• Youngoh Kim, Youngoh Kim, Seoul National University
• Dongmin Bang, Dongmin Bang, Seoul National University
• Bonil Koo, Bonil Koo, Seoul National University
• Jungseob Yi, Jungseob Yi, Seoul National University
• Changyun Cho, Changyun Cho, Seoul National University
• Jeonguk Choi, Jeonguk Choi, Seoul National University
• Sun Kim, Sun Kim, Seoul National University

Drug–Target Affinity (DTA) prediction is an important regression task for drug discovery, which can provide richer information than traditional drug-target interaction prediction as a binary prediction task. To achieve accurate DTA prediction, quite large amount of data is required for each drug, which is not available as of now. Thus, data scarcity and sparsity is a major challenge. Another important task is `cold-start' DTA prediction for unseen drug or protein. In this work, we introduce MixingDTA, a novel framework to tackle data scarcity by incorporating domain-specific pre-trained language models for molecules and proteins with our MEETA (MolFormer and ESM-based Efficient aggregation Transformer for Affinity) model. We further address the label sparsity and cold-start challenges through a novel data augmentation strategy named GBA-Mixup, which interpolates embeddings of neighboring entities based on the Guilt-By-Association (GBA) principle, to improve prediction accuracy even in sparse regions of DTA space. Our experiments on benchmark datasets demonstrate that the MEETA backbone alone provides up to a 19% improvement of mean squared error over current state-of-the-art baseline, and the addition of GBA-Mixup contributes a further 8.4% improvement. Importantly, GBA-Mixup is model-agnostic, delivering performance gains across all tested backbone models of up to 16.9%. Case studies shows how MixingDTA interpolates between drugs and targets in the embedding space, demonstrating generalizability for unseen drug–target pairs while effectively focusing on functionally critical residues. These results highlight MixingDTA’s potential to accelerate drug discovery by offering accurate, scalable, and biologically informed DTA predictions. The code for MixingDTA is available at https://github.com/rokieplayer20/MixingDTA.

• Nicoleta Siminea, Nicoleta Siminea, INCDSB; Doctoral School of Computer Science
• Victor Bogdan Popescu, Victor Bogdan Popescu, Revvity Finland Oy
• Ion Petre, Ion Petre, Department of Mathematics and Statistics
• Andrei Păun, Andrei Păun, INCDSB; Faculty of Mathematics and Computer Science

Glioblastoma is an aggressive cancer with a poor survival rate, and standard treatments often yield limited results. To explore personalized treatment options, we employed network-based analyses. Our aim was to investigate how drug recommendations might vary for individual patients compared to the conventional treatment approaches for glioblastoma.

We began by identifying patient-specific proteins from glioblastoma cases in The Cancer Genome Atlas (TCGA). Using these, we constructed protein–protein interaction networks that incorporated not only the patient-specific proteins but also those encoded by cancer-related genes and known targets of antineoplastic or immunomodulatory drugs. We then applied controllability analysis using NetControl4BioMed to identify proteins that could potentially be targeted with drugs. For each case, we compared the drugs identified through in silico analysis with those predicted to help restore the disease state to a healthy condition.

We also examined the differences between using personalized versus generic networks. Notably, 12% of drugs identified via the generic network appeared in fewer than half of the individual networks. We also found some drugs that, while absent in the generic network, were predicted to offer therapeutic value in individual patients.

This approach enables the construction and analysis of novel individual networks based on proteins identified in new glioblastoma samples. Moreover, the methodology can be adapted for other diseases. For conditions with poor prognoses, enhancing individualized network analyses is essential to improve treatment outcomes.

