Biomedical Network Science Lab
Welcome to the Biomedical Network Science Lab!
The Biomedical Network Science (BIONETS) lab investigates molecular disease mechanisms using techniques from network science, combinatorial optimization, and artificial intelligence. We develop algorithms and tools to mine multi-omics data for such mechanisms and to individuate novel strategies for mechanistically grounded drug repurposing and causally effective treatments of complex diseases. We also develop privacy-preserving decentralized biomedical AI solutions, which enable cross-institutional studies on sensitive data. Finally, we are interested in meta-scientific questions such as reproducibility and the impact of data bias on biomedical AI systems.
In our paper "Deep learning models for unbiased sequence-based PPI prediction plateau at an accuracy of 0.65" we show that usage of ESM-2 embeddings boosts performance in out-of-distribution PPI prediction to around 0.65 independently of model architecture.
In our paper "Inference of differential kinase interaction networks with KINference", we present the R tool KINference, which can be used to identify kinase-substrate links that are differentially active between two conditions.
In our paper "Data splitting to avoid information leakage with DataSAIL", we present an algorithm and Python package that facilitates leakage-reduced data splitting to enable realistic evaluation of ML models that are intended to be used in out-of-distribution scenarios.
In our paper "Emergence of power-law distributions in protein-protein interaction networks through study bias", we show that biased research interest in proteins and aggregation of interactions from multiple studies can explain why node degree distributions in PPI networks follow a power law.
In our paper "Spatial cell graph analysis reveals skin tissue organization characteristic for cutaneous T cell lymphoma", we present the Python tool SHouT to quantify tissue heterogeneity based on spatial omics data and use it to identify skin tissue patterns separating CTCL from benign conditions.
