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

AcrSeek achieves a significant increase in AUPRC from 0.017 to 0.355 under extreme class imbalance in anti-CRISPR protein detection.

• Why it matters: Accurate identification of anti-CRISPR proteins is crucial for advancing genome editing and understanding phage-host interactions, but current methods struggle with highly imbalanced data and overestimated performance.
• What they did: The approach employs a metric-learning framework using a frozen protein language model with hybrid negative mining and a joint triplet-focal loss, validated through nested 5-fold cross-validation on datasets with 1:383 class imbalance.
• The result: AcrSeek demonstrates robust performance improvements, reaching an AUPRC of 0.508 with ESM-2 3B, enabling more reliable detection of scarce positive samples and potential application to other protein-function prediction tasks.