We aimed to identify progression-relevant biomarkers in Parkinson's disease (PD) using an integrative machine learning framework applied to multi-omic cerebrospinal fluid (CSF) and plasma data. PD is a progressive neurodegenerative disorder lacking robust molecular biomarkers for early diagnosis and longitudinal monitoring. Computational biology approaches that integrate heterogeneous omics layers across biofluids provide a strategy to capture disease signals and improve biomarker discovery. Untargeted proteomic and metabolomic CSF and plasma data were obtained from the Parkinson's Progression Markers Initiative (PPMI). We implemented a comparative machine learning pipeline incorporating regularized regression, random forest, support vector machines, XGBoost, and neural networks to benchmark disease classification performance across omics layers, biofluids, and integrated settings. Feature importance and interaction-aware statistical modeling were used to prioritize biomarkers with consistent predictive value across cohorts and biofluids. Metabolomic profiles were further interrogated using pairwise statistical testing and longitudinal slope estimation to characterize non-linear disease trajectories. We identified 116 proteins significantly associated with PD, of which 27 were consistently prioritized by machine learning models. Regularized multinomial regression achieved the highest classification accuracy (86% in CSF and 79% in plasma), followed by neural networks. Interaction modeling confirmed significant cohort- and biofluid-dependent effects for prioritized proteins (P < 0.05). Metabolomic analyses identified 124 plasma and 71 CSF metabolites with significant disease-stage differences. Longitudinal analyses revealed non-linear molecular trajectories, with prodromal individuals exhibiting distinct metabolic inflection patterns. Overall, integrative machine learning of multi-fluid, multi-omic data identifies biomarkers that are discriminative and progression-relevant, underscoring the value of computational integration for PD biomarker discovery.

Accurate Enzyme Commission (EC) number (i.e., a hierarchical numerical classification of enzyme-catalyzed reactions) assignment is essential for protein function annotation and downstream applications, including metabolic engineering and genome-scale metabolic modeling. EC prediction remains challenging due to multi-functional enzymes, heterogeneous annotation quality, and limited coverage for understudied enzyme families. Moreover, it remains unclear how robust current EC prediction approaches are when applied beyond standard benchmarks, and to what extent they are suitable for metabolic modeling applications.

Here, we present a systematic benchmark of widely used EC prediction methods, including a new lightweight machine learning pipeline combining enzyme detection and sequence embedding-based classification. Rather than focusing solely on peak accuracy, we evaluate robustness and generalization across multiple evaluation settings, including homology-controlled splits and performance stratified by EC class frequency. We further assess model behavior at different levels of the EC hierarchy to better understand how different approaches learn functional information from protein sequences.

Our results highlight substantial differences between EC predictors in terms of robustness, sensitivity to sequence similarity, and coverage of underrepresented enzyme classes. These factors are critical for reliable metabolic reconstruction. While our pipeline achieves competitive performance with existing tools, the benchmark reveals that high overall accuracy does not necessarily translate into suitability for downstream metabolic modeling.

Together, this work provides a practical framework for evaluating EC prediction methods beyond conventional metrics and offers guidance for selecting and applying EC annotation tools in metabolic modeling and related applications.

Large Language Models (LLMs) are transforming biomedical research, yet their integration into reproducible bioinformatics workflows remains challenging. We present {llumen}, a novel pure-R package developed by the Oncology Data Science function at Merck, designed to empower bioinformaticians to build, deploy, and manage agentic AI workflows directly within the R ecosystem.

{llumen} transcends simple chat interfaces by providing a robust framework for autonomous agents capable of complex reasoning strategies running seamlessly on diverse infrastructure, ranging from researchers' local laptops to high-performance Linux compute servers. These agents utilize a library of 30+ tools to programmatically access omics knowledge bases (e.g., PubMed, PubChem, OpenTargets), analyze large-scale biomarker and patient data via SQL and graph queries, and process multi-modal documents and histopathology images.

Key differentiators of {llumen} are its ability to facilitate hybrid model orchestration, allowing researchers to execute secure local foundation models in conjunction with commercial frontier models, as well as its focus on enterprise-grade security and data privacy, essential for pharmaceutical R&D. The framework supports fully local execution using quantized models and integrates GDPR-compliant PII guardrails, ensuring sensitive patient data remains protected. Furthermore, it offers rigorous validation features, including agent-as-judge benchmarks and reasoning traces, to ensure scientific reliability.

By bridging the gap between state-of-the-art generative AI and traditional bioinformatics infrastructure, {llumen} enables researchers to automate tasks ranging from systematic literature reviews and statistical analysis plan drafting to hypothesis generation and molecular property prediction. Here, we demonstrate {llumen}'s utility through real-world case studies in oncology target discovery and automated machine learning.

Although large-scale interactome maps serve as an important framework for studying biological processes, they provide only limited insight into how interaction networks are reconfigured under different biological conditions. Characterizing such differential interaction behavior is critical for understanding disease mechanisms and their clinical relevance.
In this work, we present GOLIATH, a supervised deep-learning framework for inferring context-dependent alterations in protein interaction networks from paired molecular data. Applying GOLIATH across multiple experimental settings, including human cell lines, mouse tissues, and four TCGA cancer cohorts, we show that interaction-level
changes capture coherent and biologically meaningful signatures of cellular states. Predicted interaction perturbations are significantly enriched for known disease-associated pathways, even when considering a small subset of the interaction space, indicating that GOLIATH prioritizes highly informative network alterations. We further demonstrate that interaction-based representations improve prediction of clinically relevant outcomes, including patient mortality and tumor stage, compared to gene-level features and existing interaction-based baselines. Collectively,
our findings underscore the importance of studying disease through interaction-level network rewiring rather than focusing solely on individual genes. GOLIATH offers a scalable framework for identifying context-specific network alterations and connecting them to underlying biological mechanisms and clinical outcomes.
Our code and data are available at https://github.com/Kuper994/GOLIATH.

Coronary artery disease (CAD) and chronic kidney disease (CKD) are in part genetically determined and are associated with various omics layers. Methods for integrating genomics data with omics profiles remain to be standardised. This study evaluates biological data transformations to optimise the integration of genomics with other omics for comorbidity prediction in people with HIV (PWH). We trained linear and deep learning models in single-omic and multi-omic settings on two cohorts of PWH with genotype plus another omics data available. We evaluated two multi-omic integration strategies—feature concatenation and encoder-based architectures. 436 CAD and 166 CKD cases were evenly split across training, validation, and test sets using fixed patient splits. Performance was estimated via five-fold cross-validation, and we report mean accuracy with standard errors. Genotype data was represented in four ways: (i) raw SNP genotype matrices; (ii) principal component (PCA) embeddings; (iii) polygenic risk scores (PRS); and (iv) AlphaGenome-derived gene-level impact scores. Each genotype representation was compared individually and when integrated in a multi-omics model. The results demonstrate that biologically informed genomic transformations improve prediction in multi-omics models. In both classification tasks, integrating raw SNPs (CAD accuracy=0.55±0.03; CKD accuracy=0.63±0.01) or genotype PCs (CAD accuracy =0.54±0.03; CKD accuracy =0.62±0.03) with other omics reduced performance relative to the best corresponding single-omics models. By contrast, PRS (CAD accuracy=0.61±0.03; CKD accuracy=0.65±0.02) and AlphaGenome (CAD accuracy=0.57±0.03; CKD accuracy=0.67±0.02) improved accuracy. As multi-omics analyses become more prominent, methods that integrate genomics effectively without requiring large cohorts will become increasingly valuable; here, we highlight two such approaches.

Background:
Microbial interactions with therapeutic compounds are central to pharmacomicrobiomics and
precision medicine. Although MASI provides curated microbe–substance and microbe–disease
associations, computational modeling of these heterogeneous relationships under realistic
generalization settings remains limited.
Methods:
We constructed a heterogeneous graph integrating microbe–substance and microbe–disease
associations from MASI, enriched with microbial taxonomic annotations, probiotic/abundance
indicators, and substance category features. We employed a Heterogeneous Graph Transformer
(HGT) to model multi-relational interactions across microbial taxa, substances, and diseases
and to learn relation-aware node embeddings. To reduce optimistic bias, we adopted a leave-
microbe-out evaluation strategy, ensuring that all interactions of held-out microbial taxa were
excluded during training. Additionally, we applied degree-biased hard negative sampling to
simulate realistic link prediction scenarios. Model robustness was evaluated across five random
seeds.
Results:
Under leave-microbe-out evaluation with hard negatives (1:1 ratio), the model achieved ROC-
AUC = 0.9026 ± 0.0240 and PR-AUC = 0.8692 ± 0.0321, demonstrating strong generalization
to previously unseen microbial taxa. The model identified several high-confidence novel
candidates, including predicted associations between Roseburia hominis and anticholinergic
agents such as dicyclomine hydrochloride, as well as between Streptococcus thermophilus and
dicyclomine. While direct experimental validation is limited, existing studies suggest that
antispasmodic and anticholinergic compounds can modulate gut microbial composition,
supporting the biological plausibility of these predictions.
Conclusion:
This heterogeneous graph-based framework enables robust and unbiased discovery of candidate
microbe–drug interactions and provides a scalable computational foundation for hypothesis
generation in pharmacomicrobiomics.

Bone mineral density (BMD) is a measure of bone strength and a determinant of fracture risk, and measuring BMD is recommended for women approaching menopause due to estradiol deficiency and the higher risk of low bone mass and osteoporosis, which are linked to increased risk for cardiovascular and other age-related diseases. Many mathematical models have been developed in an effort to predict bone behavior and to simulate the effect of drug treatments.
However, these models are provided with only a single set of parameter values, thus neglecting between-patient variability. In this paper, we demonstrate how cross-sectional data on BMD from UK Biobank can be used to re-estimate the distribution of model parameters in a mechanistic model of bone remodelling for two different patient groups, namely post-menopausal women who are on hormone replacement therapy (HRT) and post-menopausal women who are not. Differences in the parameter distributions between the two groups result in differences in treatment outcome when simulating administration of a sclerostin antibody. The results indicate that the treatment is more efficient in women with HRT since virtual patients generated from the parameter distribution based on HRT tend to have a higher delay in returning to the baseline BMD level after the end of drug administration.
This work demonstrates how cross-sectional data, e.g. from biobanks, can be used to enrich existing mechanistic dynamic models to enable group-specific predictions.

Motivation
The rapid advancement of technology has led to an exponential increase in the number of bioinformatics tools available for data analysis and biological discovery. However, keeping up with newly developed tools has become a major challenge, particularly for researchers without programming expertise. Tasks such as installing dependencies and executing analysis workflows often require proficiency in Unix command-line operations, resulting in a steep learning curve. Although graphical workflow platforms help lower these barriers, researchers must still learn how to operate these systems and understand their configuration and execution procedures. Furthermore, incorporating newly published tools into curated workflow ecosystems can be slow, delaying the adoption of the latest computational methods. To address these challenges, we developed AIBioAgentTools Repository, a repository of bioinformatics Model Context Protocol (MCP) servers that dynamically connect large language models with bioinformatics tools, enabling efficient execution of analyses on local computers.

Results
Using natural language commands, we demonstrate end-to-end automated workflow execution for Hi-C analysis as a case study, lowering the barrier to entry for complex multi-step analyses. AIBioAgentTools Repository is designed to support a diverse range of command-line tools across genomics, transcriptomics, and other omics disciplines, enabling scalable, reproducible, and AI-assisted bioinformatics analyses.

Synthetic data (SD) is a valuable resource in the life sciences, offering new opportunities to address challenges including limited data availability, privacy constraints and restricted access to sensitive datasets. By generating artificial data, SD enables researchers to develop, test and benchmark computational methods in controlled and reproducible settings. Despite these advantages, the adoption of SD depends on reliable evaluation frameworks capable of assessing the fidelity, utility and trustworthiness of generated datasets. However, evaluation practices remain fragmented and inconsistently applied across different domains.

To investigate the current landscape of SD evaluation, the ELIXIR Machine Learning Focus Group conducted a systematic review of the scientific literature following the PRISMA guidelines. The review examined evaluation strategies across six major life science domains, aiming to identify commonly used metrics, methodological trends and existing gaps in assessment practices. Our analysis reveals that although methods for generating SD are rapidly advancing, systematic and standardized evaluation approaches are often lacking. In many cases, evaluation relies on ad hoc metrics or domain-specific practices, making it difficult to compare results across studies and limiting confidence in the quality and applicability of SD.

