Using Retrieval-Augmented Generation to Capture Molecularly-Driven Treatment Relationships for Precision Oncology

Area of Science:

• Oncology
• Artificial Intelligence
• Bioinformatics

Background:

• Precision oncology relies on expert review of extensive literature for treatment decisions.
• Manual literature review is time-consuming and can be a bottleneck in clinical practice.
• Generative artificial intelligence offers potential solutions for streamlining information retrieval.

Purpose of the Study:

• To evaluate the efficacy of retrieval-augmented generation (RAG) in supporting precision oncology treatment discussions.
• To assess the ability of an untrained large language model (LLM) to extract treatment relationships using RAG.
• To determine if RAG can reduce the labor involved in evidence-based treatment selection.

Main Methods:

• A RAG pipeline was implemented to provide relevant text chunks from publications to a large language model (LLM).
• The system utilized a trusted data source (OncoKB) for retrieving cancer treatment information.
• An off-the-shelf, untrained Llama 2 model was queried with simple questions to test information retrieval capabilities.

Main Results:

• The RAG pipeline successfully retrieved treatment relationships from the OncoKB data source.
• The untrained Llama 2 model reproduced over 80% of the known treatment relationships.
• This demonstrates the potential of RAG for efficient information extraction without model fine-tuning.

Conclusions:

• Retrieval-augmented generation (RAG) shows significant promise for enhancing precision oncology by accelerating access to treatment evidence.
• LLMs integrated with RAG can effectively answer clinical questions by leveraging curated scientific literature.
• This AI-driven approach has the potential to optimize clinical decision-making in cancer care.