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

This paper introduces a provenance-aware Text-to-Target framework that integrates schema-constrained large language model retrieval with numeric omics data analysis to generate interpretable, audit-ready disease targets and drug discovery strategies, demonstrating significant validation in Alzheimer's disease and pancreatic ductal adenocarcinoma.

The Gist

Imagine you are a detective trying to solve two very complex medical mysteries: Alzheimer's disease and a specific type of pancreatic cancer. To crack the case, you need two kinds of clues: hard numbers (like a spreadsheet of genetic data from patients) and stories (what scientists have already written in books and articles about how these diseases work).

The problem is that these two types of clues don't usually talk to each other. The numbers are too specific, and the stories are too general. If you just ask a super-smart AI (a Large Language Model) to read the stories, it might give you a vague answer that doesn't fit the specific numbers you have. If you just look at the numbers, you might miss the bigger picture of why those numbers matter.

This paper introduces a new "detective team" called Text-to-Target. Here is how it works, using a simple analogy:

The Detective Team's Strategy

Think of the AI as a Librarian who knows every book ever written about medicine, and the data analysis as a Forensic Accountant who crunches the specific numbers from your patient samples.

1. The Meeting (Fusion): Instead of letting the Librarian and the Accountant work separately, this new framework forces them to sit at the same table. The AI reads the books to find potential suspects (genes or drugs), but it must check its findings against the Accountant's hard numbers.
2. Sorting the Suspects: The system sorts the potential suspects into three groups:
   • The Anchors: These are the "super-suspects" who appear in both the books and your specific data. They are the most reliable leads.
   • The Hidden Hubs: These are suspects mentioned in the books but not explicitly in your data yet. The system keeps an eye on them as "hidden" possibilities.
   • The Novelty Nodes: These are brand new ideas that pop up when you connect the dots between the books and the data in a specific way, like a new theory that no one thought of before.
3. Building the Case: Once the suspects are sorted, the system builds a "strategy portfolio." It doesn't just guess; it creates a step-by-step plan for how to test these suspects, ensuring every step can be traced back to a specific book or a specific number.

The Results: Solving the Mysteries

The team tested this method on the two diseases mentioned:

• For Pancreatic Cancer (PDAC): The system narrowed down thousands of possibilities to a manageable list of 75 genes and created 23 specific strategies to test them. When they checked these against a massive database of real-world cancer cell tests (DepMap), the results were strong and supported their choices.
• For Alzheimer's (AD): They used stricter rules to be extra careful. This resulted in a tighter list of 34 genes and 14 strategies. When they checked these against a specialized brain research database (CRISPRbrain), the results were also statistically significant and well-supported.

The Bottom Line

The most important part of this paper isn't just that they found new suspects; it's that the whole process is transparent.

Imagine if a detective wrote a report where every single conclusion had a "receipt" attached to it, proving exactly which book or which number led to that idea. That is what this framework does. It ensures that every final suggestion for a drug or a target can be traced all the way back to the original evidence.

In short, this paper shows a way to combine the "wisdom of the crowd" (all the medical literature) with "hard evidence" (your specific patient data) to find the best leads for new treatments, without losing track of where the ideas came from. It creates a reproducible, auditable path from reading a book to finding a potential cure.

Technical Summary

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

Problem Statement
In biomedical discovery, interpreting numerical omics data for disease target identification and drug discovery requires synthesizing prior knowledge from scientific literature. While Large Language Models (LLMs) can rapidly retrieve disease mechanisms from biomedical text, relying on text-only outputs is insufficient for prioritizing targets and drugs. Such outputs often lack the necessary cohort-specific quantitative evidence, rendering them too general and unreliable for rigorous decision-making. The core challenge lies in effectively coupling the expansive mechanistic search space of LLMs with the constraints of numeric omics data to ensure both breadth and evidentiary validity.

Methodology
The authors propose a provenance-aware Text-to-Target framework designed to bridge this gap. The architecture integrates schema-constrained multi-model LLM retrieval with numeric omics data analysis through a specific modality-aware fusion step. This process operates as follows:

1. Candidate Partitioning: Retrieved candidates are categorized into three distinct groups based on their evidence support:
   • Overlap-supported anchors: Candidates supported by both literature retrieval and omics data.
   • Retrieval-only hidden hubs: Candidates identified via text but lacking immediate quantitative support.
   • Network-emergent novelty nodes: Candidates emerging from network topology analysis.
2. Staged Generation: These partitioned candidates are propagated into hypothesis and strategy generation stages, strictly governed by topology constraints.
3. Provenance Tracking: The system ensures full provenance closure, maintaining a traceable link from initial retrieval artifacts to final validation outputs.

Key Results
The framework was evaluated in two distinct disease contexts: Alzheimer's disease (AD) and pancreatic ductal adenocarcinoma (PDAC).

• Pancreatic Ductal Adenocarcinoma (PDAC): The workflow generated a balanced candidate universe of 75 genes and a portfolio of 23 strategies. These outputs demonstrated significant support from the DepMap (Cancer Dependency Map) database at both the target level and the strategy level.
• Alzheimer's Disease (AD): Under stricter candidate controls, the system produced a compact universe of 34 genes and 14 strategies. When evaluated against an expanded CRISPRbrain registry, both target-level axes showed statistical significance, accompanied by strong strategy-level enrichment.

Across both disease models, the final strategies preserved full provenance closure, enabling end-to-end auditability from the initial retrieval phase to the final validation outputs.

Significance and Claims
The paper claims to establish a transferable discovery architecture that addresses the limitations of using LLMs in isolation. Its primary significance lies in the demonstrated ability to:

• Constrain biological activity using omics evidence.
• Expand the mechanistic search space through LLM retrieval.
• Preserve interpretability through network-aware fusion.

The framework provides a reproducible basis for dual-disease target prioritization. Furthermore, it motivates the implementation of continuous literature-mechanism concordance through agentic evidence-refresh loops, ensuring that discovery processes remain aligned with the latest scientific evidence. The work does not claim to replace human expertise but rather to provide a structured, auditable method for integrating quantitative data with textual knowledge to enhance target and drug discovery.