Chemical language models (CLMs) use molecular strings such as SMILES as direct model inputs. However, the same molecule can be represented by multiple valid strings, and even “canonical” SMILES are not uniquely defined across software toolkits. This raises the possibility that notation variance, when not harmonized, affects learned representations and benchmark evaluation.
We investigated how inconsistent SMILES influence CLM behavior. First, in a survey of 264 CLM-related papers indexed in PubMed, about half did not explicitly report their canonicalization procedure, indicating limited transparency in handling notation variance. We then examined public benchmark datasets and found substantial heterogeneity in molecular notation, including redundant aromatic forms and frequent omission of stereochemical annotations. Using a molecular translation model trained on RDKit-standardized SMILES, we compared dataset-provided Raw SMILES with standardized inputs. Translation accuracy was consistently lower for Raw SMILES, indicating impaired structural recovery. In addition, Levenshtein distance between paired input strings positively correlated with L2 distance between their latent vectors, showing that character-level differences in notation were associated with divergence in representation. In several property-prediction tasks, performance degradation was limited, suggesting that downstream models can partially absorb unstable latent components. However, in ClinTox and BBBP, Raw SMILES yielded higher AUROC, and further analyses indicated that this apparent improvement was driven by notation-dependent confounding rather than improved chemical learning.
These results show that unharmonized notation variance can influence learned representations and, in some cases, distort benchmark evaluation in CLMs. Careful control and explicit reporting of molecular notation are therefore important for reproducible molecular machine learning.

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a complex metabolic disease associated with multiple systemic comorbidities and poses a major public health challenge. Although growing evidence suggests alterations in the gut microbiome and its derived metabolites contribute to MASLD development, the mechanisms underlying microbe-metabolite-host interactions remain incompletely understood. Here, we present an integrative multi-omics workflow for linking gut microbial composition and functional potential with microbial metabolites and disease-related host phenotypes.
We applied this workflow to the FLORINASH cohort, comprising 662 non-diabetic individuals with MASLD from Spain and Italy, spanning a range of obesity levels and liver disease severity. For each participant, matched shotgun metagenomic sequencing, untargeted UHPLC-MS metabolomics, and clinical phenotyping data were available. Metagenomic data were used to generate species-level taxonomic profiles and KEGG Orthology-based functional annotations, while gutSMASH was applied to predict microbial metabolic gene clusters across the identified gut species. These results were then integrated with metabolomic and clinical data to characterise microbe-metabolite-host associations.
Using this workflow, we identified a candidate phenol-associated host-microbiome interaction in MASLD. Phascolarctobacterium was implicated in hydroxybenzoate-to-phenol metabolism, with its abundance positively associated with marker metabolites of phenol metabolism. Mediation analysis further suggested that both Phascolarctobacterium abundance and related KEGG Orthology gene counts were linked to kidney function-related clinical features via phenol-associated metabolites in MASLD.
Together, these findings highlight the utility of our multi-omics workflow for uncovering microbe-metabolite-host interactions and suggest its potential for biomarker discovery and microbiome-targeted therapeutic strategies.

Clinical machine learning models are highly vulnerable to confounding bias, often exploiting systemic variables rather than intrinsic pathological signals. Sjögren's Syndrome (SS) exemplifies this challenge as physiological aging and polypharmacy can create a confounding profile, rendering elderly controls clinically similar to autoimmune patients. We hypothesized that systematic computational disentanglement could expose this bias and recover a biologically interpretable disease-related signature. In this study, we propose a confounding-aware multi-stage computational framework and apply it to an observational cohort of 196 women. The framework integrates Principal Component Analysis (PCA), Random Forest with SHAP, and multivariable regression. Unsupervised PCA revealed a latent metabolic and pharmacological phenotype in controls that simulates autoimmune xerostomia. Strikingly, standard classifiers (AUC 0.92) relied heavily on these comorbidities, indicating shortcut learning. By ablating this systemic signal, we isolated a de-confounded oral model that retained substantial diagnostic performance (AUC 0.82), primarily driven by cumulative dental damage and stimulated salivary flow. Multivariable regression showed that SS diagnosis and xerogenic medication use were independently associated with lower stimulated salivary flow, whereas age showed a smaller negative association. The magnitude of the disease effect was comparable to 28 years of age-related decline, although this comparison should be interpreted as a coefficient-based approximation. Age-stratified analysis revealed reduced diagnostic performance in patients older than 70 years, suggesting increasing phenotypic overlap between physiological aging and disease. Overall, unadjusted algorithms may overestimate diagnostic performance in complex clinical settings. These findings highlight the need for interpretable, confounding-aware workflows to capture disease-related signals and support reliable, age-aware clinical decision-making.

In prokaryotes, especially bacteria, signal transduction is often carried out by two components systems (TCS). The classical TCS is made of an histidine kinase (HK) and a response regulator (RR). Upon signal recognition, the HK autophosphorylate on a specific histidine residue and then transfers its phosphate group to its partner RR, who will regulate gene expression through various means. It has been shown that TCS are involved in response to many environmental stimuli, like light sensing or biochemical changes. Studying these systems can help understand the different types of signals a cell can perceive and how it will adapt to them. Several classifications of HKs have been proposed, based on conserved motif in their C-terminal end. One such motif is the H-Box that contains the phosphorylable histidine. These classifications where made in the early 2000s and does not include the diversity of sequenced organisms (and their TCS) that are available nowadays. We propose here a new way of classifying HKs, based on the neighborhood of the phosphorylable histidine, using protein Language Models (pLM). We use Kernel Principal Component Analysis (Kernel PCA) to reduce the dimension of the embeddings produced by the pLMs, followed by an unsupervised clustering. With this approach we are able to group HKs accordingly to the existing classification schemes and to identify clusters of HKs that could form new groups previously unseen. We plan to extend this classification method to RRs, opening new perspectives for motif-centered exploration of signaling proteins and beyond.

New plant breeding technologies rely on identifying genes associated with key agronomic traits, such as yield, fruit size, or disease resistance. Quantitative trait loci (QTL) studies have been important in linking genomic regions to these complex traits. However, QTL regions can contain hundreds of genes, which hinders further experimental validation. With QTL data available for over 400 species and more than 1,000 plant genomes, there is an unprecedented opportunity to integrate cross-species information through Comparative QTLomics. Here, we present a phylogeny-aware tool, currently under development, that integrates QTL data with phylogenomic relationships and functional evidence to improve candidate gene prioritisation. As a proof of concept, we applied this approach to the trait fruit size using QTL data from tomato, bell pepper, melon, and watermelon. Preliminary results show that Comparative QTLomics can refine candidate gene lists, recovering genes with known functional roles while also identifying novel candidates for future validation. This work further emphasises the importance of using a consistent ontology to standardise QTL data and complementary functional information to improve gene prioritisation. Overall, the framework provides an alternative strategy to use existing QTL information to accelerate the identification of genes underlying complex traits, including those lacking prior annotation.

Topic models have emerged as a popular paradigm to analyse and interpret complex single-cell and spatial data. Yet, current implementations are usually data-type specific and rely on different modelling and estimation approaches, hindering usability and interoperability. In this work we introduce TopOmics, a library to perform efficient and flexible topic modeling with any combination of -omics data, including novel spatial multi-omic data. The framework leverages standard libraries of the Python ecosystem, guaranteeing seamless integration with existing pipelines, and shows competitive performance against state-of-the-art methods while preserving interpretability. We provide several examples of TopOmics on diverse data sets, demonstrating the usefulness of a unified framework for modern bioinformatic analyses.

Integrating metabolomics with other omics and computational models
provides a holistic view of biological systems. MetaNetX
(https://www.metanetx.org) plays a key role by reconciling metabolites
and biochemical reactions across major databases (ChEBI, HMDB,
KEGG, MetaCyc, Reactome, SwissLipids, LIPID MAPS, etc.), creating a
unified namespace for annotation and cross-referencing.
Despite challenges such as incomplete stereochemistry and database
inconsistencies, MetaNetX uses molecular structures and reaction
context to improve accuracy, with manual curation for critical cases.
The resource is freely available as raw files, via SPARQL, and through an
ID mapping tool, supporting systems biology and metabolomics
research.

Background
Prostate cancer (PC) is one of the most common cancers in men. Understanding how genomic profiles map to transcriptional programs and clinical phenotypes is central to patient stratification. Homologous recombination deficiency (HRD) defines a clinically relevant subset of metastatic (m) PC; however, DNA-based HRD markers (including BRCA1/2 alterations) have limited discriminatory power.

Methods
Matched WGS and RNA-seq data from >300 mPC patients (Hartwig Medical Foundation) were analysed using differential expression and GSEA to test whether HRD can be resolved through integrated genomic and transcriptomic features. To investigate whether similar transcriptional profiles arise via distinct evolutionary pathways, we developed a computational framework to estimate tumour clonality by integrating somatic SNVs, SVs, and CNVs.

Results
We identified 31 HRD cases, including 19 with BRCA2 alterations. GSEA revealed enrichment of an RB1-loss signature in HRD tumours; however, this signal was also present in TP53-altered tumours (n = 135), limiting specificity. HRD tumours showed limited subclonal diversity (clonality >0.9; 1 = fully clonal), consistent with early genomic instability and clonal fixation. In contrast, non-HRD tumours exhibited subclonal heterogeneity (<0.9), indicative of ongoing evolutionary diversification. TP53 and RB1 alterations were associated with increased subclonal burden. Differences remained significant after adjusting for tumour purity and treatment exposure.

Conclusions
HRD tumours show restricted subclonal diversity, whereas non-HRD tumours with TP53 or RB1 alterations exhibit greater subclonal complexity, consistent with poor prognosis. Integrating subclonality with driver alterations and transcriptomic profiles improves patient stratification beyond genomic alterations alone.

Single-cell RNA-sequencing (scRNA-seq) offers powerful insights into cellular responses across tissues and disease states. We present an integrated atlas of 390,000 mouse cells across multiple tissues (heart, carotid, lung, and brain) and disease models, including heart disorder, stroke, and lung injury. To harmonize technical and biological variation across datasets, we applied scVI to integrate raw read counts and effectively removed technical batch effects.

Cell type annotation was performed using curated marker genes from the literature. We compared disease-associated gene signatures with gene–disease associations from DisGeNET, enabling the identification of overlapping and condition-specific genes.
The atlas provides a versatile framework for downstream investigations.

It enables exploration of shared and unique transcriptional programs across diseases, identification of disease-specific cellular states, and analysis of conserved patterns across conditions. Importantly, the learned latent representations from this atlas can be applied to spatial transcriptomics data, allowing inference of disease presence. Another application is the inference of dynamic transcription factor (TF) networks from time-series gene expression data, allowing the identification of stage-specific regulators across disease progression.

To ensure broad accessibility, we developed a user-friendly web-based platform, the VDA app, which enables intuitive exploration and analysis without requiring programming expertise. Users can explore the atlas and its rich metadata, perform differential expression analyses, and investigate gene functions across specific cell types and disease conditions.

Single-cell sequencing (scRNA seq) has transformed our understanding of biological systems by revealing extensive cellular heterogeneity and diverse cell states across time. However, these measurements only provide us snapshot observations of the system. Mapping these discrete snapshots onto continuous trajectories requires computational modeling. Optimal Transport (OT) has emerged as a useful framework to predict cell-state transitions by creating a coupling between two distributions across time-points. In particular, by using the unbalanced OT formulation it is possible to account for cell growth and death by applying a penalty for the imbalance of mass between time-points.
However, the impact of this relaxation and its related parameters have not been systematically analyzed. We created a simulated dataset with known ground truth to investigate how modeling assumptions, including prior estimate of birth-death rate, distance between cells and relaxation parameters for OT, impact the resulting transition maps. Our computational experiments revealed that in developing systems with evolving cell-type composition, OT successfully identifies cell-type transitions. However, in a stationary system where cell-type composition stays the same across time-points, OT performs well in capturing cell growth and death but often fails to reliably capture transitions to other cell types.
To address this problem we introduced biologically informed weights using a Reference Measure. This results in better disentanglement between cell-type transitions and growth dynamics while preserving the convex structure of the problem. Our findings provide a critical evaluation of the current OT approach and offer practical solution to improve the performance of OT in trajectory inference.