These findings highlight the urgent need for more robust and standardized evaluation methodologies to support the responsible use of SD in life sciences research. Establishing clearer evaluation frameworks will not only improve reproducibility and comparability across studies but also strengthen trust in SD-driven approaches. This review outlines key directions for developing evaluation standards that can support the broader integration of SD in scientific discovery and biomedical applications.

Current methods for integrating single-cell datasets often struggle when cell types are unevenly distributed or missing across batches. To address this challenge, we developed Coralysis, a multi-level integration algorithm implemented in R/Bioconductor that integrates imbalanced single-cell datasets by progressively identifying cellular identities through multiple rounds of divisive clustering. Coralysis performs consistently well across a wide range of single-cell transcriptomic integration tasks. It outperforms state-of-the-art methods in imbalanced integration scenarios, particularly when datasets do not share similar cell types, thanks to its intrinsic feature extraction procedure based on L1-regularized logistic regression. Beyond transcriptomics, Coralysis can integrate heterogeneous single-cell proteomics datasets, including CyTOF and CITE-seq, enabling the identification of rare populations such as basophils (0.5%). Coralysis can also use trained models to predict cellular identities in new single-cell datasets through reference mapping, allowing the detection of rare cell populations previously missed in peripheral blood mononuclear cells, such as CD16+ monocytes and natural killer cells. Finally, Coralysis provides cell-specific probability scores that help identify dynamic cellular states along with their associated differential expression programs. Overall, Coralysis facilitates the study of subtle biological variation by improving the integration of imbalanced cell types and states, providing a more faithful representation of the cellular landscape in complex single-cell experiments. Coralysis is available as an R/Bioconductor package at: https://bioconductor.org/packages/release/bioc/html/Coralysis.html.

In systems biology, complex mechanistic models are widely used to understand biological processes. Recently, multi-objective optimization has provided a useful framework for studying whether multiple biological objectives can be optimized simultaneously or whether improving one objective necessarily degrades another, i.e., whether an intrinsic trade-off exists. This question is characterized by the Pareto frontier. In this context, Pareto analysis can help biologists identify shared mechanisms across species by revealing which objectives can be jointly optimized and which cannot. However, computing the Pareto frontier is often computationally expensive. Existing approaches typically rely on brute-force exploration of the feasible region, which is inefficient and may still fail to recover the frontier under limited sampling budgets and the curse of high dimensionality. To address this challenge, we propose a physics-guided learning-based method for efficient and accurate Pareto frontier approximation from limited data. Rather than exhaustively searching the feasible space, our method directly optimizes inputs toward the frontier boundary, substantially reducing computational cost while improving frontier coverage. We apply our method to data-driven analysis of BMP signaling. The resulting Pareto structures are consistent with established biological findings, demonstrating the practical utility of our approach for biological process understanding.

Motivation: Cytometry by time of flight (CyTOF) is increasingly used to interrogate signalling in single cells, yet appropriate analytical tools remain lacking. Unlike immune profiling data, which is bimodal and amenable to gating, signalling data is quantitative and requires distinct approaches. Existing conditional variational autoencoders, while successful for immune profiling and single-cell RNA sequencing integration, lack key features for signalling analysis, including the representation of interpretable signalling modules and correction for unwanted covariance introduced by cell size.
Results: Here we describe a linear conditional variational autoencoder (linear CVAE) tailored for CyTOF signalling data. The model identifies and quantifies engagement of signalling modules, where latent variables reflect per-cell module usage and decoder weights represent module loadings. It efficiently corrects for variation induced by cell size and staining without requiring explicit markers, and performs batch integration, enabling robust, interpretable analysis of single-cell signalling data.
Availability: https://github.com/molsysbio/LinearCytoVAE

The exponential growth of data in biomedicine has created an urgent need for intuitive visualization tools. These tools must be able to effectively represent complex biological networks and remain accessible to domain experts without extensive computational training. Current network visualization approaches often require specialized programming skills and/or cannot handle the scale and complexity of modern biomedical datasets, which creates significant barriers to biological discovery. We develop RingNet, a web-based interactive visualization tool that integrates computational efficiency with flexible, user-driven exploration. The tool meets the community's need to visualize multi-modal datasets within a single, compact network representation. RingNet uses an R backend for network computation and coordinate optimization. This generates JSON data structures that feed into a JavaScript and HTML frontend, which provides real-time, interactive visualization functions. It offers dynamic layout adjustments, node and edge filtering, and customizable color schemes for representing data. It can export reproducible, publication-ready figures in SVG and PDF formats. In our case studies, we use RingNet to visualize breast cancer patients' omics profiles in a gene regulatory network and a cell-to-cell communication network in atopic dermatitis. This demonstrates RingNet's ability to reveal biological relationships across multiple data modalities. RingNet lowers the barrier to exploring, analyzing, and communicating data-driven findings, thereby accelerating research.

Alternative splicing generates protein isoforms with distinct regulatory and functional properties, yet the biological consequences of many splice variants remain poorly understood. The cytoskeletal scaffold protein ADD3 exists in two major isoforms, including a long isoform (ADD3-L) containing exon 14 that has been associated with metastatic cancer phenotypes. Given the established regulation of ADD3 by calcium–calmodulin signaling, we developed a probabilistic framework to investigate how calcium dynamics influence the activation behavior of ADD3 isoforms. The model integrates stochastic Ca²⁺–calmodulin encounters, Gamma-distributed activation kinetics, and Markov state transitions to characterize calcium-dependent activation and localization patterns. Simulations suggest that inclusion of exon 14 may enhance activation persistence and reduce stochastic variability, particularly in calcium-rich environments. Extending the framework to spatial calcium gradients predicts preferential activation near membrane-proximal regions and a calcium-dependent redistribution of ADD3-L toward cytoplasmic and nuclear compartments. These findings indicate that calcium-enriched tumor microenvironments may stabilize ADD3-L activity and promote cytoskeletal plasticity associated with metastatic behavior. More broadly, this study demonstrates how probabilistic modeling can be used to link alternative splicing events to context-dependent cellular phenotypes and generate mechanistic hypotheses for experimental validation.

The anti-viral and inflammatory responses of host cells are activated through several signalling pathways, such as NF-κB, IRF-3 and IRF-7 in response to viral infections. These transcription factors regulate the expression of antiviral genes that encode type I interferons (IFNα/β). However, the exact mechanism underlying the complex context-dependent temporal activation of IFNα/β remains unclear. Here, we have developed a probability-based mathematical model to study the co-activation of the IFNβ gene by NF-κB and IRF-3. Our model results suggest that the rapid and elevated as well as a delayed but persistent expression of IFNβ would be observed due to a temporal bias in the binding of NF-κB and IRF-3, highlighting the need for co-regulation of IFNβ by two different signalling pathways. In addition, the model predicts that the binding affinity of NF-κB to the promoter of IFNβ would be equal to or greater than that of IRF-3 for a quick, strong and persistent antiviral response.

Contextualized protein-protein interaction networks provide crucial insight into diseases and other biological processes, but for a profound understanding of such processes and their distinct effects on individuals, the protein-protein interactions within individual samples must be investigated. A straightforward approach to estimate the PPI network of a sample is to restrict a general network of known PPIs to the proteins that are found in the sample. Although proteomics methods are becoming more accessible and precise, large-scale and single-cell studies still mainly target characterizing the transcriptomics profile of the samples, which is then often used as an approximation of the protein activities. The correlation of gene expression and protein abundance is addressed by several studies, but information about the deviations of the different omics-based estimates of the PPI networks is still lacking. In this study, we performed a comparative analysis of transcriptomic-based and proteomic-based sample-specific PPI network estimates to fill this gap. We created a framework for a comprehensive and transparent comparison of the two omics levels in two independent datasets. We found that the size-adjusted characteristics of the different omics-based networks are very similar; the overall trend of how they change with age is also often the same, but the rate of the changes typically differs. These results shed light to the properties of PPI network estimations and advise caution in interpreting them appropriately.

Spatial transcriptomic (ST) technologies enable the measurement of gene expression directly within tissue sections while preserving spatial context. Many ST platforms additionally generate paired histological images alongside spatially resolved transcriptomic profiles. However, most existing computational approaches only incorporate histology images as auxiliary features in representation learning models and typically produce latent embeddings that are difficult to interpret. We present InSTaPath (Integrating Spatial Transcriptomics and Histopathology images), a multimodal topic modeling framework that links transcriptional programs with tissue morphology. InSTaPath converts token-level embeddings extracted from pretrained histology foundation models into discrete image words through vector quantization, enabling histological morphology to be represented in a count-based form analogous to gene expression. InSTaPath then jointly analyzes image-word and gene expression counts to infer shared latent topics that are interpretable through both topic-gene and topic-image-word associations. Across multiple ST datasets, InSTaPath improves spatial domain identification and uncovers biologically meaningful relationships between gene programs and tissue morphology through pathway enrichment and in silico perturbation analyses.

Single-cell perturbation screening is crucial for deciphering the mechanisms of cellular responses to gene perturbation. However, its scalability is hindered by high costs and noise levels of the labor-intensive experiments. To address this, active learning (AL) iteratively prioritizes informative perturbations to guide experimental selection, thereby enhancing model performance while minimizing costs. However, existing AL methods often struggle with the strict budget constraints typical of biological experiments. Moreover, the limited biological robustness of these methods makes it difficult to distinguish true biological signals from the noise caused by batch effects. Their biological stability across runs also requires further improvement. Therefore, it is essential to develop both efficient and robust AL strategies for budget-constrained perturbation screens. We introduce PertAL, a novel active learning framework to guide low-budget single-cell perturbation screening. To intelligently prioritize perturbations, PertAL integrates three key scoring modules: LLM-driven scoring of biological reasoning, multi-view diversity assessment, and gradient-based sensitivity quantification. Our experiments show that PertAL consistently outperforms ten baseline methods across three single-cell perturbation datasets. Notably, PertAL exhibits exceptional data efficiency. Using only 1/3 of the training data, it achieves performance comparable to that of the model trained on the full dataset. Moreover, it shows robust performance against batch effects and maintains its superiority even under extreme budget constraints. Finally, we demonstrate that PertAL yields biologically stable and meaningful results across runs. The code is available at \url{https://github.com/JieZheng-ShanghaiTech/PertAL}.

Accurate prediction of synergistic drug combinations can accelerate the discovery of cancer combination therapy, but existing deep learning
methods often generalize poorly to unseen drugs and remain computationally expensive for large-scale screening. We present DanSyn, a domain-
adaptive framework built around a hybrid structural-functional design. Instead of treating synergy prediction as a monolithic structural matching
problem, DanSyn learns transferable drug–cell interaction patterns from substructure-aware molecular sequences and unsupervised multi-omics
cellular representations, while injecting LLM-derived drug functional priors through an independent semantic branch and reducing unseen-drug
distribution shift with lightweight adversarial adaptation. Extensive experiments on DrugCombDB and DrugComb show that DanSyn achieves
state-of-the-art or highly competitive performance, provides clear gains in generalizability to novel drugs, and delivers up to 5.8× faster training
than strong deep-learning baselines through efficient fusion and pre-computed representations. Case studies further corroborate the biological
relevance and context-awareness of the top-ranked predictions. These results highlight DanSyn as a practical and scalable method for large-scale
drug synergy prediction.

Motivation: Sequential mass spectrometry metabolomics data are highly susceptible to run-order drift and batch effects. Existing correction approaches depend on dense interleaved quality-control (QC) samples and apply correction broadly across metabolites without statistical evidence for feature-specific correction, increasing the risk of overcorrection and removal of genuine biological signal.
Results: We present White Noise Normalization (WiNN), a framework for selective metabolite-wise technical correction. WiNN is built on a simple inferential principle: under adequate sample randomization, correction should be applied only when residual technical structure is detectable. The method first tests each metabolite for run-order dependence within batch or inferred run segments, fits generalized additive models only for metabolites with evidence of drift, and then restricts residual batch correction to metabolites with significant batch effects, followed by conservative dilution normalization. WiNN does not require dense QC coverage and can infer contiguous run segments when batch labels are unavailable. Across simulated, benchmark, and clinical serum metabolomics datasets, WiNN recovered ground-truth profiles, enabled biological marker discovery while maintaining strong class separation under low residual batch structure, and improved sample-identity structure, reference precision, and cross-plate reproducibility. These results show that WiNN achieves a strong balance between removal of technical artefacts and preservation of biological signal.
Availability: WiNN is implemented as an open-source R package at https://github.com/ratschlab/winn.
Contact: raetsch@inf.ethz.ch, odemler@bwh.harvard.edu
Supplementary Information: Supplementary data are available online.

