Reproduce “Patient-specific Biomolecular Instruction Tuning” on OpenResearch

About this paper

Irsyad Adam, Zekai Chen, David Laub, Shaun Porwal, Arda Pekis, Kevin Brown

Standard Model Biomedicine introduces KRONOS, a graph-LLM framework that integrates patient-specific proteomics with protein-protein interaction networks for precision oncology. Coupled with CPTAC-PROTSTRUCT, a novel proteomics instruction-tuning dataset, KRONOS achieved an AUC of 0.857 for mortality prediction and 0.849 for cancer type classification, demonstrating enhanced clinical prognostic capabilities.

Key insights

• Explicitly incorporating protein-protein interaction (PPI) network structure is crucial for enabling LLMs to accurately interpret patient-specific proteomics and achieve superior performance in clinical prognostic tasks.
• Graph-based approaches, particularly those utilizing Graph Attention Networks (GAT), provide richer contextualization for proteomics data in precision medicine applications compared to methods relying on patient similarity networks or individual protein abundances.
• The presented framework bridges existing gaps by providing both a specialized dataset and an architectural solution, advancing precision oncology through better integration of complex molecular profiles with AI reasoning.

Summary by alphaXiv.