Active Learning Dynamics for Targeted Enzyme Discovery and Overall Enzyme Function Prediction in a Practical Human-in-the-Loop Laboratory Setting

Most of the sequenced protein universe is functionally unannotated, and the enzymes of greatest value are often those least similar to anything already characterized. Active learning (AL) offers a route to spend a limited experimental budget well, but its behavior under real laboratory constraints – small batches, a pool orders of magnitude larger than the labeled training set, and assays returning a single target label – is poorly characterized. We study AL dynamics for overall enzyme function prediction and for targeted discovery of specific enzyme classes on a 30% amino acid identity-stratified benchmark, and execute the first round of human-in-the-loop laboratory characterization of AL-selected glycosyl hydrolase enzymes. Learning is dominated by batch size and fine-tuning procedure rather than acquisition strategy, often producing catastrophic forgetting at resource-limited laboratory scales, which rehearsal and larger batches alleviate. In targeted discovery of four EC classes, acquisition recovers most target sequences from the pool, but high pool recall does not translate into generalizable test performance for three of the four targets; whether it does is strongly governed by the embedding geometry of the target class and by how well labeled seed examples cover test embedding space. Finally, we report the first wet-lab round-trip for targeted discovery of marine glycosyl hydrolases: of eight AL-selected candidates, six expressed sufficiently for functional assay. We note this submission is a work-in-progress; we note ongoing work throughout, and look forward to catalyzing discussion and collaborations.