Network (GNN) architecture designed for organ-level hazard prediction. Our approach leverages a large-scale Biomedical Knowledge Graph (BKG) containing ~700k nodes and ~3.1M edges, integrating drugs, proteins, diseases, and biological processes to capture the mechanistically relevant relationships between chemical compounds and organ-specific disease endpoints. Using this framework, we developed four specialised models targeting liver, heart, kidney, and nervous system injuries. By initialising drug nodes with Morgan fingerprints, we developed a fully inductive architecture that integrates explicit chemical structural data with the BKG's multiscale biological context, enabling the assessment of novel chemical entities. The models achieved robust performance on independent validation sets, yielding AUC-ROC values of 0.81 for Liver, 0.81 for Nervous System, 0.79 for Heart and 0.74 for Kidney. CEP has been demonstrated to be useful through its application to a selection of reference drugs and chemicals, yielding promising results in characterising their organ-specific safety profiles. This work represents a significant step towards animal-free chemical testing.

Tuberculosis (TB) is currently the deadliest infectious disease globally. Despite this, the only widely accepted vaccine for TB (BCG) has variable efficacy and its mechanisms are not fully understood. We sought to understand the immune response to TB under various doses of BCG vaccination using a dose-challenge study in macaques. In this system, we fit a partial correlation network learning protective relationships within a rich panel of immunological features, including serology, immune cell quantities, and cytokine levels. Using this framework, we applied in silico knockouts and identified features involved in vaccine-mediated protection from TB. While classically known features such as interferon-gamma level and CD4 cell quantity were highlighted, our analysis also revealed that antibody titer and phagocytic activity for LAM (a TB antigen involved in immune evasion) are especially important for vaccine-mediated protection. To characterize the role of complete pathways in vaccine-mediated protection we used pathway paired subscore (PPS) revealing the most important pathways controlling the relationship between vaccine dose and measures of both early and late stages of TB infection. We then validated these pathways in unseen single-cell RNA sequencing data collected from the same macaques, confirming the presence of individual interactions in an independently fit cell-cell communication network. Our analysis sheds light on differences between early and overall protection from TB, as well as novel features and paths critical for vaccine-mediated protection. These results are promising for guiding future vaccine development and hypothesis-driven discovery of disease mechanisms.

Accurate assignment of gRNA perturbations remains a central challenge in single-cell CRISPR screens, especially for ambiguous cells with noisy guide counts or potential multiplets. Existing callers rely on guide count heuristics and provide limited information about uncertainty in final assignments. We present a latent generative model that integrates guide count evidence with RNA-derived compatibility within a unified framework for perturbation calling. The model infers per-guide latent activity probabilities, models guide counts with an on/off negative binomial process, derives perturbation cardinality through a Poisson-binomial posterior, and decodes perturbation sets through a structural posterior over candidate assignments.

We benchmarked the model across three datasets: an internal pooled CRISPR screen, a public Perturb-seq dataset, and a public arrayed benchmark approximating ground truth. The model achieved competitive performance relative to 11 gRNA-calling methods, including 0.885 exact-match and 0.896 type-match accuracy on the internal dataset, 0.939 exact-match accuracy on the arrayed benchmark, and 94.95% agreement with a strong reference caller on the Perturb-seq dataset. The structural posterior yielded meaningful uncertainty estimates, with lower assignment entropy for correct than incorrect calls on both the internal and Perturb-seq datasets.

These results support joint modeling of RNA and gRNA counts as a practical route to more reliable perturbation calling. Future work will focus on disentangling transcriptional perturbation effects from correlated variability between guide counts and RNA profiles, and extending the model to capture guide-to-gene hierarchy.

Background: Autism Spectrum Disorder (ASD) and Attention-Deficit/Hyperactivity Disorder (ADHD) are common, frequently co-occurring neurodevelopmental conditions, both linked to synaptic dysfunction. Here we investigate ASD–ADHD comorbidity using a network-based analysis of synaptic protein interactions, integrating single-cell transcriptomic data to add cell-type resolution and brain region context.

Methods: We curated ASD- and ADHD-associated gene sets and mapped them onto synaptic protein–protein interaction (PPI) networks. Community detection was used to identify functional modules, and disease labels (ASD, ADHD, comorbid) were projected onto the network to quantify their distribution across modules. To refine biological context, we incorporate single-cell transcriptomic data by aggregating expression profiles across hierarchically defined cell types and using these to weight network edges, enabling cell-type-informed network analyses.

Results: Network analysis identifies distinct synaptic protein modules enriched for ASD-, ADHD-, and comorbidity-associated genes. Ongoing integration of single-cell transcriptomic weights is expected to refine module structure and highlight cell-type-specific vulnerability. Comparing weighted and unweighted networks will test whether cell-type context alters module membership and separation of disease-associated clusters. We also anticipate the approach will support prediction of regional brain differences.

Conclusion: This network framework aims to clarify shared and distinct synaptic mechanisms underlying ASD and ADHD, and to identify cell types and circuits most relevant to comorbidity. These insights may inform more targeted mechanistic hypotheses and therapeutic strategies.

Brain metastases represent a critical clinical challenge, affecting approximately 20% of cancer patients and conferring a median survival of less than one year. Current treatment strategies, largely dictated by the primary tumor of origin, fail to account for brain metastasis specific vulnerabilities, resulting in limited intracranial control and frequent therapeutic failure. Here, we present a comprehensive multi-omics characterization of 381 brain metastases from 360 patients across 19 primary tumor types, integrating RNA sequencing, whole-exome sequencing, and magnetic resonance imaging data. Clustering based on inferred drug sensitivity profiles revealed three distinct Therapeutic Clusters (TCs), notably independent of primary tumor origin. TC1 is characterized by high proliferation, replicative stress, and sensitivity to cell-cycle and DNA damage response inhibitors; TC2 is driven by PI3K-AKT-mTOR signaling and high metabolic activity, and shows sensitivity to MAPK/RTK pathway inhibitors; and TC3 exhibits an invasive phenotype characterized by EMT, angiogenesis, and sensitivity to SRC/VEGFR blockade. These clusters display distinct genomic features, tumor microenvironment compositions, and transcriptional subtypes, with key functional traits validated by immunohistochemistry and multiplex tissue microarrays. To facilitate clinical translation, we developed a drug prioritization pipeline that identified 40 candidate compounds, currently under ex vivo validation in patient-derived organotypic cultures. Additionally, a machine learning classifier based on a minimal four-gene qPCR signature achieved 76% cross-validated accuracy in TC prediction. Together, these findings provide a framework for stratifying patients with brain metastases and identify novel candidate therapies, paving the way toward more precise and effective treatment strategies.

The extracellular matrix (ECM) critically shapes tumor fate and treatment outcome, serving as a potent prognostic factor. Yet, its compositional heterogeneity across tumors makes it difficult to assess its impact on tumor dynamics. To address this, we introduce an ECM-guided patient stratification pipeline through integration of multi-omic data in lung cancer patients. We obtained four patient groups, representing ECM-grades that showed distinct clinical features, mutation profiles, and cellular heterogeneity. Investigation of patient-specific ECM-induced intracellular signaling via network modeling revealed strong enrichment of pathways and transcriptional regulators related to epithelial-mesenchymal transition (EMT) and cancer stemness in higher ECM-grades. Drug proximity analysis on ECM-grade specific networks predicted olaparib as an ECM-grade dependent therapeutic while erlotinib to be ECM-insensitive which were validated experimentally on lung tumor cells with distinct mutational profiles in response to differing ECM microenvironments. Overall, our ECM-mediated stratification approach is a robust system for capturing ECM heterogeneity and identifying patient groups that can be selectively targeted by distinct therapeutic strategies.

Cancer is characterized by intra-tumor heterogeneity, with distinct subpopulations driven by diverse molecular programs influencing disease progression and treatment response. Multi-omics profiling provides complementary views of tumor biology, yet many computational approaches rely on shared representations that obscure modality-specific signals and limit disentangling of regulatory contributions.
We previously introduced CAVACHON, a hierarchical multi-omics framework using a variational ladder autoencoder (VAE) that leverages prior biological knowledge through a directed acyclic graph (DAG) to encode dependencies between modalities. It preserves hierarchical relationships in latent space, enabling identification of regulatory programs and subpopulations, with prior validation in single-cell multi-omics integration.
Here, we extend CAVACHON to improve clustering stability and interpretability, focusing on the reconstruction-regularization balance, a key challenge in variational models. We introduce three methodological improvements: (i) continuous Kullback-Leibler divergence annealing, enabling smooth transition from vanilla VAE to Gaussian Mixture Model-based clustering, (ii) K-means++ prior initialization, and (iii) per-sample dispersion modelling to separate data variability from regularization.

We evaluated these improvements on large-scale bulk RNA-seq datasets (TCGA), showing reliable performance on bulk data, the most abundant type in cancer research. We assessed clustering stability across 12 cancer types over 10 independent runs, demonstrating strong reproducibility (mean pairwise ARI = 0.75, std = 0.05). Across cancers, 6/12 showed high consistency (≥70%), including 4/12 with perfect reproducibility (std = 0). The remainder consistently formed single or multi-cluster patterns, while cancers with established molecular subtypes (BRCA, KIRC, KIRP, LGG) exhibited multi-cluster organization, indicating that the approach provides stable, reproducible clustering while preserving biologically meaningful heterogeneity.

Background: Transcription factors (TFs) driving the divergence between ""proliferative"" and ""stem-like"" hybrid epithelial/mesenchymal (E/M) states in triple-negative breast cancer (TNBC) remain poorly understood.
Methods: PPARD was identified as a top-ranked candidate driver through a CRISPRa screening in MDA-MB-231 cells. Screening data were processed using the MAGeCK-RRA. For ChIP-seq data, paired-end reads were aligned via Bowtie2, and reproducibility was quantified using the Irreproducible Discovery Rate (IDR) framework. Only high-confidence MACS2 peaks passing the IDR threshold were retained for differential occupancy analysis. Finally, MEME-CentriMo was utilized to quantify motif centrality and spatial binding architectures. We integrated de novo ChIP-seq data with ATAC-seq and histone modification landscapes, ensuring all tracks were processed through a unified computational pipeline.
Results: Phenotypically, PPARD-overexpressing MDA-MB-231 represents a proliferative hybrid state, whereas PPARD-overexpressing HCC1937 exhibits a stem-like hybrid state. De novo motif discovery confirmed significant PPARD recruitment in both models but revealed a distinct topological shift. In MDA-MB-231, unique peaks displayed a centralized, unimodal distribution, indicating direct DNA binding at promoter-proximal regions. In contrast, HCC1937 unique peaks exhibited a bimodal motif architecture, suggesting indirect tethering at distal enhancers. Epigenetically, PPARD occupies an ""epigenetic braking"" zone (H3K27me3/H3K4me2) at the SNAI1 promoter in MDA-MB-231 to sustain proliferation, while in HCC1937, it coincides with ""Identity Peaks"" (H3K27ac) at GATA3 enhancers to favor stemness.
Conclusion: PPARD acts as a context-specific molecular switch that rewires the TNBC transcriptional landscape through topological reprogramming. By shifting between direct promoter binding and indirect enhancer tethering, PPARD differentially recruits epigenetic modifiers to key loci, orchestrating functional heterogeneity.

