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

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Synthetic lethality (SL) is a key strategy for targeted cancer therapy.
• Existing computational methods for SL prediction often lack the ability to capture complex gene dependencies from raw data or generalize to new genes.

Purpose of the Study:

• To develop a novel multimodal deep learning framework, MuSL, for accurate and generalizable synthetic lethality prediction.
• To overcome limitations of existing methods by integrating diverse data modalities.

Main Methods:

• MuSL utilizes a multimodal approach combining transcriptomic histograms, statistical expression features, and protein-protein interaction (PPI) network information.
• Gene expression profiles are converted into 2D histograms for convolutional neural network analysis.
• A graph neural network branch leverages PPI networks with protein embeddings, while a statistical branch incorporates explicit expression descriptors.
• Contrastive learning, cross-attention, and adaptive gating are employed for modality alignment and integration.

Main Results:

• MuSL consistently outperforms existing computational methods across various evaluation settings (random, transductive, inductive).
• The framework demonstrates robustness in predicting SL for gene pairs with partially or entirely unseen genes.
• The study validates the effectiveness of integrating raw expression data with network priors.

Conclusions:

• Multimodal integration of gene expression landscapes and sequence-informed network priors is a powerful strategy for generalizable synthetic lethality prediction.
• MuSL offers a significant advancement in identifying selective anticancer targets.
• The developed framework has the potential to accelerate the discovery of novel cancer therapies.