Interpreting Omics Data Analysis with Large Language Models for Disease Target and Drug Discovery

Area of Science:

• Biomedical Informatics
• Computational Biology
• Drug Discovery

Background:

• Synthesizing prior knowledge is crucial for interpreting omics data in disease target identification.
• Large language models (LLMs) can retrieve disease mechanisms but lack specificity for prioritization without quantitative evidence.

Purpose of the Study:

• To develop a provenance-aware Text-to-Target framework coupling LLM retrieval with omics data analysis.
• To enable reliable and auditable disease target and drug prioritization.

Main Methods:

• Schema-constrained multi-model LLM retrieval integrated with numeric omics data analysis.
• A modality-aware fusion step partitions candidates (anchors, hubs, nodes) for staged hypothesis generation.
• Topology constraints and provenance closure ensure end-to-end auditability.

Main Results:

• In pancreatic ductal adenocarcinoma (PDAC), the framework identified a 75-gene universe and 23 strategies with DepMap support.
• For Alzheimer's disease (AD), it yielded a 34-gene universe and 14 strategies, showing significant enrichment with CRISPRbrain registry data.
• Final strategies maintained full provenance, linking retrieval artifacts to validation outputs.

Conclusions:

• The Text-to-Target framework offers a transferable architecture for omics-driven biological discovery.
• It balances LLM's broad mechanistic search with omics data's quantitative constraints for interpretable prioritization.
• The approach provides a reproducible basis for dual-disease target prioritization and continuous evidence-refresh loops.