Various chemicals can influence living organisms in both beneficial and harmful ways by interacting with their metabolic processes, targeting cellular metabolic machinery, genetic information, or other molecular components. To reach these targets, compounds often need to cross biological membranes, either through passive diffusion or active transport. Understanding these interactions is valuable for designing novel drugs and assessing the potential toxicity of chemical substances. The MolMeDB database provides unified and systematic data on interactions of small molecules with membranes and membrane transporters.
We have recently developed an RDF version of MolMeDB that offers a detailed, machine‑readable semantic graph representation of membrane and transporter assays. This version is integrated with major bioinformatics resources through shared identifiers. At the compound level, MolMeDB RDF connects to RDF versions of ChEBI, ChEMBL, PDB, and PubChem; at the transporter level, it links to UniProt.
MolMeDB RDF can be queried via the SPARQL endpoint provided by IDSM, which supports molecule substructure searches through SACHEM functionality as well as federated querying. Federated querying enables users to treat multiple RDF datasets as a single graph, allowing complex questions to be addressed using interconnected data from diverse areas of biology and chemistry.

Introduction: Characterising the phenotypic landscape of rare genetic disorders requires integrating heterogeneous data, including genomic variants, curated gene-disease associations, and structured clinical phenotypes. Existing approaches often treat these sources independently, limiting patient-driven, scalable modelling of disease space. We propose a graph-based framework to unify these data for integrated genotype–phenotype inference.
Methods: We constructed a multi-partite knowledge graph (KG) integrating pathogenic and likely pathogenic single nucleotide variants from DECIPHER, gene-disease associations from Gene2Phenotype, and Human Phenotype Ontology (HPO) terms. The KG encodes relationships between patients, variants, genes, diseases, and phenotypes, providing a network representation of genotype-phenotype structure. HPO terms are weighted by information content (IC) to prioritise diagnostically informative phenotypes.
Preliminary Results: The graph comprises 5,247 patients, 1,338 diseases, 1,228 genes, 3,060 HPO terms, and 124,467 edges. Patients have a median of 6 HPO terms versus 28 for diseases, reflecting differences between clinical and curated phenotyping depth. Approximately half of patient-associated terms are leaf nodes, indicating frequent use of specific phenotypic descriptors. Frequency and IC analyses highlight shared neurodevelopmental phenotypes alongside patient-enriched terms not fully captured in curated annotations. For example, 51 patients with variants in ARID1B, associated with Coffin–Siris syndrome, show strong concordance with curated disease phenotypes alongside additional patient-specific variability.
Conclusion: This framework integrates clinical and genomic data to provide interpretable phenotype structure across rare diseases. Results highlight concordance with curated knowledge and systematic differences in real-world phenotyping. Future work will apply graph neural networks for disease embedding, patient-to-disease inference, and phenotypic clustering.

For most species we know only a few places where they have been seen, not where they live. Yet conservation assessments, including IUCN Red List range measures, must infer whole ranges from these few records. Standard species distribution models treat each species alone and give a single map with no measure of uncertainty, unreliable when only five or ten records exist.
We present MetaDiffusion, a conditional denoising-diffusion model that reconstructs many species' ranges at once from sparse records. Instead of one answer it generates an ensemble of plausible maps, showing the spread of possibilities. It is trained on Lotka–Volterra metacommunity simulations, where every species' true range is known, and conditioned on the local environment and on interactions among species.
At inference it sees only five to ten presences per species and returns full reconstructed ranges. We test it on thirty simulated worlds built with environmental settings absent from training.
With ten observations the predicted ranges match the true distributions of range size, fragmentation and spatial spread (Kolmogorov–Smirnov distances 0.134, 0.231 and 0.095, all within our 0.30 bar). We report the limits: the widest-ranging species and the sparsest budget stay only partly recoverable, which we show is an information limit of the data, not the model. Against a simple smoother on identical data, the smoother recovers more individual cells, but only MetaDiffusion reproduces range geometry and gives a diverse ensemble whose pooled samples recover about twice as much as any single map.

A Knowledge Graph (KG) is a data structure where entities are encoded as nodes and relations between entities as edges holding semantic meaning. KGs contain massive, complex, and heterogeneous data that require specific methods for exploration, such as graph representation learning.

Knowledge Graph Embedding (KGE) is a graph representation learning technique that learns a vectorial representation of KG nodes and edges. The most common architecture for KGE models is the autoencoder: an encoder transforms the KG into a vectorial representation and a decoder attempts to reconstruct the original graph from the latent space. However, current KGE libraries are limited as they do not fully adopt the autoencoder architecture and lack interoperability.

We propose here KGATE (Knowledge Graph Autoencoder Training Environment), designed to address the limitations of current KGE libraries. KGATE assembles each part of an autoencoder as fine-tunable modular blocks. As a general-purpose KGE library, KGATE can easily benchmark and fast prototypemany encoder-decoder combinations in a reproducible manner. In addition, KGATE implements the building blocks for temporal embeddings of KGs, and handles the most common graph temporal representations.

We illustrate KGATE with different tasks relevant for biomedical KGs. For instance, we compared various encoder-decoder for the task of link prediction, which is relevant for therapeutic target identification or drug repurposing. In addition, using a temporal biomedical KG, KGATE unveils gene-aging disease relationships by performing link prediction and node classification contextualized by age.

KGATE's source code is available at https://github.com/BAUDOTlab/KGATE

Imaging mass cytometry (IMC) enables high-dimensional, spatially resolved characterization of the tumor microenvironment (TME). Its broader adoption, however, is limited by fragmented computational workflows and reduced accessibility, particularly for scientists with limited computational expertise.

Here, we present AgentIMC, a reproducible and user-oriented framework for IMC data analysis, developed in the context of sarcoma microenvironment studies and adaptable to other cancer types and antibody panels. Starting from ROI-level TIFF channels, the workflow integrates channel inspection, segmentation-oriented marker selection, nuclei segmentation, object-level marker quantification, phenotype assignment, spatial nearest-neighbor analysis, and phenotype-level summarization within a single pipeline.

To enhance usability, the analytical backend is coupled with an interactive interface that supports both single-ROI and batch-ROI processing. The interface allows configurable marker parameters, provides visual quality control outputs, enables downloading of result tables, and supports automated report generation based on derived analyses.

Applied to a sarcoma IMC dataset, AgentIMC identified biologically meaningful cellular populations, including myeloid-like, T cell–like, B cell–like, plasma-like, stromal-like, and endothelial-like compartments, while maintaining a consistent and auditable analysis structure across all processing stages.

Overall, AgentIMC reduces technical barriers in exploratory and translational IMC studies by combining standardized workflows, interpretable spatial summaries, and presentation-ready outputs within a unified environment. This framework supports reproducible characterization of TME organization in spatially resolved data and facilitates collaborative analysis between computational and experimental researchers.

This study is supported by Ricerca Finalizzata 2021 by Italian Ministry of Health—Giovani Ricercatori (GR)— “Change promoting,” project code GR-2021-12373209.

Tumors are characterized by diverse biological processes reflected across tissue morphology and molecular profiles. While deep learning models that integrate these data sources are increasingly more accurate, their complexity often reduces interpretability, limiting insights into the complementary roles and influence of whole-slide images (WSIs) and gene expression in cancer survival prediction.
To address this, we propose DIMAFx, an explainable framework that creates interpretable, disentangled multimodal representations for cancer survival prediction. Specifically, DIMAFx uses gene expression pathways and morphological, frequency-aware WSI prototypes, which are combined using separate attention mechanisms to form modality-specific and modality-shared representations, with feature contributions quantified using SHapley Additive exPlanations.
DIMAFx enables systematic analysis of key multimodal interactions and the biological information encoded in the disentangled representations, while maintaining state-of-the- art performance and improved representations disentanglement across four TCGA cohorts (breast, bladder, lung, kidney). The in-depth interpretability analysis on breast cancer showed that the most predictive features contain modality-shared information, including KRAS signaling contextualized by tumor morphology, and solid tumor morphology contextualized primarily by late estrogen response. The latter interaction showed that higher-grade morphology is associated with pathway upregulation and elevated predicted risk, consistent with known breast cancer biology. Among modality-specific features, interacting adipose and stromal morphologies were most influential, capturing complementary micro-environmental signals from the WSIs.
This analysis shows that DIMAFx captures biologically relevant patterns, revealing key interactions and trade-offs between modalities while maintaining state-of-the-art performance. These results demonstrate that multimodal models can overcome the traditional trade-off between performance and explainability, supporting their application in precision medicine.

Neuroblastoma (NB) is a pediatric cancer arising from the developing sympathoadrenal lineage. It shows remarkable clinical and molecular heterogeneity, reflected in profound transcriptional plasticity that may underlie tumor relapse. High-risk neuroblastoma remains difficult to treat and is responsible for approximately 15% of pediatric cancer deaths.

To characterize the NB transcriptional landscape, we systematically analyzed single-cell/nucleus transcriptomic datasets from human NB tumors across multiple studies, NB cell lines, and healthy sympathoadrenal lineage cells. Using regularized non-negative matrix factorization (NMF), we decomposed NB-gene expression into robust transcriptional programs which are associated with biologically meaningful processes.

Many programs are shared between malignant and healthy cells but display differential activity. Comparison with established NB transcriptional signatures such as ADRN and MES suggests that these larger signatures can be represented as combinations of multiple transcriptional programs. We observed links between inferred copy number variation (CNV) states and transcriptional program activity and variability, which cannot be explained by dosage-effects alone. Clustering cells by CNV state reveal distinct patterns of program activity, suggesting that genomic alterations constrain transcriptional programs at the single-cell level.

Together, our results reveal that NB transcriptional heterogeneity is organized into a set of transcriptional programs, which are partly shared with normal sympathoadrenal development and whose activity is constrained by tumor genetic alterations.

