Abstract

High-throughput sequencing has generated vast genomic repositories that remain under-annotated, hampering enzyme discovery. We present an integrated pipeline that (i) builds a high-resolution, cross-kingdom phylogenetic database, (ii) mines candidates via multilocus phylogeny, (iii) predicts activities using an evolutionary-scale protein language model, and (iv) removes false positives through multilevel residue–atom contact rescoring. When applied to the r-BOX pathway, this approach uncovered numerous previously undocumented FadB, BktB, Ter, and YdiI homologues. Our activity model achieved R2 = 0.68 and reduced the RMSE on high-value targets by 11% compared to the prior SOTA (UniKP). Contact scoring improved early enrichment (EF1%) by 16-fold. Experimental validation targeting FadB increased titers from 0.65 g/L (shake flasks) to 1.7 g/L, reaching 10.2 g/L in a fermentation process. Together, these results establish a robust, generalizable framework for discovering scarce, high-value enzymes and prioritizing functional variants at scale.