Translational research routinely requires integrating clinical records, follow-up measurements, treatment histories and molecular data. These layers are typically stored as independent tables with implicit temporal links, limiting cross-domain querying and forcing study-specific schemas to be rebuilt for each new cohort. Currently, although each type of data has well established representation formats, there is no standard data schema supporting their integration, which leads to redundant engineering developments.
We are designing a generic, modular RDF schema that decouples the core data structure from domain-specific ontology choices. The backbone models four reusable entities: subjects, clinical events, biological samples and molecular measurements, linked by relations that make the time ordering of events explicit and queryable. Disease-specific vocabularies are kept in a separate ontology layer, allowing the same backbone to be reused across pathologies and data types.
TSV files structured according to the schema are imported into AskOmics, enabling visual SPARQL querying. The framework is being instantiated on a longitudinal Crohn's disease cohort, integrating four clinical tables spanning nearly 3,000 patients and over 40,000 follow-up records. Microbiome-derived functional data are being added through the same modular mechanism.
The resulting knowledge graph enables queries combining semantic and temporal constraints, for example retrieving samples collected within defined treatment windows and stratifying them by disease activity, which would require complex multi-table joins in a relational approach. Decoupling the data backbone from domain ontologies reduces schema redesign effort for new cohorts, supports integrative queries across clinical and molecular layers and favours reproducibility across translational studies.

Learning meaningful representations of single-cell resolution histology (H&E) images remains limited by small, heterogeneous datasets and inconsistent annotations, constraining the development of robust image-based cell representations. Recent work has proposed diverse architectures for training models on histology paired with spatial transcriptomics data. We address these challenges by constructing a large-scale, standardized resource and systematically assessing how training strategies impact downstream performance.
We curated and harmonized public datasets to assemble over 20 million individual cells with paired H&E image patches and spatial transcriptomic profiles. Cells were annotated with hierarchical cell type labels at multiple levels of granularity. We trained models using contrastive learning objectives to align image and transcriptomic embeddings and benchmarked them against established cell-level image embedding approaches across tasks including cell type classification and gene expression prediction.
While patch-level histology foundation models such as UNI2 and Virchow2 generalize well across samples, specialized models such as CellViT exhibit strong sample-specific effects. Contrary to expectations, integrating data across diverse tissue types did not improve general-purpose single-cell image embeddings and in some cases reduced performance. In contrast, tissue-specific training consistently improved cell type classification (+0.45 ARI) and gene expression prediction (+0.1 PCC), indicating that morphological–molecular relationships are context-dependent.
These findings highlight the importance of tissue context for learning effective image embeddings and provide a foundation for scalable, tissue-aware multimodal learning in computational pathology.

Type 2 diabetes (T2D) is a highly heterogeneous disease that requires a personalised approach to treatment for each patient. Several attempts have been made to identify subgroups within T2D; one of the most common is based on clinical data. However, to enhance subgroup analysis, additional molecular layers need to be incorporated. Multi-omics data, which combines genomics, proteomics and metabolomics data, is available from cohort studies and may provide insights into T2D subgroups that are not captured by clinical data alone. In this study, our aim is to identify T2D subgroups using multi-omics data from a population-based cohort, and to compare the performance of different integrative multi-omics clustering methods. We applied three algorithms: Similarity Network Fusion (SNF), Multi-Omics Factor Analysis (MOFA2) and iCluster+. For MOFA2, latent factors were further clustered using k-means clustering. The clustering results were evaluated based on structure (number and size of clusters), quality metrics such as the silhouette score, and computational performance, including runtime. Additionally, we assessed the biological interpretability of the results by identifying features associated with each cluster.

Cell-cell communication is crucial for the development of complex multicellular organisms. In systems biology, computational biology, and biomedical research, elucidating cell-cell interactions requires a comprehensive reference database of robust, physiologically relevant ligand-receptor interactions. However, existing biological resources suffer from fragmented coverage, lack pathway and cellular context, and show poor cross-resource consistency. To address this, we have begun a targeted biocuration effort using the Gene Ontology Causal Activity Model (GO-CAM) framework. GO-CAMs integrate gene product activities into causally connected, machine-readable models by using three complementary ontologies: Molecular Function (MF), Biological Process (BP), and Cellular Component (CC) from the Gene Ontology (GO). Applied to receptor-ligand biology, GO-CAM connects interactions to downstream signalling pathways, specifies the cell types in which they occur, and captures microenvironmental context such as subcellular localisation and co-interacting proteins. Of the more than 1,600 human receptors catalogued in UniProtKB/Swiss-Prot, those with known ligand interactions will be systematically represented as GO-CAM models, using LLM-assisted literature screening to identify and extract the relevant experimental evidence. To date, approximately 400 receptor-ligand pairs have been curated and integrated into the GO-CAM framework.

Single-cell transcriptomics are scalable and widely used, but proteins remain closer to cellular function and phenotype. This project aims to bridge that gap by inferring protein abundance directly from RNA data, approximating protein-level insights whilst significantly reducing the associated experimental costs.

We present the second generation of our previously published method, named scLinear2, that translates transcriptomic profiles into predicted protein signatures and evaluate its ability to capture biologically meaningful variation. ScLinear2 performs at state-of-the-art accuracy levels across different datasets, while employing an efficient and interpretable linear regression approach. We demonstrate that predicted protein profiles can support downstream analysis tasks such as automated cell-type annotation by training tissue specific models. Moreover, we utilize the interpretable nature of scLinear2 to characterize the most informative genes for each protein across datasets, revealing interesting mechanistic insights. Of note, ScLinear2 is developed as both R and Python package and is directly compatible with the most widely used single-cell analysis workflows, allowing for its easier adaptation.

Taken together, we present scLinear2: an efficient and interpretable cross-modality prediction method. Our results highlight the potential downstream applications of conversion between different omics modalities whilst conserving much of the functional differences between cells. Future work includes the application of scLinear2 in clinical data, which will allow the translational exploration of inferred protein levels.

Spatial omics enables the analysis of gene expression and cell states while preserving tissue architecture, but high costs make comprehensive profiling across whole tissue sections difficult. In practice, regions of interest (ROIs) for spatial assays are often selected by visual inspection of H&E slides. However, some functionally relevant regions, including those associated with intratumor heterogeneity and the tumor immune microenvironment, cannot be identified from morphology alone. Here, we propose a framework that integrates pathology foundation models with supervised regression to identify and recommend biologically informative functional ROIs for spatial transcriptomics from H&E images. Using paired H&E and spatial transcriptomics data from HEST-1k, we define transcriptomics-derived functional scores and train models to predict regional scores from H&E embeddings extracted by OmiCLIP. In our experiments, our method showed higher correlation with transcriptomics-derived targets than a gene-query-based OmiCLIP zero-shot baseline. Moreover, ROI selection based on predicted scores recovered regions close to the optimal regions defined from ground-truth scores. These results suggest that the proposed framework identifies functionally informative hotspots beyond morphological appearance and may improve ROI design for costly spatial omics experiments.

Spatial transcriptomics technologies record genome-wide measurements of gene expression with high spatial resolution. These technologies generate large and high dimensional datasets requiring efficient automated methods for their analysis. In this study we introduce joint spatial PCA (jsPCA), a novel, fast, scalable and interpretable method for the automatic identification of spatial domains and variable genes in multi-slice and multi-sample spatial transcriptomics data. jsPCA relies on a simple mathematical formulation of a spatial covariance defined as the product of the gene expression covariance with the spatial autocorrelation. The principal components of this spatial covariance yield a biologically meaningful low-dimensional representation. From this representation, we can derive spatial domains by simple clustering. In addition, spatially variable genes can be identified directly from the principal components coefficients. Moreover, this approach enables the joint representation of multiple slices and samples, a frequent experimental setting. This joint representation is obtained without spatial alignment by computing common principal components via joint diagonalization of the set of spatial covariance matrices obtained for each slice. By leveraging data sparsity and non-convex optimization on manifold, jsPCA leads to computing time in the order of seconds to minutes, substantially outperforming state-of-the-art approaches. We benchmarked jsPCA on the Visium 10x dataset of human dorsolateral prefrontal cortex and the Stereo-seq MOSTA dataset of mouse embryonic development against 10 state-of-the-art methods. Our approach demonstrated excellent performances, comparable or better than state-of-the-art methods, such as SpatialPCA, BASS, GraphPCA or Stagate, while being much faster, interpretable, and scalable to very large datasets.

Understanding molecular function requires accurate modeling of interactomes within specific cellular contexts, yet experimental observations of context-dependent interactions remain sparse and difficult to generate at scale. Computational approaches are therefore essential for inferring missing interactions and uncovering the principles that shape context-specific protein–protein interactions (PPIs). Here, we introduce ProtScape, a novel multi-scale framework for context-specific representation learning across proteins, cells, and tissues. We first construct contextwise PPI networks for 207 cell types, using a biologically grounded gene-selection strategy that integrates transcriptomic profiles from diverse human tissues, neuronal populations, and specific cellular states in amyotrophic lateral sclerosis (ALS). ProtScape then learns representations across contexts and scales, combining context-free protein foundation model embeddings, a novel hierarchical graph neural architecture tailored to the resulting heterophilic graphs, and a masked graph autoencoding learning strategy. To reliably evaluate the biological information captured by our representations despite the absence of context-specific protein annotations, we introduce a weakly supervised multiple-instance learning framework for protein contexts, rather than assuming that protein-level labels are uniformly valid across cellular contexts.
ProtScape achieves substantial improvements in PPI prediction, while revealing context-dependent interaction rewiring, including in RNA-binding proteins implicated in ALS. Moreover, our curated downstream evaluation protocol shows that ProtScape captures strong biological signals related to protein complexes and therapeutic targets, improving the prioritization of disease-relevant candidates as validated in Parkinson’s, while enabling robust identification of relevant cellular contexts. while enabling robust identification of relevant cellular contexts. Together, these results position ProtScape as a scalable foundation for context-aware interactome modeling. Future extensions toward multimodal integration, including RNA-protein interactions, will further enable the study of regulatory mechanisms underlying complex diseases such as neurodegeneration.

Background: Predicting transcription factor (TF) interactions in biological systems presents a major computational challenge critical for deciphering regulatory logic. This is particularly important in contexts like aggressive brain tumors (e.g., Glioblastoma, GBM), where tumor subtype-specific regulatory networks orchestrate therapeutic resistance. Advanced computational modeling is needed to accurately predict cooperative TF-DNA binding in specific cellular contexts.This study details an enhanced computational tool to model cooperative binding of multiple TFs to target regulatory regions (promoters and enhancers).

Materials and methods: We significantly enhanced the Composite Module Analyst (CMA) algorithm, a method originally designed to predict TF binding compositions to promoter regions. The primary computational enhancement involves a redesigned fitness function incorporating gene expression levels (RNA-seq) and modeling factor-factor interaction dynamics. The new CMA was trained on a specific GBM dataset to identify tumor subtype-specific TF composite modules, which serves as a proof-of-concept application for its ability to correlate computational predictions with RNA-seq data.

Results: Our enhanced CMA demonstrated a marked ability for identifying cooperating TFs binding to their composite sites. Application of this robust model to the GBM subtype data successfully revealed distinct sets of key TFs (e.g., SP100, TP53, MITF, MeCP2) that are active in different tumor microenvironment contexts.

Conclusions: This work presents an advanced computational tool for modeling complex transcription regulatory landscapes. Our model provides a novel, mechanistic approach to identifying key cooperating TFs in subtype-specific regulatory networks . This advancement is applicable for dissecting the transcriptional logic of any disease context and highlights the identified cooperating TFs.

Despite advances in treatments, cancer remains the leading cause of mortality among children in Switzerland. The implementation of precision oncology for pediatric malignancies has been hindered by the lack of standardized analytical frameworks to process high-throughput multi-omics data within actionable timeframes. Here we present the tumor profiling and analytical infrastructure of KidsCan, a national pediatric precision oncology initiative.

Within the KidsCan-01: Swiss Personalized Oncology for Children flagship study, we develop reproducible bioinformatics pipelines to profile pediatric tumors and provide decision support. Our framework integrates a whole-genome sequencing pipeline to prioritize oncogenic single nucleotide variants, insertions/deletions, structural variants, and copy number alterations. This is paired with an bulk RNA-sequencing pipeline for gene expression quantification, tumor-type classification, and fusion detection.

To ensure reproducibility required for clinical decision support, we implement version control and containerize all pipeline modules using Docker, guaranteeing consistent execution across systems. Within this robust framework, we implemented a unified quality-control pipeline that evaluates sample integrity and analytical performance across sequencing modalities. We validated and refined our pipelines using a series of benchmarking experiments with multi-omic data and tumor profiling reports processed by the INFORM initiative. Furthermore, to facilitate translational application, we engineer a reporting pipeline that aggregates multi-omics findings for the pediatric molecular tumor board, streamlining interpretation of molecular vulnerabilities.

By harmonizing molecular outputs with structured clinical metadata via FHIR standards, KidsCan not only optimizes clinical decision tools for immediate patient care but also pioneers a scalable, interoperable data ecosystem to propel future pediatric cancer research in Switzerland.

