An LLM Agent System for Reducing Hallucinations in Drug Discovery Tasks

Artificial Intelligence has revolutionized drug discovery, significantly accelerating the process of identifying new therapeutics. Large Language Models (LLMs) can also play a crucial role in this process, but they are prone to hallucinations, generating inaccurate or misleading content. In this thesis, AI techniques such as Fine-Tuning and Retrieval-Augmented Generation (RAG) are employed in order to overcome this issue. Through the integration of RAG and other tools, an LLM Agent system is designed to retrieve information from external sources and generate an automatic target dossier in the context of pancreatic cancer. This system can significantly assist researchers by providing precise information on specific targets. The results show that fine-tuning and RAG enhance the LLM's expertise in the biomedical domain, resulting in more accurate and comprehensive responses. The Agent successfully generates an automatic target dossier with precise and relevant information in a minimal amount of time, showing its potential to streamline the drug discovery process. The integration of additional tools and models, along with the application of multimodal machine learning, might further improve the system, enabling it to generate a more informative and complete target dossier.