Understanding RNA virus evolution remains challenging due to their high mutation rates, compact genomes, and often scarce datasets that limit robust inference. This is particularly true for understudied arboviruses such as the Venezuelan equine encephalitis virus (VEEV), where intra-host evolutionary dynamics over time are not well understood. Investigating how host environment, temporal progression, and genomic context shape viral diversity requires approaches operating effectively under scarce data conditions. We deployed a mixed-effects modelling framework for investigating spatiotemporal mutational patterns in VEEV-infected culex mosquitoes. Intra-host diversity was modelled as a function of time since infection, tissue compartment, and protein identity, while accounting for host-specific variability. We observed substantial heterogeneity in viral diversity across hosts, indicating pronounced variation in viral population structure. Gene length and protein identity were associated with differences in diversity, suggesting heterogeneous evolutionary constraints across the viral genome. Comparisons across tissue compartments suggest differences between non-disseminated and disseminated samples, although these effects may be partially confounded by sampling time. Our results highlight the interplay between host-specific effects, genomic context, and mutation-level dynamics in shaping intra-host viral evolution. Given the inherent data scarcity associated with emerging or understudied viruses, we propose that a hybrid analytical framework combining classical statistical modeling with machine learning approaches may enable us to better capture complex patterns in such small, noisy datasets. This work showcases a foundation for developing such frameworks studying viral evolutionary dynamics for small-N viral genomic data

Post-translational modifications (PTMs) of the axonal microtubule-associated protein tau have been implicated in the development of tauopathies, a group of diseases characterized by the formation of pathological tau aggregates, including Alzheimer's disease (AD). Among the PTMs, hyperphosphorylation is of particular importance, characterized by an abnormal increase in phosphorylation at specific sites. However, the downstream pathological consequences of tau hyperphosphorylation remain unclear. In this study, we investigated the effects of pseudohyperphosphorylated tau (PHP-Tau) in model neurons, mimicking a disease-like hyperphosphorylation of tau at specific residues. We used a multi-omics approach, including proteomics, phosphoproteomics and lipidomics, to analyze molecular changes associated with tau hyperphosphorylation. In parallel, 3D reconstruction of super-resolution microscopy images was used to assess changes in axonal microtubule organization. Our integrative analysis reveals that tau hyperphosphorylation is associated with alterations across multiple molecular layers and impacts cytoskeletal organization. Notably, changes were observed in STMN1, a protein involved in microtubule assembly and disassembly, and in GM3, a sphingolipid associated with AD. Furthermore, multi-omics data indicate that tau hyperphosphorylation affects pathways associated with AD and is linked to structural alterations in microtubule organization. These findings provide new insights into how pathological tau may contribute to neuronal dysfunction by coupling molecular dysregulation with structural changes in microtubule architecture.

Combination drug therapies hold significant promise for treating complex diseases such as cancer, as they can target multiple pathways, overcome resistance, and reduce toxicity compared to monotherapies. Artificial intelligence offers powerful tools for predicting synergistic combinations; however, current approaches typically rely on a single drug descriptor, usually chemical structure, which insufficiently captures the complex biology of drug action individually and in combination. While additional modalities could help, missingness in descriptor coverage poses a fundamental challenge for multimodal integration.
To address these limitations, we propose a metric-agnostic, missing-aware, attention-based deep learning architecture. that integrates pharmacologically and biologically informed drug representations including bioassay activity profiles, structural descriptors, morphological profiles, and genetic signatures alongside cell line gene expression profiles. The model simultaneously predicts four established synergy scoring frameworks (ZIP, Loewe, Bliss, HSA), while dynamically attending to whichever modalities are present per sample, enabling robust learning under severe missingness.
Relative to the single-modality structural baseline (R² = 0.73, MSE = 120.81), our full multimodal integration improved R² by 9.78% and reduced MSE by 7.1% for synergistic samples. These gains were consistent across five cross-validation folds and statistically significant under a paired t-test (p = 0.034) and exceeded those of state-of-the-art synergy baselines.
To move beyond predictive accuracy toward interpretable biological insight, we apply SHAP-IQ interaction analysis to quantify how drug descriptors and gene expression features jointly influence predictions, revealing cross-modal interactions that highlight both drug–drug and drug mechanism–gene pairs driving synergy, and ultimately uncovering why specific combinations are synergistic in cellular contexts.

Glioblastoma multiforme (GBM) is an aggressive primary brain tumor characterized by rapid and extensive invasion into surrounding brain tissue. GBM invasion is known to involve tumor cell proliferation and remodeling of the extracellular matrix (ECM), yet the relative contributions of these processes over the course of tumor growth in spheroid models remain poorly understood. In this study, a spheroid model was used to investigate how proliferation and ECM deposition influence GBM spheroid growth. U87-MG spheroids were cultured for 14 days and assessed for changes in size, metabolic activity (Alamar Blue), proliferative behavior (Ki-67), and ECM remodeling at different timepoints. ECM remodeling was analyzed by immunofluorescent staining and confocal microscopy of hyaluronic acid and collagen types I, VI, and IX, which regulate matrix architecture and mechanical properties. Spheroid size increased throughout the culture period (Day 3: 116.8±57.7µm; Day 7: 155.6±90.77µm, p=0.0019). Early spheroid expansion was associated with high proliferative activity, low metabolic activity, and matrix deposition. Later timepoints exhibited reduced proliferation (Day 3: 51.5±17.9%; Day 7: 28.1±22.1%, p<0.0001) while demonstrating increased metabolic activity (relative increase in metabolic activity from day 3 to 7: 1.89±0.76) and ECM remodeling. These findings indicate that GBM spheroid growth is governed by mechanisms at different stages, with early growth dominated by proliferation and later growth increasingly supported by ECM deposition. This shift likely reflects adaptive strategies that allow GBM to meet the demands of its microenvironment, supporting its survival under changing conditions.

Tumor stratification remains a significant challenge due to the complex molecular landscape of cancer, as intertumoral heterogeneity can lead to different clinical outcomes and treatment responses. In this study, we propose an unsupervised Graph Neural Network (GNN)-based framework, applicable across omic modalities, to stratify cancer patients and uncover mechanisms underlying tumor heterogeneity. Our approach integrates network propagation via Random Walk with Restart (RWR) and permutation testing to construct patient-specific networks, which are then embedded using InfoGraph to obtain patient-level representations. To ensure robustness, we evaluate dimensionality reduction methods (PCA, t-SNE, UMAP) and clustering strategies (HDBSCAN, DBSCAN, affinity propagation), selecting the optimal configuration based on stability metrics (silhouette score, AMI, ARI). We then identify cluster-specific molecular patterns. In a case study of Glioblastoma Multiforme (GBM), we analyzed mutation profiles from 356 TCGA patients to construct patient-specific networks. We identified 11 clusters, four showing significant survival differences (p = 0.0018 and p = 0.036). Transcriptomic analysis revealed 221 differentially expressed genes (|logFC| > 2, p < 0.05) between identified clusters. To investigate underlying mechanisms, we applied a three-layer Graph Isomorphism Network (GIN) with PGExplainer, identifying cluster-specific subgraph motifs and pathways. One cluster was enriched in ion and membrane transport processes, while another showed enrichment in immune and signaling pathways. Overall, this framework combines unsupervised GNN-based representation learning, network propagation, and explainable AI to stratify tumors and assign biological meaning to clusters, providing insights into intertumoral heterogeneity.

Clonal hematopoiesis (CH) arises when hematopoietic stem and progenitor cells acquire somatic mutations that expand during aging. Although CH has been extensively studied through bulk-sequencing variant allele fractions (VAFs), including longitudinal VAF trajectories, the latent determinants of clonal behaviour, fitness, clonal age, and lineage bias, remain difficult to resolve from cross-sectional data.

We present a Bayesian multi-compartment model that infers these quantities from snapshot single-cell data by coupling a hierarchical model of hematopoietic differentiation to binomial and multinomial observation models for clone-resolved cell counts across stem and progenitor compartments. Our model explicitly represents compartment-specific residence times and branch allocation probabilities, enabling joint inference of selective advantage, clone age, and clone-specific deviations from a wild-type differentiation program.

We evaluated our framework using simulations with known ground truth across multiple differentiation hierarchies. Snapshot single-cell data carried strong information on lineage bias and supported precise inference of fitness and clonal age with informative priors, slow downstream compartments, and more cells. Applied to two independent human single-cell hematopoietic datasets with clonal labels, the model recovered clone-specific branch allocation shifts consistent with lineage skewing and yielded plausible posteriors for fitness and age. Relative to naive clone frequencies, inferred branch effects removed many shifts, consistent with sampling and compositional noise in the raw data. Across cohorts, the dominant signal was altered differentiation propensity rather than differences in proliferative advantage.

Our framework provides a basis for disentangling selection and fate bias in CH from cross-sectional data and defines the kinetic regimes in which snapshot-based inference is informative.

Background
Extrachromosomal DNA (ecDNA) drives oncogene amplification and tumour heterogeneity, yet its regulatory impact at single-cell resolution remains poorly understood. Integrating transcriptomic and chromatin accessibility data enables characterisation of the functional consequences of ecDNA in tumour cells.

Methods
We performed single-cell multiome (RNA + ATAC) analysis on two lung cancer samples (baseline and relapse) from the TRACERx/PEACE cohort. Due to poor RNA quality, the relapse sample was excluded. After stringent quality control, high-confidence RNA and ATAC profiles were obtained. Copy number inference distinguished tumour and diploid cells. scAmp was applied to identify genomic regions harbouring focal amplifications consistent with ecDNA. Based on these loci, we developed an integrative framework combining copy number, gene expression and chromatin accessibility into a unified ecDNA-like score to classify cells.

Results
We identified 1,329 tumour cells and 1,057 diploid cells, capturing malignant and microenvironmental compartments. scAmp identified focal amplification at the chr20q13 locus, consistent with ecDNA regions detected by matched WGS data. Cells classified as ecDNA-like exhibited coordinated increases in chromatin accessibility and gene expression at amplified loci, including oncogenic drivers such as ZNF217, a regulator of tumour progression promoting proliferation, metastasis, and therapy resistance through chromatin remodelling. These cells were enriched for MYC targets, TNFα signalling and apoptosis. Motif analysis revealed enrichment of AP-1 transcription factors, consistent with MAPK-driven regulatory programs.

Conclusions
Single-cell multiome integration enables identification of ecDNA-associated regulatory states. This framework links structural genomic alterations with transcriptional programs at single-cell resolution and provides insights into tumour heterogeneity and regulatory plasticity.

