Sequora: Agentic Sequence Analysis for Biosecurity and Genome Curation

A fast-arising risk of artificial intelligence (AI) is the large-scale, democratized ability to design new-to-nature DNA, RNA, or protein sequences, with the added danger that such sequences could evade existing biosecurity screening. Conversely, AI can help address this risk with thorough, evidence-based analysis of these sequences and their putative function. Here, we introduce $\textbf{Sequora}$, a specialized agentic system for predicting the function and potential pathogenicity of biological sequences. Sequora uses a large language model (LLM) to select from 76 bioinformatics tools exposed via six Model Context Protocol (MCP) servers, reasons through their outputs, then returns biological claims with supporting evidence and confidence scores synthesized from empirically-derived evidence tiers. On a benchmark of AI-redesigned protein variants that had evaded commercial synthesis-order screening, Sequora reached $90.7\%$ sensitivity at $98.9\%$ specificity ($F_1 = 94.9$, MCC $= 0.865$, $n = 2{,}151$), exceeding all seven configurations of the four screening providers evaluated, including the tools they patched after disclosure. It further named the pathogenic parent protein on $84.9\%$ of its true positives. On an emerging WHO critical-priority pathogen, $\textit{Candidozyma auris}$, Sequora surfaced new functional annotations for 111 proteins which were labeled ``hypothetical'' in the most recent NCBI release. A professional Candida Genome Database curator rated $94.6\%$ of Sequora annotations acceptable under a six-point study rubric, compared with $17.1\%$ for the base LLM model (Wilcoxon $p = 7.1 \times 10^{-18}$). One assignment identified a homolog of GWT1, a known antifungal target, linking a previously unannotated protein to a potential direction for antifungal research. These applications illustrate how specialized agentic sequence analysis can support biosecurity screening and broader biological research.