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

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

Model Architecture

The MuSL framework integrates multimodal biological data through a tri-branch deep learning architecture to predict synthetic lethal (SL) gene pairs. As illustrated in the figure above, the model jointly learns from transcriptomic images, statistical features, and protein interaction networks.

1. Feature Learning Pathways

MuSL processes input data through three complementary pathways:

Image-Based Expression Branch (CNN):

Transforms the transcriptome profiles of gene pairs into two-dimensional joint density distribution maps (32×3232×32 histograms).

Utilizes a Convolutional Neural Network (CNN) to automatically extract spatial features, capturing complex patterns such as functional decoupling and mutual exclusivity directly from raw data.

Statistical Feature Branch (MLP):

Extracts 35 handcrafted statistical features capturing explicit expression patterns (e.g., correlation, variation) from the same transcriptomic data.

Projects these features through a dedicated fully connected layer (MLP) to form a representation that complements the deep image features.

Graph-Based Network Branch (GNN):

Models the Protein-Protein Interaction (PPI) network using a Graph Neural Network (GNN).

Crucial Innovation: Initializes node features using ESM2 protein language model embeddings, integrating evolutionary and sequence-level priors to generate robust topological relationship features.

2. Adaptive Fusion & Prediction

Cross-Modal Integration: The framework employs a cross-attention and gating fusion mechanism to dynamically weight and integrate features from all three pathways (CNN, Statistics, and GNN) based on their contextual importance.

Multi-Head Prediction: Predictions are generated from both modality-specific heads and the final fused representation using four specialized classifiers.

3. Training Strategy

Composite Loss Function: The model optimizes a joint objective that combines:

Classification losses from all four prediction heads.

Contrastive learning components to enforce semantic alignment and improve cross-modal consistency.

Required Data Files

The following H5AD files need to be downloaded and placed in the processed_data/ directory:

a549_cell_line_imputed.h5ad - A549 cell line single-cell RNA-seq data

k562_cell_line_imputed.h5ad - K562 cell line single-cell RNA-seq data

tcga_all.h5ad - TCGA bulk RNA-seq data

protein_embeddings.pt - Embeddings from ESM2

all_emb_genept.pkl - Embeddings from GenePT