AcrSeek: Metric Learning with Hybrid Negative Mining for Anti-CRISPR Protein Detection under Extreme Class Imbalance

Anti-CRISPR (Acr) proteins inhibit CRISPR-Cas immunity and are key targets for precise control of CRISPR-based genome editing and phage-host coevolution research. Their experimental identification is costly and low-throughput, so computational predictors are used to prioritise candidates for testing. In this setting validated positives number in the hundreds while phage-derived putative negatives reach tens of thousands, producing extreme class imbalance. Yet existing classification-based predictors train and evaluate on 1:1 balanced datasets and allow high-similarity sequences between training and test, overestimating real-world performance.

We present AcrSeek, a metric-learning framework that trains a projection head over a frozen protein language model (PLM) encoder with hybrid negative mining (offline global plus online semi-hard) under a joint triplet-focal objective. Under nested 5-fold cross-validation with leakage controlled by 40% identity clustering, AcrSeek improved AUPRC on the imbalanced (1:383) test set from 0.017 to 0.355 over AcrNET (matched PLM backbone), reaching 0.508 with ESM-2 3B. AcrSeek therefore remains structurally robust under extreme class imbalance and may also serve other protein-function detection problems with scarce positives.

Source code at https://github.com/jeongchans/acrseek, archived at Zenodo (DOI: 10.5281/zenodo.21946803); data and model checkpoints at Zenodo (DOI: 10.5281/zenodo.20115442).

Supplementary data are available at Bioinformatics online.