The gut microbiome, the complex community of microorganisms inhabiting the human gastrointestinal tract, is inherently high-dimensional and compositional. Identifying taxa associated with host phenotypes is a central challenge with important diagnostic and therapeutic implications. Differential abundance (DA) analysis addresses this task, but compositional constraints, sparsity, and inter-taxa dependencies complicate statistical inference and limit interpretability. Moreover, most existing DA methods produce unstructured lists of taxa, without quantifying relative association strengths or capturing relationships among them.
We present BLINCHESS, a novel computational framework that reformulates DA analysis as a ranking problem. BLINCHESS employs an iterative pairwise ratio comparison strategy, circumventing the need for explicit normalization. By treating taxa as competitors in a tournament-style framework, the method leverages the Elo rating system, commonly used to rank players in zero-sum games, to assign each taxon a continuous score reflecting its differential signal. This approach yields a prioritized ranking of taxa associated with the phenotype and naturally groups taxa with similar DA behavior. The resulting rankings further enable unsupervised clustering to identify modules of taxa with coherent differential patterns.
We evaluate BLINCHESS on both synthetic and real-world datasets, including inflammatory bowel disease (IBD), colorectal cancer (CRC), and gastric cancer (GC). Across simulations, BLINCHESS matches or outperforms established DA methods in detection accuracy and robustness, while providing interpretable rankings and structure-aware groupings of taxa. These findings highlight the potential of pairwise comparison “based ranking models as a scalable and interpretable approach for DA discovery in microbiome data.

The human gut microbiome plays a vital role in host health, and alterations in microbial composition and function have been linked to diseases such as inflammatory bowel disease (IBD), colorectal cancer (CRC), and metabolic disorders. Metagenomic studies typically analyze microbiomes through either taxonomic profiling to identify species, or functional profiling to characterize gene families. However, these approaches are usually performed independently, limiting our ability to link specific microbial taxa to their functional contributions. Furthermore, current methods lack the capacity for large-scale, accurate functional copy number estimation directly from complex metagenomic samples.
Here, we present TEMPEH (Taxonomic and Enzymatic Metagenomic Profiling with Enhanced Hybrid Annotation), a computational pipeline designed to bridge this gap. By employing both k-mer and alignment-based annotation methods against a unified taxonomic-functional reference database, TEMPEH achieves coupled, read-level annotation. Crucially, it generates highly accurate matrices of species-function pairs, complete with precise copy number estimations, enabling the identification of combinatorial microbiome features that drive differences between case and control samples.
To interpret these high-resolution profiles, we generated genome-scale functional profiles for a large set of reference genomes to construct a comprehensive, low-dimensional functional embedding. By projecting TEMPEH-derived species profiles into this continuous reference space, we demonstrate the utility of this framework in distinguishing normal functional variation in healthy microbiomes from the metabolic deviations characteristic of dysbiosis.
Ultimately, TEMPEH provides a scalable computational framework to integrate taxonomic and functional dimensions, offering deeper insights into microbiome-mediated disease mechanisms and informing potential targeted therapeutic strategies.

H&E-stained histological images remain the gold standard in clinical pathology, providing cost-effective, rapid, and robust visualization of tissue architecture and cellular morphology. Recent advances in machine learning have substantially expanded their utility, enabling representation learning, cross-modal alignment, and prediction of spatial omics directly from H&E images. Multiple studies have been published, demonstrating that such predictions are feasible for metabolomics, transcriptomics, and proteomics at the slide, spot, and single-cell levels. Here, we focus on models that predict transcriptomics at single-cell resolution.

Despite rapid methodological development, rigorous and objective evaluation of these approaches remains limited. To address this, we benchmarked three methods: SpatialEx, a hypergraph-based model compatible with multiple pretrained image encoders; GHIST, a deep learning framework using a UNet3+ backbone trained from scratch on paired H&E and spatial transcriptomics data; and Pixel2Gene, which uses the HIPT pathology foundation model to extract histological features that are fed into a simple MLP. As controls, we included simple linear regression models trained on embeddings from multiple foundation models, totalling 16 models evaluated in this study.

We evaluated all models on in-house Xenium 5K panel data from acute myeloid leukemia skin tissue. To further validate our findings, we are extending this pipeline to publicly available Xenium skin datasets, including a 5K and a 300-gene panel dataset. Our results show that predictive performance of published models was limited for the majority of genes and, surprisingly, comparable to that of a simple linear regression model trained on foundation model embeddings.

Molecular tumor signatures vary with age across cancer types, leading to distinct phenotypic and clinical outcomes. This study investigates early-, middle-, and late-onset network- and pathway-level signatures in multi-omic data from The Cancer Genome Atlas. Patients were divided into age groups: early-, mid-, and late-onset. In pan-cancer, these groups differ in survival probability (p<0.0001) and share mutations in 234 driver genes and 380 common alterations at the same amino acid residues. At the tissue level, the uterus and the large intestine have the highest number of mutations, with the uterus having the highest number of mutated residues shared across all groups. The medians of the frequency of patients that are early-, mid-, and late-onset with mutations in shared drivers differ (p<0.05). Additionally, the frequency of early-onset patients with mutations in TP53, MYC, and CCND3 is higher (p<0.05). In pan-cancer, all groups differ in tumor mutation burden (increases with age) and tumor purity (decreases with age) (both p<0.0001). Tissue-level exceptions include breast and kidney for tumor mutation burden, and brain and adrenal gland for tumor purity. Using Virtual Inference of Protein-activity by Enriched Regulon analysis (VIPER), we infer differentially active transcription factors, observing that signaling strength based on pathway expression scores is higher when comparing early- with mid-onset patients than when comparing early- with late-onset patients for the PI3K-Akt pathway in the brain and the HIF-1 pathway in adrenal gland cancer. Overall, these findings highlight age-specific molecular changes across tissues in multi-omic data that might enable personalized therapies.

Tuberculosis is the world's deadliest infectious disease. Asymptomatic tuberculosis (aTB) comprises approximately 50% of prevalent TB cases and is challenging to diagnose due to clinical and molecular heterogeneity. A systems-level approach is required to untangle this variability and the underlying disease biology. In a community-wide TB screening cohort, we investigated whether we could identify sub-phenotypes of aTB based on integrated clinical, microbiological, radiological, and inflammatory (CMRI) features, and whether these phenotypes are associated with distinct biomolecular profiles.

We applied Factor Analysis of Mixed Data and k-means clustering to stratify aTB individuals, validating cluster stability through silhouette analysis and bootstrapped resampling. Differential expression analysis of a 46-analyte Luminex panel identified proteomic markers characteristic of each cluster. We identified three aTB subtypes stratified by disease severity: Cluster A comprised younger, female-predominant participants with higher BMI and minimal inflammation, lung pathology, and bacterial burden. Cluster B represented an intermediate phenotype with significant lung pathology without cavitation. Cluster C was skewed towards older males with lower BMI, elevated inflammation, higher bacterial burden, and cavitary disease. Proteomic profiling reinforced disease heterogeneity, revealing molecular differences between clusters, including high expression of immune regulators PD-L1 and MIP-3β in Cluster C. Cluster B demonstrated elevated cytokine levels compared to Cluster A, consistent with its intermediate phenotypic severity. These findings reveal aTB endotypes with differential proteomic expression, possibly indicating distinct pathways contributing to disease heterogeneity. Combined with multi-omics analyses, this stratification framework enables mechanistic understanding of aTB and facilitates development of targeted diagnostics and therapeutics.

Glutamatergic synapses mediate the major excitatory signaling in the central nervous system and are critically implicated in neurodevelopmental and neurological disorders. However, how genetic variation across glutamatergic synapse genes translates into disease vulnerability and phenotypic diversity remains poorly understood.

Here, we present a systems-level analysis of 626 human glutamatergic synapse genes integrating genetic variation from gnomAD and ClinVar with tissue expression, protein–protein interaction networks, Gene Ontology annotations, and Human Phenotype Ontology data. To quantify gene-level disease association, we introduce a pathogenesis ratio that captures the relative enrichment of pathogenic variants while accounting for mutational burden.

We show that disease-associated variation is highly non-uniform and concentrated in a small subset of genes, with approximately 10 genes accounting for ~55% of pathogenic variants. These highly vulnerable genes are preferentially expressed in cortical regions and occupy central positions within synaptic interaction networks, forming hubs enriched in postsynaptic density and NMDA receptor signaling. In contrast, low-pathogenicity genes display sparse connectivity and reduced network integration.

At the clinical level, glutamatergic synapse genes exhibit strong phenotypic convergence, predominantly affecting neurodevelopmental functions despite substantial genetic heterogeneity.

Together, our results reveal a hierarchical and network-driven architecture of disease vulnerability at the glutamatergic synapse. To support exploration and interpretation, we provide GluSynDB, an integrated resource linking genetic variants, molecular networks, and clinical phenotypes, enabling systematic prioritization of disease-associated genes.

Spatial transcriptomics (ST) has enabled the quantification of mRNA directly within its tissue context, thereby preserving the spatial organization of gene expression. In a typical ST workflow, consecutive tissue slices are collected from a sample, but they are usually analysed as separate two-dimensional sections rather than being aligned into a unified three-dimensional (3D) representation. Aligning slices into a three-dimensional space enables downstream analyses, such as 3D domain clustering and spatial differential expression. More than 24 computational methods have been proposed for this task, creating a need for systematic benchmarking.
Out of these 24 methods, we were able to run 14 algorithms, spanning approaches from deep learning to diffeomorphic registration and optimal transport, across multiple ST acquisition technologies. We considered three alignment settings: pairwise tissue alignment within a single donor, 3D reconstruction of multiple slices, and alignment across different sequencing technologies. All tasks were carried out on both biologically derived ST data and in silico generated ST data. In the pairwise alignment task, SPACEL and INSPIRE, both deep learning-based approaches, achieved the best performance based on the accuracy of cell type matching across slices. In addition, on the in silico data, we quantified runtime and compute requirements for all methods on comparable datasets.
In this benchmark, we show that no single algorithm is universally optimal in terms of accuracy. Instead, we find that performance depends on the alignment objective and the ST technology, and there can be large differences in computational efficiency between algorithms.

Phenotypic heterogeneity within isogenic microbial populations poses a clinical challenge by enabling cells to respond differently to drug treatment. In Saccharomyces cerevisiae, the Pdr5 efflux pump, encoded by the PDR5 gene, displays bimodal expression giving rise to a drug-tolerant phenotype. Despite its clinical relevance, analogous behaviour happens in Candida species, the mechanisms by which this heterogeneity is passed across generations remain poorly understood.
We performed live-cell imaging to reconstruct lineage-resolved Pdr5-GFP dynamics in both untreated and drug-treated yeast cell populations with multigenerational resolution. Semi-automated single-cell tracking resulted in 4707 single-cell time series and ~ 300k Pdr5-GFP measurements. Significant mother-daughter correlation in Pdr5-GFP levels (r=0.80) pointed to Pdr5 protein inheritance, raising a key question: does this inheritance reflect the passive partitioning of proteins at division or the active transmission of the underlying regulatory state?
We therefore formulated a dynamical systems model of immature and mature Pdr5-GFP governed by a stochastic gene synthesis rate. Using Kalman filtering and smoothing, we inferred single-cell synthesis rates by decoupling them from maturation, dilution, and degradation. A high mother-daughter correlation of inferred synthesis rates (r=0.60 vs. r=0.02 for random pairs) suggested an epigenetic basis to Pdr5-GFP heterogeneity. Under drug treatment, this regulatory inheritance was reshaped, indicating that tolerance can emerge and propagate across lineages rather than being restricted to pre-existing high-Pdr5 cells.
Our results indicate that Pdr5 heterogeneity is maintained through heritable regulatory dynamics, linking epigenetic memory to the propagation of drug-tolerant phenotypes.