• Terence Egbelo, Terence Egbelo, University of Sheffield
• Zeyneb Kurt, Zeyneb Kurt, University of Sheffield
• Charlie Jeynes, Charlie Jeynes, Evotec UK
• Mike Bodkin, Mike Bodkin, University of Dundee
• Val Gillet, Val Gillet, University of Sheffield

Drug discovery faces high clinical failure rates due to adverse events (AEs) from both on- and off-target interactions.
Biomedical knowledge graphs (KGs) integrate domain knowledge in network form. KG completion is the classification task of predicting new relations based on the existing graph.
Our study predicts novel target-AE associations as KG completion using a large-scale biomedical knowledge graph. We incorporated ground truth target-AE associations from multiple sources, including the T-ARDIS database (Galletti et al 2021), into the Drug Repurposing Knowledge Graph (DRKG) by Ioannidis et al (2020).
Rather than ""black-box"" deep learning approaches to KG completion, we employ interpretable ""metapath""-based predictive features that maintain direct reference to domain semantics, following precedents set by Fu et al (2016) and Himmelstein et al (2017).
We introduce a novel approach to address the problem of topological imbalance in KGs, a type of graph data bias. This bias occurs when high-degree entities (nodes in the KG) are overrepresented in the positive ground truth associations, leading to poor prediction performance on less-studied (and thus low-degree) entities—precisely where good inference is most critical. Our bias mitigation method transforms degree sparsity into a useful signal when learning associations of sparsely connected protein targets.
Our approach demonstrated significantly improved AE prediction for the least-studied targets (accuracy increase of ~15% for the bottom 15% of targets by number of AE associations) compared to the well-known Degree-Weighted Path Count (DWPC) method by Himmelstein et al (2017).
Finally, we demonstrate prediction interpretability, including in cases where alternative methods produce errors.

• Johannes Kersting, Johannes Kersting, Technical University of Munich
• Lisa Spindler, Lisa Spindler, Technical University of Munich
• Joaquim Aguirre-Plans, Joaquim Aguirre-Plans, STALICLA SL
• Chloé Bucheron, Chloé Bucheron, University of Vienna
• Quirin Manz, Quirin Manz, Technical University of Munich
• Tanja Pock, Tanja Pock, Technical University of Munich
• Mo Tan, Mo Tan, Technical University of Munich
• Fernando Delgado-Chaves, Fernando Delgado-Chaves, University of Hamburg
• Cristian Nogales, Cristian Nogales, University of Vienna
• Jan Baumbach, Jan Baumbach, University of Hamburg
• Jörg Menche, Jörg Menche, University of Vienna
• Emre Guney, Emre Guney, STALICLA SL
• Markus List, Markus List, Technical University of Munich

Disease modules provide unique insights into the mechanisms of complex diseases and lay the foundation for mechanistic drug repurposing. Algorithms for their identification leverage biological networks to extend an initial set of disease-associated genes (seeds) into subnetworks reflecting biological processes likely to be integral components of the investigated disease. These subnetworks can unveil causal pathways and provide drug repurposing efforts with promising new targets for therapeutics.
Various computational methods have been developed for disease module identification. Since these methods differ in their modeling assumptions and techniques, evaluating various tools across different parameters to optimize for a specific use case is advisable. However, this can be tedious since the individual tools require specific installation and input preparation procedures. Moreover, identifying the best modules is not straightforward and requires topological and biological validation strategies.
To mitigate this, we developed a comprehensive pipeline for disease module identification, evaluation, and subsequent drug prioritization utilizing the workflow system Nextflow. Our pipeline automatically deploys software dependencies using Docker, making installation easy. It prepares the inputs for and runs six popular module detection tools, including DIAMOnD, DOMINO, and ROBUST. The generated outputs are annotated with biological context information, converted into a unified BioPAX format, and extensively evaluated. The latter includes assessing the biological relevance based on overrepresentation analysis and the tool DIGEST, as well as robustness and consistency analyses.
With our contribution, we allow the community to systematically compare different approaches for disease module discovery, thus contributing to robustness and reproducibility in systems and network medicine.

