AutoCompBio: Literature-Grounded Workflow Optimization for Computational Biology

Computational biology relies on analysis workflows whose performance depends on interactions among preprocessing, representation, integration, and clustering choices. We introduce \textbf{AutoCompBio}, an AutoML system that combines scientific literature with iterative execution to optimize these workflows. AutoCompBio retrieves relevant analytical decisions, translates them into a constrained search space, and refines candidate workflows using feedback from the target dataset. Across 17 datasets spanning single-cell RNA sequencing (scRNA-seq), lipidomics, proteomics, and metabolomics, it achieves the highest mean best score among seven strategies (0.612), improving over random search by $+0.052$ (95\% bootstrap CI $[+0.035,+0.069]$) on a shared, label-guided objective. Its advantage over a generic LLM with access to the same corpus is $+0.022$ (CI $[+0.009,+0.036]$). Gains are strongest in scRNA-seq, with positive results in metabolomics and proteomics and near-neutral performance in lipidomics. Improvements on held-out datasets in three modalities further support the portability of the approach. Together, these results establish AutoCompBio as an effective framework for adapting published methods to new datasets through constrained, empirical workflow search.