LARGE LANGUAGE MODEL AGENTS FOR BIOMEDICAL RESEARCH AND DRUG DISCOVERY: FROM LITERATURE MINING TO HYPOTHESIS GENERATION

Abstract

The rapid development of large language models (LLMs), retrieval-augmented generation, multimodal artificial intelligence, and autonomous AI agents is creating new opportunities for biomedical research and pharmaceutical discovery. Conventional biomedical research requires researchers to review extensive scientific literature, identify knowledge gaps, formulate hypotheses, select relevant experimental approaches, and iteratively evaluate evidence. The rapidly expanding biomedical literature, increasing complexity of biological datasets, and multidisciplinary nature of drug discovery make these activities increasingly difficult to perform manually at scale. Large language model agents offer a potential solution by combining natural-language reasoning with information retrieval, knowledge graphs, computational tools, databases, and specialized scientific workflows. This paper examines the emerging role of LLM-based agents across the biomedical research and drug-discovery pipeline, with particular emphasis on literature mining, evidence synthesis, disease-mechanism analysis, target identification, drug repurposing, molecular design, and hypothesis generation. A conceptual agentic framework is proposed in which specialized agents perform literature retrieval, evidence extraction, knowledge synthesis, hypothesis generation, hypothesis criticism, experimental design, and validation. Recent evidence indicates that multi-agent systems can generate experimentally testable biomedical hypotheses and can support iterative scientific discovery. A 2025 study using a multi-agent LLM framework for Alzheimer's disease drug-combination discovery reported improved predictive performance over a traditional knowledge-based approach and experimentally validated one predicted combination in vitro. Recent 2026 research has further demonstrated systems that connect literature search, hypothesis generation, experimental data analysis, and hypothesis refinement within a continuous scientific workflow. Nevertheless, challenges remain concerning hallucination, evidence quality, biological validity, reproducibility, data heterogeneity, interpretability, regulatory compliance, and human oversight. The paper concludes that LLM agents should be viewed as research collaborators and evidence-synthesis systems rather than autonomous replacements for biomedical scientists. Properly designed verification, provenance, expert review, and experimental validation mechanisms are essential for responsible deployment.