CRISPR-FusionNet: A Multimodal Ensemble Learning Framework for sgRNA Efficacy Prediction

Introduction: The Clustered Regularly Interspaced Short Palindromic Repeats–CRISPRassociated protein 12a (CRISPR-Cas12a) system, commonly known as CRISPR-Cpf1, has become a next-generation genome-editing tool with substantial promise for disease treatment and synthetic biology. However, enhancing its editing efficiency relies critically on accurate prediction of singleguide RNA (sgRNA) activity, which remains a major challenge. Existing methods, which primarily employ homogeneous encoding or single-model frameworks, are fundamentally limited in their ability to integrate symbolic, physicochemical, and semantic information, thereby constraining both predictive performance and generalizability. Methods: In this work, we present CRISPR-FusionNet, an ensemble learning framework designed for sgRNA activity prediction by leveraging multimodal feature fusion. Specifically, we developed dedicated deep spatio-temporal networks, each tailored to a distinct encoding strategy, to extract high-level sequence features. Furthermore, we incorporated the physicochemical properties of RNA sequences into the predictive model through a dynamic programming (DP) approach. Furthermore, we provide an online predictor for estimating the on-target editing efficiency of candidate sequences: http://www.csmu.edu.cn:8501 results: Experimental results demonstrated that CRISPR-FusionNet outperformed state-of-the-art methods on Cpf1 benchmark datasets. Results: Experimental results demonstrated that CRISPR-FusionNet not only outperformed stateof- the-art methods on CRISPR-Cpf1 benchmark datasets, but also exhibited strong performance on CRISPR-associated protein 9 (CRISPR-Cas9) test sets; it also yielded improved outcomes in classification tasks. Discussion: To validate robustness and generalizability, we extended evaluations to both classification tasks and CRISPR-Cas9 datasets, demonstrating consistent performance across distinct CRISPR-Cas systems. Feature importance analysis was conducted using Random Forest (RF) and SHapley Additive exPlanations (SHAP). Furthermore, a perturbation-based approach was employed to quantify the functional impact of specific nucleotide positions. Conclusion: To achieve accurate prediction of sgRNA on-target activity, multi-dimensional feature extraction is essential. This can be effectively accomplished by integrating deep learning and machine learning within an ensemble framework, which significantly boosts the model’s predictive capability.