RNA subcellular localization is a fundamental biological phenomenon that underlies processes such as cell polarity formation and differentiation. Importantly, localization patterns change in response to intracellular and extracellular environments as well as cellular states, and their disruption is known to contribute to neurological disorders and cancer. Therefore, elucidating the mechanisms of RNA localization is an important challenge for understanding complex biological regulatory networks and overcoming disease. To date, methods for predicting RNA localization have primarily relied on RNA sequence information. However, these approaches have mainly focused on predicting population-averaged (bulk-level) RNA localization across multiple cells, making single-cell-level prediction difficult. As a result, they have been unable to fully capture cell-state-dependent variations in RNA localization. In this study, we developed a deep learning method to predict RNA localization at the single-cell level by integrating RNA sequence information, cellular imaging data from Cell Painting, a multiplexed morphological profiling assay, and spatial transcriptomics data. Compared with a baseline model that outputs the average localization proportion in the training data, our method predicted RNA localization proportions more accurately. In addition, the model predicted with high correlation the localization proportions of individual cells averaged across the gene dimension, suggesting that it can capture cell-specific average RNA localization tendencies reflective of cellular state. This study demonstrates the potential of integrating RNA sequence and cellular state information to enable single-cell-level prediction of RNA localization and contribute to a better understanding of cell-state-dependent mechanisms of RNA localization control.

Single-cell RNA sequencing (scRNA-seq) has greatly advanced our understanding of cellular heterogeneity, and recent advances in long-read scRNA-seq have made it possible to resolve isoform expression at single-cell resolution. In parallel, deep generative models have provided a powerful framework for learning latent cell-state representations from high-dimensional single-cell data. Because transcript isoforms can alter coding potential, RNA stability, subcellular localization, and regulatory interactions, isoform diversity is a key layer of gene regulation underlying cell identity, differentiation, and disease. However, existing methods still lack a unified framework for integrating isoform-level expression variation and sequence differences, making it difficult to determine how sequence-defined isoform usage contributes to cell-state-specific regulation and biological function. Here, we present a deep generative framework that jointly learns cell-specific isoform expression and sequence embeddings of full-length transcripts derived from a nucleotide language model. By linking latent cell-state representations with isoform sequence embeddings, our method enables integrated analysis of sequence-associated isoform variation across cells. Applied to long-read scRNA-seq data, our framework learned biologically meaningful cell-state representations while enabling estimation of isoform heterogeneity linked to sequence features. Furthermore, gradient-based attribution analysis and in silico mutagenesis showed that the model could identify nucleotide sequence regions important for predicting isoform variation. These results provide a computational foundation for linking isoform-defining sequence features to cell-state-dependent regulation at single-cell resolution.

Astrocyte reactivity is a central contributor to neurodegeneration, yet the morphological heterogeneity of reactive states and their relationship to transcriptional programs remain poorly understood. A key challenge is that biological conditions rarely correspond to a single, homogeneous phenotype, a complexity obscured in aggregate population analysis.
Here, we address this by generating a multimodal dataset of human iPSC-derived astrocytes from ALS patients and controls, profiled by high-content fluorescence imaging and bulk RNA sequencing under basal and controlled pro-inflammatory conditions. We develop SI-SimCLR, a spatially informed contrastive learning framework that learns biologically meaningful representations from microscopy images, without segmentation or predefined labels. SI-SimCLR outperforms standard baselines in capturing disease- and inflammation-associated morphological variation across experimental batches.
Unsupervised analysis of SI-SimCLR embeddings revealed a structured morphological landscape composed of twelve distinct substates. Using Optimal Transport to construct a morphological transition graph, we found VCP-mutant astrocytes occupy a constrained genotype-specific region under basal conditions, while high-dose inflammatory stimulation partially shifts these states toward control-like configurations. Substate-resolved analysis further highlighted untreated VCP-mutant astrocytes as cell-autonomous morphological states overlapping with inflammation-induced reactive phenotypes.
Integration with bulk RNA sequencing revealed a striking dissociation: while inflammatory stimulation dominates transcriptional variation, the ALS mutation primarily drives morphological organization. This indicates morphological and transcriptional responses to disease represent partially independent axes of astrocyte dysfunction.
Together, these results establish a scalable, annotation-free framework for high-resolution characterization of phenotypic heterogeneity, providing a principled foundation for substate-resolved analysis of astrocyte biology in neurodegeneration and beyond.

Translation initiation is an essential regulatory stage of protein synthesis, preceding elongation. This stage operates through two mechanisms, primary and secondary, involving complex interactions among multiple eukaryotic initiation factors (eIFs). In the primary mechanism, eIF2-GDP, eIF2B, and eIF5 function analogously to clutch, brake, and accelerator, regulating GDP-GTP exchange and the formation of the active eIF2-GTP complex. Further, the binding of Met-tRNA forms the ternary complex (TC), which serves as the clutch in the secondary mechanism. Understanding the operation of this clutch-brake-accelerator molecular system under the integrated stress response (ISR) is essential for studying the translational control. Despite extensive biochemical insights, it remains unclear how these interactions generate decision-making dynamics under both normal and stress conditions. Here, we develop a mechanistic mathematical model based on experimental observation to investigate how transitions between translation initiation and termination occur at the initiation stage. The model incorporates phosphorylation-dephosphorylation (PdP) reactions to model ISR as a fail-safe mechanism. Because the reaction network contains several unknown kinetic parameters, we use Chemical Reaction Network Theory (CRNT) to perform structural analysis and identify hidden positive feedback loops embedded in the initiation mechanism. Bifurcation analysis reveals ultrasensitivity and bistability under normal conditions. Under ISR, the system exhibits both bistability and tristability for physiologically relevant parameter regimes. We associate bistability with switching between initiation and termination, and tristability with recovery and attenuation during stress. Together, these results demonstrate that translation initiation operates as a threshold-driven switching system rather than a graded process, enabling rapid, robust decisions across varying cellular conditions.

Coding agents can author interpretable single-cell embedding models, here called blueprints, directly from the scientific literature, without ever training a computational model or being exposed to gene expression data. Such a blueprint is auditable by construction, but being purely prior-driven it captures only what the literature already describes and cannot, on its own, correct mismatches with a given dataset or surface structure the literature has not named.

Here, we let a coding agent refine the blueprint against data with structured feedback from a boosting autoencoder that performs structured dimensionality reduction with implicit feature selection. Its latent dimensions are interpretable, each linked to a small gene module, matching the gene-selected, named modules of the agent-authored blueprint. Trained on the data with the blueprint's axes as a prior, it learns additional axes that capture complementary structure the blueprint does not explain. Because the newly learned axes share the blueprint's gene-module vocabulary, contrasting them with the blueprint-anchored axes yields quantitative metrics and qualitative descriptors of cell-group gene programs and latent-space topology.

We first show that, across multiple datasets, agent-authored blueprints yield embeddings whose named axes faithfully discriminate the cell types they name and reach quality competitive with conventional, foundation-model, and program-informed baselines, while remaining batch-robust by construction. Building on this, we demonstrate the refinement on mouse cortex data, where the blueprint misses cell subtypes. The boosting autoencoder surfaces these as structure beyond the literature prior, and a coding agent folds them back into the blueprint as new axes that recover the missed subtypes.

LLMs encode large amounts of gene expression literature knowledge, offering a promising resource for downstream inference and the integration of single-cell RNA sequencing data. However, leveraging this knowledge remains challenging due to instability and limited interpretability in prompt-based LLM approaches. We suggest that providing a simulation task for synthetic single-cell data forces an agentic LLM system to encode structural relationships from the literature into quantitative models, thereby improving stability and interpretability. These models enable generation of synthetic data that can be contrasted with real-data patterns.

Specifically, our approach builds on an agent interacting with a modular simulation script in a two-step process. The simulation script contains a cell blueprint, defining the action space of the agent, and a statistical model generating synthetic cells represented as unranked lists of highly expressed genes. The agent is first tasked with updating the blueprint to reflect biological knowledge from retrieved literature. Next, the agent updates the blueprint based on patterns from real data. The resulting blueprints and their differences are directly interpretable, exposing modules, gene-level probabilities, temporal trends, and knockout effects.

We demonstrate the framework on a simulation task for microglia in the developing mouse brain with a HexB knockout. This simulation task allows to extract concise descriptions of literature knowledge. Assessing the variability of the derived blueprints across multiple agent runs highlights how uncertainty can be decreased by restricting the action space of the agent. Moreover, the variability quantifies how real data contains additional or diverging patterns compared to the literature.

Primary Immune Regulatory Disorders (PIRDs), which includes Common Variable Immunodeficiency (CVID) and Combined Immunodeficiency (CID), affect about 77,500 patients across Europe. They are characterized by immune dysregulation. Although 25% of cases have a monogenic cause, the majority lack a molecular diagnosis, leaving clinicians to rely on empirical, trial-and-error treatment selection that prolongs disease activity and increases healthcare burden.

We propose a multi-omics integration framework that combines genomics, transcriptomics, proteomics, epigenomics, microbiome, and clinical data for patient stratification and therapy prioritization. This approach uses trained models, like HyenaDNA and RNABERT to encode each type of data. Then it combines them through a fusion module to produce a patient-level embedding and contextualizes this embedding with a biomedical knowledge graph to output ranked drug candidates.

Classical multi-omics integration methods such as MOFA or DIABLO identify patterns in the data but are not designed to use external biological knowledge, which is crucial when patient cohorts are small. Coupling modality-specific embeddings with a knowledge graph helps compensating for this scarcity by injecting curated priors on gene-gene, gene-drug, and pathway relationships while also capturing interactions across immune layers that no single omic can reveal on its own.

By coupling multi-omics patient embeddings with a biomedical knowledge graph, this design aims to leverage scarce rare-disease data using structured biological priors. The approach is intended to enable precision medicine in PIRDs by generating interpretable, patient-specific therapy rankings.

The human microbiome is a key determinant of health and disease, but most reported associations remain descriptive and lack a system-level view. Multi-omic profiling can provide such a view, yet integration is challenging because omics differ in scale, structure, and biological meaning. Existing embedding-based approaches often either have limited predictive power or collapse all omics into a single latent space, losing omic-specific information.
We introduce PAPRICA (Phenotype-Aware Parallel Representation for Integrative omiC Analysis), a parallel encoder-decoder framework that embeds each omic in its own latent space while jointly modeling cross-omic and phenotype-associated relationships. PAPRICA is trained to reconstruct each omic, align samples across omic-specific latent spaces, and capture variation related to continuous phenotypes such as fecal calprotectin in IBD.
Across three multi-omic microbiome datasets, PAPRICA outperformed alternative integration models in both cross-omic prediction and phenotype prediction. These results suggest that PAPRICA effectively preserves omic-specific signals while capturing shared and phenotype-relevant structure, providing a flexible framework for microbiome multi-omic integration.

Recent advances in cross-species RNA-seq analysis have enabled the identification of conserved and species-specific features, contributing to both evolutionary biology and biomedical research. However, alignments based solely on RNA expression are limited in their ability to capture regulatory logic. With the emergence of multi-omics data, particularly chromatin accessibility measured by ATAC-seq, it has become possible to investigate upstream regulatory states and transcription factor activity. Integrating RNA and ATAC enables more direct characterization of transcriptional regulatory modules, which are often more conserved than gene expression patterns across species. Despite these advantages, multi-omics integration faces several challenges, including technical heterogeneity and the limited availability of comparable datasets. To address these issues, various methods have been proposed to enable integration across modalities that are not measured in the same cells (diagonal integration). However, these approaches do not adequately model species-specific differences or evolutionary changes in regulatory relationships. In this study, we propose a novel framework, Multi-omics Unified embedding across Species (MUSE), which constructs a cross-species graph by introducing inter-species edges based on orthologous relationships and protein sequence similarity, while preserving the regulatory graph structure within each species. Based on this integrated graph, MUSE aligns omics measurements across species, enabling biologically meaningful cross-species comparisons at both the omics and regulatory module levels.

