Design and Evaluation of Network Algorithms and Deep Learning Models in Systems Biology and Biomedicine

Abstract

We present novel solutions to four interrelated problems in network biology and computational biomedicine. The first two address methodological challenges in network biology, while the latter two focus on applications of network-based computational modeling in therapeutics.

(i) We introduce a provenance tracing framework for random walk-based network diffusion algorithms. By quantifying the contribution of individual paths to a node's diffusion score, the framework captures the influence of global network topology and identifies critical mediators of information flow. This approach enhances the interpretability of diffusion-based predictions, a key requirement for their reliable application in biomedical contexts.

(ii) Recognizing that the utility of network diffusion and other network-based algorithms depends on the quality of the underlying networks, we present ICoN, an unsupervised coattention-based graph neural network model for integrating heterogeneous protein-protein interaction networks. ICoN learns joint embeddings across multiple networks and captures complementary biological evidence. ICoN surpassed individual networks across three downstream tasks: gene module detection, gene coannotation prediction, and protein function prediction. Compared to existing unsupervised network integration models, ICoN exhibits superior performance across the majority of downstream tasks and shows enhanced robustness against noise. This work introduces a promising approach for effectively integrating diverse protein-protein association networks, aiming to achieve a biologically meaningful representation of proteins.

(iii) Shifting our focus towards therapeutic applications of computational models, we systematically evaluate deep learning models for drug synergy prediction, an important challenge in combination therapy design. Using the SynVerse framework, we assess 16 models across diverse feature types and deep learning based architectures and demonstrate that models often exploit dataset-specific shortcuts rather than biologically meaningful drug or cell line features. Robust evaluation, as enabled by SynVerse, may improve the likelihood that model predictions hold up when tested in experimental and clinical settings.

(iv) Finally, SynVerse's findings motivate the need to build a database of synergistic drug combinations annotated with their underlying synergy mechanisms so that one may build mechanism-aware predictive models. Hence, in the final project, we construct a structured database of synergy mechanisms from biomedical literature using a large language modelbased pipeline. This approach integrates literature retrieval, information extraction, biomedical entity recognition, and knowledge graph construction to extract both free-form and structured representations of mechanisms. The resulting resource, encompassing around 3,000 drug combinations, lays the foundation for developing mechanism-aware predictive models for drug synergy.