• Ying-Ju Lai, University of Pittsburgh, United States
• Tianyuzhou Liang, University of Pittsburgh, United States
• Po-Yuan Chen, Academia Sinica, Taiwan
• Yu-Che Tsai, National Taiwan University, Taiwan
• George C. Tseng, University of Pittsburgh, United States
• Yufei Huang, University of Pittsburgh, United States
• Yu-Chiao Chiu, University of Pittsburgh, United States

Motivation: Accurate prediction of drug-target interactions (DTIs) is fundamental to drug discovery and mechanistic understanding. While deep learning has advanced computational DTI prediction, most existing methods rely primarily on molecular structural representations, including drug structures and protein sequences, while overlooking cellular phenotypes that reflect downstream biological effects. Cell Painting enables high-content morphological profiling that captures systems-level responses to chemical and genetic perturbations but remains underutilized in DTI modeling. Integrating molecular information with cellular phenotypes offers an opportunity to improve both predictive performance and biological interpretability.
Results: We propose a two-stage contrastive learning framework integrating drug structures, protein sequences, and Cell Painting morphological profiles into a unified embedding space. Stage 1 learns modality-specific representations independently from structure-based and image-based data; Stage 2 aligns these via multi-positive contrastive learning to bridge molecular structural information with cellular phenotypes. Cross-modal retrieval achieves median Recall@10 values of 0.77 (random split) and 0.33 (scaffold split), outperforming bilinear and random baselines. In external DTI prediction on the BIOSNAP dataset, our model achieves an AUC of 0.92 with image-based representations and 0.90 under structure-only settings, surpassing existing methods. Model interpretation via integrated gradients reveals pathway-specific morphological signatures associated with drug targets, providing biologically interpretable insights into drug mechanisms.

• Li-Ting Chen, University Medical Center Utrecht, Princess Máxima Center for Pediatric Oncology, Oncode Institute, Netherlands
• Jeroen de Ridder, University Medical Center Utrecht, Princess Máxima Center for Pediatric Oncology, Oncode Institute, Netherlands
• Myrthe Jager, University Medical Center Utrecht, Princess Máxima Center for Pediatric Oncology, Oncode Institute, Netherlands

Motivation: Cell-free DNA (cfDNA) whole-genome sequencing (WGS) is a promising approach for detecting cancer recurrence. It enables cancer detection by identifying all tumor-derived cfDNA (ctDNA) molecules carrying somatic single nucleotide variants (sSNVs). While, ideally, a sequencing platform should be highly accurate for reliable ctDNA detection, in reality, all sequencing platforms introduce sequencing errors that generate false positives indistinguishable from true SNVs. Understanding how sequencing parameters influence ctDNA detection sensitivity at low tumor fractions (TFs) in cfDNA samples is essential for guiding sequencing strategies in clinical contexts. To model cfDNA sequencing for tumor detection which contains asymmetric noise, and multiple interacting parameters, analytical modeling is intractable, motivating large-scale parallelized simulation.

Results: We developed a simulation framework to generate in silico cfDNA data across ten cancer types. In total, 480 million cfDNA samples were simulated from tumor WGS profiles. Overall, the lowest detectable TF differs substantially between cancer types under identical sequencing conditions due to variations in mutational load. For cancers with high mutational load, 3x coverage with low-error techniques reliably detects TFs below 0.1%. In contrast, cancers with low mutational load require at least six-fold higher coverage to achieve comparable detection thresholds. Increasing sequencing quality scores from Q30 to Q55 at 30x coverage further enhances sensitivity, enabling detection of TFs as low as 1x10⁻⁵. This study provides a comprehensive framework for optimizing sequencing parameters, offering valuable guidance for tailoring future technology development for specific cancer types and clinical applications.