Biomedical knowledge required for understanding disease mechanisms and developing effective therapeutics is dispersed across hundreds of databases, ontologies, and publications in heterogeneous and non-standardised formats. Knowledge graphs (KGs) offer a principled framework for integrating such data; however, most existing biomedical KGs remain constrained by limited scope, lack of systematic update pipelines, sparse metadata, and interfaces that are inaccessible to non-technical users. We present CROssBARv2, a large-scale heterogeneous biomedical KG system designed to support systems biology research and drug discovery. CROssBARv2 integrates data from 34 curated sources into a Neo4j graph database comprising 2,709,502 nodes and 12,688,124 relationships across 14 biologically meaningful node types, including proteins, genes, drugs, compounds, diseases, pathways, phenotypes, GO-terms, and more. The system features fully automated pipelines for regular data retrieval and standardisation, ensuring long-term maintainability and scalability. Node embeddings encoding biological features are incorporated to enable vector-based similarity searches and downstream predictive tasks. A key contribution is CROssBAR-LLM, a natural language interface that translates user-submitted biomedical questions into Cypher graph-db queries, executes them against the KG, and returns contextually accurate responses, enabling graph exploration without programming expertise. Systematic benchmarking across multiple datasets demonstrated that CROssBAR-LLM substantially outperforms web-search-augmented LLMs in biomedical question-answering accuracy, effectively eliminating hallucinations through grounding in structured data. Deep learning models trained on CROssBARv2 for protein function prediction achieved state-of-the-art performance, further demonstrating the KG's utility for biological inference. CROssBARv2 is publicly accessible at https://crossbarv2.hubiodatalab.com/llm, with full source code and datasets available at https://github.com/HUBioDataLab/CROssBARv2.

The extracellular matrix (ECM) is a dynamic regulator of tumor–microenvironment interactions, influencing both tumor progression and heterogeneity. Matrisome-associated gene expression profiles can capture transcriptional programs that extend beyond discrete subtype classifications. This study investigates transcriptional matrisome dynamics across pancreatic cancer progression using an integrative multi-omic framework. Tumors from The Cancer Genome Atlas (TCGA) pancreatic adenocarcinoma (PAAD) cohort (n = 178) were stratified based on matrisome-associated gene expression scores, revealing variation in program activity across samples. Higher activity is associated with increased inflammation, elevated hypoxia, and poorer survival. We also observed that high-scoring tumors exhibit increased stromal content, prompting us to assess whether this program is tumor-intrinsic, revealing heterogeneous expression across cell lines. To assess whether this program reflects coordinated cell-cell communication, we use single-cell transcriptomic data to localize the program in PAAD cell populations, revealing fibroblast and stellate populations as major contributors, with heterogeneous activation across cell types. Network modeling is used to further define coordinated gene interactions underlying ECM remodeling. Together, these findings suggest a matrisome-associated transcriptional program that links tumor-intrinsic activity with stromal interactions across cancer progression.

Fragile X Syndrome (FXS) is a
neurodevelopmental disorder caused by mutations
in the FMR1 gene, resulting in the loss of FMRP,
an RNA-binding protein regulating translation
of hundreds of mRNAs. The Fmr1 knock-out
mouse models this deficiency and is used
to study molecular perturbations in the FXS
brain. Omics analysis shows that FMRP loss
disrupts coordination between transcriptomic and
translatomic layers. But limited sample availability
and dataset heterogeneity hinder detection of
subtle, coordinated multi-omic dysregulations.
To address this, we trained a Multi-Channel
Variational Autoencoder (MCVAE) on wild-type
samples to learn a shared latent representation
of transcriptomic and translatomic modalities
via cross-modal reconstruction. Testing MCVAE
on Fmr1-knock-out samples revealed deviations
from wild-type as anomalies, uncovering known
and novel perturbations. Compared to alternative
methods, MCVAE shows stronger enrichment
for FMRP mRNA targets and improved genotype
discriminative power in post-hoc tests. Translatomic
anomalies exhibited coordinated relationships
with transcriptomic anomalies, as supported by
publicly available databases exploration. Moreover,
these anomalies mapped to validated FMRP
regulators and neurodevelopmental pathways,
establishing MCVAE as a framework to uncover
coordinated molecular perturbations underlying
the FXS pathophysiology and guide biomarker and
therapeutic target identification.

Introduction
Over the past decade, several machine learning methods have been developed to predict patient responses to anti-cancer drugs. Despite methodological advances, they present limited translational relevance for personalized medicine. The main limitations of current state-of-the-art methods can be summarized by their inflated performance, limited transparency and scarce validation in patient-derived material.

Methods
Cancer drug response prediction models use gene expression and chemical information to learn patterns associated with treatment sensitivity. For training, we used gene expression data from Cell Model Passport and pharmacogenomic data from Genomics of Drug Sensitivity in Cancer.

To improve transparency, we developed NELLY, a deep learning model incorporating a dynamic weighting mechanism (DWM). This module enables extraction of patient-specific gene-level attributions associated with predicted patient response.

Models were trained and evaluated using a cell-line-blind 10-fold cross-validation to assess performance in unseen cell lines. Additionally, we applied recommendation-based methods, including Precision@K and NDCG@K to assess models' ability to rank the most effective drug for a given patient.

For independent validation with patient material, we have assembled a patient-derived-cancer-organoid pharmacogenomic atlas. The use of this fully independent dataset allowed us to evaluate generalization performance to clinically relevant samples.

Conclusion
We demonstrated that NELLY outperforms state-of-the-art models in generalization to cancer organoid data, achieving the best performance for patient-specific drug prioritization. Additionally, DWM-derived gene attribution analysis uncovered resistance-associated programs related to cancer cell plasticity, thereby adding mechanistic interpretability beyond drug ranking.

The emergence of new pathogenic bacteria is frequently driven by horizontal gene transfer (HGT), which enables rapid adaptation to novel ecological niches. The emetic lineage of Bacillus cereus carries the pXO1-like megaplasmid pCER270, encoding the non-ribosomal peptide synthetase genes ces required for cereulide toxin biosynthesis. Here, we combine quantitative proteomics and protein–protein interaction (PPI) network analysis to dissect how pCER270 and its ces toxin cluster rewire cellular functions, adaptation and virulence in the emetic reference strain F4810/72 (aka AH187).
We employed two mutants: a plasmid-cured derivative lacking the entire pCER270 and a ces knockout variant retaining the remaining plasmid backbone, and compared them to the wild type using TMT-based high-pH offline 2D-RPLC-MS/MS (Exploris 480/FAIMS). Differentially abundant proteins were projected onto a whole-genome PPI network to quantify changes in network topology, complemented by in-depth phenomic profiling. Network analysis uncovered functional coupling between plasmid-encoded and chromosomal modules, showing that acquisition of pCER270 reshapes global cellular organisation.
Plasmid loss disrupted protein clusters involved in cell surface integrity, transcriptional regulation, sporulation and central metabolism, indicating broad systems-level remodelling. In contrast, targeted deletion of ces caused more local, module-restricted perturbations in the PPI network and phenotype space. Together, our data illustrate how HGT-driven acquisition of a cereulide-encoding megaplasmid differentially tunes network architecture and functional adaptation, providing a systems biology framework for studying pathogen emergence in B. cereus and other Gram-positive bacteria.

Kinetic models of the Calvin–Benson cycle, parameterized with species- or condition-specific rate constants, offer a mechanistic foundation for understanding how photosynthetic metabolism responds and adapts to environmental variation. Such models are essential for predicting carbon assimilation under fluctuating nitrogen, light, and CO2 conditions. However, accurate parameter estimation remains a major bottleneck, particularly in large-scale models with highly nonlinear relationships and often underdetermined parameters given the available data. We benchmark four state-of-the-art high-dimensional Bayesian optimization (HDBO) methods for their ability to estimate kinetic parameters from photosynthetic CO2 response (A/Ci) curves. These curves characterize the dependency of the net CO2 assimilation rate (A) on the intercellular CO2 concentration (Ci) and are widely used to infer photosynthetic capacities and limitations in vivo. We simulate synthetic A/Ci datasets from two large-scale kinetic models of photosynthesis endowed with mass-action and Michaelis-Menten kinetics, respectively. Each HDBO method is tasked with recovering the original parameter set by minimizing a chi-square cost function between simulated and target A/Ci curve. All the methods exploit the inherent structure of biochemical networks—where topological and kinetic dependencies constrain the effective degrees of freedom—to accelerate convergence. Across 10,000 iterations per method, we observe that each method demonstrates its capability to adaptively focus the search within the relevant subspaces, achieving reasonable fits. These results highlight the effectiveness of HDBO techniques in biochemical model calibration. Our findings demonstrate that HDBO is a promising tool to significantly improve the precision and computational efficiency of kinetic parameter estimation in large-scale photosynthetic models.

Background
Pheochromocytomas and paragangliomas (PPGL) are rare neuroendocrine tumors characterized by molecular heterogeneity and limited therapeutic options. To identify actionable upstream regulators beyond gene-centric associations, we implemented an integrative strategy focused on an emerging drug target, phosphodiesterase 3A (PDE3A), combining transcriptomic modeling with functional validation in patient-derived samples.

Materials and Methods
RNA sequencing data from TCGA PCPG (n=179) were stratified by median PDE3A expression to PDE3A-high and PDE3A-low expression groups. Protein expression was validated by immunohistochemistry in an independent FFPE cohort (n=151). To investigate differentially activated pathways in PDE3A high and low expression groups, a workflow of PROGENy, DecoupleR and CARNIVAL was implemented. Computational predictions were evaluated by ex vivo drug screening of 83 compounds in an independent patient-derived ex vivo samples (n = 21).

