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

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• CRISPR/Cas9 genome editing relies on single-guide RNA (sgRNA) activity.
• Predicting sgRNA activity is crucial for efficient genome editing.
• Existing deep learning methods lack optimal strategies for integrating multiple biological features.

Purpose of the Study:

• To develop a deep learning framework, CrisprFusion, for accurate sgRNA activity prediction.
• To introduce a novel Multi-Grain Cross Attention Fusion Module for effective multimodal feature integration.
• To enhance the interpretability of sgRNA activity prediction models.

Main Methods:

• Developed CrisprFusion, a deep learning framework with a four-branch input structure for four biological features.
• Implemented a Multi-Grain Cross Attention Fusion Module for branch-level gating and token-level alignment.
• Evaluated the model on seven high-throughput datasets and four functional screens.

Main Results:

• CrisprFusion achieved superior average performance across multiple datasets compared to six baseline methods.
• The model demonstrated competitive performance in cross-cell-line validation.
• Ablation studies and attention visualization confirmed the effectiveness of the fusion module and feature importance.

Conclusions:

• CrisprFusion provides an effective and interpretable approach for multimodal biological feature integration in sgRNA activity prediction.
• The novel fusion module significantly improves prediction accuracy.
• This work advances the field of genome editing by enhancing the prediction of sgRNA efficacy.