Estimating the response of tumor cells to specific perturbations is crucial for identifying effective treatments that selectively target cancer cells
while sparing healthy ones, enabling personalized medicine approaches. Large-scale initiatives, such as DepMap, have profiled cancer cell line
responses to various drug treatments and gene knockouts, facilitating the development of computational models that predict sensitivity of cancer
cells to different perturbations. Existing models utilize diverse methods for encoding perturbations, including various chemical fingerprints and
types of gene-gene relationships. They also rely on different architectures and are often trained on distinct datasets. This variability makes it
unclear which chemical, genetic, or cell line encoding is most informative for predicting cancer cell viability following perturbation treatment.
To address this gap, we systematically evaluated various approaches to encode chemical and genetic perturbations and cell lines on the tasks of
predicting cell viability and gene dependency. We found that for genetic perturbations, STRING- based encodings yield the highest performance,
considerably outperforming GO-term and protein language model based encodings, which showed promising results in previous perturbation
prediction studies. For chemical perturbations, while most encoders showed comparable performance, those pre-trained on other bio-assay data
yielded the highest performance. Finally, we found that for cell line encodings, raw gene expression features outperformed more sophisticated
approaches, such as transcriptomics foundation model embeddings, as well as genotype-based encodings. Together, our results identify promising
approaches for encoding chemical and genetic perturbations and enable virtual screening for perturbations with selective toxicity.

Gene selection in high-dimensional biological data remains challenging due to the instability of standard feature selection methods, limited incorporation of biological knowledge, and poor generalizability across datasets. Methods focusing solely on statistical significance or predictive performance often identify different gene sets across studies, undermining reproducibility and biological interpretation. We present GeneSelectR 2.0, a framework integrating three complementary dimensions of gene importance: selection stability across cross-
validation folds, predictive utility combining coefficient magnitude and mutual information, and biological relevance via Gene Ontology semantic similarity. In fully nested cross-validation benchmarks on two clinical RNA-seq datasets: atopic dermatitis (SOS-ALL; n=149) and anti-PD-L1 immunotherapy response in bladder cancer (IMvigor210; n=192) GeneSelectR 2.0 outperformed all compared feature selection methods, yielding
the highest AUC for sample classification across all tested feature set sizes. This advantage was most pronounced at smaller feature set sizes: with only 100 genes, GeneSelectR 2.0 achieved an AUC of 0.68 for atopic dermatitis and 0.75 for the immunotherapy cohort. For the latter, no other method reached comparable performance even with up to 500 genes.

Cell-cell communication orchestrates tissue homeostasis and disease progression through dynamic ligand-receptor signaling. Yet no method exists to infer executable multicellular dynamical models from temporal data. We introduce here a Boolean network inference framework that enables the reconstruction of such models. With this approach, multiple distinct cell types are modeled jointly, each represented through its receptor and ligand nodes, with a bipartite dependency structure reflecting the two processes of intercellular communication: receptors respond to ligands secreted by any cell in the system, while ligands encode the intracellular response to a cell's own receptors. The timescale separation between fast receptor activation and slow ligand secretion is formalized as pseudo-steady-state constraints, chaining successive observed states into a dynamical sequence. From this sequence, the framework utilizes BoNesis to exhaustively infer all Boolean networks whose dynamics are consistent with the observations, producing an ensemble that explicitly identifies which regulatory interactions are determined by the inputs and which reflect the intrinsic redundancy of biological signaling. On a controlled two-cell toy model with known ground truth, we show that the inferred ensemble contains the ground-truth network, that the interaction structure is preserved under partial observations, and that receptor states are recoverable even when unobserved. Applied to a 50-node three-cell-type model of the CLL tumor microenvironment, the framework recovers the expected sequential niche establishment order and generates biologically coherent predictions, including for perturbations.

Motivation: Tumor-educated platelets (TEPs) represent a pivotal resource for liquid biopsy, reflecting transcriptomic alterations induced by the tumor microenvironment. Current analytical methods often focus on single-gene differential expression, overlooking high-order regulatory dynamics, and, in the case of co-expression analysis, the ""signed"" nature of co-expression relationships is not adequately exploited. This work proposes a computational framework based on Structural Balance Theory (SBT) to identify driver genes in glioma by evaluating the structural instability (frustration) of co-expression networks.
Results: By leveraging the Friendship-Like Differential Co-expression Network (FLDCN) framework, we applied the Local Balance Index to quantify individual gene contributions to network imbalance across various topological configurations. The analysis identified a consistent gene signature associated with platelet activation, glioma-specific pathways, and immune system modulation. The robustness of the proposed approach was further validated through cross-dataset analysis on independent cohorts, demonstrating that tumor-induced molecular rewiring generates stable and reproducible topological signals.

Motivation: In network medicine, the protein-protein interaction network, or interactome, is an essential resource for identifying candidate proteins underlying diseases and other phenotypic traits. Indeed, it can be leveraged via the guilt-by-association (GBA) paradigm, which asserts that interacting proteins are likely to participate in the same molecular processes. Network propagation that combines the topology of the interactome with prior knowledge about disease genes is a promising strategy to identify new candidate genes contributing to diseases. However, existing network propagation algorithms are often biased toward highly connected proteins, called “hubs”.
Results: Here, we present Guilt-by-association centrality, a novel network propagation algorithm designed to avoid inflated scores for hubs. We tested GBA centrality across four human phenotypes: two male infertility phenotypes, a cardiomyopathy and dyschromatopsia. For each phenotype, we assessed whether GBA centrality could recover known phenotype-associated genes and whether new candidate genes were enriched in relevant tissues. Depending on the phenotype, GBA centrality outperformed or matched the state-of-the-art network propagation methods. Furthermore, the results confirmed that GBA centrality is free of bias toward high-degree nodes. Therefore, GBA centrality is a powerful method for identifying new genes across diverse phenotypes and constitutes a strong alternative to existing network propagation algorithms.
Availability: GBA centrality is implemented in Python and C. The code is available under the GNU GPL (v3.0) on: https://github.com/jedrzejkubica/GBA-centrality

Predicting single-cell responses to perturbations from destructive assays is challenging because true pre- and post-perturbation measurements
of the same cell are unavailable. Existing methods show promise, but in the unpaired setting they still face three recurring difficulties: simple
priors that are poorly matched to multimodal cell-state geometry, noisy pseudo-trajectories induced by weak couplings, and unstable dynamics
near transitions between local state regions. We present a geometry-aligned flow matching framework for unpaired single-cell perturbation
prediction. The key design choice is training-inference consistency: pseudo-paired paths are constructed from observed control cells during
training, and test-time ODE integration also starts from the observed control state. A chart-aware geometry conditioner extracts atlas-guided
features from the control cell and perturbation context, while transport-aware pseudo-pair construction and a soft tangent-space multi-scale
vector field regularize trajectory learning. We evaluate the framework on the Norman CRISPR benchmark and the sciPlex3-K562 chemical
perturbation benchmark under standard, perturbation-out-of-distribution, and continuous-dose settings. Because the assays are destructive
snapshots, evaluation emphasizes condition-level and distribution-level agreement rather than literal one-to-one cell matching. Across these
two K562-centered public benchmarks, the clearest gains appear in the harder generalization regimes, especially perturbation OOD and dose
extrapolation. A direct consistency study further shows that anchoring both pseudo-supervision and rollout at the observed control state improves
accuracy, smoothness, and stability relative to sampled-rollout alternatives. Together, these results identify control-anchored geometry-aware
dynamics as an effective strategy for unpaired perturbation generalization.

Morphological alterations, including shortened neurites and impaired network formation, are a recurrent hallmark of amyotrophic lateral sclerosis (ALS). Observed in vivo and in iPSC-derived motor neuron (iPSC-MN) models, they hold potential as biomarkers for disease susceptibility and drug efficacy. Accurate quantification remains challenging: high-content imaging requires staining, while phase-contrast microscopy (PCM) is label-free and higher-throughput but complex iPSC-MN morphology has remained inaccessible to generic segmentation tools. MONDRIAN (Motor Neuron Disease Real-Time Imaging and Analysis) was purpose-built to overcome this barrier.

Images from 29 iPSC-MN lines (C9orf72- and TDP-43-mutated ALS, healthy controls) were acquired by live-cell imaging on the IncuCyte S3 platform. Cellpose, a widely adopted deep learning segmentation framework in cell biology, and a dedicated object classifier were fine-tuned on manually curated iPSC-MN images, enabling discrimination of soma, neurite arbours, growth cones, and debris. Per-object extraction yielded 167 morphological features fed into an XGBoost classifier optimised by Bayesian hyperparameter search (Optuna).

The XGBoost classifier distinguished ALS from control cultures with an AUC exceeding 0.94 under stratified 5-fold cross-validation. Unsupervised UMAP embedding followed by Leiden clustering identified disease-enriched subpopulations, with top features concordant across both approaches. Neurite texture and spatial organisation emerged as the most discriminative morphological signatures.

MONDRIAN enables high-throughput, staining-free morphometric assays for disease and rescue phenotypes in patient-specific models. Available as open-source software, identified morphological features will guide multi-omics integration, connecting morphological features to transcriptomic and proteomic signatures to uncover molecular drivers of motor neuron degeneration, providing a reproducible phenotypic readout for therapeutic screening.

Early segmentation strategies for spatial transcriptomics relied on morphological stains that delineate cell boundaries within a single focal plane. This approach ignores several technical challenges in spatial transcriptomics data:

1. Cells may be sectioned without their nuclei, leading to undersegmentation.
2. Overlapping cells in the z-plane can appear as one cell in 2D projections, leading to mixed expression profiles.
3. Transcript diffusion can occur during tissue processing, contaminating neighboring cells.

To address these limitations, transcript-informed segmentation methods have been developed that leverage spatial co-expression patterns of transcripts. However, evaluating the quality of the segmentation is difficult due to the high-dimensional and sparse nature of the data and the lack of manually curated datasets.

We introduce SegTraQ, a Python-based framework for segmentation and transcript assignment quality control in spatial transcriptomics data. SegTraQ computes quantitative metrics designed to highlight regions or samples with poorly segmented cells, and guide the choice of appropriate segmentation methods. Next to basic descriptors such as the number of genes, transcripts or cell morphology, we also provide metrics to evaluate how effectively the segmentation captures underlying biological structure. Spatially aware metrics quantify expression similarities between intracellular compartments and the local cellular neighborhood, and also assess the spatial distribution of transcripts within cells. For supervised QC, SegTraQ supports label transfer from scRNA-seq references and computes expression purity and signal spillover scores. Finally, we examine the ability of the segmentation to resolve transcripts belonging to overlapping cells across the z-dimension.

Pseudomonas aeruginosa PAO1 is classified as a high-priority multidrug-resistant pathogen by the World Health Organization, particularly in healthcare-associated infections, due to the limited availability of effective antibiotics. Its pathogenesis is strongly linked to quorum sensing (QS), a bacterial communication system that coordinates gene expression associated with virulence factors, including biofilm formation.
To provide a systems-level understanding of these processes, the most up-to-date and curated genome-scale metabolic model (GEM) for this strain, iJD1249, was reconstructed. The model was developed following established protocols by integrating and refining data from existing GEMs (iPae1146 and CCBM1146) with updated biochemical reactions and their gene-protein-reaction associations from BioCyc and KEGG. Validation was performed using Flux Balance Analysis to assess carbon source utilisation across well-defined in silico media, including Luria–Bertani medium, minimal medium, and synthetic cystic fibrosis medium.
The resulting model comprises 1,249 genes, 1,051 proteins, 1,208 biochemical reactions, 205 exchange reactions, and 1,178 metabolites. Notably, iJD1249 features a three-compartment structure (cytoplasm, periplasm, and extracellular space) and integrates QS-related pathways, representing a key advancement over previous models. The model produced growth rate and doubling time predictions consistent with experimental data and demonstrated improved accuracy in simulating complex nutritional environments compared with previous GEMs.
Overall, iJD1249 provides a robust platform for exploring the metabolic basis of P. aeruginosa PAO1 virulence and offers a foundation for identifying novel metabolic targets for antimicrobial strategies.

