CrisprFusion: A feature fusion model with multi-type input features for sgRNA activity prediction

The CRISPR/Cas9 system enables precise and efficient genome editing, but its efficacy heavily relies on sgRNA activity. Although deep learning has been widely applied to sgRNA activity prediction, existing methods often integrate multiple biological features without a well-designed fusion strategy. To tackle this issue, we present CrisprFusion, a deep learning framework that explicitly encodes four biological features through a four-branch input structure. The core of our model is a novel Multi-Grain Cross Attention Fusion Module, which performs fusion at two levels: branch-level gating for adaptive reweighting of different modalities, and token-level alignment for capturing position-specific interactions along the 23-nt sgRNA sequence. We evaluate CrisprFusion on seven high-throughput datasets with six representative baselines. Our method achieves consistent and superior average performance across all datasets and remains competitive in cross-cell-line validation on four functional screens. Ablation experiments verify the effectiveness of the proposed fusion module, and attention visualization reveals the importance of individual biological features. Overall, CrisprFusion offers an effective and interpretable approach for multimodal biological feature integration in sgRNA activity prediction.