• Santiago Noto, School of Applied Mathematics, Fundação Getulio Vargas, Brazil
• Santiago Ferreyra, School of Applied Mathematics, Fundação Getulio Vargas, Brazil
• Ruben Jimenez, School of Applied Mathematics, Fundação Getulio Vargas, Brazil
• Diego Galeano, School of Applied Mathematics, Fundação Getulio Vargas, Brazil
• Alberto Paccanaro, Department of Computer Science, Centre for Systems and Synthetic Biology, Royal Holloway, University of London, Brazil

Motivation: Identifying drug targets is fundamental in drug development, both for discovering new therapies and for ensuring effective and safe treatments. Drug-target interactions (DTIs) have been predicted using machine learning approaches that integrate heterogeneous data; however, often these models are complex and lack interpretability.
Results: We investigate whether a simple, fully interpretable linear model can achieve competitive performance for DTI prediction. We propose LI-DTI (Linear Interpretable Drug-Target Interaction), a prediction model inspired by recommender systems. LI-DTI learns from different drug-drug and target-target similarity matrices and provides interpretable predictions as a linear combination of these similarity measures. We show that LI-DTI can recover drug-target interactions even when drugs or targets have no previously known interactions, across multiple cross-validation settings. We further evaluate performance while mitigating potential bias arising from high chemical similarity between drugs or sequence similarity between targets. Finally, we assess LI-DTI in a prospective evaluation, training on DTIs present in DrugBank from 2011 and testing on interactions added through 2022. Across all evaluations, LI-DTI achieves state-of-the-art performance while producing interpretable predictions. For practical use, we provide a web-based tool that enables users to visualize individual LI-DTI predictions for DrugBank (2025) and inspect the biological evidence underlying them. Our results indicate that simple linear models with well-curated similarity features can deliver robust and interpretable DTI predictions, facilitating hypothesis generation and downstream experimental prioritization.
Availability and Implementation: Code and data available at https://github.com/paccanarolab/LI-DTI. Web tool available at https://paccanarolab.org/lidtiweb/.
Contact: alberto.paccanaro@rhul.ac.uk
Supplementary information: Supplementary data are available at Bioinformatics online.

• Pia Rautenstrauch, Berlin Institute of Health at Charité – Universitätsmedizin Berlin, Germany
• Uwe Ohler, Max-Delbrück-Center for Molecular Medicine in the Helmholtz Association (MDC) & Humboldt-Universität zu Berlin, Germany

Single-cell data integration is a central challenge in modern genomics, enabling the joint analysis of datasets across experiments. Integration methods aim to preserve genuine biological variation while removing technical batch effects. As downstream conclusions depend on integration methods, rigorous evaluation and independent benchmarking are essential. Quantitative metrics are central to these efforts.

Among these metrics, adaptations of the silhouette score have emerged as a popular choice. Originally developed to assess unsupervised clustering results by comparing within-cluster cohesion to between-cluster separation, it is now frequently applied to evaluate integration performance.

However, we show that silhouette-based metrics suffer from fundamental and largely overlooked limitations in this context. We formalize the silhouette score and its adaptations for single-cell integration and demonstrate through simple simulations that their underlying assumptions are violated under basic conditions. As a result, these metrics can misleadingly reward poor integration outcomes. We further corroborate these findings across multiple real-world datasets, where silhouette-based metrics fail to reliably assess batch effect removal and biological signal conservation.

To address these limitations, we outline alternative evaluation strategies that enable more robust assessment of integration success. This is critically important, as it guides method selection, which directly impacts analysis outcomes. More broadly, our results not only demonstrate shortcomings of silhouette-based metrics in single-cell integration benchmarking but also highlight the need for metric validation in computational biology.

• Yimin Fan, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Yixuan Liu, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Yunhua Zhong, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Yue Wang, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Kin Hei Lee, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Yixuan Wang, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Xinyuan Liu, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Jiayi Li, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Xuesong Wang, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Ziqian Lin, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
• Lei Li, Canchen Technology, China
• Yu Li, Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong

Bulk RNA-seq deconvolution methods rely on single-cell reference data to estimate cell-type proportions in heterogeneous tissues, but certain cell types may be systematically absent from single-cell references due to technical limitations such as poor dissociation efficiency, low capture rates, or cell fragility. While recent studies have shown that signatures of missing cell types persist in deconvolution residuals, existing approaches cannot automatically determine how many cell types are missing, estimate their specific proportions, or reconstruct their expression signatures. Here, we present DeconX, a computational framework that addresses these limitations by generating ``pseudo-cells" from deconvolution residuals, enabling estimation of the number, proportions, and expression signatures of missing cell types. Through comprehensive evaluation on simulated datasets, we demonstrate that DeconX accurately recovers missing cell-type proportions and expression profiles across varying conditions, and systematically identifies key factors affecting performance, including expression similarity between missing and reference cell types and missing cell-type proportions. Application to high-grade serous ovarian cancer (HGSOC) samples successfully identifies and quantifies adipocyte populations that are absent from matched single-cell references, recovering biologically interpretable expression signatures consistent with known adipocyte markers. We further demonstrate that DeconX can determine the number of missing cell types in an unsupervised manner, enabling fully automated analysis of unknown tissue compositions. DeconX transforms residual-based missing cell-type inference from exploratory analysis into a complete computational framework, enabling accurate characterization of tissue heterogeneity from archival bulk RNA-seq data even when single-cell references are incomplete. Codebase is available at \url{https://github.com/Seniorious123/DeconX/}.