Fluorescence-activated droplet sorting (FADS) has become a vital tool in high-throughput biological assays, capable of sorting droplets. Traditional FADS platforms, however, are generally limited to 2-way sorting, which does not scale to the high-throughput, multiplexed sorting capabilities of FACS. To address this limitation, we present a pneumatic microvalve-based multiplexed fluorescence-activated sorting system, designed to simultaneously sort up to five different sample populations. This system can be applied to a variety of sample types, including single cells, hydrogels, and droplets. We evaluate its performance, focusing on throughput and accuracy across various sorting scenarios, including both intensity-based single-color sorting and more complex dual-color sorting of mixed sample pools. Achieving a maximum sorting throughput of 55 Hz with an accuracy exceeding 97%, our system demonstrates superior precision and the ability to effectively manage intricate sorting tasks.

Genetic mutations can alter protein structure, function, and interactions, leading to disruptions in cellular processes and contributing to cancer development. In this study, we propose a novel graph-based framework for cancer prediction that jointly models genetic mutations and protein–protein interactions. We construct a heterogeneous graph in which patients and proteins are represented as nodes, protein–protein interactions define edges between proteins, and patient–protein connections are established based on observed DNA mutations. Patient nodes are encoded using feature vectors derived from gene mutations, weighted through an information retrieval-inspired scheme that reflects the relative importance of each gene in disease progression.

To capture the varying effects of mutations, we employ attention-based graph neural networks that enable adaptive node representation learning by integrating both mutation information and interaction topology. The proposed approach is evaluated on whole exome sequencing data from the UK Biobank, focusing on four prevalent cancer types: breast, prostate, lung, and colon cancer. Experimental results demonstrate that the model effectively discriminates between cancer and control groups.

Furthermore, we extend our framework with an explainability module that identifies genes contributing most to model predictions. Notably, several of these genes are consistent with previously reported cancer-related genes, highlighting the biological relevance of the approach. Overall, our results show that integrating mutation data with protein interaction networks in a graph-based framework improves cancer classification performance and supports the discovery of potentially causal genes.

Carbapenem-resistant Acinetobacter baumannii (CRAB) is a critical antimicrobial resistance threat with a facultative, often intracellular lifestyle and strikingly heterogeneous virulence: some strains are essentially opportunistic, non-pathogenic, whereas others cause fulminant disease. WHO listed CRAB as prime threat pathogen, needs immediate attention. The phenotypic diversity, creates a key unresolved question: which regulatory and metabolic programs distinguish highly virulent CRAB from less virulent or commensal-like relatives under clinically relevant conditions?
Here, we present an integrative multi-omics framework to dissect condition-specific virulence mechanisms across antibiotic stress, biofilm formation, and planktonic states. Public RNA-seq datasets were systematically processed using DESeq2 and edgeR to identify robust differentially expressed genes across conditions. Transcriptomic signatures were integrated with regulatory network inference to reconstruct gene regulatory interactions and pinpoint key regulators such as the AdeRS and BfmRS two-component systems that link efflux, biofilm formation, and outer-membrane remodeling. Functional and gene set enrichment analysis revealed pathway-level rewiring of iron acquisition (including the acinetobactin cluster bauA–basD) and multidrug efflux (AdeABC) under antibiotic exposure, while weighted gene co-expression network analysis identified condition-specific gene modules associated with virulence and resistance phenotypes.
Our analysis highlights critical shifts in iron uptake, efflux regulation, and biofilm-associated pathways that may underlie why some CRAB strains successfully exploit intracellular niches whereas others do not. By prioritizing these context-dependent modules rather than isolated genes, this framework supports rational design of future anti-virulence and resistance-breaking therapeutics and offers a mechanistic roadmap to resolve the “mystery” of divergent pathogenic behavior in A. baumannii and other ESKAPE pathogens.

Animal and in vitro culture models have been essential in developing and evaluating human therapeutics and vaccines. However, observations rarely generalize across different biological systems (i.e., cells, species, etc.), resulting in a major need to translate observations from pre-clinical models to patients, as many drugs and vaccines ultimately fail in the clinic.
Translating between domains with non-overlapping feature spaces (e.g. species) and no paired samples is a fundamentally hard problem across biology and machine learning, and most existing methods cannot handle this scenario, including contrastive learning-based approaches.
We developed FlowTransOP to translate biological observations across such domains without requiring one-to-one feature mappings and paired data. We leverage a generative deep learning model that learns a mapping between two domains by directly aligning their distributions via flow matching; thus, learning a transformation that carries the entire feature profile by matching the underlying probability flows. Critically, a regularization term constrains the velocity field so that structurally similar conditions remain proximate after transformation (guided by a pre-aligned latent space), enabling conditional generation along an embedded manifold.
Evaluated on the L1000 dataset across 4 tasks using 5-fold CV, FlowTransOP outperforms naive and random baselines and remains competitive with gold-standard approaches (e.g. AutoTransOP) when sufficient pairing exists. However, when pairs are scarce (<35 pairs) or absent, or when cross-domain features are only moderately correlated (r<=0.58), only FlowTransOP maintains its performance in translating omic profiles. Overall, FlowTransOP can translate perturbations between pre‑clinical models and patients when direct correspondences are unavailable, enabling reliable inference for therapeutic development.

Biomolecules form complex interaction systems at various levels of biology. The dynamics of these interactions result in cellular phenotypes and responses. One example of a biological interaction system is a signal transduction network. Boolean models characterise the components of these systems with binary state variables and their interactions as governed by rules that determine their influence on each other. This approach does not require kinetic and abundance information about these components. Tools are available to define, analyse, and manage these models, but there is a lack of overview and evaluation of such tools' quality and functionality.

This work aims to facilitate the selection of software tools for Boolean modelling of signal transduction and gene regulatory networks, by developing and applying benchmark and evaluation criteria to the available tools. Twenty-five Boolean modelling tools were identified and assessed for compliance with FAIR4RS. Additionally, usability was assessed using a set of 24 sub-criteria developed for this purpose. Furthermore, the tools' functionalities were analysed and validated, and their ability to simulate and analyse Boolean models of different sizes was benchmarked.

Our comparison revealed that, although the results produced by the 25 tools were consistent, their analytical functionalities and ability to simulate larger models varied considerably. The size constraints for the tools' functionalities ranged from 25 to over 25,000 nodes. Each tool covered only part of the available functionalities and many exhibited room for improvement in terms of FAIR4RS compliance and usability.

In cancer therapy, combining multiple drugs is a common strategy to circumvent treatment resistance and reduce side effects. Existing machine learning (ML) approaches for estimating drug combination responses mainly focus on predicting synergy scores that measure the synergistic or antagonistic potential of two drugs. While synergy scores can support drug development and repurposing efforts, they rely on theoretical assumptions unlikely to hold in vivo. Moreover, drug synergy does not inherently guarantee treatment effectiveness. These factors heavily impede the applicability of synergy scores for deriving personalized treatment recommendations.
To overcome these limitations, we pioneer ML models that predict treatment responses directly, without relying on synergy scores: Our models predict the inhibition of cell growth at specific concentrations of one or more drugs defined in the model input. Notably, our approach is the first to provide dose-specific predictions of drug (combination) sensitivity for cell lines and drugs not used during model training. Thereby, we effectively mimic the challenge of predicting responses for previously unseen cancer patients or newly developed compounds. Furthermore, any conventional measure of drug sensitivity or synergy can be reconstructed from the predictions, making our approach highly flexible and suited for various tasks. In a large-scale benchmarking of the DrugComb database, we compare different ML algorithms and input representations. As an application in personalized treatment recommendation, we showcase that our models can accurately rank the most effective mono- and combination therapies for a given cell line using an extension of our recently developed sensitivity measure (CMax viability) to two-drug combinations.

In recent years, brown adipose tissue (BAT) has garnered interest as a potential target for obesity-related therapies due to its energy expenditure capacity and its positive response to exercise. One such study has been the ACTIBATE randomized controlled trial, which seeks to study BAT activation in young sedentary adults as a function of exercise. The cohort consists of 145 individuals split into a control group (no exercise), a moderate-intensity exercise group, and a high-intensity exercise group. Measurements are taken pre- and post-intervention, with multiomic information including RNA-seq from skeletal muscle and subcutaneous adipose tissue, and metabolomics and lipidomics from plasma. While early investigations centering on neck adipose tissue (NAT) volume could confirm the effect of exercise decreasing body weight, fat mass, and visceral adipose tissue mass, no evidence was found for significant increases in NAT volume or BAT activation in general. These early studies however only considered the clinical data together with the yet unpublished RNA-seq to assess global differences between groups of individuals. Our work aims to integrate the additional metabolomic and lipidomic measurements to find sub-groups of individuals showing a distinct response to exercise intensity and duration, and to use the multi-omics signature to predict personalised exercise response. Furthermore, we aim to develop a multiomic pipeline and incorporate machine learning techniques in order to stratify the cohort and determine potential molecular markers and clinical parameters. Our goal is that these findings will ultimately be used to develop personalized exercise routines to improve metabolic health.

Polyploid giant cancer cells (PGCCs), typically arising from whole-genome duplication, are major drivers of therapeutic resistance and tumor recurrence. Building on our recently published high-throughput single-cell drug screening platform and ongoing efforts to profile PGCC responses across multiple cell lines representing diverse cancer lineages, we developed an interactive web-based platform that enables integrative analysis and visualization of these data. The platform integrates high-throughput PGCC drug screening results with multi-omic features curated from public cancer data resources, including somatic mutations, copy number alterations, gene expression profiles, and pathway activity scores. It supports two primary analysis modules: (1) a gene- or pathway-centric module that identifies compounds whose anti-PGCC efficacy is influenced by specific molecular features, and (2) a compound-centric module that identifies genes or pathways associated with the efficacy of a selected compound. Statistical comparisons, interactive visualizations, and annotations from external knowledge bases enable users to explore drug-feature relationships and generate new hypotheses. The platform facilitates systematic investigation of PGCC vulnerabilities through data-driven exploration. Example analyses demonstrate that compounds targeting oxidative stress response and cytoskeletal remodeling pathways preferentially suppress PGCC-enriched populations, consistent with their structural plasticity and adaptive signaling. Integration with baseline pharmacogenomic datasets further distinguishes PGCC-selective inhibitors from broadly cytotoxic agents, supporting the prioritization of therapeutic candidates. In summary, this interactive web resource provides an accessible and scalable framework for analyzing the molecular and pharmacologic landscape of polyploid giant cancer cells. We anticipate that it will accelerate the discovery of biomarkers and therapeutic strategies to overcome treatment resistance driven by PGCCs.

Understanding multi-way interactions among biological entities—such as genes, proteins, diseases, pathways, and drugs—is essential for elucidating disease mechanisms and advancing targeted therapeutics. However, traditional graph-based models rely on pairwise relationships and often fail to capture higher-order structural semantics in heterogeneous biomedical networks, leading to information loss.
To address this limitation, this study proposes a novel deep learning framework based on Hypergraph Neural Networks (HGNNs) to model complex, high-order biomedical relationships derived from the large-scale CROssBARv2 knowledge graph. The CROssBARv2 system is a comprehensive resource that integrates heterogeneous biological and biomedical data from multiple databases into a unified knowledge graph, comprising millions of nodes and relationships across diverse entity types such as genes, proteins, drugs, and diseases. The dataset is further enriched with literature-derived associations from PubTator 3.0, improving both coverage and biological relevance. Disease-centric heterogeneous hypergraphs are constructed, where each hyperedge represents a disease and its 1-hop multi-relational neighbours, enabling explicit modelling of multi-entity interactions.
For robust binary classification, an advanced negative sampling strategy is employed at a 1:1 ratio, allowing the model to learn highly discriminative features while avoiding biases introduced by random sampling. The learned node embeddings are subsequently aggregated using an attention-based feature fusion mechanism, which dynamically weights the contribution of each node to estimate interaction probabilities.
The proposed framework enables accurate prediction of gene–disease–pathway–drug associations and facilitates large-scale inference of previously unknown relationships. Overall, it provides a scalable and effective computational pipeline for disease-specific relation prediction and novel drug repurposing applications.

