AcrPLMEvo: A Two-Stage Framework Integrating Evolutionary Profiles with Protein Language Models for Anti-CRISPR Prediction.

Abstract

Anti-CRISPR (Acr) proteins are natural inhibitors of CRISPR-Cas systems and are important regulators for controllable genome-editing applications. However, their computational identification remains challenging because Acrs are sequence-diverse, weakly conserved, and supported by limited labeled data. Here, we present AcrPLMEvo, a two-stage framework that integrates protein language model (PLM) representations with PSSM derived evolutionary profiles for low-homology Acr prediction. We systematically compared four representative PLM backbones, parameter-efficient adaptation strategies, and alternative PSSM-coupling routes. Evolutionary profiles were not universally beneficial; instead, their effects depended on both PLM backbone and the stage at which they were incorporated. A key finding was that evolutionary information was more consistently beneficial when retained at the downstream decision stage than when used only during PLM adaptation. Guided by this observation, AcrPLMEvo combines PSSM-aware DoRA adaptation of ESM-2 with frozen feature extraction and final-stage evolutionary feature reintroduction. In the matched benchmark comparison, AcrPLMEvo achieved the best overall performance among competing Acr predictors, with an AUC of 0.965 and an AUPRC of 0.778. Its predictive reliability was further supported on an independently curated external set of 44 proteins, where it correctly classified 41 proteins and produced no false positives. These results indicate that stage-consistent integration of evolutionary profiles can improve PLM-based Acr prediction and support the prioritization of low-homology Acr candidates.