PharmaMap-LLM: Fine-Tuning Large Language Models for Drug-Drug Interaction (DDI) Analysis

Abstract

Adverse Drug-Drug Interactions (DDIs) pose a critical challenge under polypharmacy. Traditional methods (rules, classical ML) struggle with scale and novelty in free-text literature. The proposed framework PharmaMap-LLM a DDI extraction framework that instruction-tunes a biomedical LLM via QLoRA for memory-efficient do main adaptation on a single RTX 3090. The system uses an agentic flow: a Predictor (MedLLaMA-3-8B) with two Validators and a clinician-facing Risk agent so that predictions, confidence, and rationales are cross-checked for explainability. Training leverages the DDIExtraction 2013 corpus and an instruction dataset derived from DrugBank pairs (∼20,000), with entity normalization (SciBERT tokenizer). The work reports hardware and training arguments (learning rate, LoRA ranks, batch/sequence sizes) to support reproducibility. On DDIExtraction 2013, PharmaMap-LLM attains F1 (0.87), Precision (0.89), and Recall (0.85), surpassing prior text-mining baselines. Human evaluation by a clinical pharmacologist (200 cases) rated 83% of responses as accurate or better (mean 4.2/5); an LLM-as-judge audit (GPT-4, 2,000 outputs) found 92.3% correctness. Against GPT-4o on 500 shared prompts, validity was similar; the difference w as not significant by McNemar's χ2 (continuity-corrected), p ≈ 0.40. Brief error analysis highlights under-detection of pharmacodynamic patterns and missing dose/time qualifiers, which inform our mitigation strategy (PD-focused exemplars, retrieval of threshold cues, calibrated confidence). Overall, the proposed system achieves an explainable, reproducible, and deployment-friendly DDI tool for clinical decision support.