A SHAP-Guided Grouped L1 Regularization Method for CRISPR-Cas9 Off-Target Predictions

CRISPR-Cas9 has emerged as a remarkably powerful gene editing tool and has advanced both research and gene therapy applications. Machine learning models have been developed to predict off-target cleavages. Despite progress, accuracy, stability, and interpretability remain open challenges. Combining predictive modeling with interpretability can provide valuable insights into model behavior and increase its trustworthiness. This study proposes a group-wise L1 regularization method guided by SHAP values. For the implementation of this method, the CRISPR-M model was used, and SHAP-informed regularization strengths were calculated and applied to features grouped by relevance. Models were trained on HEK293T and evaluated on K562. In addition to the CRISPR-M baseline, three variants were developed: L1-Grouped-Epigenetics, L1-Grouped-Complete, and L1-Uniform-Epigenetics (control). L1-Grouped-Epigenetics, using penalties split by on- and off-target epigenetic factors, moderately improved mean precision, AUPRC, and AUROC relative to the baseline, as well as showing reduced variability in precision and AUPRC across seeds, although its mean recall and F-metrics were slightly lower than those of CRISPR-M. L1-Grouped-Complete achieved the highest mean AUROC and Spearman correlation and presented lower variability than CRISPR-M for recall, F1, and F-beta, despite reduced recall and F-metrics relative to CRISPR-M. Overall, this approach required only minor architectural adjustments, making it adaptable to other models and domains. While results demonstrate potential for enhancing interpretability and robustness without sacrificing predictive performance, further validation across additional datasets is required.