Multi-omics technologies are essential for advancing systems-level understanding of biological processes through the integration of genomics, transcriptomics, proteomics, and metabolomics data. However, their adoption in research and industry remains limited by insufficient computational infrastructure, lack of standardized workflows, and limited bioinformatics expertise. In addition, the growing scale and complexity of multi-omics datasets require robust, reproducible, and scalable analytical solutions.
The MultiOmics Centre (MOC), a joint initiative of Laimburg Research Centre and Eurac Research, addresses these challenges by combining advanced experimental and computational capabilities. The centre provides the laboratory infrastructure and trained personnel for omics data generation, alongside a dedicated computational environment. This includes high-performance computing, data management systems, and standardized pipelines that support reproducible and FAIR-compliant multi-omics analyses.
To support genomic analyses, I developed a web-based application that streamlines and standardizes the execution of the in-house Nextflow pipeline for genome-wide association studies (GWAS). The application provides a user-friendly interface to specify input data and analysis settings, automatically handling configuration and pipeline execution. This reduces manual intervention, minimizes errors, and ensures consistent, reproducible analyses across studies. The underlying pipeline, nf-gwas, is designed for biobank-scale GWAS analyses and it automates pre- and post-processing steps, integrates regression modelling via the REGENIE package, and supports single-variant, gene-based, and interaction testing. It also provides comprehensive reporting, enabling the exploration of results across thousands of phenotypes.
By streamlining the use of the nf-gwas pipeline, the web application enables fast, scalable, and reproducible GWAS analyses, contributing to the MOC's efforts to accelerate the adoption of multi-omics approaches.

Oxidative stress, characterized by an imbalance between reactive oxygen species (ROS) and antioxidants, plays a pivotal role in inflammatory responses associated with both chronic diseases and acute injuries. In this study, OXidative Stress PREDictor (OxSpred), a supervised learning model tailored to accurately annotate the oxidative stress state of innate immune cells at the single-cell level, is introduced. Compared to the traditional gene-set-variation-analysis-based enrichment method, OxSpred demonstrates superior accuracy with an area under the receiver operating characteristic curve of 0.89 and offers interpretable embeddings with significant biological relevance. Using the predicted ROS states, precise elucidation and interpretation of the roles of novel innate immune cell subtypes can be achieved. Overall, OxSpred enhances the utility of single-cell transcriptomic datasets by providing a robust in silico method for determining intracellular oxidative stress states, thereby enriching the understanding of innate immune cell functions during inflammation.

We have found an activating mutation (M207T) in the early cell-cycle regulator cyclin dependent kinase 4 (CDK4) in a patient with severely disturbed adipose tissue distribution and metabolic syndrome. To understand the underlying mechanisms, we overexpressed the mutant CDK4 or pharmacologically inhibited the kinase, and conducted proteomic, phosphoproteomic and metabolic analyses during the first two days of adipogenic differentiation of human mesenchymal stem cells. Together, our data covering over 8000 proteins, 20000 phosphosites and 400 metabolites suggest that the mutant CDK4 induces major changes in the flux of glucose, including altered use of the TCA cycle. Here, we present two distinct computational approaches to dissect how CDK4 activity controls signal propagation during early adipogenesis. First, we portray conventional multi-omics integration methods, including matrix factorization and constrained based reconstruction to derive the downstream effectors of active CDK4. Second, we identify possible signal propagation pathways that explain how the observed metabolic changes are controlled by CDK4, by utilizing the SIGnaling Network Open Resource (SIGNOR) database to search for potential shortest paths starting from an activated CDK4 and leading to the observed effectors. We follow the binary activation/deactivation rules of the database to identify the expected activation of proteins along a path and match these expectations to our observed data at both protein and phosphorylation levels. Our approach shows the potential of integrating longitudinal multi-omics data with curated databases to reveal dynamic, intracellular signal propagation pathways that control the early steps of cellular differentiation.

Kinetic models of the Calvin-Benson cycle (CBC) are used to study photosynthesis, but their behavior depends strongly on parameter values that are often uncertain. In this study, we performed a sensitivity analysis on several published CBC models to assess how changes in parameters influence steady-state metabolite levels. Each kinetic parameter and metabolite concentration was perturbed individually by ±5% to ±25%, and the resulting effects on the system were quantified. In addition, selected pairwise perturbations were analyzed to examine interaction effects. Across all models, we observed that metabolic control is not evenly distributed. Instead, a relatively small number of parameters consistently exert a stronger influence on system behavior, while most parameters affect only one or a few metabolites. Key processes affecting model simulations include phosphate balance, adenosine triphosphate (ATP) synthesis, and specific Calvin cycle reactions such as 3-phosphoglycerate (PGA) reduction, sedoheptulose-1,7-bisphosphatase (SBPase) activity, and ribulose-5-phosphate kinase (Ru5P kinase), although the extent of this control differs between models. Overall, these results suggest that only a subset of parameters is critical for determining system behavior. Comparison of correlation structures across models reveals clear model-specific differences, with certain parameter pairs showing positive correlations in one model but weak, absent, or negative correlations in others. Correlation analysis further shows that parameters are not independent but organized into coordinated groups, reducing the effective dimensionality of the parameter space. Identifying these key parameters and dependencies can help simplify models and improve parameter estimation when experimental data are limited.

The UniProt Knowledgebase (UniProtKB) is a central reference resource for protein sequences and functional annotation supporting multiomics analyses. To improve integration of protein and small molecule knowledge, we completed a first-pass systematic reannotation of all human enzymes and transporters in UniProtKB/Swiss-Prot using Rhea, a curated resource of biochemical reactions structured on the ChEBI ontology.
All human enzymes and transporters were systematically reassessed and updated, resulting in reaction-based annotations for more than 4,100 human proteins. These annotations include balanced reactions where all molecules are linked to the ChEBI ontology, and atoms are mapped between reactants and products. This first phase establishes the foundation for comprehensive reaction-based annotation of enzyme and transporter functions and will support expanded analyses of metabolic processes and multiomics data integration in future releases.

* The UniProt Knowledgebase (UniProtKB; www.uniprot.org) is a central resource for protein sequences and functional annotations. However, protein annotation remains largely protein-centric, often limiting integrated and system-level pathway representations required for mechanistic understanding of biological processes.
* Here, we present a pathway-centric curation approach for the complete functional annotation of biological processes, using the visual phototransduction pathway as a model.
* Using a combination of UniProtKB queries, AI-based search strategies, and the identification of proteins associated with vision disorders, we identified 71 proteins mapped to the visual pathway network, spanning phototransduction in photoreceptors to retinal processing. Among these, 54 proteins are associated with vision disease variants. We are currently annotating each protein in the phototransduction signaling pathway with up-to-date protein functions, including standardized biochemical reactions (Rhea) within UniProtKB, and disease associated variants. The Gene Ontology Causal Activity Models (GO-CAM) framework is used to formally represent causal relationships between molecular activities, enabling machine-readable and computable pathway models.
* This work provides a standardized, computable representation of the phototransduction pathway and establishes a scalable approach for complete pathway annotation, supporting integrative analysis of molecular mechanisms and disease phenotypes.

Purpose: Anaplastic thyroid cancer (ATC) is among the deadliest malignancies, yet the spatial co-evolution of dedifferentiation and microenvironment remodeling during progression remains poorly understood.
Methods: We integrated bulk RNA-seq (n=1,634), scRNA-seq (n=106), and spatial transcriptomics (Xenium: 82 cores, 2.35M cells; Visium: 48 slides, ~94k spots). A 25D embedding of cell-type fractions and functional signatures was resolved by spatial Leiden micro-communities then KMeans (k=7), transferred to Visium via Elastic Net, and linked to genotype, IHC, and survival with patient-grouped cross-validation.
Results: Seven niches emerged — Perivascular, Quiescent, Immune effector, Immune suppressed, ECM remodeling, TAN infiltration, EMT transition — arranged on a monotonic low-to-high thyrocyte ATC-score axis. Perivascular and Quiescent showed mutual boundary enrichment (O/E 2.8–4.0); at the opposite pole, TAN infiltration enveloped EMT transition at the highest adjacency (O/E≈8.2, FDR<0.001), with cores ~88% ATC. Immune suppressed was 3.3-fold enriched in BRAF V600E and correlated with PD-L1 IHC (p<0.05); NRAS favored EMT/TAN. TAN and Immune suppressed conferred the highest Cox hazards (HR=6.71, FDR<10⁻⁶); an Elastic Net classifier reached AUROC=0.95 (patient-grouped CV) for ATC vs others. Within Quiescent (Q), dedifferentiation was clonal (Moran's I 0.15–0.39, FDR<0.05), driven by TACSTD2/MUC1 gain and EPCAM/CDH1 loss. TROP2/MUC1 retained epithelial identity (PTC>PDTC>ATC) and topped Q intra-heterogeneity drivers (Stouffer Z>17); their bulk composite stratified DFS in DTC (HR≈2.2, p=0.003) but not ATC, flagging DTC recurrence risk.
Conclusion: Niche-resolved mapping yields actionable stratification: Immune suppressed identifies PD-L1-high BRAF V600E potential ICI candidates; EMT/TAN mark NRAS-associated aggression for TKI or neutrophil targeting; Perivascular/Quiescent identify indolent disease with TROP2/MUC1 surveillance.

Amyloid positron emission tomography (PET) is commonly used to classify individuals as amyloid-β (Aβ) negative or positive, but binary approach may overlook intermediate stages of pathology. Here, we aimed to identify data-driven subgroups along the Alzheimer's disease (AD) continuum using Aβ PET imaging to enable a more nuanced characterization of amyloid accumulation.
We analyzed 3,110 Aβ PET scans from the ADNI and A4 cohorts using petVAE, a variational autoencoder trained to reconstruct two-dimensional PET slices without diagnostic labels or predefined regions of interest. The model generated 11,648-dimensional latent representations per scan, which were used for exploratory analyses and clustering across the AD spectrum.
We identified four clusters that differed significantly in standardized uptake value ratio (p < 1.64×10⁻⁸) and cerebrospinal fluid (CSF) Aβ levels (p < 0.02), indicating that petVAE effectively positions scans along a continuous Aβ trajectory. Two clusters (Aβ−, Aβ−+) were largely amyloid-negative, while two (Aβ+, Aβ++) were predominantly amyloid-positive. The extreme clusters (Aβ−, Aβ++) aligned with conventional classifications and showed marked differences in cognition, APOE ε4 carrier frequency, and CSF Aβ, Tau, and phosphorylated Tau (p < 3×10⁻⁶). Intermediate clusters exhibited increased odds of APOE ε4 carriership (p < 0.026). Individuals in Aβ+ and Aβ++ clusters had a higher risk of progression to AD over 6 years (hazard ratios 2.42 and 9.43; p < 1.17×10⁻⁷).
Overall, petVAE reconstructs PET images accurately while learning biologically meaningful representations of Aβ pathology. This data-driven framework enables detection of subtle disease stages and supports investigation of preclinical AD.

Organelle morphology provides a sensitive and integrative readout of cellular state, reflecting changes in metabolism, signaling, and environmental perturbations such as stress. While three-dimensional (3D) fluorescence microscopy enables detailed observation of organelle structure, translating volumetric image data into biologically meaningful conclusions remains challenging. Existing approaches often address only individual analysis steps and rely on summary statistics of morphometric descriptors, limiting their ability to capture population-level heterogeneity and perturbation-dependent remodeling.
Here, we present Morphomapper, a FAIR end-to-end workflow for quantitative analysis of organelle morphology from 3D confocal fluorescence imaging. Morphomapper integrates automated deep learning-based volumetric segmentation, extraction of geometric features, dimensionality reduction, and statistical inference into a reproducible and FAIR-compliant pipeline. We benchmark automated 3D segmentation against expert human annotations, demonstrating robust reconstruction of complex organelle shapes. Using the resulting segmentations, we systematically compare two-dimensional and three-dimensional morphometric descriptors and show that 3D features substantially improve the discrimination of stress perturbations. Importantly, Morphomapper treats organelles as populations rather than isolated objects. We introduce an optimal transport-based framework to compare distributions of organelle morphologies across perturbations, enabling principled quantification of stress-induced shifts in population structure. By operating on full morphological distributions instead of mean descriptors alone, Morphomapper captures heterogeneity as a primary signal and links structural remodeling to environmental perturbations. This framework bridges the gap between volumetric segmentation and biologically actionable inference, providing a scalable approach for identifying morphology-based biomarkers of stress perturbation states from 3D imaging data.

Drug discovery faces significant challenges — 90% of drugs entering Phase 1 clinical trials never reach the market. Open Targets, a public-private partnership, integrates human genetics and genomics data to systematically address these challenges through the Open Targets Platform (platform.opentargets.org).

The Platform integrates data from over 20 diverse sources to build and score gene-disease associations, supporting evidence-based target prioritisation for drug discovery. In its 10th year, the Platform underwent a major expansion incorporating the most comprehensive set of genetic associations to date, requiring significant advances in both data infrastructure and user-facing interfaces.

This update includes ancestry-specific fine-mapping and colocalisation of GWAS and molecular QTL data from the GWAS Catalog, eQTL Catalogue, FinnGen, and UK Biobank Pharma Proteomics Project. Over 2.6 million credible sets were generated, yielding more than 400,000 gene-disease associations supported by over 1 million evidence items.

