Large Language Models

General-purpose large language models, used without any fine-tuning, match or complement specialized predictors on three tumor genomic interpretation tasks, with the largest gains on data from new assays and institutions.

Without any task-specific training, general-purpose large language models approach the accuracy of specialized supervised models at separating tumor mutations from clonal hematopoiesis variants, flagging oncogenic variants and predicting cancer type from tumor genomic profiles, and ensembling them with a supervised classifier improves generalization across institutions.

CodeWearBook is a retrieval-grounded large language model agent that answers natural-language questions about wearable health datasets through controlled SQL and Python execution layers, so the model’s access to raw data is restricted and the workflow can be adapted to new data sources and contexts.

Benchmarking zero-shot large language models against supervised, tailor-made predictors on cancer-type, somatic-vs-CHIP and variant-oncogenicity classification across 100,000+ samples from 19 institutions, and combining them into LLM–specialist ensembles that raise cross-institution accuracy by up to 55%.

Retrieval-grounded LLM agents that let researchers query large wearable health datasets in natural language, with privacy preserved by restricting the model to controlled SQL and Python execution layers instead of raw data access.