Boosting GPT models for genomics analysis: generating trusted genetic variant annotations and interpretations through...

Area of Science:

• Genomics
• Artificial Intelligence
• Bioinformatics

Background:

• Large language models (LLMs) possess broad knowledge but lack specialized domain expertise, such as in genomics.
• Variant annotation data is critical for interpreting genetic variants and prioritizing disease-related findings from large-scale sequencing.
• Current LLMs require enhancement to effectively process and interpret complex genomic data.

Purpose of the Study:

• To improve the performance of LLMs in the genomics domain.
• To integrate variant annotation data into LLMs using retrieval-augmented generation (RAG) and fine-tuning techniques.
• To enhance the interpretation and prioritization of genetic variants.

Main Methods:

• Implemented retrieval-augmented generation (RAG) to integrate 190 million variant annotations into GPT-4o.
• Utilized fine-tuning techniques on GPT-4 with variant annotation data.
• Compared the effectiveness of RAG and fine-tuning for knowledge injection.

Main Results:

• RAG successfully integrated a large volume of accurate variant annotations, enabling users to query specific variants for interpretation.
• Fine-tuning improved performance in some annotation fields, but overall accuracy remained suboptimal compared to RAG.
• RAG demonstrated superior performance over fine-tuning in terms of data volume, accuracy, and cost-effectiveness for factual knowledge integration.

Conclusions:

• The study successfully enhanced LLM capabilities in genomics through RAG and fine-tuning.
• RAG is a more effective method than fine-tuning for injecting factual genomic knowledge into LLMs.
• This pioneering work paves the way for advanced AI systems in genomics for clinical diagnosis and research.