Deep Prior Framework: integrating functional specificity with general plausibility for targeted protein evolution.

The efficiency of directed protein evolution largely relies on computational methods to enrich mutants with high fitness. Traditional strategies, such as zero-shot approaches based on Protein Language Models (PLMs), primarily leverage general "plausibility" priors learned from natural sequences. However, in the absence of experimental feedback, their ability to guide evolution toward specific functions ("specificity") remains limited. Here, we introduce the Deep Prior Framework (DPF), a novel paradigm that integrates universal structural plausibility with task-oriented specificity priors. DPF incorporates an innovative Bernoulli-Attention (BATT) module within a Mixture of Experts architecture, enabling efficient screening of high-fitness mutants. Benchmarking on nine deep mutational scanning datasets, DPF outperforms existing methods in terms of PLM-based method. More importantly, our method also shows high performance on in silico directed evolution of Blastobotrys adeninivorans xanthine dehydrogenase (BaXD) without intermediate experimental feedback. Experimental validation of the top-ranked mutants showed an average activity enhancement of over four-fold compared with WT, with the best mutant achieving more than a nine-fold improvement. Furthermore, we applied DPF to a large-scale annotation of unreviewed sequences in UniProt Knowledgebase (UniProtKB). Of the 42 913 366 predicted samples (~21.56% of the total), 90.07% (38 934 581 proteins) were assigned high-confidence functional labels. In summary, this study demonstrates that DPF, by incorporating specificity-aware functional priors, can significantly advance efficient and targeted protein engineering.