MuSL: Multimodal deep learning for generalizable prediction of synthetic lethality from sequence, transcriptomic, and network data

Synthetic lethality (SL) offers a promising paradigm for identifying selective anticancer targets. How ever, many computational SL prediction methods rely on handcrafted expression summaries or single data modalities, limiting their ability to capture fine-grained genegene dependency patterns directly from raw expression profiles and to generalize to unseen genes. We propose MuSL, a multimodal deep learning framework that integrates transcriptomic histograms, statistical expression features, and protein-protein interaction (PPI) network in formation for SL prediction. In MuSL, gene-pair expression profiles from either bulk or single-cell data are converted into two-dimensional histograms, allowing a convolutional neural network to learn distributional patterns such as co expression loss and mutual exclusivity directly from raw expression landscapes. A graph branch operating on a PPI network initialized with ESM2 protein embeddings captures complementary topological and sequence-informed priors, while a statistical branch provides explicit low-dimensional descriptors of the same transcriptomic profiles. These modalities are aligned through contrastive learning and integrated by cross-attention and adaptive gating. Across random, transductive, and inductive evaluation settings, MuSL consistently outperforms strong baselines and remains robust when test pairs contain partially or entirely unseen genes. These results support multimodal integration of raw expression landscapes and sequence-informed network priors as an effective strategy for generalizable SL prediction. The source code and datasets are available at (https://github.com/JieZheng-ShanghaiTech/MuSL).