About KG2ML-RAG

What is KG2ML-RAG?

KG2ML-RAG is an integrated platform for exploring and prioritizing novel biomedical associations. It combines predictions generated by the KG2ML framework, literature-based evidence retrieval and analysis using retrieval-augmented generation (RAG), and knowledge subgraph visualization through CondensedKG.

The platform integrates three key components:

• Positive-Unlabeled (PU) learning (PULSNAR): Identifies likely novel associations between diseases and biomedical entities (e.g. gene-disease associations) that are not explicitly represented in the Data Distillery Knowledge Graph (DDKG). PULSNAR is the PU-learning algorithm used within the KG2ML framework to identify candidate associations.
• CondensedKG: Provides a focused representation of relationships among diseases, genes, compounds, and other biomedical entities. CondensedKG is a subgraph generated from the DDKG and enables users to explore the biological context surrounding predicted associations.
• Retrieval-Augmented Generation (RAG): Retrieves relevant scientific literature and synthesizes evidence to generate literature-backed explanations for candidate associations. The explanations include citations to the underlying publications, allowing users to trace the evidence back to the original sources.

Together, these components create a workflow that moves from prediction → evidence → interpretation. KG2ML-RAG is designed to help researchers prioritize novel hypotheses generated by KG2ML, understand the biological context of predicted associations, and efficiently identify relevant supporting or conflicting evidence without manually searching through thousands of publications.

Resources

KG2ML-RAG utilizes the following resources:

• KG2ML — ML framework to identify disease-associated genes.
• PULSNAR — Class proportion estimation under SNAR.
• ProteinGraphML — Disease-protein associations.
• CFDE — NIH Common Fund Data Ecosystem.
• UMLS — Unified Medical Language System.
• DDKG — Data Distillery Knowledge Graph.
• DrugCentral — Online drug information resource.
• Pharos — Online resource for drug target information.
• TIN-X — Target importance and novelty explorer.
• NIH PubMed — Biomedical literature used for retrieval-augmented generation.
• EU PMC — Biomedical literature used for retrieval-augmented generation.

Team

Contact

For questions, feedback, or collaboration, Contact Us.