Locally deployed AI agents for CRISPR guide discovery from outbreak reports

Abstract

Translating outbreak reports into computational work supporting diagnostic development requires biological interpretation, literature review, data retrieval, and specialized analysis. We developed a domain-specific agent harness to coordinate a CRISPR guide-discovery pipeline using locally hosted open-weight models. The agent interprets requests, reviews literature, assembles inputs, and initiates analysis; the pipeline assesses candidate guides for researcher inspection. Evaluation with Gemma 4 31B, Nemotron 3.5 Super VL, and Qwen 3.8 27B on an NVIDIA DGX Spark demonstrated completion of the computational workflow. Disease and pathogen identification were consistent across runs, while gene selection, strain coverage, and candidate yield varied. Nemotron 3.5 Super VL had the shortest reported times to evaluation. Qwen 3.8 27B applied pipeline evaluation criteria more consistently, but required longer evaluation times. The runs demonstrated the utility of smaller language models when coupled with structured bioinformatics workflows. These findings support the feasibility of locally deployed agent coordination for diagnostic research as frontier models add additional restrictions on biological tasks.