• Narein Rao, Indian Institute of Technology Kanpur, India
• Tanush Kumar, Department of Mathematics and Statistics, Indian Institute of Technology Kanpur, India
• Dina Kazemi, Translational Genomics Program, Garvan Institute of Medical Research, Darlinghurst, Australia, Australia
• Shaozhi Hou, Translational Genomics Program, Garvan Institute of Medical Research, Darlinghurst, Australia, Australia
• Atefeh Khakpoor, Translational Genomics Program, Garvan Institute of Medical Research, Darlinghurst, Australia, Australia
• Merrin Mary Eapen, Translational Genomics Program, Garvan Institute of Medical Research, Darlinghurst, Australia, Australia
• Rhea Pai, Curtin Medical School, Curtin University, Perth, Western Australia, Australia
• Liang Qiao, The Westmead Institute for Medical Research and Westmead Hospital, University of Sydney, Sydney, Australia, Australia
• Archita Mishra, University of Sydney, Sydney, Australia, Australia
• Florent Ginhoux, Gustave Roussy Cancer Campus, Villejuif, France, France
• Jerry Chan, KK Research Center, KK Women’s and Children’s Hospital, Singapore, Singapore
• Jacob George, The Westmead Institute for Medical Research and Westmead Hospital, University of Sydney, Sydney, Australia, Australia
• Ankur Sharma, Translational Genomics Program, Garvan Institute of Medical Research, Darlinghurst, Australia, Australia
• Hamim Zafar, Indian Institute of Technology Kanpur, India

The advancement of single-cell RNA sequencing and spatial transcriptomics has enabled the inference of cellular interactions in a tissue microenvironment. Despite advances in cell-cell interaction inference, methods capable of mapping the influence of ligands on downstream target genes across spatial niches harboring specific cell type composition, crucial for resolving niche-specific relationship between ligands and their downstream targets are still lacking. Here, we present Renoir for
charting the ligand-target activities across a spatial topology, delineating spatial communication niches harboring specific ligand-target activities and spatially mapping pathway-level activity of genesets. Across spatial datasets with varying resolution (spot to single-cell) ranging from development to disease, Renoir infers cellular niches with distinct ligand-target interactions, spatially
maps pathway activities, and identifies context-specific cell-cell interactions, including hepatocyte-macrophage interactions in fetal liver and interactions between onco-fetal and bi-potent cells in hepatocellular carcinoma. Renoir uncovers biological insights and therapeutically-relevant cellular crosstalk from spatial transcriptomics data.

• Aurelien Pelissier, X Google Moonshot, United States
• Miroslav Phan, ETH Zurich, Switzerland
• Didier Le Bail, Centre de Physique Théorique (CPT), Aix-Marseille University, CNRS, France
• Niko Beerenwinkel, ETH Zurich, Switzerland
• Maria Rodriguez Martinez, Yale University, United States

Stochastic simulation is foundational to computational biology, yet existing frameworks face a fundamental tension: classical algorithms such as the Gillespie stochastic simulation algorithm (SSA) are efficient but assume memoryless, homogeneous agents, while agent-based approaches capture biological diversity at prohibitive computational cost. Real biological systems often exhibit both memory effects and heterogeneity at the level of individual agents, features that existing methods cannot handle efficiently.

We introduce MOSAIC (Modeling of Stochastic Agents with Individual Complexity), a rejection-based simulation framework that extends the Gillespie SSA to systems with agent-specific dynamics and arbitrary non-Markovian waiting-time distributions. Rather than maintaining an exhaustive list of reaction propensities, MOSAIC tracks a single global maximum rate and evaluates individual rates on demand, yielding constant-time updates per accepted event, independent of system size. This design preserves Gillespie-like scalability while enabling each agent to follow distinct inter-event distributions (Gamma, Pareto, Weibull, log-normal) with rates that adapt dynamically to the evolving system state.