• Samuele Firmani, Samuele Firmani, Helmholtz Center Munich
• Valter Bergant, Valter Bergant, Technical University of Munich
• Corinna Grünke, Corinna Grünke, Institute of Virology
• Yang An, Yang An, Helmholtz Munich
• Alexander Henrici, Alexander Henrici, TUM
• Francesco Paolo Casale, Francesco Paolo Casale, Helmholtz Munich
• Andreas Pichlmair, Andreas Pichlmair, Technical University of Munich
• Annalisa Marsico, Annalisa Marsico, Institute of Computational Biology

Recent outbreaks of COVID‑19 and monkeypox (Mpox) underscore the need for scalable tools that can disentangle complex host‑virus‑drug interactions and reveal potential therapeutic targets. Although multi‑omics technologies and high‑throughput screens generate rich datasets, integrating these heterogeneous signals to prioritise disease genes remains difficult, especially for poorly understood viruses.
We present GATE, a graph message‑passing neural network (MPNN) that ranks host factors and candidate antiviral drug targets across viral infections. Thanks to the pre‑training phase, GATE learns how extracts maximum value from sparse, weakly labelled data by learning to recognise disease‑related proteins and drug targets within a protein–protein interaction (PPI) network enriched with Gene Ontology functional embeddings, PPI‑derived positional encodings (PEs) and ESM2 language‑model embeddings. The model is then fine‑tuned with virus‑specific multi-omics data.
On both SARS‑CoV‑2 and Mpox tasks, GATE outperformed state‑of‑the‑art models and simpler baselines. Of the architectures evaluated, the Principal Neighbourhood Aggregation (PNA) layer propagated protein features most effectively. Positional encodings were the most informative inputs, while for SARS‑CoV‑2 the protein–viral interactome and effectome data also proved effective. Pre‑training further boosted performance, particularly in the data‑sparse Mpox setting. Predicted host‑factor genes significantly overlapped hits from independent CRISPR‑KO screens and matched validated antiviral targets for both viruses. In addition, GATE’s explanations show underlying biological mechanisms and help prioritise candidates for experimental validation.
GATE is task‑agnostic, scalable and accommodates future omics modalities through modular input feature sets, accelerating discovery and repurposing of antiviral therapeutics.

• Piotr Sliwa, Piotr Sliwa, University of Oxford
• Heather Harrington, Heather Harrington, University of Oxford
• Gesine Reinert, Gesine Reinert, University of Oxford
• Julian C. Knight, Julian C. Knight, University of Oxford

Integrative analyses of multi-omic patient datasets are crucial to uncover disease subtypes, yet challenges arise from modality‑specific variability and missing data. We propose MLModNet, a multilayer network framework for robust patient stratification. MLModNet employs an extended resampling‑based method (Pareto‑COGENT) to build stable, informative, and modality‑specific patient similarity networks, integrates them into a multiplex network including patients missing individual assays, and detects patient stratification via multiplex‑adapted Leiden clustering.

We applied MLModNet to the COVID‑19 Multi‑omic Blood Atlas (COMBAT) dataset, integrating proteomics, transcriptomics, and cytometry. MLModNet discovered five patient endotypes that refine clinical WHO severity categories, each exhibiting distinct immune‑metabolic signatures involving IL‑33, TREM1, interferon response pathways, and shifts in cell proportions. Clinically, MLModNet clusters significantly stratified ICU‑free survival, and early cluster assignment probabilities predicted subsequent clinical markers (CRP, D‑dimer, Acuity scores). External validation on an independent Olink dataset confirmed the reproducibility of these endotypes and their prognostic relevance. Extensive ablation analyses further supported the robustness of the identified clusters.

MLModNet thus provides a scalable strategy to translate heterogeneous, incomplete multi‑modal data into biologically meaningful, clinically actionable patient stratifications.

• Chad Myers

• Fiona Brinkman, Fiona Brinkman, Simon Fraser University

Microbiome analysis is increasingly becoming a critical component of a wide range of health, agri-foods, and environmental studies. I will present case studies showing the benefit of integrating very diverse metadata into such analyses - and also pitfalls to watch out for. The results of one such cohort study will be further presented, illustrating the need for analyses that allow one to flexibly view metadata in the context of microbiome data. The results support the multigenerational importance of “healthy