Multi-omics Sampling-based Graph Transformer for Synthetic Lethality Prediction

Xusheng Zhao, Hao Liu, Qiong Dai, Hao Peng, Xu Bai, Huailiang Peng

TL;DR

Researchers have developed MSGT-SL, a novel Graph Transformer framework that predicts Synthetic Lethality (SL)—a genetic state where the simultaneous mutation of two genes kills a cancer cell. By combining shallow multi-view Graph Neural Networks (GNNs) with a structure-aware "cross-omics" sampling strategy, the model captures both local biological patterns and long-range gene dependencies, significantly outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Synthetic Lethality is the "holy grail" of targeted cancer therapy. If Gene A is already mutated in a tumor, and Gene A and Gene B are "synthetic lethal" partners, then a drug blocking Gene B will selectively kill cancer cells while sparing healthy ones.

However, identifying these pairs in a lab is slow and expensive. Computational methods, specifically Graph Neural Networks (GNNs), have become popular for predicting SL pairs by treating genes as nodes and relationships as edges. But GNNs face two major hurdles:

  1. Structural Limitations: As GNNs get deeper to capture more information, they suffer from "over-smoothing" (features becoming indistinguishable) and "over-squashing" (information bottlenecks).
  2. Data Mismatch: There is a massive scale difference between known SL data (the "core" genes) and large-scale multi-omics data (gene expression, physical interactions, etc.). Integrating these without drowning out the specific SL signals is a non-trivial challenge.

Methods

The authors propose MSGT-SL (Multi-omics Sampling-based Graph Transformer), which operates in four distinct steps:

  1. Multi-view GNN (MVGNN): Instead of a deep GNN, they use a shallow, one-layer GNN across multiple "views" (different types of biological data). This captures immediate, local structural patterns without the risk of over-smoothing.
  2. Cross-omics Sampling: This is the "secret sauce." Starting with a batch of "core" genes from SL data, the model performs parallel random walks across multiple omics graphs. This identifies a balanced set of "non-core" genes that are structurally relevant, preventing the model from being overwhelmed by the sheer volume of general omics data.
  3. Self-attention (Graph Transformer): The sampled gene features are fed into a standard Transformer self-attention layer. Unlike GNNs, which are restricted by "hops" on a graph, the Transformer allows every gene in the sample to "communicate" with every other gene, capturing long-range dependencies.
  4. SL Prediction: Finally, gene features are concatenated in pairs and passed through a classifier to predict whether the relationship is lethal or viable.

Datasets

The model was evaluated on two real-world cancer cell line datasets:

  • K562 (Chronic myeloid leukemia)
  • Jurkat (Acute T-cell leukemia)

To enrich these, the authors integrated four categories of multi-omics views:

  • Physical & Genetic interactions from BioGRID.
  • Gene Expression profiles from the Cancer Cell Line Encyclopedia (CCLE).
  • Co-Essentiality data from DepMap (CRISPR-Cas9 screens).

Key Insights and Ideas

  • Random Walks as a Bridge: Using random walks for sampling isn't just about reducing data size; it’s about "structure-aware" integration. It ensures that the auxiliary omics data included is biologically related to the core genes being studied.
  • Transformers Solve GNN Depth Issues: By using a Graph Transformer, the authors bypass the need for deep GNN layers. The Transformer provides "global" awareness in a single layer, avoiding the performance degradation seen in deep GNNs.
  • Natural Structural Encoding: The authors found that concatenating features from different omics views acts as a "natural structural encoding," meaning they didn't need to add complex positional encodings typically required by Transformers.

Key Results

MSGT-SL consistently outperformed several state-of-the-art baselines (including GCN, GAT, and specialized SL models like KG4SL and PiLSL).

  • Performance Gain: On average, MSGT-SL showed a 4.5% improvement in the "leave-gene-pair-out" setting and a 3.8% improvement in the more difficult "leave-gene-out" (inductive) setting.
  • Robustness: Ablation studies showed that removing the MVGNN module caused the most significant drop in performance, proving that local structural information is still a critical foundation for global predictions.
  • Efficiency: Despite using a Transformer architecture, the sampling strategy keeps the computational cost competitive with simpler GNN models.

Limitations and Future Work

  • Wet-lab Validation: While the model shows high predictive accuracy, the authors note that these "dry-lab" predictions still require biological validation in a laboratory setting.
  • Knowledge Graph Integration: While the model uses multi-omics data, it doesn't yet fully leverage the complex, typed relationships found in massive Knowledge Graphs (KGs), which could be a future avenue for even richer feature learning.
  • Generalization: Future work aims to apply MSGT-SL to identify SL pairs across a wider variety of rare cancer types where data is even more sparse.

Summary updated

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.
BibTeX
@inproceedings{zhao2023multi,
  author = {Xusheng Zhao and Hao Liu and Qiong Dai and Hao Peng and Xu Bai and Huailiang Peng},
  title = {{Multi-omics Sampling-based Graph Transformer for Synthetic Lethality Prediction}},
  booktitle = {BIBM},
  year = {2023},
  url = {https://openreview.net/forum?id=iYInAYhT1Q},
}

Loading related content...

Figure