We demonstrate MOSAIC across three biological domains. First, in a model of clonal B-cell affinity maturation, MOSAIC reproduces experimentally observed clonal dominance patterns with near-linear runtime scaling where standard Gillespie methods become prohibitive. Second, in a Hes1 transcription model with delayed negative feedback, MOSAIC captures state-dependent, non-exponential elongation times that classical delay-based algorithms cannot dynamically accommodate. Third, applied to empirical face-to-face interaction data, MOSAIC-TN reproduces heavy-tailed interaction statistics and higher-order network features beyond the reach of existing temporal network models. MOSAIC is openly available and establishes a practical, unified framework for heterogeneous stochastic simulation in biology.

• Bar Rozenman, Department of Biomedical Engineering, Columbia University, United States
• Kevin Hoffer-Hawlik, Columbia University, United States
• Elham Azizi, Columbia University, United States
• Nicholas Djedjos, Columbia University, United States

Motivation: Cellular states are strongly influenced by spatial context, but single-cell RNA sequencing (scRNA-seq) loses information about local tissue organization, while spatial proteomic assays capture limited marker panels that constrain transcriptomic inference. Integrating these modalities can elucidate how spatial niches shape transcriptional programs, yet existing approaches depend on either feature-level correspondence such as gene–protein linkage or cell-level barcode pairing, which is often unavailable.
Results: We present ARCADIA (ARchetype-based Clustering and Alignment with Dual Integrative Autoencoders), a generative framework for cross-modal integration that operates without cell barcode pairing and does not assume direct feature-to-feature correspondence. ARCADIA identifies modality-specific archetypes, i.e., convex combinations of cells representing extreme phenotypic states, and aligns these anchors across modalities by minimizing the discrepancy between their cell-type composition profiles. The aligned archetypes define a shared coordinate system that anchors dual variational autoencoders (VAEs) trained with cross-modal geometric regularization, preserving archetype structure and spatial neighborhood information while enabling bidirectional translation between modalities. On semi-synthetic CITE-seq data, ARCADIA outperforms existing weak-linkage methods. Applied to independent human tonsil scRNA-seq and CODEX data, ARCADIA reconstructs known tissue architecture and reveals spatially dependent transcriptional programs linking B-cell maturation and T-cell activation or exhaustion to microenvironmental niches.
Availability and Implementation: Source code is accessible at https://github.com/azizilab/ARCADIA_public. Reproducibility scripts and data are available at https://github.com/azizilab/arcadia_reproducibility.

• Bi-Rong Wang, Biomedical Engineering, Eindhoven University of Technology (TU/e); Institute for Complex Molecular Systems, TU/e, Netherlands
• Chen-Yi Liao, Leiden Academic Center for Drug Research, Leiden University, Leiden, the Netherlands, Netherlands
• Erik Danen, Leiden Academic Center for Drug Research, Leiden University, Leiden, the Netherlands, Netherlands
• Elsa Neubert, Leiden Academic Center for Drug Research, Leiden University, Leiden, the Netherlands, Netherlands
• Federica Eduati, Biomedical Engineering, Eindhoven University of Technology (TU/e); Institute for Complex Molecular Systems, TU/e, Netherlands

Spatial computational models such as agent-based models (ABMs) offer powerful in silico tools to study tumor dynamics, yet imaging data are still rarely used to inform these models directly. We present an ABM optimization framework that leverages convolutional encoders to compare spatial patterns between experimental imaging data and ABM-generated outputs within a shared latent space. This quantitative comparison was used to estimate ABM parameters across three datasets, ranging from synthetic data to 3D tumoroid-T cell co-culture microscopy and histopathology images from The Cancer Genome Atlas skin cutaneous melanoma samples. Estimated parameters were evaluated using data-derived features and experimental knowledge, including experimental conditions and gene expressions. Simulations using optimized parameters reproduced key spatial features of the training images, such as tumor boundary complexity and tumor-tumor neighborhood structure. Together, these results demonstrate a flexible framework for ABM parameter optimization using spatial data across modalities, enabling systematic investigation of how spatial architecture influences tumor progression and immune interactions.

