Generalist large language models complement tailor-made predictors for tumor genomics interpretation

Area of Science:

• Genomics
• Artificial Intelligence
• Computational Biology

Background:

• Large language models (LLMs) possess broad knowledge but their utility in specialized medical tasks is uncertain.
• Clinical genomic interpretation involves complex data analysis for accurate diagnosis and treatment.

Purpose of the Study:

• To evaluate the performance of general-purpose LLMs on three key clinical tumor genomic interpretation tasks.
• To determine if LLMs can replace or augment existing task-specific predictive models.

Main Methods:

• LLMs were assessed on distinguishing tumor vs. non-tumor mutations, driver vs. passenger mutations, and inferring cancer type from sequencing reports.
• Performance was compared against benchmark tailor-made models and evaluated on out-of-distribution data.

Main Results:

• General-purpose LLMs matched benchmark performance for tumor mutation identification.
• Ensembling LLMs with tailor-made models enhanced performance for mutation classification.
• LLMs showed superior or supplementary performance for cancer type inference on novel data.

Conclusions:

• Current LLMs offer valuable complementary expertise for clinical genomic interpretation without fine-tuning.
• LLMs can augment state-of-the-art predictors, improving the accuracy and scope of genomic analysis in oncology.