PriMPT: Prior-Informed Multiscale Paired-Token Representation Learning for CRISPR/Cas9 Off-Target Cleavage Activity Prediction

Abstract Accurate prediction of CRISPR/Cas9 off-target cleavage activity requires molecular representations that explicitly capture the interaction compatibility between guide RNA and target DNA. This study presents PriMPT, a prior-informed multiscale paired-token representation learning framework that redesigns guide-target tokenization using 1-, 2-, and 3-gram paired tokens and injects mechanism-informed pair-prior descriptors into the corresponding embeddings through scale-specific gated injection, thereby incorporating RNA–DNA interaction information before contextual modeling and facilitating interaction-aware learning of guide-target compatibility. Across guide-disjoint cross-validation on six public data sets and a cross-data set external test, PriMPT achieved the numerically highest mean area under the precision-recall curve (AUPRC) among the evaluated models in each of the seven scenarios, with ablation studies identifying input-level pair-prior injection as the most consistent contributor and Integrated Gradients analysis yielding attribution patterns consistent with established aspects of Cas9 target recognition. These results suggest that PriMPT captures mechanism-relevant sequence-pair patterns and offers a mechanism-informed representation learning strategy for CRISPR/Cas9 off-target risk assessment.