Feature: add LLM script for gene-disease directional relations (inhibit/activate/enhance) from literature

[AR-Shicheng]

Summary

Add an LLM-based script that screens batches of PubMed abstracts and extracts gene-disease relationships with directional labels.

Motivation

Manual review of literature for gene-disease directionality is slow and hard to scale. A reusable script would convert abstracts into structured gene-disease interaction maps that can support downstream analysis and prioritization. For KnowledgeGraph, directional gene-disease labels would enrich the structured claims and knowledge-graph edges already derived from PubMed literature.

Proposed behavior

Given a batch of PubMed abstracts (or PMIDs), the script should:

1. Extract gene and disease entities from each abstract
2. Infer the directional relationship between them
3. Return structured outputs with supporting evidence PMIDs and a short rationale

Directional relationship labels to support

• Activity: inhibit, activate, enhance, suppress, stimulate, promote, reduce, increase, decrease, modulate, regulate, upregulate, downregulate, potentiate, attenuate, block, amplify
• Receptor pharmacology: agonize, antagonize, partial agonism, inverse agonism, allosteric activation/inhibition
• Causal disease roles: cause, contribute to, drive, protect against, predisposes, risk-increasing/decreasing, associated with, exacerbate, ameliorate, prevent, trigger
• Genetic mechanisms: loss-of-function, gain-of-function, haploinsufficiency, dominant-negative, overexpression, underexpression, knockout/knockdown protective or pathogenic
• Molecular interactions: bind, phosphorylate, dephosphorylate, ubiquitinate, cleave, induce/repress expression
• Pathway effects: induce, repress, mediate, activate/inhibit pathway, synergize
• Clinical phenotype direction: worsen/improve phenotype, delay/accelerate onset, increase/decrease severity or susceptibility

Desired output format

gene | disease | directional_relationship | confidence | evidence_pmids | rationale

Acceptance criteria

• Accepts a batch of PMIDs or abstracts as input
• Extracts gene-disease pairs with directional relationship labels
• Returns evidence PMIDs and a brief rationale for each finding
• Outputs results in a structured table (CSV/JSON)
• Does not invent relationships not supported by the abstract text