Results
In TCGA-PCPG sample groups, CARNIVAL identified a MAPK1-RPS6KA3 and GSK3B-CCND3-CDK6 regulatory axis specific to the PDE3A high subgroup and PROGENy inferred activated hypoxia and PI3K signaling. Ex vivo screening revealed sensitivity to PDE3A modulators 5/21 (anagrelide) and 7/21 (BAY 2666605) samples. Drug sensitivity k-means clustering stratified compounds into three clusters: apoptosis causing drugs (e.g. navitoclax, cisplatin, niclosamide); a subgroup-selective cluster targeting more specific pathways such as mTOR and MEK (e.g. binimetinib, trametinib, vistusertib, AZD8055); and a low-efficacy cluster (e.g. alectinib, phenformin, olaparib) comprising agents with heterogeneous molecular targets.

Conclusions
PDE3A represents a therapeutic vulnerability in a molecularly defined subset of PPGL and integrative transcriptomic network modeling enables identification of clinically relevant drug dependencies in heterogeneous cancer types.

Personalized cancer vaccines require neoantigen selection strategies that consider not only peptide presentation but also the biological importance of the altered gene within the tumor. We developed Epitope Generation Gateway (EGG), a modular Snakemake workflow for end-to-end neoantigen discovery and prioritization from raw DNA and RNA sequencing data. EGG identifies candidate epitopes arising from somatic single nucleotide variants, indels, gene fusions, and alternative splicing events, while also supporting HLA typing and downstream ranking. What distinguishes EGG from existing pipelines is the integration of patient-specific biological context into neoantigen prioritization. Using RNA sequencing data, EGG constructs personalized gene co-expression networks and combines network centrality with protein interaction information, gene essentiality data, expression, and peptide binding-related features. This allows prioritization of neoantigens derived from genes more likely to be functionally necessary for tumor fitness and less likely to be lost through immune escape. As a case study, we applied EGG to TCGA skin cutaneous melanoma cases and obtained prioritized neoantigen candidates. Current work extends these analyses to a cohort of Finnish melanoma patient data. In parallel, this work connects to development of the In Vitro Immune Systems Resource for Precision Medicine, a translational resource integrating neoantigen analyses from approximately 2,000 cell lines with HLA typing and immune features, including KIR, TCR, and BCR information, supporting neoantigen studies and guide selection of suitable cell combinations for downstream co-culture experiments Overall, EGG provides a reproducible framework for biologically informed cancer vaccine target discovery while supporting the identification of more durable neoantigen candidates.

Signaling pathways sense and propagate information from the extracellular environment to dictate a cell's response, governing a range of essential functions. However, signaling pathway activity is difficult to decipher due to the vast combinatorial space of possible interactions, the nonlinearity of these interactions, and pathway crosstalk.

Current data-driven approaches often lack mechanistic interpretability, while mechanistic models typically fail to scale to genome-wide, multi-context settings. Here, we present scLEMBAS, a mechanistically grounded artificial intelligence framework for modeling signaling pathway activity at single-cell resolution. scLEMBAS enables genome-scale inference of transcription factor (TF) activity from extracellular or intracellular perturbations. To do so, we integrate recently established biologically informed neural network approaches with compositional deep learning techniques from single-cell perturbation modeling. The model captures nonlinear signal propagation of perturbation through a learnable adjacency matrix representing the protein-protein interaction (PPI) signaling network.

We demonstrate that scLEMBAS can accurately predict transcription factor activity in diverse datasets representing a variety of tissue and perturbation contexts. Additionally, we show that the model captures cell subtype specific perturbation responses despite being agnostic to such labels. Beyond predictive accuracy, the mechanistic nature of scLEMBAS provides distinct utility: learned PPI weights provide a meaningful representation of the network beyond topology alone, and the model can “self-prune” by downweighting spurious interactions irrelevant to the biological context. Additionally, the learned parameters enable identification of perturbation-specific subnetworks that capture the core proteins driving cell type-specific responses. Together, these capabilities establish scLEMBAS as a framework for dissecting context-specific signaling mechanisms at single-cell resolution.

Spatial transcriptomics (ST) is an emerging technology that combines sequencing-based molecular profiling with high-resolution histological imaging to enable spatially resolved gene expression analysis within tissues.
The convergence of these two modalities is central to the interpretation of high-resolution datasets such as 10x Genomics Visium HD, where gene expression data are overlaid onto microscopy images of tissue sections to integrate molecular and spatial information.

To address the fragmentation and integration challenges of current spatial transcriptomics analysis workflows, we developed nfdata-omics/spatialomics, a reproducible and scalable pipeline for ST data analysis that unifies molecular and imaging workflows.
The pipeline is implemented in Nextflow [1] following nf-core [2] standards, ensuring modularity, portability across computational infrastructures, and traceability.

Starting from raw sequencing data and slide-associated microscopy images, the pipeline performs quality control, spatial alignment, quantification via Space Ranger, and generation of spatial objects with automated quality assessment.
A central feature of the pipeline is the integration of image segmentation, implemented using Cellpose [3], enabling identification of cellular and tissue structures from histological images. Segmentation outputs are integrated with transcriptomic data by assigning bins to segmented cells, allowing a shift from bin-level to cell-level resolution.
The pipeline also provides an integrated reporting layer that aggregates quality metrics, execution metadata, and analytical summaries across samples, supporting interpretation, reproducibility and scalability.
Here we show the application of this workflow to a Visium HD dataset of metastatic renal cell carcinoma, comparing patient cohorts stratified by their blood serum cholesterol levels, demonstrating its applicability to clinically relevant, heterogeneous biological conditions.

Chronic inflammatory sequelae following viral encephalitis results in potentially lifelong deficits in cognitive function. The drivers for this sequelae originate from alterations in cell states of post-mitotic central nervous system (CNS) cell types that propagate aberrant signaling beyond the course of the infection. We used a novel infection fluorescence tracing mouse model to identify that CNS cells types, including both glial and neuronal classes, display durable shifts in transcriptomic signatures following clearance of infection from individual cells, rare cells we refer to as survivors. Given the unclear role of infectious disease in neurodegeneration, and the challenges of studying rare cell populations, we are developing a computational approach to generalize these findings to a broader scope of neuroinfection and neurodegeneration models for hypothesis generation. This approach will leverage a combination of spatial transcriptomics data and survivor enriched snRNAseq data to build a two-step classifier. Prior to classification, public scRNAseq or snRNAseq will be annotated using the MapMyCells tool from the Allen Institute. For classification, step one will use cell type specific summary signatures from neighborhood analyses to identify cells associated with an inflammatory microenvironment, then the associated cells will be further classified by step two into survivor or neighbor cell states using a supervised machine learning model trained for each CNS cell type using rare state enriched snRNAseq data. We expect improved performance in identifying this rare and important infection relevant cell state through this combination of steps versus an RF model alone.

Background: Hepatocellular carcinoma (HCC) remains a major clinical challenge due to late-stage diagnosis, pronounced tumour heterogeneity, and the absence of reliable pathognomonic biomarkers. To overcome these limitations, we developed an integrated framework to identify and prioritise clinically relevant biomarkers systematically.
Methods: Publicly available RNA-seq datasets were analysed to identify differentially expressed genes associated with HCC, followed by pathway enrichment and protein-protein interaction assessment to determine hub genes. An explainable AI-driven predictive modelling approach was employed to classify tumour samples and prioritise biomarkers with significant patient survival outcomes. Their biomarker potential was validated across cell line models, immunohistochemical profiles, single-cell transcriptomic landscapes, immune infiltration patterns, and genomic alteration profiles. Finally, their expression patterns were contextualised across distinct HCC etiologies, and therapeutic potential was assessed via drug-gene interaction networks.
Results: The XGBoost-based classifier demonstrated outstanding diagnostic performance in independent tissue cohorts (AUROC 0.998) and serum-derived exosomal datasets (AUROC 0.990). Five biomarkers were prioritised, including established mitotic regulators, CDK1, CDC20, and E2F1, which exhibited sustained overexpression in malignant hepatocytes and novel HCC markers, ADRA1D and GLP2R, which showed prominent enrichment within stromal and endothelial compartments and demonstrated tumour-suppressive characteristics through consistent downregulation, respectively. Their expression was pronounced across HCC comorbidities, varying between conditions, thereby reflecting disease heterogeneity and highlighting their potential to target diverse pathological contexts.
Conclusion: Collectively, this framework defines context-specific molecular signatures and establishes a mechanistically grounded foundation for precision diagnostics and individualised HCC management.

Computational methods are essential tools for modern microbiome research. Yet, challenges such as lack of standardization, reproducibility, and transparency often limit reliable analysis and interpretation. Bioconductor addresses these challenges through a global, community-driven network that provides a robust, open-source ecosystem of high-quality tools.

The presentation highlights how collaborative development supports microbiome research. In this domain, a growing ecosystem of packages builds on shared Bioconductor data structures such as TreeSummarizedExperiment, which extends the widely adopted SummarizedExperiment class. By relying on interoperable classes and consistent design principles, tools developed by different groups can work seamlessly together. For users, this reduces time spent on data wrangling, broadens access to well-established statistical methods, and facilitates the application and benchmarking of approaches developed across different fields. The collaborative nature of the ecosystem ensures that methods and workflows continue to advance through contributions and shared expertise from a diverse community of developers and users.

This common foundation enables robust and evidence-based analysis workflows while supporting methodological innovation. It lowers barriers for collaboration between package authors, encourages method reuse, and accelerates cross-disciplinary exchange. As microbiome research progresses toward multi-omics integration, longitudinal studies, and other increasingly complex study designs, interoperable infrastructure supported by an engaged community becomes increasingly critical for scalable and reproducible analyses.

Treatment responses in oncology vary substantially across patients, and identifying actionable biomarkers predictive of differential treatment benefit remains a central challenge in precision medicine. Unlike traditional Machine Learning (ML), causal ML enables the quantification of individual-level outcome shifts attributable to therapy, providing a principled approach to uncover patient characteristics associated with heterogeneous treatment effects. However, systematic frameworks for biologically informed discovery of predictive biomarkers and clinically relevant subgroups remain limited. Here, we present a robust computational framework based on causal ML for predictive biomarker discovery and subgroup refinement in randomized controlled trials (RCTs). Using state-of-the-art causal ML methods, individual treatment effects are estimated to identify clinical and molecular features associated with differential treatment benefit. Biostatistical analyses support biological interpretation of molecular patterns associated with heterogeneous treatment effects. The framework combines two complementary strategies: (1) bottom-up data-driven discovery of predictive biomarkers, and (2) top-down evaluation and refinement of domain-informed hypotheses and clinical guidelines. We demonstrate our approach using the randomized phase III trial FIRE-3 (AIO KRK-0306) in metastatic colorectal cancer (mCRC), integrating genomic, transcriptomic, and clinical data to identify biomarkers of differential sensitivity or resistance to anti-EGFR (cetuximab) versus anti-VEGF (bevacizumab) therapy. This analysis enables systematic evaluation of clinically established subgroups in mCRC, including RAS mutation status and primary tumor sidedness, while identifying additional candidate biomarkers that refine patient stratification. Overall, this work demonstrates how causal ML enables biologically informed predictive biomarker discovery and improved patient stratification, thereby providing a generalizable approach to characterize heterogeneous treatment effects in precision oncology.