Motivation: The rapid increase of high-throughput biomedical assays has improved patient phenotyping but also amplified analytic challenges. Heterogeneous study protocols, batch effects, large numbers of features, modality-specific distributions and frequent missing values make straightforward aggregation and interpretation of multi-omics cohorts difficult. These challenges complicate unsupervised tasks such as patient stratification, which remain critical for discovery of disease subtypes and actionable biomarkers. Thus, there is an acute need for computational methods that can integrate heterogeneous omics layers, handle modality-specific data characteristics, learn non-linear cross-modal structure, and produce outputs that are both useful and interpretable for domain experts. We present a framework that fuses heterogeneous gut microbiome multi-omics profile (metagenomics and metabolomics) for unsupervised patient stratification.
Results: Applied to cohorts composed of Crohn's disease, ulcerative colitis and healthy individuals, our approach yields coherent and well-separated patient clusters and outperforms traditional multi-omics integration based solely on graph fusion. The model identifies a set of discriminative features that drive stratification. Healthy clusters are distinguished by Alistipes shahii and deoxycholic acid, suggesting functional bile-acid and indole metabolism. In contrast, inflammatory clusters are defined by Ruminococcus gnavus and tryptophan accumulation, pointing to impaired anti-inflammatory processing. Additionally, the metabolite ADMA separates these subgroups through divergent associations with E. coli and F. prausnitzii. Statistical analyses and literature evidence support the biological relevance of these microbial-metabolic signatures. All together, this graph-based strategy provides a robust solution for integrating heterogenous multi-omics data from gut microbiome, highlighting molecular signature and potential actionable candidate biomarker for IBD medicine.

Urban forest fragments worldwide experience chronic anthropogenic disturbance, habitat fragmentation, and widespread biological invasions, yet the biochemical mechanisms that enable invasive species to persist and dominate in such environments remain poorly understood. Lantana camara, one of the world's most widespread woody invaders in tropical and subtropical forests, is known for prolific production of volatile organic compounds (VOCs) that mediate plant–plant, plant–microbe, and plant–herbivore interactions. In this work, we examine essential oil composition of L. camara across three urban forest fragments of the Delhi Ridge (India) representing a quantified disturbance gradient. Disturbance intensity was classified using decadal (2016–2025) changes in vegetation structure derived from NDVI analysis, corroborated by field-based observations.
Essential oil yield and compound richness increased with disturbance intensity, with 68, 81, and 90 compounds detected at low-, moderate-, and high-disturbance sites, respectively. Sesquiterpenes dominated all sites, and β-caryophyllene, α-humulene, and bicyclogermacrene formed a conserved chemotypic core. However, class-level shifts were evident along the gradient; monoterpene abundance (notably sabinene) declined under higher disturbance (from 10.2% to 4.0%), whereas oxygenated monoterpenes (e.g., eucalyptol) increased (from 5.1% to 8.2%). A bipartite network analysis identified 22 compounds consistently shared across sites alongside disturbance-associated unique emissions, indicating coexistence of chemotypic stability and biochemical plasticity, thereby underscoring metabolic flexibility as a hallmark of urban invasion.
These findings demonstrate that L. camara maintains a conserved defensive chemical framework while modulating volatile expression in response to urban disturbance.

Multi-omics data are typically represented as collections of matrices and analyzed by learning a shared representation of samples. However, this formulation collapses structured biological variation arising from factors such as individuals, cell types, and molecular features into a single axis, limiting the ability to capture interactions across these dimensions. Tensor-based approaches address this limitation by explicitly modeling multi-axis structure, but are predominantly restricted to linear decompositions.
Here, we introduce MANTIS (Multi-Axes Nonlinear Tensor Integration System), a framework for nonlinear modeling of tensor-structured multi-omics data. MANTIS represents data as a tensor and learns low-dimensional embeddings for each biological axis, enabling explicit modeling of interactions across individuals, cellular contexts, and molecular features. By combining axis-specific representations with Gaussian process-based nonlinear interactions, MANTIS captures higher-order dependencies beyond linear tensor factorization.
Importantly, MANTIS provides interpretability through relevance maps, which quantify the contribution of molecular features to specific regions of the latent space, enabling direct linking of feature programs to biological axes without post-hoc clustering.
We apply MANTIS to single-cell datasets from cancer patients, demonstrating its ability to capture variation across individuals and cell types. The learned embeddings separate healthy and disease samples, recover expected cell type structure, and group genes into functionally coherent programs supported by enrichment analysis.
MANTIS provides a flexible and interpretable framework for modeling multi-axis biological data, unifying tensor-based representations with nonlinear modeling.

Insertion sequences (IS) are mobile genetic elements that shape prokaryotic evolution, yet their diversity and ecological roles remain poorly characterized in environmental DNA. The Prokaryotic Atlas of Insertion Sequences (PAIS) provides public access to over 4.1 million IS elements identified across Earth's biomes, including detailed sequence information, homologous sequence clusters, cargo gene annotations, and links to major repositories such as JGI and the SRA. While PAIS enables large-scale exploration of IS diversity, leveraging this resource for ecological inference requires methods that link sequence-level information to environmental context. Here, we present PAIS-NN, a sequence-based machine learning framework to predict IS biome labels and make PAIS usable for downstream ecological inference. We extracted a representative subset of IS clusters for supervised model training and encoded sequences using k-mer frequency embeddings. A feedforward neural network with calibrated outputs was trained using class balancing and regularization strategies to ensure robust generalization. Model performance was assessed on a held-out test set, demonstrating reliable prediction of IS biome labels from sequence alone. To translate these predictions into ecological insights, we developed a multinomial expectation-maximization (EM) approach that infers ecosystem composition from distributions of predicted IS labels. This enables probabilistic estimation of ecosystem mixtures and the detection of gradients between environments. We applied PAIS-NN to metagenomic samples from a river estuary and recovered smooth transitions in ecosystem composition consistent with environmental gradients. These results demonstrate that IS sequence diversity contains informative ecological signals that can be quantitatively resolved.

Interoperability in healthcare remains challenging due to inconsistent terminologies and data standards across systems, hindering data exchange, clinical decision support, and research. Existing synonym resources often miss the messy realities of clinical data, including inconsistent abbreviations and local conventions. Without benchmarking datasets that reflect this variability, interoperability solutions cannot be adequately tested in practice. This preliminary study focuses on laboratory terminology, examining the 100 most frequent LOINC terms from MIMIC-III, a publicly available ICU dataset. Concept descriptions are enhanced through a unified framework that integrates ontology-based lexicons, rule-based variant generation, and large language model–based expansion. Comparative evaluation against existing methods demonstrates improved diversity of representations without compromising semantic fidelity.

Background: Cardiovascular diseases (CVDs) and their cardiometabolic comorbidities frequently co-occur and share risk factors, indicating common molecular mechanisms. However, these pathways remain poorly characterized at the transcriptomic level and are often studied separately.
Objective: To identify shared gene modules linking cardiometabolic traits and fatty liver disease through gene co-expression network analysis.
Materials and Methods: Data from the Young Finns Study (YFS), including gene expression profiles (25,950 genes) and clinical data from 993 individuals, were analyzed. A signed weighted gene co-expression network (WGCNA) was constructed to identify modules of co-expressed genes. Module–trait associations were evaluated using correlation analysis for cardiometabolic traits and logistic regression for fatty liver status, adjusting for age and sex. Differential expression analysis was performed on the full dataset, and pathway enrichment analysis was conducted for genes within significant modules.
Results: A signed co-expression network analysis identified 20 gene modules. Several modules showed significant associations with key cardiometabolic traits, including BMI, waist circumference, blood pressure, lipid levels, and glucose (FDR < 0.05). Six modules were significantly associated with fatty liver, with one module showing the strongest positive association (OR = 1.43, FDR = 1.47 × 10⁻³). Both positive and inverse associations were observed. Modules linked to fatty liver also showed consistent relationships with cardiometabolic traits, indicating shared transcriptomic patterns.
Conclusion: These findings reveal shared transcriptomic mechanisms linking cardiometabolic traits and fatty liver disease and highlight key molecular pathways for future investigation.

Per- and polyfluoroalkyl substances (PFAS) are persistent and ubiquitous environmental contaminants associated with diverse adverse health outcomes. Despite prior extensive molecular analyses, mechanistic understanding of PFAS hepatotoxicity remains incomplete. In recent years, human liver spheroids have emerged as a useful tool in this space for investigating human-relevant responses to PFAS exposure across longer-term exposure periods. Here, we reprocessed publicly available PFAS RNA-sequencing data from human liver spheroids using the R-ODAF pipeline to generate a high-confidence set of differentially expressed genes (DEGs) and applied an integrated multimodal platform combining cheminformatic target prediction and transcription factor (TF) activity inference within a liver-specific prior knowledge network. While conventional pathway enrichment recapitulated known PFAS effects on SREBF- and PPAR-driven lipid metabolism, it provided limited insight for compounds with fewer DEGs, such as PFBS. Multimodal integration revealed compound-specific regulatory programs, with PFOS and PFOA exhibiting a time- and dose-dependent suppression of a conserved TF module comprising SREBF, CEBP, and HNF family members. Cheminformatic protein target predictions for the various PFAS compounds highlighted several common targets, including MAP3K5 and PLCG1, as well as several compound-specific target predictions. Incorporating these predictions with network biology analyses elucidated upstream cascades differentiating PFOA- versus PFOS-mediated regulation of lipid metabolism, with the PFOA-specific inhibition of PTPN7 being a discriminating feature, highlighting potential molecular bases for compound-specific toxicological profiles. Collectively, our findings demonstrate that integrating structural, transcriptomic, and network-biology approaches enhances mechanistic resolution beyond traditional enrichment analyses, providing testable hypotheses for PFAS-induced perturbations and guiding experimental validation of key targets.

Understanding how microorganisms interact with their environment is a central challenge in systems biology. Reverse ecology aims to infer the minimal or most plausible environmental conditions (i.e., culture media) that enable specific metabolic functions. This is computationally challenging due to the combinatorial explosion of possible media compounds in genome-scale metabolic networks, limiting scalability, and because of the lack of biological realism of existing approaches.
We introduce MEDORA, a combinatorial framework designed to predict biologically relevant culture media from genome-scale metabolic networks. The main feature of MEDORA is to take into account phenotypic and functional data such as Biolog (nutrient usage), TnSeq experiments (gene/reactions essentiality), and metabolomic profiles (expected compounds) in the design of culture media.
To that goal, MEDORA first includes a ranking module that evaluates candidate media based on biological relevance. Compounds and reactions can be categorized by the user into distinct levels (e.g., required, curated, hypothetical, forbidden), allowing a structured biological interpretation of both inferred media and the pathways they activate. Second, MEDORA integrates heterogeneous biological constraints to directly guide media inference. By incorporating these data directly into the prediction process, MEDORA filters out topologically valid but biologically implausible solutions, substantially reducing the candidate space.
We applied MEDORA to the metabolic network of R. leguminosarum to assess the biological relevance of inferred culture media. Results show that MEDORA prioritizes ecologically meaningful media while preserving targeted metabolic functions, providing a scalable solution adaptable to non-curated networks.

Drug combination therapy is an established clinical strategy aimed at improving treatment outcomes in cancer, where multiple pathways contribute to disease progression and resistance. Computational approaches such as DeepSynergy have shown promise by integrating chemical structure and gene expression to predict drug synergy. However, these modalities do not capture how drugs alter cellular phenotype, which provides mechanistic insight into treatment responses.

To address this, we incorporated morphological profiles from Cell Painting, a high-content imaging assay that quantifies ~1,500 features per cell following chemical perturbation. We curated 51,793 drug combinations across 166 cell lines from DrugComb, integrating gene expression, Morgan fingerprints, and Cell Painting features. We evaluated classical machine learning baselines, including XGBoost, and a multi-layer perceptron (MLP) based on DeepSynergy, and observed that whilst XGBoost could leverage morphological features, the MLP could not effectively integrate them. We subsequently developed a cross-attention deep learning architecture that integrates the three modalities through separate encoding branches, employing Feature-wise Linear Modulation (FiLM) to condition morphological features on gene expression.

The FiLM-conditioned model achieved an average R2 of 0.777 ± 0.067 across three synergy metrics (Bliss, HSA, and ZIP), outperforming both the best classical model (XGBoost, 0.699 ± 0.066) and MLP baselines (0.697 ± 0.065). Ablation studies confirmed that FiLM conditioning is essential for exploiting morphological information, as removing it rendered morphological and structural features indistinguishable in predictive value. We are now developing interpretability methods to identify gene-morphology interactions driving predictions, with the goal of generating generalisable morphological representations for datasets lacking imaging data.