• Jiahang Li, Nankai University, China
• Steffen Waldherr, University of Vienna, Austria
• Wolfram Weckwerth, University of Vienna, Austria

Physical inactivity and low fitness have become global health concerns. Metabolomics, as an integrative approach, may link fitness to molecular changes. In this study, we analyzed blood metabolomes from elderly individuals under different treatments. By defining two fitness groups and their corresponding metabolite profiles, we applied several machine learning classifiers to identify key metabolite biomarkers. Aspartate consistently emerged as a dominant fitness marker. We further defined a body activity index (BAI) and analyzed two cohorts with high and low BAI using COVRECON, a novel method for metabolic network interaction analysis. The method can automatically construct a simplified model of the metabolic network and calculate the changes in the biochemical regulations using a novel algorithm. COVRECON can help to understand the complex dynamics and mechanisms of biological systems and diseases (1,2). Through this approach，we identified causal molecular dynamics in multi-omics data. Aspartate-amino transferase (AST) was among the dominant processes distinguishing the groups. Routine blood tests confirmed significant differences in AST and ALT. Aspartate is also a known biomarker in dementia, related to physical fitness. In summary, we combine machine learning and COVRECON to identify metabolic biomarkers and molecular dynamics supporting active aging.
References
1，Li J, Waldherr S, Weckwerth W. COVRECON. Bioinformatics, 2023, 39(7): btad397.
2，Li J, Weckwerth W, Waldherr S. NPJ Systems Biology and Applications, 10(1), 137.

• Ljubisa Miskovic, Ecole Polytechnique Fédérale de Lausanne, Switzerland
• Subham Choudhury, Ecole Polytechnique Fédérale de Lausanne, Switzerland
• Ilias Toumpe, Ecole Polytechnique Fédérale de Lausanne, Switzerland
• Ousaama Gabouj, Ecole Polytechnique Fédérale de Lausanne, Switzerland
• Jakob Behler, Ecole Polytechnique Fédérale de Lausanne, Switzerland
• Vassily Hatzimanikatis, Ecole Polytechnique Fédérale de Lausanne, Switzerland

Dynamic metabolic models provide a mechanistic framework for studying how cellular systems respond over time to genetic and environmental perturbations. By linking biochemical mechanisms to transient cellular behavior, such models are becoming increasingly important in computational biology for predictive analysis, hypothesis generation, and integration of heterogeneous biological data. Recent generative machine-learning approaches have substantially accelerated the parameterization of large-scale and near-genome-scale kinetic models, making high-throughput dynamic studies increasingly feasible. However, adapting trained generators to new physiological regimes or organisms often still requires retraining, and the biological meaning of their latent representations remains poorly understood.
Here, we present a latent-space exploration framework that repurposes pretrained generative networks to construct kinetic models with targeted dynamic properties without retraining from scratch. Rather than modifying network parameters, the approach operates directly in latent space to identify directions that systematically tune dynamic behavior while preserving biochemical feasibility and consistency with the underlying mechanistic model.
We demonstrate the framework in Escherichia coli by tuning response times, identifying enzyme-level dynamic bottlenecks, and repurposing generators trained in one setting to produce models with distinct dynamic properties under anaerobic growth. We further test robustness and generalizability in Saccharomyces cerevisiae, showing that latent-space control remains effective across generators at different training stages and across different latent-space regions.
Overall, this work establishes latent-space control as an interpretable and computationally efficient strategy for reusing generative models in mechanistic systems biology, enabling scalable analysis of metabolic dynamics across physiological states and supporting more reusable, data-constrained dynamic modelling workflows in computational biology.

• Efrat Muller, Tel Aviv University, Israel
• Shiri Baum, Tel Aviv University, Israel
• Elhanan Borenstein, Tel Aviv University, Israel

Motivation: A central goal in human gut microbiome research is to identify disease-associated functional shifts, an objective increasingly pursued through metagenomic and metabolomic assays. However, common differential abun-dance analyses of genes or metabolites often yield long and difficult-to-interpret feature lists. Aggregating features into predefined pathways can improve interpretability but relies on fixed pathway boundaries that may not reflect context-specific functional changes. Moreover, even when paired metagenomic-metabolomic data are available, they are often analyzed separately or linked only through simple statistical associations.

