CRISPR-FMC: A Dual-Branch Hybrid Network for Predicting CRISPR-Cas9 On-Target Activity

Chuxuan Li¹Jian Li^1*Quan Zou²Hailin Feng^1*

• ¹Zhejiang Agriculture and Forestry University, Hangzhou, China
• ²University of Electronic Science and Technology of China, Chengdu, China

Accurately predicting the on-target activity of sgRNAs remains a key challenge in CRISPR-Cas9 applications, largely due to the limited generalization ability of existing models across diverse datasets, small-sample conditions, and complex sequence contexts. Many current approaches rely on shallow architectures or unimodal encodings, hindering their ability to capture the intricate dependencies underlying Cas9-mediated cleavage. To address these limitations, we propose CRISPR-FMC, a dual-branch hybrid neural network that begins with a data preprocessing stage combining One-hot encoding and contextual embeddings from a pre-trained RNA-FM model. This dual-input scheme enables the model to represent both explicit nucleotide composition and highlevel semantic features. The architecture incorporates multi-scale convolution (MSC), BiGRU, and Transformer blocks to extract hierarchical sequence features, and employs a bidirectional cross-attention mechanism with a residual feedforward network to enhance multimodal fusion and generalization. Evaluations on nine public CRISPR-Cas9 datasets show that CRISPR-FMC consistently outperforms existing baselines in both Spearman and Pearson correlation metrics, with particularly strong performance under low-resource and cross-dataset conditions. Ablation studies validate the contribution of each component, while base substitution analysis highlights the model's sensitivity to the PAM-proximal region, aligning with biological evidence. Overall, CRISPR-FMC provides a robust and interpretable framework for sgRNA activity prediction across heterogeneous genomic scenarios.