Abstract
Synthetic lethality (SL) prediction is used to identify if the co-mutation of two genes results in cell death. The prevalent strategy is to abstract SL prediction as an edge classification task on gene nodes within SL data and achieve it through graph neural networks (GNNs). However, GNNs suffer from limitations in their message passing mechanisms, including over-smoothing and over-squashing issues. Moreover, harnessing the information of non-SL gene relationships within large-scale multi-omics data to facilitate SL prediction poses a non-trivial challenge. To tackle these issues, we propose a new multi-omics sampling-based graph transformer for SL prediction (MSGT-SL). Concretely, we introduce a shallow multi-view GNN to acquire local structural patterns from both SL and multi-omics data. Further, we input gene features that encode multi-view information into the standard self-attention to capture long-range dependencies. Notably, starting with batch genes from SL data, we adopt parallel random walk sampling across multiple omics gene graphs encompassing them. Such sampling effectively and modestly incorporates genes from omics in a structure-aware manner before using self-attention. We showcase the effectiveness of MSGT-SL on real-world SL tasks, demonstrating the empirical benefits gained from the graph transformer and multi-omics data.
Introduction
Synthetic Lethality (SL) represents a cornerstone of modern targeted cancer therapy. The concept is straightforward yet powerful: if two genes, A and B, are in a synthetic lethal relationship, the mutation of either gene alone allows the cell to survive, but the simultaneous mutation or inhibition of both leads to cell death. In a clinical context, if a cancer cell possesses a specific mutation (Gene A), a drug that targets its SL partner (Gene B) will selectively eliminate the tumor while sparing healthy tissue.
Despite its promise, identifying SL pairs through traditional wet-lab experiments is prohibitively slow and expensive. While computational methods—particularly Graph Neural Networks (GNNs)—have emerged as a popular alternative, they face significant technical hurdles. Deep GNNs often suffer from "over-smoothing," where node features become indistinguishable, and "over-squashing," where information bottlenecks prevent the model from capturing long-range dependencies. Furthermore, integrating massive multi-omics datasets (such as gene expression and physical interactions) with sparse SL data remains a challenge. To address these gaps, we present MSGT-SL, a Multi-omics Sampling-based Graph Transformer.
Methodology
MSGT-SL is designed to bridge the gap between local structural patterns and global gene dependencies through a four-stage architecture:
1. Multi-view GNN (MVGNN)
To capture immediate biological context without the degradation associated with deep networks, MSGT-SL utilizes a shallow, one-layer GNN. This layer operates across multiple "views"—different types of biological data—to extract local structural patterns. By keeping the GNN shallow, the model avoids over-smoothing while still grounding the gene representations in their immediate neighborhood.
2. Cross-omics Sampling
A key innovation of MSGT-SL is its structure-aware sampling strategy. Rather than attempting to process the entirety of multi-omics data, which can overwhelm the specific SL signal, the model performs parallel random walks starting from a batch of "core" genes. This identifies a balanced subset of "non-core" genes that are structurally relevant across different omics graphs, ensuring that the auxiliary data included is biologically pertinent to the prediction task.
3. Graph Transformer (Self-attention)
The sampled gene features are fed into a standard Transformer self-attention layer. Unlike traditional GNNs that are limited by fixed "hops" on a graph, the Transformer allows every gene in the sample to communicate with every other gene. This architecture naturally bypasses the over-squashing problem and allows the model to capture long-range dependencies that are often missed by local message-passing mechanisms.
4. SL Prediction
Finally, the refined gene features are concatenated in pairs and passed through a classifier. This module predicts the probability of a synthetic lethal relationship versus a viable one.
Experimental Setup
Datasets
The model was rigorously evaluated on two real-world cancer cell line datasets:
- K562: Chronic myeloid leukemia.
- Jurkat: Acute T-cell leukemia.
To provide a holistic biological context, we integrated four categories of multi-omics data:
- Physical & Genetic interactions (BioGRID).
- Gene Expression profiles (Cancer Cell Line Encyclopedia - CCLE).
- Co-Essentiality data (DepMap CRISPR-Cas9 screens).
Key Results and Insights
MSGT-SL consistently outperformed state-of-the-art baselines, including GCN, GAT, and specialized SL models like KG4SL and PiLSL.
- Performance Gains: The model achieved a 4.5% improvement in the "leave-gene-pair-out" (transductive) setting and a 3.8% improvement in the more challenging "leave-gene-out" (inductive) setting.
- Ablation Studies: Removing the MVGNN module resulted in the most significant performance drop, confirming that local structural information remains a critical foundation even when global dependencies are captured by the Transformer.
- Structural Encoding: Interestingly, the study found that concatenating features from various omics views acts as a natural structural encoding. This allowed the model to function effectively without the complex positional encodings typically required by Transformers.
Conclusion and Future Work
MSGT-SL demonstrates that combining shallow GNNs with Graph Transformers and structure-aware sampling provides a robust framework for SL prediction. By effectively integrating multi-omics data, the model overcomes the traditional limitations of deep GNNs.
While the results are promising, future research will focus on validating these "dry-lab" predictions in biological settings and extending the framework to leverage more complex, typed relationships found in large-scale Knowledge Graphs. Ultimately, MSGT-SL offers a scalable and accurate path toward identifying new therapeutic targets for rare and difficult-to-treat cancers.
Summary updated