Results: We introduce MAAMOUL, a knowledge-based computational framework that integrates metagenomic and metabolomic data to identify disease-associated, data-driven microbial metabolic modules. Leveraging prior knowledge of bacterial metabolism, MAAMOUL maps disease-association scores onto a global microbiome-wide metabolic network and identifies custom modules enriched for altered genes and metabolites. Applying MAAMOUL to inflammatory bowel disease (IBD) and irritable bowel syndrome (IBS) datasets revealed significant disease-associated modules not detected by conventional pathway-level analysis. In IBD, modules reflected disrupted sulfur and aromatic amino acid metabolism and enhanced microbial nucleotide salvage, whereas in IBS they linked purine and nicotinate/nicotinamide metabolism. These results demonstrate that network-guided multi-omic integration can uncover coherent functional shifts in the gut microbiome overlooked by single-omic or purely statistical approaches.

Availability: MAAMOUL is available as an R package at https://github.com/borenstein-lab/MAAMOUL.

• Veronica Lombardi, Department of Biology and Biotechnologies "Charles Darwin", Sapienza University of Rome, Italy
• Lorenzo Di Rocco, Department of Statistical Sciences, Sapienza University of Rome, Italy
• Eleonora Meo, Department of Biology, University of Rome“Tor Vergata”, Italy
• Veronica Venafra, Department of Biology and Biotechnologies‘Charles Darwin’, Sapienza University of Rome, University of Rome Tor Vergata, Italy
• Elena Di Nisio, Department of Biology and Biotechnologies‘Charles Darwin’, Sapienza University of Rome, Italy
• Velerio Perticaroli, Department of Biology and Biotechnologies‘Charles Darwin’, Sapienza University of Rome, Italy
• Mihail Lorentz Nicolaeasa, Ph.D.Program in Cellular and Molecular Biology, Department of Biology, University of Rome‘Tor Vergata’, Italy
• Chiara Cencioni, Institute of System Analysis and Informatics“Antonio Ruberti”,National Research Council (IASI-CNR), Italy
• Francesco Spallotta, Department of Biology and Biotechnologies‘Charles Darwin’, Sapienza University of Rome; Istituto Pasteur Italia, Italy
• Rodolfo Negri, Department of Biology and Biotechnologies‘Charles Darwin’, Sapienza University of Rome; (IBPM); (CNR), Italy
• Francesca Sacco, Department of Biology, University of Rome“Tor Vergata”, Italy
• Livia Perfetto, Department of Biology and Biotechnologies‘Charles Darwin’, Sapienza University of Rome, Italy

A crucial challenge in oncology is deciphering the intricate mechanisms underlying reprogramming in cancer cells to advance the ability of diagnosis and treatment of cancer patients. Consequently, whole-patient multi-omic analysis has become increasingly common in clinical practice, but, thus far, no efficient strategy for the identification of molecular pathways dysregulated at the patient level has emerged, nor have these advances been effectively employed to define personalized therapeutic regimens.
The main shortcoming is the lack of a robust computational framework to exploit such information by integrating and interpreting the available multidimensional data to drive translational solutions.

To fill this gap, we developed PatientProfiler, a computational workflow that allows for multi-omic data analysis and standardization, generation of patient-specific mechanistic models of signal transduction, and extraction of network-based prognostic biomarkers.
To benchmark it, we retrieved genomics, transcriptomics and (phospho)proteomics derived from 122 treatment-naïve breast cancer biopsies and identified patient-specific mechanistic models that recapitulate oncogenic signaling pathways. Network topology-based stratification revealed seven subgroups of patients associated with unique transcriptomic signatures and distinct prognostic values. The workflow successfully recovered established Basal-like 1 and Basal-like 2 breast cancer subtypes, while also highlighting distinct mechanistic drivers with potential clinical relevance.

In conclusion, PatientProfiler provides a generalizable workflow for decoding cohort-level multi-omic data into interpretable mechanistic models to identify deregulated signaling pathways and prognostic biomarkers. It can be applicable across different cancer types and other diseases, leading to the optimization of personalized therapies based on dysregulated signaling.