To surface this complexity, new interactive interfaces were designed and built: variant-centric pages display credible sets, colocalisation results, and clinical and pharmacogenomic annotations for over 6.5 million variants. New study and credible set pages allow researchers to explore GWAS and QTL evidence in detail. Enhanced machine learning–based Locus-to-Gene (L2G) assignments further refine gene prioritisation, identifying more than 358,000 credible sets with L2G score >0.5.

These frontend-driven enhancements unify evidence from common and rare variant studies within a single explorable interface, enabling researchers to build stronger causal links between genes and diseases and supporting data-driven drug target identification.

Biologically informed neural networks (BINNs) embed pathway, ontology, or protein interaction structure directly into neural networks, promising interpretable disease prediction in which hidden nodes map to named biological entities. Yet BINNs have been difficult to train at biobank scale, and the reliability of their biological interpretations remains largely untested.
Here, we present a fast BINN implementation to train on UK Biobank genotype and plasma proteomics data from ~450,000 individuals across six common diseases. BINNs achieve competitive predictive performance, but we uncover two major limits to their interpretability. First, attribution scores are strongly biased by graph topology, so node degree and layer position explain most of the variance in the scores. A simple normalization reduces this bias, but can weaken enrichment for known disease genes. Second, BINNs exhibit substantial predictive multiplicity: independently trained models with identical architectures and data reach similarly accurate solutions while prioritizing different genes and pathways. We show that this multiplicity persists in highly predictive models even after reducing model parameters, network complexity and input correlation.
Although this multiplicity makes single-model explanations unstable, the range of different interpretations can also reveal complex disease biology. In 100 replicate BINNs for type 2 diabetes prediction from proteomics data, we find distinct solution clusters that prioritize either inflammatory or hepatic-metabolic pathways, mirroring known disease heterogeneity. Thus, training and analyzing large ensembles of BINNs can turn multiplicity into a tool for studying complex disease mechanisms in human cohorts.

Development is a complex process driven by coordinated cell proliferation, differentiation, and spatial organization. Classically, two ways to specify cell types during development have been hypothesized: the mosaic and regulative modes. In the mosaic mode, the fate of a cell rely on lineage-inherited factors. In contrast, in the regulative mode, the fate of a cell depends on space-dependent factors. The relative contributions of both modes remain poorly quantified. We present a novel approach to measure these contributions from single-cell data from C. elegans development. The invariant lineage of C. elegans allows the integration of spatial positions, lineage relationships, and protein expression data. Using single-cell protein expression profiles as a readout of cell state, we define two quantifiable metrics: 1) a proxy for the contribution of the mosaic mode, computed as the strength of the relationship between the cell-cell lineage distance and the cell-cell expression distance, 2) a proxy for the contribution of the regulative mode, computed as the strength of the relationship between the cell-cell context distance - capturing spatial neighborhood similarity - and the cell-cell expression distance. To validate these metrics, we compared empirical results from C. elegans to artificial models with defined developmental rules. Our analysis reveals the coexistence of mosaic and regulative modes, with their relative contributions varying across tissues and developmental stages. For example, in skin tissue, the mosaic mode dominates in early development, while the regulative mode prevails later. Our approach offers a quantitative, unbiased, and perturbation-free method to study fundamental principles of developmental biology.

Longitudinal microbiome data are key to understanding the dynamics of microbial communities and their relationships with the host and environment. However, analysis of such data is challenging due to high dimensionality, compositionality, irregular sampling and temporal dependencies on external covariates. Existing analytical approaches typically address only subsets of these challenges, limiting their ability to yield biologically interpretable insights. We introduce LGTM, a probabilistic modeling framework that combines flexible non-linear longitudinal modeling with interpretable topic-based representations of the microbiome. LGTM simultaneously discovers coherent microbial subcommunities (""topics"") and models how their abundances change over time and in relation to host and environmental covariates. Using multiple longitudinal human gut microbiome datasets, we demonstrate that LGTM identifies diverse and stable microbial topics while achieving competitive performance in imputation and forecasting tasks. A key strength of the framework is its interpretability: LGTM discovers biologically coherent microbial topics and directly quantifies associations between covariates and microbial dynamics. LGTM is available at https://github.com/yuanx749/lgtm.

Transcriptomic profiling provides a global, top-down view of cellular responses to environmental conditions, perturbations, and disease. Weighted gene co-expression network analysis (WGCNA) is among the most widely used methods to model these responses, grouping co-expressed genes into modules that can be related to sample traits and summarized by hub genes. However, because modern transcriptomic datasets quantify tens of thousands of genes, WGCNA’s quadratic scaling with gene number becomes prohibitive. This is especially problematic for resampling-based workflows, which rerun WGCNA across many bootstrapped datasets to assess module robustness or stabilize recovered modules.
A common way to reduce this cost is blockwise WGCNA, which partitions the gene set and analyzes it in chunks. Across the Tabula Muris Senis dataset, we find that blockwise WGCNA reproduces full-dataset WGCNA in its module–trait correlations while differing substantially in module gene content. With a block size of half the genes, the average Jaccard similarity between corresponding modules is only 0.6.
Here, we present tensor-wgcna, a GPU-accelerated PyTorch/Triton implementation that, on a single consumer GPU (RTX 4090), runs up to 10 times faster than the R implementation and uses roughly half the peak memory, while retaining a module Jaccard similarity of 0.9-1.0 across the Tabula Muris Senis and GTEx adult datasets.
Lastly, we demonstrate the practical relevance of tensor-wgcna on a hand-curated cold-stress dataset from Arabidopsis thaliana leaves (90 samples, 23,668 genes). Using bootstrap-based module stabilization, we not only recover functional modules but also align them with physiologically relevant phases of the cold response.

Whole-genome sequencing (WGS) has improved rare-disease diagnosis, yet 50–60% of patients remain undiagnosed, underscoring persistent challenges in causal variant prioritization. Single-omics outlier analyses, particularly RNA-seq, can increase diagnostic yield, but are often performed separately for each omics layer or combined only through late integration. We hypothesized that integrating phenotypic and multi-omics evidence within a unified network framework would improve causal gene prioritization.

We developed an integration framework that jointly models HPO-encoded phenotypes and patient-specific omics signals to prioritize causal genes. The method constructs a heterogeneous, multi-layer knowledge graph linking diseases, phenotypes, genes, and biological processes, with optional patient-specific omics subnetworks connected through gene nodes. For each patient, random walk with restart (RWR) is initiated from phenotype seed nodes to compute gene-prioritization scores. We first evaluated whether applying RWR on the knowledge-network backbone alone, without omics profiles, could serve as an effective gene-prioritization strategy. Performance was benchmarked against Phen2Gene and GADO using MyGene2 (N=146) and a subset of the Deciphering Developmental Disorders cohort (DDD; N=164). We then assessed a multi-omics extension in the SwissPedHealth cohort (N=37).

Our RWR framework improved causal-gene prioritization, ranking the causal gene in MyGene2 within the top 10 for 26.0% of cases, compared with 17.4% for Phen2Gene, and 2.1% for GADO. Performance was higher in the DDD subset (RWR: 86.6% top 10, Phen2gene: 73.0%, GADO: 39%). Adding multiomics evidence in SwissPedHealth further improved ranking beyond phenotype-only inference, highlighting the potential of RWR on heterogeneous graphs for propagation of phenotype and multi-omics signals in patient-specific gene prioritization.

Inflammatory bowel disease (IBD) remains a therapeutically challenging, molecularly heterogeneous condition for which most approved agents have been developed as pan-disease therapies, overlooking the distinct pathogenic drivers of ulcerative colitis (UC) and Crohn's disease (CD). To uncover subtype-specific therapeutic opportunities, we integrated bulk RNA-seq from seven independent cohorts — three CD cohorts (317 patients vs. 91 controls) and four UC cohorts (175 patients vs. 94 controls) — into an enhanced LINCS L1000–based drug repurposing pipeline. This analysis nominated two distinct, subtype-selective mechanisms of action: cholesterol-uptake inhibition (ezetimibe) for UC and histone deacetylase (HDAC) inhibition (vorinostat and pyroxamide) for CD. To systematically map the on- and off-target landscape of these candidates, we developed a graph convolutional network (GCN) model for drug–target interaction prediction, achieving area-under-the-precision-recall curves (AUPRC) of 0.89 and 0.95 on validation and test sets, and 74–80.5% top-100 accuracy when benchmarked against the Pabon reference dataset. Among the top 30 GCN-predicted targets for each candidate, 13 (ezetimibe), 6 (vorinostat), and 14 (pyroxamide) emerged as novel — neither annotated as direct drug targets in pharmacological databases nor previously linked to the corresponding IBD subtype. Together, these results nominate pharmacologically tractable, subtype-resolved therapeutic candidates and uncover previously uncharacterized targets supporting a stratified-medicine framework for IBD. Experimental validation of the prioritized drug–target pairs is ongoing.

Clinical microbiome databases document thousands of microbe–disease associations, yet most lack a curated molecular mechanism, and the evidence needed to trace one is scattered across resources with incompatible identifiers. Researchers have no automated way to determine which reported associations already have mechanistic support.

We present MicroKnow, a Neo4j knowledge graph integrating clinical microbe-disease evidence with mechanistic evidence across microbes, metabolites, host genes, and diseases (46,911 nodes, 738,096 edges), designed as a hypothesis-generation tool for experimental microbiome research. For each of 12,902 reported microbe-disease pairs, we ask whether the same microbe reaches the same disease through a metabolite it produces and a host gene that metabolite regulates.

Of 12,902 pairs, 17.4% recover such a path, a 1.11-fold enrichment over a degree-preserving null (z = 10.4, p < 0.001). However, restricting the comparison to pairs where a path could form shows reported and unreported pairs are equally likely to carry a mechanism (62%, p = 0.45), indicating the overlap reflects annotation density rather than biological concentration. MicroKnow maps the gap bidirectionally: 10,661 reported pairs lack any mechanism, while 13,394 mechanism-grounded pairs lack clinical validation. Only 5.8% of metabolites carry any documented gene interaction, making metabolite-to-gene curation the main bottleneck.

A Dash web interface (microknow.de) supports pair- and disease-level queries and returns a ranked shortlist of clinically unvalidated microbe–disease pairs as targets for experimental follow-up.

Macrophages represent a major immune cell type in tumor microenvironments, they exist in multiple functional states and are of strong interest for therapeutic reprogramming. While signaling cascades defining proinflammatory macrophages are better characterized, pathways that drive polarization in immunosuppressive macrophages are incompletely mapped. We exposed primary human macrophages to a range of stimuli, which is abundant in the tumor microenvironment, and profiled the induced transcriptome, proteome and phosphoproteome changes. We rank-normalized the gene expression levels and aligned the in vitro transcriptomes to a pan-cancer macrophage atlas built from single-cell RNA sequencing profiles collected from over 30 patient studies. This showed that in vitro macrophages were able to recapitulate different functional aspects of the in vivo states. For instance, exposure to adenosine resulted in the upregulation of a dozen of metallothionein genes and downregulation of the MHC-II antigen presentation system, which strongly resembled transcriptional state of Metallo Macrophages – poorly studied and tumor associated macrophages linked to bad prognosis. Furthermore, we mapped several under-appreciated kinase signaling cascades that play a role in the establishment of immunosuppressive states, such as those involving the PAK2 kinase, and we integrated the generated multi-omics profiles by mapping the significant hits to a knowledge-based interaction network, which allowed us to find modules of highly connected elements. The latter approach is available through the NOODAI web platform. Overall, this study contributes to in-depth multi-omics characterizations of macrophage phenotypic landscapes, which can be of relevance for assisting future interventions that aim to therapeutically alter immune cell compartments.

Constraint-based metabolic modelling tools — COBRApy, COBRA Toolbox, KBase — uniformly require local installation, cloud accounts, or queued HPC jobs, creating friction that impedes exploratory analysis and classroom use. Browser-based alternatives such as Escher-FBA and Fluxer exist but rely on server-side solvers or GLPK.js executing on the main UI thread, blocking interaction for models beyond ~500 reactions.

We present MetaboViz, an open-source React application implementing a three-tier compute architecture that automatically routes analyses to the fastest available solver: (i) a HiGHS WebAssembly Worker running off-thread in the browser, requiring zero installation; (ii) a local Python kernel (pip install metaboviz-kernel) communicating via WebSocket JSON-RPC, which loads the model once and keeps it resident across repeated interactive solves; and (iii) a stateless FastAPI edge service as fallback.