Systems of intracellular biochemical reactions are highly complex, usually involving parts that cannot be directly measured. Representing these systems as networks, with nodes for biochemical species and edges, their reactions help to quantitatively characterize their function and the effects of dysregulation. Causal discovery methods can uncover interactions within these networks from observational data, detecting hidden effects from partial observations.
We benchmark state-of-the-art temporal causal discovery methods on time series data from simulations of biochemical kinetics models. Our results demonstrate good performance on toy models for this task, particularly when data is sampled in a way that is consistent with the timescales of the system. By omitting data, we consider the problem of reconstructing these networks in the presence of latent confounders and unobserved species participating in reactions. Causal discovery indicates time-uncorrelated confounders with bidirected edges and unobserved species through time-delayed edges, locating hidden effects, and estimating their typical timescales. Finally, we extend these benchmarks to the reconstruction of a model of the epidermal growth factor receptor signalling network, a well-studied system frequently dysregulated in cancer.
Altogether, our work showcases the feasibility and usefulness of causal discovery methods as part of the data-driven mathematical modelling pipeline for systems of biochemical reactions.

Genomic profiling is unavailable for most patients in routine clinical practice, severely limiting the use of molecular data for patient stratification. However, longitudinal electronic health records (EHR) offer an untapped resource. By effectively integrating this rich clinical information, healthcare providers can deliver highly personalized medical care, make precise diagnoses, and offer optimal treatments based on the latest medical knowledge. To bridge the gap between widely available clinical histories and scarce molecular profiles, we introduce EHR-to-omics latent imputation. This approach learns a direct mapping from coded longitudinal EHR sequences to a low-dimensional omics latent state. We propose a novel architecture, Marker-Aware Dual-Stream Encoding (MADE), which explicitly isolates biomarker tokens via a dedicated pooling stream, demonstrating that biomarker-token subspace selection is critical for robust cross-modal prediction.
To benchmark our approach, we evaluated eight models across five independent splits, utilizing 32-dimensional PCA omics embeddings from 1,097 TCGA-BRCA tumour samples paired with synthetic longitudinal EHR sequences. On our synthetic TCGA-BRCA benchmark, oracle marker subspace selection recovers a 5.7% mean absolute error improvement over standard full-vocabulary neural encoders. Because clinical EHR vocabularies lack predefined biomarker boundaries, we developed a blind, three-stage data-driven discovery pipeline. This approach recovers 56% of the oracle cosine gain without requiring token-level labels. Importantly, the imputed latent vectors retain meaningful downstream biological utility, preserving 43% of PAM50 breast cancer subtype classification accuracy and 89.5% of the overall survival concordance index relative to omics-supervised baselines. These results establish a robust, generalizable framework for inferring molecular states directly from routine clinical data.

Coronary artery disease (CAD) involves the formation of atherosclerotic plaques in the lumen of coronary arteries. This narrowing effect (called stenosis if > 50% occlusion, sclerosis < 50%) can lead to acute cardiac events. In this study, we combined microbiome and metabolomics data to identify biomarkers and molecular mechanisms involved in CAD. A total of 48 plasma samples from patients with stenosis (N = 23), sclerosis (N=12) and disease controls (N=13) were analyzed. More than 1000 metabolites and lipids were quantified using MxP Quant 1000 kit from biocrates life sciences gmbh. Microbiome data was obtained by mapping microbial reads using an in-house built pipeline to obtain microbe abundances. A Multi-Omics Factor Analysis (MOFA) was performed to find multi-omic factors associated with CAD. Then, the most relevant metabolomic and microbiome features were integrated using Group Aggregation via UMAP Data Integration (GAUDI) method. Finally, regularized Canonical Correlation (rCC) was deployed to evaluate the associations between single metabolites and microbiome genera. Our analysis found three multi-omics clusters: one dominated by stenosis, one dominated by disease controls and one including patients from all the groups. The multi-omic interaction network created showed two modules. The first one showed negative and positive interactions of different triglycerides (TG) with Capnocytophaga and Cryptococcus, respectively, whereas the second displayed negative associations between phospholipids and Corynebacterium, and positive ones with Citrobacter and Clostridium, among others. This study suggests potential biomarkers and mechanisms involved in the progression from subclinical sclerosis to coronary stenosis.

While immunotherapies show notable clinical success, predicting durable patient outcomes remains challenging, and biomarkers like PD-L1, MSI/dMMR, and tumor mutational burden show strong clinical limitations. At the same time, costs and efforts associated with single-cell analyses are still considerably high. Here, we build on the previous work that used bulk transcriptome deconvolution approaches to predict immunotherapy response. However, we implement more granular deconvolution and utilize scRNA-seq references to resolve rare cell subtypes and more accurately navigate intratumoral heterogeneity. We used this data-driven framework to analyze bulk RNA sequencing samples from patients treated with the immune-checkpoint therapy (n=1,523) across 15 studies and 6 cancer types. For this, we leveraged high-resolution references from 715 scRNA-seq samples collected across 18 studies. To study synergistic effects between cell populations, we employed random forest machine learning models. We observed that, individually, fractions of cytotoxic T-cells, fibroblasts, and tumor-associated macrophages (TAMs) had major roles in dictating the therapeutic response. Crucially, this approach also highlighted critical interdependencies: In a melanoma cohort treated with anti-PD-1 therapy, the presence of cytotoxic T-cells alone was associated with a response rate of 73%. However, the combined presence of TIMP1+ TAMs and cytotoxic T-cells significantly increased the response to 95%. Previous work showed that TIMP-1 can activate MHC-I expression in myeloid cells and drive CD8+ T-cell infiltration in melanoma. Thus, evaluating the synergistic co-occurrences of specific TME cells and T-cells will be able to predict immunotherapy response with far greater precision than isolated cell abundances alone.

Inherited metabolic disorders (IMDs) affect approximately 1 in 800 newborns and present with diverse, often non-specific symptoms that frequently delay diagnosis and compromise patient quality of life. Addressing this diagnostic challenge is a central goal of the Horizon Europe project Recon4IMD, which applies personalized genome-scale metabolic modeling to improve patient stratification and clinical decision-making. The utility of such models, however, is currently limited by incomplete coverage of both lipid metabolism and IMD-associated proteins.
To address these gaps, we generated a comprehensive network of theoretically possible lipid reactions by combining pathway knowledge from Rhea with structural data from SwissLipids and carried out targeted curation covering all human enzymes and transporters in UniProtKB/Swiss-Prot linked to IMDs. These enhanced datasets are integrated into the ReconX Knowledge Graph (ReconXKG) - a next-generation human metabolic network that builds on prior reconstructions such as Recon3 and Human1 while drawing on resources including UniProt, Rhea, the Virtual Metabolic Human, and MetaNetX.
By improving both the completeness and predictive capacity of metabolic models, this integrated strategy supports more accurate diagnosis and paves the way toward personalized treatment for IMD patients.

Background: The microtubule-associated protein TAU (MAPT) plays a central role in the pathogenesis of Alzheimer's disease (AD). While the loss of Tau protein has been linked to neuronal resistance against amyloid-beta (Aβ) toxicity, the coordinated molecular landscape underlying this resilience remains poorly understood.
Methods: In this study, we employed an integrative multi-omics approach, combining re-analyzed transcriptomic data with newly processed proteomic layers from human iPSC-derived neurons. We utilized a rigorous bioinformatics pipeline, including DESeq2 for transcriptomics and Limma for proteomics, followed by cross-layer integration to identify convergent molecular signatures. Transcription factor (TF) enrichment analysis was performed to identify upstream regulators of the resilient phenotype.
Results: Our analysis revealed a striking ""transcriptomic stability"" in MAPT-KO neurons; while wild-type (WT) neurons exhibited widespread dysregulation of synaptic and apoptotic genes upon Aβ exposure, KO neurons maintained a homeostatic state nearly identical to controls. Integration of the two omics layers identified 27 core consensus hits, including CISD2, ERBIN, and ELOVL4, which were consistently preserved or upregulated in resistant cells. Functional enrichment highlighted a significant enhancement in ER-proteostasis and lipid metabolism. Furthermore, TF analysis identified KLF1 and EGR1 as potential master regulators orchestrating this protective molecular network.
Conclusion: These findings demonstrate that Tau deletion induces a proactive molecular reprogramming rather than a passive resistance. By identifying a coordinated multi-omics signature of neuroresilience, this study provides novel therapeutic targets that could be leveraged to mimic the protective effects of Tau reduction in the treatment of AD and other tauopathies.

Acute myeloid leukemia (AML) treatment usually requires toxic chemotherapy. Differentiation therapy is a compelling alternative, as seen in APML with all-trans retinoic acid (ATRA). Extending this paradigm to other AML subtypes requires improved mechanistic understanding of regulatory networks governing cell fate decisions. We therefore applied our cSTAR approach (cell State Transition Assessment and Regulation) that combines machine learning with mechanistic modelling to identify novel differentiation strategies.


Six single-cell RNA-seq datasets from human bone marrow and blood (>120,000 cells) were integrated, generating a high-resolution map of haematopoiesis. We mapped discrete states spanning from haematopoietic stem cells to mature myeloid populations. Using cSTAR, we constructed multiple State Transition Vectors (STVs) describing differentiation trajectories, and Dynamic Phenotypic Descriptors (DPDs) quantified progression along defined paths. Validation across independent human, mouse and AML datasets further confirmed robust cell state assignment and trajectory reconstruction.



To infer causal signaling networks governing AML state transitions, we used perturbation transcriptomic data from the LINCS database, applying Bayesian Modular Response Analysis (BMRA) to identify key regulators predicted to drive myeloid differentiation, including ATRA, mocetinostat (HDAC inhibitor), and Ro-3306 (CDK2 inhibitor). AML cells were treated with these agents as monotherapies and combinations in vitro. Immunophenotyping showed that expression of myeloid differentiation markers was most pronounced with combination treatments, in line with cSTAR predictions. Transcriptional analysis revealed that combinations both amplified ATRA-induced expression, and triggered additional alternative differentiation programmes.


In summary, we provide a quantitative reconstruction of human haematopoietic differentiation and show that cSTAR can predict strategies to overcome AML differentiation arrest.

