Antimicrobial resistance is a critical global health challenge, yet systematic computational frameworks for discovering drug–target interactions (DTIs) across diverse bacterial species remain lacking. Existing knowledge graph (KG)-based drug repurposing systems are centered on human disease biology and do not model pathogen-specific proteomes or cross-species molecular relationships. We present PathogenKG, a heterogeneous KG embedding framework for cross-species antimicrobial drug repurposing. It integrates protein–protein interaction networks from 31 bacterial pathogens, orthology relationships, functional annotations, and validated drug–target associations into a unified graph of over 3 million triples, enabling drug–target evidence to propagate between phylogenetically related organisms. To handle the extreme class imbalance inherent in this setting—only 302 known DTI edges among millions of triples—we design a training pipeline combining focal loss, type-constrained negative sampling, and targeted oversampling, validated on an established biomedical KG benchmark. We compared heterogeneous GNN encoders and evaluated the framework through a compound-centric repurposing protocol that ranks all pathogen proteins as candidate targets. Domain-expert analysis of top-ranked predictions in nine case study compounds confirms that the model recovers known interactions and generates mechanistically grounded cross-species hypotheses, with 22.2% of the top-20 predictions rated biologically plausible. PathogenKG is the first multi-pathogen KG framework specifically designed for cross-species antimicrobial drug repurposing via link prediction.
PathogenKG: Cross-Species Drug Repurposing via Heterogeneous Knowledge Graph Link Prediction / De Filippis, G.M., Rinaldi, A.M., Tommasino, C.. - 16948:(2027), pp. 601-618. (European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD 2026 ita 2026) [10.1007/978-3-032-37679-4_34].
PathogenKG: Cross-Species Drug Repurposing via Heterogeneous Knowledge Graph Link Prediction
Giovanni Maria De Filippis;Antonio Maria Rinaldi;Cristian Tommasino
2027
Abstract
Antimicrobial resistance is a critical global health challenge, yet systematic computational frameworks for discovering drug–target interactions (DTIs) across diverse bacterial species remain lacking. Existing knowledge graph (KG)-based drug repurposing systems are centered on human disease biology and do not model pathogen-specific proteomes or cross-species molecular relationships. We present PathogenKG, a heterogeneous KG embedding framework for cross-species antimicrobial drug repurposing. It integrates protein–protein interaction networks from 31 bacterial pathogens, orthology relationships, functional annotations, and validated drug–target associations into a unified graph of over 3 million triples, enabling drug–target evidence to propagate between phylogenetically related organisms. To handle the extreme class imbalance inherent in this setting—only 302 known DTI edges among millions of triples—we design a training pipeline combining focal loss, type-constrained negative sampling, and targeted oversampling, validated on an established biomedical KG benchmark. We compared heterogeneous GNN encoders and evaluated the framework through a compound-centric repurposing protocol that ranks all pathogen proteins as candidate targets. Domain-expert analysis of top-ranked predictions in nine case study compounds confirms that the model recovers known interactions and generates mechanistically grounded cross-species hypotheses, with 22.2% of the top-20 predictions rated biologically plausible. PathogenKG is the first multi-pathogen KG framework specifically designed for cross-species antimicrobial drug repurposing via link prediction.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