Systemic Lupus Erythematous (SLE) is an autoimmune disease characterized by highly diverse clinical manifestations, ranging from asymptomatic disease to severe pathology such as kidney failure. Uncharacterized disease trajectory dynamics thus far have prevented a complete understanding of the implicated heterogeneous biological mechanisms. In this work, we aim to develop a model that captures the disease progression and stratifies patients based on distinct longitudinal disease trajectories. The patient data is from the Swiss SLE Cohort Study, comprising ~3000 longitudinal clinical visits, collected from ~630 patients. The developed approach stratifies SLE patients using masked denoising autoencoders trained on tabular clinical data. Latent space embeddings were extracted to stratify patients into robust disease-course clusters. These clusters enabled stratification of patients with similar longitudinal disease dynamics including distinct remission and progression patterns. A simple nearest-neighbour regressor on the learned embeddings predicted future disease activity more accurately than baselines using raw clinical features, suggesting the representation captures clinically meaningful severity information. Overall, our work, leverages deep learning models to identify distinct disease trajectory clusters using a well-characterized SLE longitudinal cohort, which can aid in disease diagnostics, patient-stratification and better understanding of the disease heterogeneity. As a next step, we will integrate extensive flow cytometry data from SLE patients of the cohort and healthy controls, to investigate underlying disease trajectory-defining biological processes.

Good target discovery and validation are key tasks in pharmaceutical research to mitigate risk of failure in clinical phases. We have developed causal methods based on molecular regulation (directed, signed, weighted) graphs, using graph-flow and shortest-path approaches. As a proxy for the disease phenotype, we use differential expression profiles.

We present here the comparison in performance with other commonly used methods, leveraging three different publicly available data resources: (1) Drug Perturbation: LINCS1000 dataset (2) Single-cell CRISPR knockout: Replogle et al. 2022, Perturb-Seq (3) Outcomes in Phase 2 Clinical trials across disease groups defined by common MESH terms. Our results show consistently that methods based on graph flow outperform path-based and other approaches.

The different nature of the three sets of data we are using provides complementary views on different aspects of the performance of the methods aimed at solving the problem of target discovery and validation.

Understanding protein mechanisms in health and disease requires characterizing the functional roles of individual amino acid residues at the proteome scale. Existing resources typically address either sequence annotation, structural visualization, or contact analysis in isolation, leaving a critical gap for integrated, residue-centric investigation. To bridge this gap, we developed Atlantis, a freely accessible database and web platform that unifies structural, interactional, and functional information across the entire human proteome.
Atlantis integrates residue-level annotations from UniProt, InterPro, Pfam, IntAct, PhosphoSitePlus, PDBe, AlphaFill, ClinVar, and AlphaMissense into a hybrid MySQL/Neo4j architecture, enabling both structured queries and graph-based traversals. The database covers over 11 million residues across 20,000 human proteins, with pre-computed intra- and inter-protein contacts from 55,000 PDB structures and 20,000 AlphaFold models, plus hundreds of AlphaFold-Multimer complexes for GPCRs and the LRRK2 interactome.
A key distinguishing feature is the Pathfinder module, which enables 12 levels of graph-based structural analysis directly within the web interface — including shortest allosteric communication paths via Dijkstra's algorithm, betweenness centrality-based hub identification, PTM crosstalk mapping, variant-of-uncertain-significance resolution, and pan-family evolutionary hotspot detection. By cross-referencing contact networks with functional annotations, Pathfinder transforms static residue data into mechanistic, actionable insights.
Complementing this, the Contact Calculation module extends the analytical framework to user-provided structures, enabling on-the-fly computation of residue contact networks with immediate overlay of Atlantis functional annotations onto novel complexes.
Atlantis (https://atlantis.bioinfolab.sns.it/) is a free, active analytical platform that mechanistically interprets protein function and disease variants, offering more than data storage.

Genome-scale metabolic models provide a critical link between predictive modeling and experimental bioprocess design. However, options for rigorous model validation are currently limited. Here, we propose an approach for validating models against experimentally measured exchange fluxes, previously published 13C metabolic fluxes, and experimentally measured protein abundances.
We analyzed two genome-scale CHO cell models (iCHO1766 and iCHO3K) constrained with experimentally measured, condition- and time-specific exchange fluxes, and applied flux balance analysis and parsimonious FBA with biomass maximization. Model validation was performed using three complementary approaches: comparison of predicted and measured exchange fluxes via selective constraint relaxation, validation of intracellular fluxes against published 13C metabolic fluxes, and evaluation of predicted fluxes using condition- and time-specific proteomics data.
Using the selective relaxation approach, iCHO1766 predicted 9 out of 22 measured exchange fluxes with R2 ≥ 0.75, exhibiting strong correlation with measured fluxes. Application of parsimonious FBA substantially improved prediction performance, with 16 out of 22 exchange fluxes achieving the same threshold. Nevertheless, specific reaction fluxes (lactate, glycine, and glutamic acid exchange fluxes) were poorly predicted across both approaches and were consistently predicted as zero, indicating limitations in model structure or objective formulation.
In summary, genome-scale CHO cell models provide a powerful framework for predicting metabolic behavior and guiding bioprocess design. iCHO1766 achieved strong correlation with measured exchange fluxes, and parsimonious FBA further improved predictions. Model-dependent performance highlights opportunities to refine larger models like iCHO3K. Ongoing validation with intracellular 13C fluxes and proteomics will enhance prediction reliability for optimizing culture conditions and metabolic interventions.

Endogenous retroviruses (ERVs) are normally silenced in mammalian genomes, however, epigenetic dysregulation can lead to their cancer-specific expression, making them promising vaccine targets. Through a pan-cancer transcriptome analysis, we identify that cancer-specific ERV expression is especially pronounced in Acute Myeloid Leukemia (AML), in line with epigenetic dysregulation being a hallmark of AML. We set out to leverage the AML-specific ERV expression to develop a therapeutic AML vaccine (EVX-04). With outset in our clinically validated AI-immunologyâ„¢ platform, we developed a model, ObsERVâ„¢, for identifying the optimal set of ERV-derived protein fragments that would elicit broad T-cell responses in AML patients. ObsERVâ„¢ takes as input ERV expression levels in AML samples, HLA population frequencies and protein-translation propensity. The protein-translational propensities were established by large-scale analysis of thousands of cancer biopsies and cell lines characterized by ribo-seq, immunopeptidomics and proteomics. From a pool of 5 million ERV-derived protein fragments expressed in AML, ObsERVâ„¢ selected 16 ERV fragments rich in HLA ligands for the EVX-04 vaccine. On an independent AML cohort, EVX-04 was benchmarked to cover 95% of the patients with at least 15 HLA ligands. In vitro validation demonstrated that each ERV fragment in EVX-04 can elicit specific T-cell responses across multiple HLA diverse subjects. A mouse surrogate of EVX-04 was found to confer complete tumor control in B16F10 and CT26 in vivo challenge experiments. Overall, ObsERVâ„¢ designed a novel ERV-based off-the-shelf AML vaccine, which we preclinically validated and are preparing for a first-in-human clinical trial.

生物硝化抑制（BNI）是改善小麦氮利用效率的有前景特征，但由于高维度和非线性特征相互作用，其化学决定因子在非靶向根-渗出物代谢组学中仍难以解析。我们分析了44个小麦样本中的根渗出代谢组，并制定了监督式二元分类任务，以预测高BNI表型。我们构建了一个可解释的深度学习框架（平衡相关性评分），利用注意力增强的多层感知器学习特征相关性，同时结合SHAP归因（LightGBM）和稳定性选择（Elastic-Net）等互补证据，提升鲁棒性。BRS框架不仅优先考虑单个代谢物，还优先考虑与BNI相关性的协同组合（对和三元）。在我们的结果中，组合的辨别力比单一代谢物更为一致，单一、对和三元组的AUC值最高可达0.82。置换测试显示，排名前列组合的预测性能超过了标签随机化下的预期。利用本研究提出的方法，我们获得了对影响BNI的关键代谢物及其协同组合的新系统见解。结果表明，若干代谢物及其潜在组合效应可能在BNI调控中发挥重要作用，从而为下游功能验证实验和智能育种提供理论基础和优先候选靶点。

Complex diseases arise from the interplay of multiple biological factors, and integrating heterogeneous data sources, including multiple omics layers, has been shown to improve prediction of future health status and incident disease risk. The increasing availability of well phenotyped longitudinal cohorts has enabled researchers to apply survival modelling frameworks to predict disease onset using high-dimensional multi-omics profiles. However, most existing risk models remain confined to a single omics layer, constraining predictive performance and biological insight, while the limited available multi-omics integration methods largely rely on simple concatenation or early fusion strategies that fail to capture cross modality interactions highlighting the need for robust integration approaches, particularly in population scale studies of the human microbiome.
IntegratedLearner is a published R package for multi-omics prediction and classification, available on GitHub. Here, we extend it to time-to-event outcomes, enabling efficient multi-omics integration for incident disease risk prediction. As a case study, we demonstrate its application to heterogeneous microbiome derived data layers, including taxonomic composition and metabolomic profiles, allowing information to be shared across omics layers for consolidated risk predictions.
We develop a unified survival framework supporting early, late, and intermediate multi-modal fusion strategies, benchmarked against single-omics and concatenation-based baselines using tree ensembles, penalized Cox models, and boosting approaches. Performance is assessed via concordance index, Brier score, and AUC, along with feature-signal outputs for biological interpretability. The framework is being implemented within a Bioconductor compatible workflow leveraging (Tree)SummarizedExperiment and MultiAssayExperiment data structures, ensuring reproducibility and scalability across microbiome and multi-omics cohort studies.

The precise discrimination between malignant and normal cells within the tumor microenvironment (TME) remains a fundamental challenge in single-cell transcriptomics. Accurate annotation is essential for characterizing tumor heterogeneity, yet existing methods struggle with the overlapping transcriptional profiles of cancer cells and their normal counterparts. We compiled single-cell RNA-seq data from 44 patient cohorts (369 tumor samples across 11 major cancer types). We evaluated the discriminative power of >1,000 pathway and signature-based features, including oncogenic signaling, metabolism, antigen presentation, DNA repair, proliferation, and epithelial–mesenchymal transition. We find that discriminative features are strongly cancer type-specific, and no single feature generalizes across all cancer types with consistently high performance. Using these features, we benchmarked multiple classifiers under independent test cohorts validation and observed improved predictive performance relative to existing malignancy annotation tools. Finally, analysis of feature divergence across cancer types highlights distinct transcriptional programs associated with malignancy, providing interpretable insights into tumor-specific biology at single-cell resolution. By analyzing which features are most divergent across different malignancies, our framework provides transparent biological insights into the unique transcriptional landscape of specific cancer types, facilitating a deeper understanding of the molecular drivers of malignancy at single-cell resolution.

Endometrial cancer (EC) co-opts immune tolerance mechanisms similar to those at the maternal–fetal interface, but their spatial organization remains unclear. We integrated bulk and spatial transcriptomics across EC, healthy endometrium, and spontaneous abortion samples. EC showed transcriptional signatures of cytotoxic suppression and reduced macrophage phagocytosis, while spatial analyses revealed tumor-localized enrichment of TIGIT and CD47 pathways. Macrophage activity increased with distance from CD47-positive tumor cells, indicating proximity-dependent immune suppression. These findings identify spatially organized checkpoint activity—particularly the CD47–SIRPA axis—as a key mechanism of immune evasion in EC.

Cardiorespiratory fitness (CRF) is a robust predictor of cardiovascular events and all-cause mortality, often outperforming traditional risk factors. Nevertheless, its routine clinical assessment is limited by the need for specialized equipment, trained personnel, and time-intensive procedures. Given the close relationship between CRF and vascular health, surrogate markers capturing vascular features may offer a practical alternative for its estimation.

Retinal color fundus images (CFIs) provide a non-invasive window into systemic microvascular health and have been widely used to predict cardiovascular risk factors and disease. However, their potential for estimating CRF remains unexplored. In this study, we introduce RetFit, a novel CRF estimator derived from CFIs using state-of-the-art vision transformer models. We assessed its clinical relevance by examining associations with cardiovascular risk factors, disease outcomes, and genetic architecture, and benchmarked its performance against submaximal exercise test–derived CRF (SETCRF).

