BIND: Large-Scale Biological Interaction Network Discovery through Knowledge Graph-Driven Machine Learning

Abstract

The complex interactions between biological entities provide valuable insights into fundamental life processes, which pave the way to a deeper understanding of disease mechanisms for the development of innovative therapeutic strategies. To enable large-scale predictions of biological interactions, multifarious AI-driven predictors have been developed. However, most of these are developed by leveraging information from only a limited subset of interaction types, and the broader interaction types landscape could facilitate AI algorithms to learn more informative patterns to predict unknown interactions. To address the need of a robust and precise interaction predictor, we introduce BIND (Biological Interaction Network Discovery), a predictor that leverages simultaneous learning across 10 biological entities and 30 interaction types. To develop BIND, we first evaluate 11 distinct Knowledge Graph Embedding Methods on the largest public biomedical interaction dataset namely PrimeKG. For each relation type, we extracted entity embeddings from the top 5 performing Knowledge Graph Embedding Models (KGEMs) and input them into 7 distinct machine learning classifiers. Rigorous evaluation of 1,050 predictive pipelines demonstrated that specific combinations of KGEMs and classifiers achieved F1 scores of 90% to 99% across various interaction types. Comprehensive evaluation across each relation type identified the top-performing predictive pipelines, which became the foundation of the BIND web application. To reveal the practical utility of our web application in identifying novel biological interactions, we conducted a case study on drug-phenotype interactions. The application gave 1,355 high confidence predictions, from which potential interactions were subsequently validated by scientific evidence found within the existing literature (Table 5). We believe BINDs web applications public access will serve as a valuable tool for biologists to identify unknown interactions that can be subsequently validated through wet-lab experiments. O_TBL View this table: org.highwire.dtl.DTLVardef@18666a0org.highwire.dtl.DTLVardef@bbb608org.highwire.dtl.DTLVardef@1ac062borg.highwire.dtl.DTLVardef@c90ea1org.highwire.dtl.DTLVardef@1df9369_HPS_FORMAT_FIGEXP M_TBL O_FLOATNOTable 5.C_FLOATNO O_TABLECAPTIONPredicted Drug-Phenotype Associations with Strong Literature Support C_TABLECAPTION C_TBL