The adverse outcome pathway (AOP) framework was developed to organize knowledge about the sequence of causally connected key events (KEs) that link molecular-level perturbations that can arise from chemical exposure to a higher-level adverse outcome (AO). As the number of suspected synthetic contributors to human disease grows, AOPs provide a structured approach to elucidate the underlying mechanisms. However, common AOP development methods depend on manual curation, which is labor intensive and time consuming. As such, there are relatively few (< 600) AOPs already developed (compared to 350,000 chemicals in commerce). To help speed up AOP development, we created a workflow for identification of putative AOPs, focused on Parkinson disease as an AO and using insecticides as prototypical stressors. Our method bridges diverse databases (e.g., The Human Proteome Atlas, Bgee) to integrate and structure these data into cause-effect chains that predict AOPs. Insecticide-gene interactions extracted from the Comparative Toxicogenomics Database represent the first KE (molecular initiating event), while Gene Ontology, Reactome, and Monarch terms were statistically linked as the KEs. This pipeline yielded over 20,000 candidate AOPs, ranked by weight of evidence to prioritize toxicity mechanisms for further study. We also integrated known genetic variants implicated in Parkinson disease from DISGENET to inform possible gene-environment interaction. Our findings highlighted a potential role for insecticides in the neuronal apoptotic process and the canonical Wnt/β-catenin pathway via the PRKN gene (a crucial gene in Parkinson disease). The proposed computational pipeline can be extended to other chemicals and adverse outcomes of interest.

The identification of circulating predictive biomarkers for treatment response in non-small cell lung cancer (NSCLC) is critical to optimize patient selection and improve therapeutic outcomes. However, while plasma-based biomarkers offer a minimally invasive and accessible approach, low abundance of disease-relevant analytes remains a major analytical and computational challenge to detect subtle but biologically relevant changes associated with therapy response.
We used plasma multi-omics, integrating targeted metabolomic profiling using the biocrates MxP Quant 1000 assay, unbiased mass spectrometry proteomics (TrueDiscovery P2 DIA-MS), and the NULISAseq inflammation panel to analyze samples from patients with NSCLC enrolled in the multicenter phase II clinical trial SAKK 17/18. Parallel profiling quantified approximately 6,000 plasma proteins, 250 inflammation-related markers, and more than 1,100 metabolites and lipids, enabling cross-platform integration of proteomic, cytokine, and metabolic signatures.
To identify predictive biomarkers of therapeutic response, we implemented and compared machine learning frameworks that integrate metabolic, proteomic, and inflammatory features. Multi-omics models outperformed single-omic approaches in stratifying responders and non responders based on samples obtained before therapy start, highlighting the complementary information captured by each dataset. Feature-level integration further enabled the identification of coordinated molecular signatures associated with treatment response. The best results were obtained with Multi-Omics Factor Analysis (MOFA)-based biomarker identification.
Overall, this study demonstrates the value of integrated plasma multi-omics data and computational modeling for predictive biomarker discovery in NSCLC. The presented workflow is scalable and clinically applicable, providing a foundation for response monitoring and systems-level characterization of treatment effects in oncology.

Understanding the interaction between genes and how regulation occurs between them is a crucial task to understand the complexity of developmental biology and complex diseases. Therefore, there is a great need to be able to model gene regulatory networks from the data we have today. To understand causality, and not just correlation, we need methods that can extract information from time series and be able to make predictions. For this, we developed a transformer-encoder-based model that, when properly trained, can predict the trajectory of the time series of knock-out experiments. Additionally, we can extract from it the underlying gene regulatory network of the system. The method has been tested with network models from the BEELINE networks and is competitive in accuracy with algorithms such as dynGENIE3 and BINGO. As we leverage the use of the transformer architecture and high computing power, our model runs in considerably less time than both other methods.

Out of the numerous existing trajectory inference methods, only a few provide unsupervised cell lineage detection and scale to large single-cell flow or mass cytometry datasets, derived from a static bone marrow snapshot. We extended our previous pseudotime computation method, TimeFlow, with an unsupervised lineage detection strategy for cytometry data. TimeFlow 2 divides the pseudotime axis into several segments and computes cell clusters within each segment. In addition, it connects clusters between consecutive segments to track differentiation pathways at a coarse cell cluster level. By using an optimal transport-based cost function, TimeFlow 2 groups these paths based on their similarity and identifies lineages from stem cells to different mature populations without explicitly hard-coding the number of lineages in the sample. We used our method to delineate cell trajectories of major cell populations such as neutrophils, monocytes, erythrocytes and B-cells using flow cytometry datasets from 5,000 to 600,000 cells. Then, we demonstrated its potential in challenging mass cytometry datasets with both major and rare cell populations and compared it to other established trajectory inference methods. Furthermore, we modeled the evolution of cytometry markers along pseudotime and examined differences in monocytic differentiation patterns between healthy individuals and patients with Acute Myeloid Leukemia or Myelodysplastic Syndromes. We found TimeFlow 2 to be modular allowing for general clustering methods such as Gaussian Mixture Models or cytometry specialized methods such as FlowSOM.

Clinical laboratory tests are essential tools to monitor overall health and can reveal early signs of serious health conditions. While considerable work has been done on applying artificial intelligence to electronic health record (EHR) data for disease risk prediction, the prediction of abnormal laboratory measurements remains underexplored. To address this gap, we developed models based on EHR data to predict future abnormal measurements for 4 clinically actionable clinical laboratory tests: HbA1C, TSH, eGFR, and LDL; and are prospectively validating the predictions with a recall study where we invite individuals for follow-up testing.
We trained XGBoost models, as well as transformer encoder-based models capturing the longitudinal trends of the lab measurements, to predict abnormal values separately for each test type. We included individuals aged 30-70 as part of the FinnGen study (N=310,526) without a history of the relevant diseases. Predictors included individuals' age and sex, BMI, education level, prior laboratory measurements, diagnoses, and medication purchase information.
The final XGBoost models reached AUCs ranging from 0.78 (TSH) to 0.91 (eGFR) and average precisions from 0.14 (TSH) to 0.30 (LDL). Prior laboratory measurements were the strongest predictors, substantially improving the models compared to baselines with age, sex, and BMI. While for eGFR, the transformer models provided significant additional improvements.
Our work demonstrates that both tabular- and longitudinal-based methods can predict future abnormal lab measurements using EHR data, highlighting their potential use in targeted screening, early detection, and preventive interventions.

Accurate molecular stratification of cancer is complicated due to substantial tumor heterogeneity. Current approaches often rely on single modalities or complex integration frameworks that are difficult to scale and interpret (Rudin et al, 2019), while systematic integration of multimodal data with longitudinal clinical outcomes remains challenging (Tran et al, 2025). Here, we report our work-in-progress towards utilizing multimodal machine learning to define novel tumor archetypes, which could provide insights into tumor characteristics relevant to patient outcomes and treatment, exhibiting the potential to guide personalized medicine.

We leverage deep molecular tumor profiling data along with comprehensive longitudinal clinical and health registry data available in the iCAN Digital Precision Cancer Medicine Flagship. Presently, our study includes 3,359 samples spanning 15 tumor types, for which exome and RNA sequencing have been performed. We derive somatic mutation representations with MuAt (Sanjaya et al, 2023) and integrate them with RNA-sequencing data and copy number aberrations using a simple and scalable framework based on principal component extraction.

We then define tumor archetypes based on clustering of this multimodal molecular view and evaluate their association with survival outcomes and treatment information. Our results indicate that combining exome and RNA sequencing provides a synergistic effect in defining tumor archetypes, enabling improved molecular stratification beyond tumor type alone, and offering a scalable approach for large clinical cohorts.

AI is transforming biomedical research by enabling analysis of complex, high-dimensional datasets for precision medicine applications. However, leveraging diverse omics data sources for predictive modelling remains challenging due to heterogeneity, and small cohort sizes. In this work, we designed AI analysis pipelines to predict Rheumatoid Arthritis (RA) flare risk upon therapy tapering/discontinuation and support treatment decisions.
Data were collected from 100 RA patients in sustained clinical and imaging remission at Fondazione Policlinico Universitario A. Gemelli IRCCS. Synovial tissue samples were profiled using multiple modalities, including histology (n=100), flow cytometry (n=53), spatial transcriptomics (n=21), and single-cell RNA sequencing (n=23). Flare occurrence within 6 months was used as prediction endpoint.
A unified pipeline was used across all datasets, incorporating feature selection (recursive feature elimination), hyperparameter optimisation (Bayesian search) and Logistic Regression, Random Forest and XGBoost as classifiers, with nested cross-validation used for pipeline training and model selection. In modalities generating multiple instances per patient these were treated as independent observations but ensuring that all rows of a patient were either in training or test sets by performing cross-validation at patient level.
Across datasets, we evaluated multiple models and sampling strategies, observing trade-offs between metrics. Performance variations highlighted influence of both data sources and algorithmic choices Explainable AI (SHAP) was used to identify biomarkers' role in models and further support decision making. In conclusion, among all models tested, cross validation AUC ranges between 0.66 to 1 across different datasets. This approach is currently being evaluated in an ongoing clinical trial (ClinicalTrials.gov ID: NCT05952440).

The endometrium exhibits substantial cyclical remodelling in response to hormonal fluctuations, especially during the inflammatory decidual reaction at the onset of the midluteal phase, which defines the time-sensitive window of implantation. Although the decidual reaction is associated with significant gene expression changes, the pathways driving this propagation and their histological correlates remain insufficiently characterised.

We analysed 320 biopsies from 162 patients collected during the luteal phase (2-13 days post-ovulation). Patients were 35 ± 4 years old with 0-4 previous miscarriages on average. Pathway activity across 50 hallmark pathways was quantified using gene set variation analysis on bulk RNA sequencing data. Distinct temporal trajectories emerged, with sequential pathway activation and suppression, including progressive activation of inflammatory pathways and decreased fatty acid metabolism and cholesterol homeostasis over time. Mixed-effects modelling revealed significant differences between miscarriage and non-miscarriage patients, with an inflammatory cluster enriched in miscarriage and driven by gland-specific genes. Age also significantly influenced 18 pathways. A heterogeneous graph neural network, designed to capture endometrial architecture, was employed to predict pathway scores directly from 265 Haematoxylin and CD56-stained whole-slide images (WSIs) using 5-fold cross-validation. Ten pathways achieved Pearson correlations > 0.5 with ground truth bulk RNA-seq pathways. Top pathways included inflammatory related pathways and allograft rejection and reached correlations up to 0.62.
This is, to our knowledge, the first study to infer pathway-level transcriptomic activity directly from endometrial histology, providing a scalable framework for linking tissue architecture to molecular function and enabling image-based characterisation of reproductive health.