We benchmark all tiers on four BiGG models spanning 95–3,942 reactions. The local kernel achieves FBA solve times of 0.4–15 ms after a one-time model load (19 ms–1.1 s), and FVA of 19 ms–4.8 s using parallel dual-simplex. The edge tier is dominated by HTTP serialisation rather than solver time: iJO1366 (2,879 KB payload) incurs 1.0 s wall time versus 82 ms solver time, confirming that model caching in the kernel tier is essential for interactive use. Results are presented as Jupyter-style notebook cells with per-reaction flux visualisation and tier provenance.

MetaboViz enables sub-second FBA on genome-scale models in any browser, with no login or installation required, filling a gap between lightweight educational tools and full platforms such as KBase.

The mechanisms by which a defined effector commensal bacterial consortium promotes CD8+ T cell-mediated anti-cancer immunity remain unknown. We developed a multimodal immune profiling and data integration framework, coupled with a bacterial antigen prediction pipeline, to resolve how microbial signals are transmitted across intestinal and systemic compartments, supported by complementary in vitro and in vivo validation.
Two mechanistic axes were investigated: (1) molecular mimicry between bacterial peptides and tumor neoantigens, and (2) epithelial reprogramming by bacterial-derived products. For the first we developed a computational antigen prediction pipeline that interrogates genomic sequences of the effector consortium to identify cross-reactive bacterial peptides, followed by in vivo validation in mouse models. For the second, we integrated bulk RNA sequencing and FC of intestinal organoids exposed to bacterial supernatants. Systemic immune responses were further characterized by serum cytokine profiling and single-cell RNA sequencing of CD8 T cells from ICI-treated mice, integrating transcriptomic, TCR, and protein data.
While in vivo studies did not support molecular mimicry, pathway and cytokine modeling identified IFN-gamma signaling and responses to bacterial-derived molecules as dominant features of the effector condition. We observed epithelial-derived cytokines propagating systemically, together with consistent upregulation of Cxcl9 and Cxcl10 in serum, and increased expression of their receptor Cxcr3 in CD8 T cells. These findings define a coordinated chemokine axis linking microbial stimulation to effector T cell recruitment.
This integrative multimodal approach provides a better understanding of the biology of the complex host-microbiome system and establishes a mechanistic basis for enhanced anti-tumor immunity.

Neuronal aging pace varies markedly between individuals, but what drives this variation remains unknown. Using cell-type-specific transcriptomic clocks applied to single-nucleus RNA sequencing data from 226 adults (ages 20-90), we quantified neuronal aging residuals as a donor- dominant phenotype. Variance decomposition revealed that microglial transcriptional programs predict inter-individual variation in neuronal aging residuals, a directional asymmetry consistent with a non-cell-autonomous relationship between microglial states and neuronal aging trajectories. This asymmetry is accompanied by an age-dependent shift from homeostatic to inflammatory microglial dominance beginning in midlife, with inflammatory dominance probability rising from 26% at age 35 to 92% by age 65, replicated in an independent cohort. IFNγ signaling emerges as the dominant microglial program associated with accelerated neuronal aging in late adulthood. Candidate regulators of microglial IFNγ activity are computationally prioritized as intervention targets warranting functional validation.

Animal models are essential for studying nociception, yet differences in cellular and transcriptional architecture complicate translation to human pain biology. To address this gap, we developed a cross-species single-cell transcriptomic integration pipeline to build a comprehensive tongue atlas from mouse, rat, marmoset, and human datasets. Tongue tissue from naïve C57BL/6 mice and common marmosets was collected and processed for scRNA-seq using 10x Genomics platforms. Human datasets were obtained from the Tabula Sapiens Consortium, and rat datasets from NCBI GEO. A unified human gene space was constructed using species-specific ortholog mapping strategies — Ensembl-derived tables for mouse and rat, and NCBI Gene E-utilities for marmoset — yielding ~17,000 mapped genes. Species-aware batch correction was applied using Harmony with conservative parameterization to preserve biological divergence. The final atlas comprises 135,736 cells spanning 7 major cell types. To systematically quantify cross-species conservation, we implemented a multi-step pipeline using FindConservedMarkers (log2FC > 0.6, Bonferroni P < 0.05 across all species), identifying 1,342 conserved genes from 4,026 candidates. Normalized rank-based standard deviation analysis classified 51.8% of conserved genes as stable, including 206 highly stable genes. Epithelial cells showed the strongest transcriptional conservation (Pearson r = 0.83–0.86), while Schwann cells exhibited the highest variability. Validation against canonical markers confirmed conservation in 17 of 24 genes. This atlas and accompanying pipeline provide a scalable computational framework for comparative and translational studies of pain-relevant biology.

Parkinson’s Disease (PD) is often diagnosed after irreversible neurodegeneration due to reliance on physical symptoms for diagnosis. Current machine learning (ML) diagnosis models often rely on a single data modality, which limits predictive performance and disease mechanisms identification. Integrating multiple molecular and imaging modalities offers an opportunity to improve diagnostic accuracy while enabling systems-level insights into PD pathogenesis. We developed a multimodal ML framework combining MRI, biomarkers, and miRNA from Parkinson's Precision Medicine Initiative (PPMI). We used a Siamese Network with triplet loss to overcome data scarcity and model complex patient-similarity representations. We used model feature importances to create therapeutic hypotheses for targeting PD, leading to the identification of IL-6 as a gene driving PD inflammation. We validated IL-6 as a target through differential gene expression and gene set enrichment analyses. We applied structure-based drug design to inhibit IL-6 by first generating fragments within computationally identified binding pockets. These fragments were developed into final candidates through iterative optimization based on random mutation and selection of high-scoring structures, mimicking an evolutionary model. Our model achieves an ROC AUC of 0.94, leading to performance superior to single modality or naive data integration baseline methods. Using CNN-based docking and scoring, we designed drug candidates with binding affinities of -11.2 and -11.5 kcal/mol. This study offers a unique end-to-end pipeline linking multi-omics integration with structure-based drug design, demonstrating how interpretable multimodal machine learning can improve PD diagnosis and uncover disease mechanisms while identifying therapeutic targets.

The demand for food is expected to grow substantially in the coming years. To address this challenge, especially in the context of climate change, a deeper understanding of genotype-phenotype relationships is crucial for improving crop yields. Recent advances in high-throughput technologies have transformed the landscape of plant science research. However, there is an urgent need to integrate and consolidate complementary data to understand the biological system.

We introduce AgroLD, a knowledge graph that uses Semantic Web technologies to seamlessly integrate plant science data. AgroLD is designed to facilitate hypothesis formulation and validation within the scientific community. With approximately 1 billion triples, it integrates and annotates data from more than 151 datasets across 19 distinct sources.

The overarching goal is to provide a specialized knowledge platform addressing complex biological questions in the plant sciences, including gene participation in plant disease resistance and adaptive responses to climate change.

In the current genomic data jungle it is tempting to integrate data not only at a multi omics level but also across multiple species. Major research fields focus on comparative genomics, where knowledge on multiple species is compared to build stronger evidence for biological mechanisms and on the concept of translational genomics, also known as precision medicine where one uses model organisms to understand specific human diseases and adapted treatments. In addition, recent developments, such as pan genomics, bring new technical opportunities to describe similarities between population and species.
In this context we focus on the understanding of the biological mechanisms underlying feed efficiency, the capacity to convert feed into production gains, using experimental populations of pig and rabbit. We first looked into orthogroups between the two species and ambitioned to detect orthoSNPs, genomic positions localised in the same genome functional unit and showing variability in both species. Based on this knowledge we will perform genome wide association studies and selection signature detection on each species separately and combining genomic and phenotypic information coming from the experimental designs on both species. We hope to identify species specific biology pathways and responses to selection for feed efficiency and using the combined set up common, universal, biological pathways explaining variability in feed efficiency.
With this study we intend to build a modelling framework that could be extended to other phenotypes and species, always considering species evolutionary history as a keep criteria to combine study designs.

The rumen microbiome plays a critical role in livestock productivity and environmental sustainability, yet our understanding of the ecological drivers governing stable-state microbial communities remains limited. Host-associated microbial communities influence digestion and health while simultaneously producing greenhouse gas emissions. The rumen represents a dynamic ecosystem, characterised by temporally distinct microbial enzymatic action during feed digestion. Temporal succession, driven by niche specialisation facilitates development of stable microbial communities. This research aims to develop and apply AI techniques to uncover the key ecological drivers of stable-state rumen microbial communities and their functional outcomes. Analysis of microbiome data is hindered by batch effects and data heterogeneity across studies. By integrating machine learning models with microbiome data, this project will identify key descriptors of niche specialisation and microbial community dynamics that influence both productivity and environmental impact. Traditional supervised methods can fail to capture universal patterns which limits their application to large scale datasets in diverse contexts. Our approach involves developing rumen-specific foundation models capable of capturing the complexity and temporal dynamics of rumen microbiome data while maintaining biological interpretability. Transformer-based language models employ self-attention and masked-language modelling to learn contextual representations of microbial community composition, enabling transfer learning in downstream tasks. Leveraging metagenomic datasets of thousands of samples collected over several years, this model will reveal which microbial profiles and community interactions drive desirable outcomes, such as reduced methane emissions and improved feed efficiency. This work addresses a gap where microbial ecology, livestock productivity, and computational biology intersect. The anticipated outcomes include AI-driven tools that can inform livestock management strategies, enabling targeted interventions to optimise rumen function while reducing greenhouse gas emissions.

High-throughput screening (HTS) resources like ToxCast are vital for regulatory research but are often hindered by their scale (~100 GB) and complex database architectures. We present the ToxCast Evidence Graph, a lightweight semantic layer that transforms massive screening data into a portable, queryable system. By filtering the database into assay-scoped SQLite builds, we reduced the footprint to ~3 GB for specific use cases (e.g., developmental neurotoxicity) while preserving concentration-response data, fitted models, biological endpoints, and quality flags.
A custom RDF projection exposes these elements as linked entities in GraphDB, enabling structured queries across chemicals, assays, and potency parameters (AC50). To bridge bioactivity with exposure context, we integrated CPDat v4.0 functional-use and product-category records. While the graph manages semantic relationships, dense curve data remains in SQLite, retrieved on-demand for drill-down analysis.
A Streamlit prototype facilitates exploration via a locally deployed, lightweight LLM that translates natural language into SPARQL. Strictly grounded by a versioned RDF schema, the LLM acts as a query generator rather than a knowledge source, ensuring data sovereignty and preventing hallucinations.
This architecture provides a transparent, auditable blueprint for regulatory agencies (e.g., BfR) where traceability and expert review are paramount. By providing a portable ""evidence-in-a-box,"" we demonstrate how semantic technologies and local AI can democratize access to complex toxicological big data.

Genome-scale metabolic models (GEMs) provide valuable insight into microbial systems. The first GEM of the model organism Escherichia coli was developed in 2000. The model has been updated several times, arriving at the most recent version iML1515 (2017). Over the years, several studies published improvements to the iML1515-model, but these improvements have not been integrated, leading to several parallel versions of iML1515. We have, therefore, combined previously reported modifications of the iML1515-model to arrive to a consensus model: Ecoli-GEM.

This consensus model is distributed in a standardized way called Standard-GEM, which is a defined github repository structure. As such, we have created an efficient way to collaborate on improving the E. coli model, as parallel versions and model duplications have been merged. This increases the model quality since all current improvements have been verified, and future alterations can be openly discussed in the repository.

We have used the Ecoli-GEM (and its enzyme-constrained version) to study the biological role of protons in E. coli, as proton fluxes are linked to the energy metabolism. We show that E. coli is not only nutrient-limited, but also energy-limited due to proton-saturation as maintaining a neutral cytosolic pH places an energetic burden on the cell. We have quantified this energetic burden for different carbon sources, in oxic and anoxic conditions, and we show that the recent insights in the dynamics of formate transporter (FocA) greatly impact the simulated growth rate. These insights in the dispersion and biological relevance of GEMs will improve their applicability in metabolic engineering studies.

Understanding how gene expression evolves along somatic cell lineages is central to processes such as development and tumor progression. Recent single-cell technologies now enable joint measurement of gene expression and lineage relationships, opening new opportunities to study temporal dynamics. Continuous trait models, such as Brownian Motion and Ornstein-Uhlenbeck processes, provide a natural framework, yet their applicability in single-cell contexts remains largely unexplored.

Here, we evaluate these models under realistic scenarios of somatic evolution and introduce an observation model that explicitly accounts for scRNA-seq noise. Using simulations based on age-dependent branching processes, we generated lineage trees across biologically motivated scenarios and assessed model recovery under different selective regimes.