• Jocelyn Garcia, Tufts University, United States
• Kevin Yu, Tufts University, United States
• Catherine Freudenreich, Tufts University, United States
• Lenore Cowen, Tufts University, United States

In Baker's yeast, there exists a comprehensive collection of pairwise epistasis experiments that, for nearly every pair of non-essential genes, measure the growth of the double-knockout strain as compared to its component single knockouts. This data can be represented as a weighted signed graph termed the genetic interaction network, and we introduce a new ILP-based method named GIDEON to search for a diverse collection of Between-Pathway Models (BPMs) in this network, where BPMs are a graph motif signature that indicates potential compensatory pathways in the genetic interaction network.

With both an improved distribution-informed edge weighting scheme and an improved ILP method, GIDEON produces BPM collections that are substantially larger and with better functional enrichment compared to previous methods. We find some interesting new BPM gene sets including one with potential insights into antifungal drug targets through ties between ergosterol and aromatic amino acid biosynthesis.

• Dezso Modos, Imperial College London, United Kingdom
• John Thomas, Imperial College London; UKRI MRC Laboratory of Medical Sciences, United Kingdom
• Johanne Brooks-Warburton, University of Hertfordshire; Lister Hospital;, United Kingdom
• Martina Poletti, Quadram Institute Bioscience; Earlham Institute, United Kingdom
• Balazs Bohar, Imperial College London, United Kingdom
• Yufan Liu, Imperial College London, United Kingdom
• Matthew Madgwick, Quadram Institute Bioscience; Earlham Institute; IBM Research Europe, United Kingdom
• Wen-Xin Kang, Imperial College London, United Kingdom
• Benjamin Alexander-Dann, University of Cambridge, United Kingdom
• Azedine Zoufir, University of Cambridge, United Kingdom
• Padhmanand Sudhakar, KU Leuven, Belgium
• Domenico Cozzetto, Imperial College London, United Kingdom
• David Fazekas, Department of Genetics, Eötvös Loránd University, Hungary
• Shamith Samarajiwa, Imperial College London, United Kingdom
• Simon R Carding, Quadram Institute Bioscience, United Kingdom
• Nicholas Powell, Imperial College London, United Kingdom
• Bram Verstockt, KU Leuven, Belgium
• Andreas Bender, University of Cambridge, United Kingdom
• Tamas Korcsmaros, Imperial College London; Quadram Institute Bioscience; Imperial BRC Organoid Facility, United Kingdom

Genome-wide association studies have identified numerous susceptibility loci in complex diseases, such as chronic immune-mediated inflammatory disorders (IMIDs), yet their impact on pathomechanisms remains poorly understood. Low effect sizes, polygenicity, and predominance within non-coding genomic regions remain major challenges to the functional interpretation of IMID-associated single-nucleotide polymorphisms (SNPs). To address this, we present a novel systems genomics approach which models the cumulative impact of non-coding SNPs on downstream cellular signalling and gene regulatory networks. Applying this to the prototypical chronic IMIDs of Crohn's disease (CD) and ulcerative colitis (UC), both forms of inflammatory bowel disease (IBD), we individually analysed 2,636 patient genomes. Signals from non-coding SNPs were found to propagate towards well-established and novel CD- and UC-associated pathogenic pathways through the signalling and gene regulatory layers. The SNP-propagated gene regulatory networks stratified CD and UC patients into distinct clusters corresponding to cell type-specific gene dysregulation and potential therapeutic response. This approach bridges the gap between genotype and phenotype, laying the foundations for accelerating precision medicine in complex diseases.

• Umur Kaya, Freie Universität Berlin, Germany
• Xhemal Kodragjini, Freie Universität Berlin, Germany
• Samuel Zambrano, Università Vita-Salute San Raffaele, Italy
• Katharina Baum, Freie Universität Berlin, Germany

