Improved precision oncology question-answering using agentic LLM

Despite the widespread application of Large Language Models (LLMs) in biomedical research, their clinical adoption faces significant challenges. These challenges stem from concerns about the quality, accuracy, and comprehensiveness of LLM-generated answers. Most existing work has focused on fine-tuning LLMs based on foundation models, which have not yet fully addressed accuracy and reliability issues. In this work, we propose an agent-based approach that aims to make LLM-based systems clinically deployable for precision oncology, while mitigating common pitfalls such as hallucinations, incoherence, and "lost-in-the-middle" problems. To achieve this, we implemented an agentic architecture, fundamentally shifting an LLM's role from a simple response synthesizer to planner. This agent orchestrates a suite of specialized tools that asynchronously retrieve information from various sources. These tools include curated document vector stores encompassing treatment guidelines, genomic data, clinical trial information, drug data, and breast cancer literature. The LLM then leverages its planning capabilities to synthesize information retrieved by these tools, generating comprehensive and accurate responses. We demonstrate GeneSilico Copilot's effectiveness in the domain of breast cancer, achieving state-of-the-art accuracy. Furthermore, the system showcases success in generating personalized oncotherapy recommendations for real-world cases.