RetFit demonstrated strong prognostic value for both cardiovascular events and all-cause mortality and showed significant associations with a broad range of disease states and risk factors, with consistent effects across two independent external cohorts. Although RetFit and SETCRF exhibited moderate phenotypic correlation (r = 0.45), their genetic associations were largely distinct. Interpretability analyses suggest that RetFit predictions are primarily driven by retinal vascular features, with attention maps highlighting vascular regions and arterial bifurcation count emerging as the most strongly associated feature.

Overall, these findings underscore the potential of retinal imaging as a scalable, cost-effective, and accessible approach for CRF estimation, supporting its application in large-scale screening and risk stratification.

Advances in high-throughput sequencing and other assay technologies have enabled the extensive molecular and phenotypic characterization of cancer cells, generating large and complex multi-omics datasets that provide a more holistic view of tumor biology. However, integrating these data across different cancer cell models remains a significant challenge due to their high dimensionality, heterogeneity, and missing omics.

Here, we propose a conditional adversarial multiview variational autoencoder (VAE) that integrates genomics, transcriptomics, methylation, drug response and CRISPR-Cas9 gene essentiality screens data from both preclinical and clinically relevant cancer models, when available. Specifically, we focus on the mosaic integration of Cancer Dependency Map data comprising cancer cell lines and organoids, together with TCGA tumor patient samples.

This approach aligns heterogeneous datasets into a shared latent space, effectively bridging the gap between preclinical models and patient tumors. The learned latent space achieves strong cross-modality alignment, with reconstruction performance reaching an average Pearson correlation greater than 0.9 across observed omics layers. Notably, the framework accurately imputes missing molecular profiles, including DNA methylation in organoids and CRISPR-Cas9 gene essentiality profiles in tumor samples, enabling in silico reconstruction of functional genomic dependencies across patient tumors.

Overall, this adversarial multiview VAE provides a unified framework for data integration and augmentation across cancer models and establishes a foundation for the identification of patient-specific vulnerabilities, ultimately supporting the prioritization of effective therapeutic strategies for clinical validation.

Understanding the phenotypic consequences of genetic perturbations is essential for uncovering gene function, mapping disease mechanisms, and prioritising drug targets. Despite the rapid growth of large-scale perturbation datasets, ranging from CRISPR knockouts to deep mutational scanning (MAVE) and single-cell expression profiles (Perturb-seq), their heterogeneous formats and fragmented storage across disparate repositories hinder systematic integration and meta-analysis.

We present the Perturbation Catalogue, an integrated, curated resource that harmonises human gene perturbation, variant analysis, and expression data. Central to the Catalogue is a highly unified metadata schema, implemented as an extensible Pydantic model, which standardises experimental designs, readout assays, and model systems across diverse paradigms. A robust data warehousing architecture powers a responsive, user-friendly portal and a versatile API. The platform features standardised Differential Expression Analysis (DEA) and Gene Set Enrichment Analysis (GSEA) for Perturb-seq datasets, alongside dynamic visualisations such as heatmaps for MAVE and comprehensive lists of significant hits.

By seamlessly linking different modalities of perturbation data, the Catalogue bridges the gap between target identification and functional validation.

Machine‑learning models in biology often achieve strong predictive performance yet remain difficult to interpret. A well‑known example is the epigenetic aging clock, which accurately predicts chronological age, but the biological relevance of the contributing CpG sites is still debated. Recent work suggests that stochastic drift, rather than active regulation, drives much of its predictive power.

We generalize this idea beyond aging and across omic layers. We show that stochastically drifting features can acquire stable model weights because biological constraints impose a consistent population-level direction of change. This “deterministic drift” is not specific to aging, instead it generalizes across biological conditions and explains why high‑dimensional omics models remain predictive even when many features lack direct functional roles.

As a proof‑of‑concept, we demonstrate this principle in both simulated and real datasets. Using tissue gene expression, we show that the model trained only on one tissue and stochastic noise addition, can predict the other tissue, for example predicting cortex from liver. This is valuable in settings where perturbed samples are difficult to obtain, such as senescence research. The model benefits from high feature dimensionality and performs well even with limited sample sizes.

While previous work explained why aging clocks work despite stochasticity, we show the same principle enables cross-state prediction broadly even in non-stochastic biological processes. Using our method, we further developed a framework to distinguish genuine regulatory signals from stochastic noise. These results suggest that deterministic drift is a widespread property of biological systems exploitable for prediction.

Understanding plant gene regulation requires decoding complex epigenetic landscapes. High-throughput sequencing technologies have generated diverse, massive datasets profiling transcription factor (TF) binding, chromatin accessibility, DNA methylation, and 3D genome organization. However, the mechanisms by which these distinct epigenetic layers coordinate gene regulation across different biological conditions remain unclear, and the accurate identification of true regulatory elements is often hindered by experimental noise. In this study, we developed an integrated multi-omics platform that integrates diverse regulatory datasets and enables high-resolution analysis of gene regulation. Our platform systematically integrates ChIP-seq, ATAC-seq, WGBS, and Hi-C data with comprehensive gene annotations through an updated suite of analytical modules and a high-resolution genome browser. By integrating these data types, the platform significantly improves the accuracy of regulatory element identification; for instance, combining ChIP-seq candidate sites with ATAC-seq and BS-seq data confirms true transcription factor binding within accessible, low-methylation chromatin. Additionally, integrated Hi-C data reveals long-range chromatin interactions, successfully connecting distal regulatory elements to specific gene promoters. Furthermore, the implementation of specialized differential analysis pipelines allows users to identify condition-specific regulatory changes in TF binding, chromatin accessibility, and DNA methylation across different experimental conditions. Together, this integrated resource empowers the plant research community to efficiently analyze multi-omics landscapes, accelerating the discovery of regulatory elements and complex networks underlying biological processes.

Mass spectrometry-based proteomics enables simultaneous quantification of thousands of proteins and offers a rich substrate for machine learning-driven diagnostics, yet translation into routine health care remains absent. Here, we present a generalizable framework for deriving clinically actionable decision support from high-dimensional proteomic data using machine learning, implemented within a hospital setting and demonstrated across two distinct disease areas.
In a cohort of patients with suspected tick-borne disease (neuroborreliosis), we applied plasma and cerebrospinal fluid (CSF) proteomics combined with machine learning to distinguish neuroborreliosis from differential diagnoses. The resulting models captured complex host-response signatures and achieved high diagnostic performance (CSF: AUC = 0.90, MCC = 0.63, plasma: AUC = 0.80, MCC = 0.48). The blood-based model outperforms existing diagnostic approaches and, unlike current strategies that are insensitive in early disease, relies on protein signatures detectable at early stages. This approach offers a minimally invasive, blood-based diagnostic strategy with the potential to fundamentally change current diagnostic workflows.
In a second application, we analyzed blood samples to develop a machine learning-based proteomic classifier for hemoglobinopathies. Despite the genetic basis of these disorders, proteomic signatures enabled accurate and scalable classification (AUC = 1.00, MCC = 0.94), reflecting downstream functional consequences of genetic variation. This is particularly relevant as current diagnostic strategies can be time-consuming, expensive, or limited in detecting all variants.
Together, these results illustrate how machine learning applied to clinical proteomics can move beyond biomarker discovery toward deployable diagnostic tools, supporting a new paradigm of data-driven, multi-protein decision support in routine healthcare.

Quantitative characterization of vascular architecture is essential for understanding tissue homeostasis and disease processes. While digital pathology platforms enable large-scale whole slide image (WSI) analysis, robust tools for automated vessel-level morphometric quantification in routine immunohistochemistry (IHC) sections remain limited. We developed VeSpA, an automated computational pipeline for segmentation and quantitative analysis of CD31-stained vessels in WSIs.

VeSpA is an open-source Python-based pipeline for batch processing of histological images. The workflow integrates CMYK color-space transformation to isolate the DAB (3,3'-diaminobenzidine) signal, Otsu automatic thresholding, and morphological refinement for vessel segmentation. Individual vessels are identified as connected components and quantified using region-based morphometric descriptors including area, major and minor axis length, eccentricity, and orientation. Segmentation performance was evaluated by comparing automated vessel masks with manually annotated ground-truth masks produced independently by two pathologists on four randomly selected sections (4000×4000 pixels). Pixel-level metrics including accuracy, precision, recall, specificity, F1/Dice score, and intersection-over-union were calculated.

VeSpA generates binary vessel masks, visualization overlays, and structured morphometric output files for each image. Across eight comparisons, it achieved a mean segmentation accuracy of 87.3%, demonstrating robust performance across slides with variable staining intensity and tissue density. Individual vessels were consistently separated and assigned unique identifiers, enabling single-vessel morphometric analysis at slide scale. A QuPath extension provides a graphical interface for direct integration into digital pathology workflows.

VeSpA provides a scalable, reproducible, and training-free approach for automated vessel segmentation and morphometric quantification in CD31-stained WSIs.

Motivation: Biocuration workflows often rely on comprehensive literature searches for specific biological entities. However, standard search
engines such as MEDLINE and PubMed Central provide an incomplete picture of the scientific literature because they do not index the increasing
amount of valuable information published in supplementary data files. Over two years, we addressed this gap by systematically extracting
text from a large proportion (85%) of these files, resulting in 35 million searchable documents. To assess the information gain provided by supplementary
data files beyond the manuscripts, we searched both for mentions of dozens of Global Core Biodata Resources (GCBRs), which are
fundamental biological databases essential for the life sciences. We searched for mentions of GCBR names and accession numbers, which
uniquely identify biological entities within these resources.
Results: The recall gain from using the supplementary data files to search for articles mentioning resource names is 6%. In addition, 97% of all
accession numbers identified were published in the supplementary data files, highlighting their increasing importance for highly specific topics
or curation pipelines. We show that the number of accession numbers published in the supplementary data files is increasing year on year, but
that 87% of these are published in Excel files. This format facilitates human readability and accessibility, but severely limits machine reusability
and interoperability. We therefore discuss alternative and complementary approaches to the publication of research data.

The UK Dementia Research Institute (UK DRI) comprises nine geographically distributed centres generating diverse datasets across Omics, imaging, behavioural, and clinical associations. However, datasets are often deposited independently, limiting discoverability and hindering integrative computational analyses.
To address this, we developed the Dementia Data Nexus (DDNexus), a unified, searchable platform for accessing UK DRI datasets linking to our analysis tools. Built on FAIR data principles, DDNexus aggregates and harmonises metadata from multiple sources to enable efficient data discovery and reuse.
Database construction combined manual curation and automated pipelines. UK DRI principal investigators were contacted directly to identify datasets, while a comprehensive publication database was assembled and mined for associated datasets using metadata from NCBI. Automated text annotation, informed by Europe PMC methods, was used to extract accession numbers from full-text articles. These were further enriched via APIs from major repositories, including PRIDE, Gene Expression Omnibus, and ArrayExpress, alongside large language models for metadata annotation.
Beyond a directory, DDNexus serves as the central hub for the UK DRI informatics data ecosystem. From this directory, users can directly select datasets to execute standardised analysis pipelines, explore via our data visualization platform Neuromics Explorer, and interrogate their gene lists through our DataMap biological knowledge graph platform.
The platform is implemented using a full-stack architecture incorporating relational database systems, scalable data indexing for efficient search, alongside containerised deployment for scalability. By integrating discovery, analysis, and visualisation, DDNexus demonstrates how a field, in this case dementia, can be brought together through data discoverability to enhance access.

Hypertension (HTN) is a complex, polygenic risk factor for cardiovascular disease, yet current genetic studies explain only a fraction of its phenotypic variance. To better capture the biological interrelations driving disease complexity, this project aims to map distinct biological endotypes of hypertension and identify a ""Resilient"" cluster: patients exhibiting the lowest comorbidity rates despite long-standing HTN. By comparing this resilient group to non-resilient clusters, we expect to isolate specific molecular resilience factors, such as protective genetic variants and unique methylation patterns.
To achieve this, we integrate clinical, lifestyle, genetic, epigenetic, and proteomic data from the Generation Scotland dataset to construct a Multi-Modal Patient Similarity Network (PSN). We apply rigorous feature selection across modalities - utilising Random Forest and Elastic Net classifiers alongside validated markers - to overcome dimensionality challenges. Individual networks are then sparsified using K-Nearest Neighbours, combined via Similarity Network Fusion (SNF), and clustered using the Louvain method.
Ultimately, the isolated molecular factors will inform the development of a predictive ""Resilience Score"" for complication-free survival. To ensure robustness, this score and the derived SNF parameters will undergo subsequent validation in the independent UK Biobank dataset.