Universal differential equations (UDEs) have emerged as a powerful tool for scientific discovery, uniting differential equations and deep learning. Yet, their application to the noisy and complex data of systems biology remains limited. Here, we demonstrate that UDEs, paired with symbolic regression, can quantify the crosstalk between NF-κB and p53 signaling pathways based on experimental data exhibiting complex, oscillatory dynamics. We validate the framework on synthetic benchmarks, showing that crosstalk recovery is robust to measurement noise and improves markedly with the number of available time series. We then apply the framework to 106 simultaneously measured single-cell p53 and NF-κB time series following DNA damage and NF-κB activation, constituting the largest-scale UDE application in systems biology so far. UDEs with both a detailed and a minimal mechanistic p53 model consistently identify a monotonically increasing crosstalk function in which elevated NF-κB levels enhance p53 synthesis. Symbolic regression distills the learned neural network output into a compact, interpretable closed-form expression, providing a quantitative, time-resolved characterization of NF-κB-driven amplification of p53 at the single-cell level. Our work thus delivers a dual contribution: methodologically, it establishes UDEs as a viable tool for gaining quantitative insights in complex, noisy biological systems at scale; biologically, it provides a data-driven characterization of NF-κB-p53 pathway crosstalk from single-cell dynamics. Source code is available at https://github.com/DILiS-lab/ude-crosstalk-discovery.

• Erva Ulusoy, Hacettepe University, Turkey
• Sevval Bostanci, Hacettepe University, Turkey
• Bora Engin Deniz, Hacettepe University, Turkey
• Tunca Dogan, Hacettepe University, Turkey

Motivation: Molecular representation learning is central to computational drug discovery. However,
most existing models rely on single-modality inputs, such as molecular sequences or graphs, which
capture only limited aspects of molecular behaviour. Yet unifying these modalities with
complementary resources such as textual descriptions and biological interaction networks into a
coherent multimodal framework remains non-trivial, hindering more informative and biologically
grounded representations.
Results: We introduce SELFormerMM, a multimodal molecular representation learning framework
that integrates SELFIES notations with structural graphs, textual descriptions, and knowledge
graph–derived biological interaction data. By aligning these heterogeneous views, SELFormerMM
effectively captures complementary signals that unimodal approaches often overlook. Our
performance evaluation has revealed that SELFormerMM outperforms structure-, sequence-, and
knowledge-based models on multiple molecular property prediction tasks. Ablation analyses further
indicate that effective cross-modal alignment and modality coverage improve the model's ability to
exploit complementary information. Overall, integrating SELFIES with structural, textual, and
biological context enables richer molecular representations and provides a promising framework for
hypothesis-driven drug discovery.
Availability: SELFormerMM is available as a programmatic tool, together with datasets, pretrained
models, and precomputed embeddings at https://github.com/HUBioDataLab/SELFormerMM.

• Ahmad Alkhan, University of Limerick, Ireland
• Michael Lynch, University of Limerick, Ireland
• Maire Lavelle, University Hospital Limerick, Ireland
• Aedin Culhane, University of Limerick, Ireland
• Elizabeth Ryan, University of Limerick, Ireland
• Elizabeth Ryan, University of Limerick, Ireland
• Elizabeth Ryan, University of Limerick, Ireland

Motivation: Perineural invasion (PNI) is an important pathological phenotype associated with poor prognosis in multiple malignancies. The primary detection method is visual inspection of whole slide images (WSIs), which is labor-intensive, time-consuming, subjective, and prone to high inter-observer variability. Developing reliable, accurate deep learning models for PNI detection is con-strained by the lack of pixel-level annotated WSIs.
Results: We evaluated three backbone architectures and two different approaches to improve PNI detection in a multi-organ dataset of colon, prostate, and pancreatic adenocarcinomas. We report three key findings. First, two pathology-pretrained foundation models, Virchow-2 and UNI, sub-stantially outperformed ImageNet-pretrained CNNs (EfficientNet-B3, ConvNeXt-2), with distinct baseline error profiles reflecting differences in pretraining-data composition. Second, a data cura-tion strategy driven by confidence-based pseudo-labelling (threshold p > 0.9) with human-in-the-loop review expanded the dataset from 262 to 352 WSIs, yielding a 12.4% relative F1 improve-ment (0.740 to 0.832) and a 55.5% reduction in false positives per slide; an ablation attributed 70% of the F1 gain to data volume and 41% of the FP reduction to benign-class curation. Third, per-organ analysis revealed that the primary driver of this improvement was not data volume alone but the targeted annotation enrichment of underrepresented morphologies in adjacent-normal and benign tissue, including desmoplastic stroma, crypts, and small blood vessels, that had been a systematic source of false positive predictions across tissue types.
Availability: Our implementation is available at https://github.com/AhmadAlkhan/PNI_SSL.