We show that inference is sensitive to transition diagram structure, cell type uncertainty, and errors in tree reconstruction, often leading to overestimation of selection. Critically, methods that ignore the observation process are strongly biased by expression noise, often overestimating the effect of selection.

Overall, our results highlight the need for caution when interpreting the results of these models in single-cell settings, where technical noise, model misspecification, and uncertainty in lineage reconstruction can strongly influence inferred evolutionary dynamics. Explicitly modeling the observation process is essential for robust phylogenetic inference from single-cell gene expression data.

Non-invasive biomarker monitoring can provide quantitative physiological data for model-assisted health assessment. Lactate is an important metabolic indicator of exercise intensity, fatigue, and training load. However, conventional blood lactate testing requires invasive sampling and is unsuitable for frequent monitoring during exercise and recovery. In this study, we developed a microneedle-based electrochemical lactate sensor for rapid salivary lactate testing and evaluated its potential for lactate-based exercise intensity assessment. The working electrode was modified with Prussian Blue, MWCNT-COOH and lactate oxidase immobilized through EDC/NHS crosslinking to enable enzymatic electrochemical lactate detection. The sensor was characterized in buffer and human saliva, including calibration performance, detection limit, response behavior, reproducibility, saliva matrix effects, and signal correction. Sensor-derived salivary lactate changes were further mapped into a lactate-based interpretation framework to classify low, moderate, and high exercise-load responses and to identify potential overreaching or overtraining-risk patterns based on elevated lactate responses and delayed recovery trends. This approach links non-invasive salivary biomarker sensing with physiological-state classification, supporting rapid point-of-care assessment of exercise intensity and recovery status. Although larger participant cohorts and additional physiological markers are required for further validation, this work demonstrates the potential of salivary lactate dynamics as an accessible biomarker input for model-assisted training-load assessment and personalized health monitoring.

Background:
Adopting a Molecular Pathological Epidemiology perspective, we present a probabilistic in silico Digital Twin of the FGFR3 pathway. Bridging pathologic epidemiological data with micro-cellular multi-omics, this framework establishes a dynamic, predictive modeling foundation for uro-oncology to execute virtual experiments.

Methods:
Developed in Python via a ""Grey Box"" network inference approach, the multi-pathway FGFR3 architecture (IP3, MAPK, JAK/STAT, mTOR) was modeled as a probabilistic electrical circuit to overcome data sparsity. Integrating systems biology and multimodal data, input voltages were derived from normalized transcriptomic expression of FGFR3. Signal routing via stochastic Markov Transition Matrices was governed by pathway conductance and endogenous inhibitors. Epidemiological constraints utilized cBioPortal mutation frequencies (NMIBC cohort), while final current outputs were predicted by formula and quantified via statistical transcriptomic footprinting (decoupleR).

Results:
The dynamic model successfully mapped signal fluxes across parallel cascades, calculating probabilistic gene activations. Virtual experiments simulating targeted interventions computed realistic signal rerouting. At higher resistances, flux diverted to PLCγ/STAT survival pathways — mathematically recapitulating adaptive bypass resistance observed clinically upon MAPK/PI3K inhibition in FGFR-driven urothelial carcinoma. These simulated flux diversions and bypass resistance mechanisms (BRAF-MEK/ERK) demonstrated strong biological validity in FGFR-targeting drug resistance, aligning with established literature of cancer genetics.

Conclusions:
This biologic framework translates multi-omics and mutational data into dynamic, predictive models. By rendering the FGFR3 network as an electrical circuit, we provide a literature-consistent proof-of-concept for systems biology in uro-oncology, offering a computational platform for anticipating bypass resistance and screening novel therapeutics in silico.

Multiple sclerosis (MS) is characterized by heterogeneous white matter lesions that vary in immune activity, myelin damage, and spatial organization. This heterogeneity complicates the systematic characterization of lesion subtypes and their role in disease progression. To address this, our group is generating a large-scale neuropathological imaging dataset in close collaboration with the Netherlands Brain Bank (NBB), a unique resource providing high-quality post-mortem brain tissue with minimal post-mortem delay. These high-resolution histological images capture the full spectrum of lesion variability; however, their analysis remains limited by the need for expert annotation.

Our core technical contributions are threefold: (1) a patch-based processing strategy that enables efficient analysis of gigapixel whole-slide images; (2) the use of a vision foundation model, whose rich and generalizable representations - learned from large-scale image data - transfer effectively to this specialized histological domain; and (3) a label-efficient pipeline that leverages these pre-trained features to achieve strong performance using only 100 annotated images from a cohort of ~2,000. Notably, the foundation model embeddings were sufficiently expressive to capture pathologically relevant structures. The model achieved 85% accuracy in lesion segmentation. Overall, this embedding-based framework provides a scalable and data-efficient approach for large-scale analysis and can be extended to other neuropathological applications.

Pacemaker neuronal identity is classically attributed to specific ion channels — most notably HCN and T-type calcium channels. Here we challenge this view through large-scale computational sampling of a 21-channel, single-compartment Hodgkin-Huxley model incorporating dynamic Nernst potentials, full calcium and sodium homeostasis, and biologically constrained parameter bounds.
From 1.5 million simulated cells, 390,022 were validated as pacemakers. Three principal findings emerge. First, calcium channel subtypes show near-zero inter-channel correlations (mean |r| = 0.004), with coefficient of variation exceeding 0.93 within any frequency band — demonstrating that calcium channel composition is degenerate with respect to pacemaker identity. Second, 29.4% of all pacemaker cells fire without any Kv1 conductance, achieving stable rhythmic output through alternative repolarization mechanisms. Third, HCN conductance correlates with firing frequency at r = 0.097 across 390,022 cells, indicating that HCN is neither necessary nor sufficient for pacemaker function.
These findings are cross-validated against a human substantia nigra single-nucleus RNA-seq dataset (n = 22,048 dopaminergic neurons; Kamath et al., 2022), where HCN2 expression variance and calcium channel co-expression patterns are consistent with predicted degeneracy profiles.
Together, these results reframe pacemaker identity as an emergent property of ion channel composition space — with direct implications for understanding differential neuronal vulnerability in Parkinson's disease and for future therapeutic target selection.

Missing observations arising from irregular sampling and participant attrition are a major analytical bottleneck in longitudinal microbiome studies. Existing imputation methods either treat temporal observations through discrete multivariate representations, without leveraging the underlying continuity of biological processes, or rely on deep generative models such as GANs and diffusion architectures, which require substantial training data and can be unstable or difficult to interpret in studies with limited replicates or sparse time points.
We propose an imputation framework that models each taxon trajectory as a smooth latent function on the centered log-ratio scale, estimated through Functional Principal Component Analysis (FPCA) in the Principal Analysis by Conditional Expectation (PACE) framework to handle sparse and irregular sampling. Imputation pools information from observed time points within each trajectory and from other replicates through the shared FPCA covariance structure. To address heterogeneity across replicates, an adaptive clustering step based on FPCA scores is performed locally for each missing observation, so that imputation borrows strength only from trajectories with comparable temporal patterns. Robustness is further enhanced through Fraiman–Muniz functional depth, which down-weights anomalous trajectories that would otherwise distort covariance estimation. Uncertainty is quantified through both analytic and nonparametric bootstrap confidence intervals.
We evaluate the method on two longitudinal datasets across multiple taxonomic resolutions and missingness rates ranging from 10% to 70%, under MCAR, MAR, and MNAR mechanisms, benchmarking against DeepMicroGen. The framework achieves lower mean absolute error at substantially reduced computational cost, with improvements most pronounced at high missingness rates.

The age at which Mendelian diseases first manifest varies widely, ranging from prenatal to adult presentation, yet the biological and clinical determinants of this variability remain poorly understood. Although individual disorders have provided valuable clues, a systematic analysis across the Mendelian landscape has been lacking.

We conducted a comprehensive analysis using curated age-of-onset annotations from the Human Phenotype Ontology (HPO). Mendelian diseases were grouped into four onset categories: congenital, neonatal, pediatric, and adult. We assessed the enrichment of 827 phenotypic features (HPO terms), 596 gene functions (Gene Ontology), and modes of inheritance across these categories. Furthermore, we trained a Random Forest classifier to predict disease onset based on clinical phenotypes, achieving an F1-score of 0.71 across the four categories.

Our results reveal distinctive phenotypic signatures: congenital onset was uniquely characterized by developmental anomalies such as Ventricular septal defect, while adult-onset diseases were enriched for reproductive features and late-stage neurological manifestations. Statistical analysis showed that adult-onset diseases were substantially enriched for autosomal recessive inheritance, whereas autosomal dominant inheritance was comparatively underrepresented. The supervised model highlighted that morphological and developmental abnormalities served as the most informative features for onset prediction.

These results establish HPO-curated onset annotations as a valuable resource for elucidating relationships among disease manifestation, clinical phenotype, inheritance patterns, and gene function. Together, our findings provide a unified framework for understanding temporal variation in Mendelian disease presentation and may support future efforts in automated disease characterization.

Cell death mechanisms, such as autophagy and apoptosis, play critical roles in cellular homeostasis and disease. Autophagy is an evolutionarily conserved mechanism that remains active under homeostatic conditions and supports cell survival by degrading damaged or obsolete components. It is closely linked to programmed cell death (apoptosis), and disruption of their crosstalk is implicated in neurodegeneration, inflammation, cancer, and aging. This work provides a computational framework to study regulatory interactions between these processes.

We introduce two new inputs to our existing, basally accurate biochemical reaction network model: glucose starvation and treatment with the autophagy inducer rapamycin. Based on mass-action kinetics, the model describes 175 interactions among 114 species using ordinary differential equations implemented in Cantera, with parameter optimization via the FOCTOPUS algorithm in the Optima++ environment. Simulations are validated against experimental data.

The starvation module reproduces glucose dynamics accurately and incorporates mechanistic and structural features of the energy sensor AMPK, enabling differential activation by distinct adenosine nucleotides. The model predicts two-fold higher ADP-mediated than AMP-mediated activation during early starvation, a distinction often not captured by comparable models. The rapamycin module implements passive transport and accounts for protein-protein interactions using literature-derived reaction rates. The refined model reproduces the dynamics between AMPK and the autophagy suppressor mTORC1 and initiator ULK1.

Overall, the new modules accurately simulate glucose starvation and rapamycin treatment, facilitating inference of real-world effects from in silico simulations. Our work therefore bridges computational systems biology and biomedical research, providing a foundation for patient-specific simulations, supporting future applications in personalized medicine.

Background. Cellular proteomes encode regulatory relationships not only through physical interactions but also via coordinated variation in protein abundance. While positive covariation reflects cooperative processes, negative covariation may capture inhibitory or compensatory constraints. However, how somatic mutations reshape these network-level dependencies remains largely unexplored.
Approach. We integrated genomic and proteomic profiles from 1,067 CPTAC tumors and analyzed mutation-associated changes in protein covariation across more than 20,000 proteins. Using regression models with mutation–interaction terms, we quantified how missense variants alter correlation structures and assessed the utility of covariation patterns for scalable AI-based interaction prediction.
Results. Positive associations were widespread (~69,000 pairs at |r|>0.5, adj. p<0.05), whereas negative associations were comparatively rare (~580 pairs). Somatic mutations perturbed both types of relationships, with negative dependencies disproportionately affected. Pathogenic variants induced a stronger attenuation of correlations (median −0.30) than benign variants (−0.16). Notably, pathogenic mutations accounted for the largest fraction of significant interaction changes among negatively correlated pairs (~21%), suggesting selective vulnerability of inhibitory links.
Conclusion. These results indicate that cancer involves a mutation-driven disruption of proteomic regulatory balance, particularly affecting inhibitory interactions. The resulting covariation signatures provide mechanistic signals that can inform next-generation AI models for inferring interaction states from large-scale proteomic data.

High-grade serous ovarian carcinoma (HGSOC) presents substantial inter-patient heterogeneity at morphological, molecular and clinical levels. No current method delivers reliable patient stratification for prognosis or treatment selection. Patient similarity networks (PSNs) combined with graph-based learning offer a way to model relational structure across patients. However, their effectiveness depends on the quality of the underlying patient representations and the strategies used to construct the graphs. Representations from generic pre-trained models often fail to capture task-relevant features, and commonly used similarity metrics do not reflect cross-modal relationships. We present a framework for multimodal PSN construction that addresses these limitations. We refine feature representations from high-dimensional modalities, including histopathology, to better capture domain-specific structure, and integrate them with complementary clinical and molecular modalities. Modalities are combined using similarity network fusion alongside alternative data-driven fusion strategies. We systematically compare similarity measures and graph construction techniques to assess their effect on downstream learning, and use the resulting networks to train graph-based models for survival prediction. We expect that improved representations and more informed graph construction will yield PSNs that better capture clinically meaningful relationships between patients, improving survival prediction in HGSOC. By moving away from task-agnostic features and heuristic similarity definitions, our approach aims to generalise to other heterogeneous disease settings where patient stratification remains a challenge.

