AutoBio: Evidence-Guided Adaptation of Biological Analysis Workflows

Selecting an analysis workflow remains a substantial part of computational biology: established methods must be assembled, adapted to a dataset, and evaluated together. We present AutoBio, a system for turning prior workflows and limited experimental feedback into executable analyses. A numerical controller learns workflow contrasts, distinguishes supported estimates from unexplored changes, and optionally uses language-model hypotheses to prioritize experiments. We evaluate this design on three single-cell annotation and two protein-recovery tasks. A four-workflow starting portfolio provides strong initial performance; numerical adaptation recovers both measured protein-bank optima. In a paired advisor ablation, numerical-only search preserves final LFQ quality without inference, whereas advice improves PBMC balanced recall by 0.853 percentage points. Against OpenCode using the same model, AutoBio matches LFQ quality at 28.6\% lower inference cost and improves PBMC recall by 0.230 points at the common evaluation budget, although PBMC inference costs more. These development results support a modular approach to workflow automation in which scientific execution, numerical search, and language reasoning have distinct, measurable roles.