Understanding complex diseases requires integrating biological knowledge across multiple scales, from molecular interactions to disease phenotypes. However, these relationships are often fragmented across data modalities, limiting the ability to interpret how entities such as genes, proteins, drugs, pathways, and traits interact within a coherent biological framework.
We present DataMap, a knowledge graph designed to model and contextualise relationships between biological entities relevant to neurodegenerative diseases. DataMap connects diverse modalities including SNPs, eQTLs, genes, proteins, pathways, and disease traits into a unified, biologically meaningful network. By explicitly structuring these relationships and preserving their biological context, the platform enables users to trace how molecular signals propagate across different levels of biological organisation and identify points of convergence across datasets.
DataMap supports multiple analytical scenarios, including gene-centric and drug-centric exploration. It also enables comparative analysis through a gene set connectivity approach, where relationships between two gene sets are visualised within the network to highlight clustering and shared biological context. In addition, DataMap incorporates an LLM-based interface, GeneTalk, allowing users to query the knowledge graph using natural language and interactively explore network-derived insights. Each query is supported by statistical frameworks, including gene set enrichment approaches such as MAGMA and background gene set calculations, enabling robust interpretation.
Built as a scalable, cloud-native platform with a microservices architecture, DataMap provides an interactive web interface for efficient querying and visualisation of complex biological networks. By combining integrative modelling, statistical analysis, and intuitive exploration, DataMap enables interpretable, data-driven hypothesis generation and target prioritisation in neurodegenerative diseases research.

Image-derived vascular phenotypes from retinal imaging provide a scalable, non-invasive window into systemic vascular health and are widely used to characterise microvascular structure and its links to disease risk. However, how vascular geometry influences haemodynamic function remains poorly understood.
We developed a computational framework that directly translates imaging-derived vascular geometry into subject-specific retinal haemodynamic simulations. Using UK Biobank fundus image segmentation masks, we reconstructed vascular geometries and simulated blood flow using COMSOL Multiphysics, a finite element-based platform for computational fluid dynamics. The framework further incorporates subject-specific physiological inputs, including blood pressure and blood viscosity. Simulations performed within seconds per image yield high-throughput, spatially resolved maps of flow dynamics, including flow velocity, wall shear stress, and pressure, at population scale. Importantly, they reveal how individual variations in vascular geometry shape heterogeneous haemodynamic environments across this microvascular system.
Together, these results establish a mechanistic link between static vascular structure and dynamic function, enabling functionally informed vascular phenotypes for large-scale genetic and epidemiological studies, with potential to provide novel insights into systemic vascular health and disease risk.

Wolbachia pipientis is a maternally transmitted bacterial endosymbiont predicted to infect more than 8 million different arthropod species. As some Wolbachia strains can also disrupt insect virus infection, it has been deployed to fight the spread of dengue and zika virus. The exact mechanisms enabling Wolbachia-mediated antiviral protection is not fully understood, but suggests metabolic competition between Wolbachia and virus. As Wolbachia cannot be genetically modified, it has been difficult to determine the genetic underpinning of most observable effects on the insect host.
Using the model insect, Drosophila melanogaster, and its Wolbachia symbiont, wMel, this work applies a computational approach with Genome-scale metabolic modelling (GEM) to infer metabolic activity from single cell transcriptomic data of Drosophila ovaries with and without wMel infection. First, a metabolic task-centric curation campaign for D. melanogaster GEM was deployed to improve model quality and predictability. We next adapted the scCellFie implementation which compresses high-dimensional transcriptomics datasets into metabolic task scores, allowing ready comparison and interpretation of metabolic shifts between cell types.
Applied to 7752 cell profiles, we reveal cell-type specific metabolic activities in Drosophila ovary tissues and perturbations in Wolbachia-infected conditions. This work shows that Wolbachia has only limited effects on metabolic processes in ovaries, in agreement with the finding that wMel does not cause reproductive fitness costs in its host. As previous work has shown that Wolbachia does have global effects to Drosophila metabolism, our current work provides evidence that these changes are localized to tissues other than the ovaries.

In clinical cancer genomics, patients are rarely profiled across all molecular data types. Multi-omics models integrating bulk mRNA expression, miRNA, and DNA methylation hold promise for survival prediction, but block-wise missingness is the norm in practice: a patient may have RNA sequencing but no methylation data, or a dataset may lack miRNA profiling entirely. Existing approaches either fuse only observed modalities without reconstructing absent ones, or learn cross-modal latent mappings without any constraint on survival signal preservation.
We present CrossPoE, a Product-of-Experts variational autoencoder that handles modality missingness through six directional cross-modal translation heads trained jointly with a survival-preserving objective. When a modality is absent, CrossPoE synthesises a pseudo-posterior from observed modalities and injects it into the PoE fusion. Critically, translation heads are trained to match task-predictive survival scores across modality pairs, not merely approximate target latent geometry. This constraint ensures pseudo-posteriors carry survival-relevant signal under missingness, and is absent from all prior latent completion methods. The directional translation structure also enables a Jacobian-based interpretability pipeline that identifies hub latent dimensions by their structural centrality across translation paths and attributes them to input features via Integrated Gradients, with stability confirmed by fold-level majority voting.
We evaluate under RNA block missingness as the most clinically relevant stress test, given RNA's dominant independent prognostic contribution. On TCGA-BRCA (1,095 samples, 145 PFI events), CrossPoE matches HEALNet-Omics on missingness robustness, and ablation confirms the translation mechanism as the primary driver: removing translation heads increases the RNA missingness C-index decline from −0.008 to −0.043. On TCGA-KIRC (535 samples, 160 PFI events), CrossPoE significantly outperforms HEALNet-Omics, CrossAE, and PoE-VAE on RNA missingness robustness, confirmed by bootstrap CIs excluding zero.

Single-cell omics analysis has unveiled the heterogeneity of various cell types within tumors. However, no methodology currently reveals how this heterogeneity influences cancer patient survival at single-cell resolution. Here, we introduce scSurv, combining a Cox proportional hazards model with a deep generative model of single-cell transcriptome, to estimate individual cellular contributions to clinical outcomes. The accuracy of scSurv was validated using both simulated and real datasets. This method identifies cells associated with favorable or adverse prognoses and extracts genes correlated with their contribution levels. In melanoma, scSurv reproduces known prognostic macrophage classifications and facilitates hazard mapping through spatial transcriptomics in renal cell carcinoma. We also identified genes consistently associated with prognosis across multiple cancers and demonstrated the applicability of this method to infectious diseases. scSurv is a novel framework for quantifying the heterogeneity of individual cellular effects on clinical outcomes.

Achieving curative outcomes remains a major challenge as drug resistance continues to hinder chemotherapy efficacy. MicroRNAs (miRNAs), as post-transcriptional regulators of gene expression, play essential roles in proliferation, apoptosis, and immune regulation, yet their integration into predictive models using clinical patient cohorts remains insufficiently investigated. In this study, DrugMiR was developed as an explainable multi-omics modeling framework for identifying miRNA signatures and their regulatory targets associated with chemotherapy response across five commonly used drugs: cisplatin, carboplatin, gemcitabine, paclitaxel, and fluorouracil. Differentially Expressed miRNAs (DEmiRs) were obtained from 1,015 patient samples in TCGA (The Cancer Genome Atlas), and predictive models were constructed using five machine learning (ML) algorithms. DEmiR-based models achieved strong predictive performance for platinum-based therapies, including cisplatin (AUC = 0.81) and carboplatin (AUC = 0.79), while integration with experimentally validated target genes improved or maintained predictive performance across all investigated drugs Comparative network analysis revealed recurrent miRNA regulators shared across multiple drug cohorts, particularly members of the miR-200 family, which were consistently identified in cisplatin, carboplatin, and gemcitabine response signatures. These miRNAs converged on key targets including GATA6, BMP2, and ITGA3, implicating pathways related to cellular adhesion, extracellular matrix remodeling, and survival signaling. Collectively, the DrugMiR framework demonstrates that integrating drug response-associated miRNAs with their validated target genes improves the interpretability and predictive modeling of chemotherapy response using clinical patient cohorts. By combining post-transcriptional regulatory information with downstream gene expression changes, DrugMiR identifies biologically relevant drug response-associated signatures and provides a systematic framework for exploring the molecular mechanisms underlying therapeutic sensitivity and resistance across cancer types.

Recent research highlights the connection between the gut microbiota and Major Depressive Disorder (MDD), however the influence of sexual dimorphism on these microbial signatures is still not clear. We investigated sex-specific compositional shifts in the gut microbiota utilizing 16S rRNA amplicon data from a cohort of 1,269 individuals (644 healthy controls; 625 clinically diagnosed with MDD). To shed light on dimorphic effects, we applied a computational pipeline combining ecological network modularity (SCNIC), linear modeling adjusted for compositional, zero-inflated data (LinDA), and explainable non-linear machine learning. Linear models explicitly incorporating a sex-diagnosis interaction term revealed that the relationship between MDD and the genera Bacteroides and Gallintestinimicrobium is sexually dimorphic. Community-wide beta diversity indicated that global microbial restructuring in MDD is subtle, but non-linear Random Forest classifiers paired with SHAP analyses clarified sex-stratified effects, and we show that male and female MDD phenotypes present distinct microbial profiles. Specifically, Eggerthella and Enterocloster are strong probabilistic predictors of MDD in females, whereas Turicibacter and Agathobacter have greater importance for classification in males. BMI and age acted as major confounding predictors in both cohorts. In summary, the combined linear and non-linear approach provides further evidence of sexual dimorphic microbial profiles in MDD gut microbiome characterization.

A tumor’s response to therapy is driven by a complex interplay of tumor-intrinsic and -extrinsic factors, with the tumor microenvironment (TME) serving as a critical extrinsic determinant. The TME is highly complex and dynamic, significantly influencing disease progression by contributing to both anti-tumor immunity and tumor-promoting immunosuppression, primarily through its diverse immune cell components. Immunotherapies, particularly immune checkpoint inhibitors (ICIs), enhance anti-tumor functions, largely by restoring exhausted T cell activation and overcoming immune evasion. However, response to these therapies remains limited and exhibits large heterogeneity among patients, highlighting the need to better understand the underlying mechanisms driving therapeutic efficacy and resistance.

To address this need, we conducted a comprehensive transcriptomic analysis integrating bulk RNA sequencing (n=5,078) and single-cell RNA sequencing (n=340) datasets of patients with solid tumors treated with ICI therapy for which we have extensive clinical information. This approach enabled us to characterize the TME landscape across large clinical cohorts. We identified specific relative immune cell compositions present in TME prior to therapy that serve as predictors of therapy outcome. Furthermore, we mapped the compositional changes within the TME induced by ICI treatment. Specifically, our investigation focused on the distinct roles of lymphocyte subpopulations and their complex cellular interactions within the TME, providing novel insights into the mechanistic drivers of ICI response. These results reveal informative biomarkers linked to the TME and highlight potential new therapeutic targets to improve patient outcomes.

Metabolomics is the comprehensive study of all metabolites and their interactions in a biological system. Pathway and enrichment analysis is frequently used in metabolomics studies to identify over-represented biochemical entities relative to a defined background, linking metabolites to their broader biological context. BiNChE is an ontology-based chemical enrichment analysis tool based on the ChEBI hierarchy. Here we present an updated version of the web tool, ChEBI-N (ChEBI eNrichment), which retains core functionality including graph-based visualisations of enriched entities and multiple pruning strategies. This allows researchers to identify the most significant chemical classes related to their metabolites of interest, for example those associated with a specific disease state. We introduce two key updates. First, input metabolites are no longer restricted to ChEBI identifiers and can instead be submitted as SMILES strings, which are automatically mapped to parent molecules if no direct ChEBI match exists. This broadens the range of supported compounds and lowers the barrier to entry for users unfamiliar with ChEBI nomenclature. Second, a narrower, curated background dataset is now available, improving the precision and biological relevance of enrichment results for specific metabolite sets. Together, these updates make ChEBI-N more accessible and analytically flexible for metabolomics research.