Recent advances in representation learning have transformed natural image analysis, yet their impact on biological microscopy remains poorly understood. In this work, we systematically investigate what deep learning models actually learn from large-scale bioimaging data, focusing on two key biological contexts: cell culture imaging and tissue histology.

We benchmark a wide range of representations, including state-of-the-art pretrained vision models, domain-specific foundation models, randomly initialized networks, and simple handcrafted features. Across multiple datasets and tasks, we show that surprisingly simple or untrained representations can achieve performance comparable to advanced pretrained models. This suggests that current benchmarks may not reliably capture biologically meaningful representations, but instead exploit low-level visual cues or dataset-specific biases.

To better understand these behaviors, we analyze representation structure through dimensionality, layer-wise contributions, and robustness across biological tasks. We find that high-performing models often rely on low-dimensional signals and that shallow features can outperform deeper representations, challenging common assumptions derived from natural image domains. Furthermore, we demonstrate that biologically interpretable structure-only baselines, constructed without cell-level information, can remain competitive, reinforcing concerns about the validity of existing evaluation protocols.

Overall, our results highlight a critical gap between benchmark performance and biological relevance in bioimaging. We advocate for more rigorous evaluation strategies and stronger baselines to ensure that learned representations capture meaningful biological mechanisms rather than spurious correlations. This work provides practical guidelines for the development and assessment of deep learning models in computational biology and high-content imaging.

Pancreatic β-cell dysfunction, characteristic of type 2 diabetes (T2D), involves transitions across distinct cellular states preceding failure, yet the mechanisms driving these transitions remain poorly understood. Although single-cell RNA sequencing (scRNA-seq) studies have enabled the identification of -cell subpopulations, they lack large sample sizes and prediabetes representation, constraining insight into disease progression across the spectrum of dysglycemia. To address this gap, we integrated five scRNA-seq datasets to (i) map β-cell state dynamics across non-diabetic, pre-T2D, and T2D, and (ii) to define molecular programs underlying state transitions. We identified eight transcriptionally distinct β-cell clusters spanning three functional axes related to insulin (INS) secretion, amylin (IAPP) secretion, and cellular stress responses. INS-low clusters were enriched in T2D, while a metalloprotein cluster was enriched in non-diabetic, consistent with early protective or compensatory activity. In addition, we identified two distinct IAPP-expressing clusters aligned with either INS-high or INS-low states, delineating divergent programs of functional maintenance versus stress adaptation and metabolic rewiring. Trajectory inference revealed a continuum from functional to dysfunctional β-cell states, highlighting progressive changes in transcriptional programs related to calcium ion channel activity and GABAergic signaling pathways. These findings suggest adaptive mechanisms that may support β-cell survival despite functional decline. Integration with genome-wide association study data showed that T2D-enriched β-cell clusters are significantly associated with genetic risk for T2D. In summary, we uncover a novel spectrum of β-cell states linking functional identity, stress adaptation, and genetic risk, refining our understanding of coordinated metabolic and signaling rewiring underlying β-cell dysfunction and persistence in T2D.

Selecting appropriate software and parameter settings for processing shotgun metagenomic data is essential for accurate analysis and interpretation. This remains challenging given the growing number of available bioinformatics tools. The Critical Assessment of Metagenome Interpretation (CAMI) is a community-driven initiative that provides standardized, unbiased evaluations of metagenomic methods, including assemblers, taxonomic profilers, binners, and pathogen detection tools. Community participation is driven by the provision of benchmarking challenges, in which participants apply their methods on novel datasets and submit their results for independent evaluation. Previous work by the community established widely adopted datasets, metrics, and formats, guiding both method applications for data analysis and their development. In 2026, CAMI III challenges will be launched, covering datasets from different high and low microbial biomass body sites, reflecting the latest sequencing technologies and introducing new evaluation categories. As before, microbial communities will cover multiple domains, i.e., Bacteria, Eukaryotes, Archaea, and Viruses, and also include plasmids. To support method development, testing of scalability, and familiarization with CAMI formats ahead of the upcoming benchmarking challenges, a preparatory ("toy") human gut dataset comprising paired short- and long-read shotgun metagenomes was recently released. These data were modeled to represent 20 longitudinal samples across 10 individuals. In contrast to the toy dataset, challenge datasets will include unpublished data, which is essential for informative benchmarking of reference-based methods. Further information is available at https://cami-challenge.org/.

Metabolism comprises the network of chemical reactions that sustains life, including nutrient processing, energy production, and the synthesis of essential cellular components. Due to its complexity, metabolism is typically represented as a network of pathways, where reactions are grouped into functional units that perform specific biological tasks. A major resource for studying metabolism is the Kyoto Encyclopedia of Genes and Genomes (KEGG), which organises metabolic knowledge into pathway maps linking genes—represented as KEGG Orthologs (KOs)—through the reactions they catalyse. Although KO annotations derived from sequencing data are essential for identifying the functional potential of organisms, they are often difficult to interpret in isolation and do not directly reveal which metabolic functions are active.

Between pathways and individual KOs, KEGG defines modules as intermediate functional units. These modules correspond to smaller, well-defined sub-pathways composed of specific combinations of KOs and associated reactions. Owing to their finer granularity and reusability, modules provide a promising framework for representing metabolism as a set of functional building blocks. However, modules remain underutilised, as current KEGG tools are largely pathway-centric and present them mainly as static diagrams rather than structured, network-based representations.

To address this limitation, this work introduces Module Graph, a web-based tool that generates module graphs from a set of KOs. In this representation, modules are nodes connected when they share active compounds, enabling the visualisation, analysis, and comparison of metabolic organisation across organisms and metagenomic samples while preserving biological interpretability.

Engineering biological systems with specified functions requires navigating an extensive design space, which is challenging to achieve with wet-lab experiments alone. To expedite the design process, mathematical modeling is typically employed to predict circuit function in silico ahead of implementation, which, when coupled with computational optimization, can be used to automatically identify promising designs. However, circuit models are inherently inaccurate, which can result in suboptimal or non-functional in vivo performance. To mitigate this, we propose combining Bayesian inference, Thompson sampling, and risk management to find optimal circuit designs. Our approach employs data from non-functional designs to estimate the distribution of model parameters and then employs risk-averse optimization to select design parameters that are expected to perform well, given parameter uncertainty and biomolecular noise. We illustrate the approach by designing adaptation circuits and genetic oscillators using real and simulated data, with models of varied complexity.

The rapid generation of heterogeneous biomedical research data during the SARS-CoV-2 pandemic demonstrated some of the challenges integrating multi-modal datasets required for rapid-response decision-making. In particular, it highlighted the need for research data systems capable of harmonising longitudinal clinical, serological, and experimental assay data to support viral and immune surveillance.

In response, we developed a modular data infrastructure that underpins our ongoing virology research for the UCLH Crick Legacy study (REC20/HRA/4717). This observational cohort began in January 2021 to characterise SARS-CoV-2 infection- and vaccine-induced immune responses. The infrastructure supports the ingestion and integration of participant-level data spanning serological measurements, clinical metadata, recorded vaccination, and experimental assay outputs conducted repeatedly over years.

Our data model enforces harmonisation across modalities, enabling linkage of longitudinal records at the participant and sample levels. Time-indexing and data alignment is performed using our publicly-available R package, chronogram. Data integration workflows standardise assay outputs and clinical variables into a unified schema, supporting version-controlled updates as new data are acquired.

Our system provides structured data outputs to support downstream analyses that investigate the ongoing effects of the pandemic on immunity, reducing the gap between hypothesis generation and computational implementation. As a result, it has supported analyses within the Legacy study and collaborating consortia, including seasonal vaccine monitoring and real-time estimates of immune response against SARS-CoV-2 variants.

Beyond Legacy, this research infrastructure continues to develop with pandemic preparedness in mind, and presents a framework for developing modular data architectures that can support future infectious disease surveillance and response.

Neoepitope prediction is central to cancer immunotherapy, yet current computational pipelines often generate large candidate sets with limited biological interpretability, largely due to the neglect of transcript isoform diversity and redundancy across highly similar gene families. Consequently, the biologically relevant neoepitope landscape remains obscured by substantial noise.

Here, we present a transcript- and isoform-aware framework for neoepitope discovery that integrates predictions from MHCflurry, NetMHCpan, and PRIME across globally prevalent HLA class I alleles. Using peptide repertoires spanning 8–11 amino acids derived from the Genotype-Tissue Expression (GTEx) Project dataset, we constructed a large-scale, transcript-aware neoepitope database and applied a consensus-based prioritization strategy.

Initial integration yielded over 53,000 candidate peptide–HLA pairs. Incorporating most-dominant transcript (MDT) information as part of the prioritization pipeline reduced this set to 58 candidates, reflecting the combined effect of transcript-level filtering and multi-model consensus. These high-confidence candidates formed a consistent, non-redundant core, predominantly originating from a single gene, CYFIP2.

In contrast, the broader candidate space was dominated by immune receptor gene families, including KIR and immunoglobulin loci, which inflated apparent neoepitope diversity through sequence redundancy and multi-transcript mapping. After removing this redundancy, we identified 3,380 uniquely mapped candidate peptides from biologically relevant genes.

Collectively, our results define a three-layer structure of the predicted neoepitope landscape: a large redundant background, an intermediate set of uniquely mapped candidates, and a minimal high-confidence core defined by transcript dominance. This framework provides a principled strategy for reducing false positives and improving the biological relevance of neoepitope prioritization.

Gene prioritization (GP) methods are increasingly used to rank candidate disease genes from heterogeneous biological data, but reported performance remains difficult to compare across studies because methods are evaluated using different feature sets, training protocols, and metrics, and may also be affected by information leakage. We present a unified, time-aware benchmarking study that evaluates classical and modern machine-learning approaches for GP, including kernel methods and deep neural networks. We further benchmark multiple gene-level feature modalities under a prospective evaluation protocol and compare data integration strategies at multiple levels across both deep-learning and kernel-based settings. To better reflect practical discovery settings, we evaluate performance using metrics that emphasize both global ranking and early retrieval, and introduce BioSim, a biological enrichment-based similarity metric that assesses the functional coherence between top-ranked unlabeled genes and known disease genes.

Across 48 diseases from 11 ICD-10 categories, our benchmark shows that classical methods such as SVMs remain highly competitive with deep neural networks, and that feature type and feature quality strongly influence model performance. Early fusion can produce mixed results, whereas intermediate fusion in both deep-learning and kernel-based settings, particularly geometric kernel fusion, yields the most consistent gains across ranking metrics. For example, in Diseases of the Blood and Certain Disorders, intermediate fusion substantially improved early retrieval, more than doubling Recall@150 relative to the best corresponding non-fused setting. Late fusion, meanwhile, more often gives stronger BioSim performance. Overall, this work provides a leakage-aware benchmark and practical guidance for evaluating GP methods in realistic discovery scenarios.

Right at the moment while we are learning or forming new information, our brain generates fascinating synaptic signals called Long-Term Potentiation (LTP), a neurophysiological process underlying learning and memory formation. During this process, the brain generates synaptic signals that can be recorded as time-series waveforms using techniques such as Microelectrode Arrays (MEA). These signals, particularly the first field excitatory postsynaptic potential (fEPSP), rapidly change in response to stimulation, pharmacological intervention, or subtle pathological alterations, thereby capturing biologically meaningful information regarding synaptic strength and temporal dynamics.
However, current LTP analysis methods suffer from major limitations. Most studies rely on oversimplified analytical approaches without validating whether waveform changes truly represent synaptic activity or drug-specific effects. Synaptic regions are often manually selected, while statistical methods are frequently applied without verifying assumptions such as data normality. Furthermore, standardized quality control procedures for detecting artifacts, noise, waveform heterogeneity, or batch effects are often lacking.
To address these challenges, we developed LTP Analysis Software, a standardized platform for LTP waveform analysis. The platform performs waveform preprocessing, synaptic region detection, noise removal, artifact detection, unsupervised clustering, feature extraction, and interpretable supervised machine learning.
We evaluated the software using electrophysiological recordings from over 50 idiopathic Normal Pressure Hydrocephalus (iNPH) patients, a cohort in which approximately 50% of cortical biopsies exhibit early Alzheimer's Disease (AD)-related pathology. Using Random Forest classification with group cross-validation, the framework achieved 92% accuracy in pathology prediction, suggesting that different pathological conditions are associated with distinct LTP waveform signatures.