Using graph-based model to identify cell specific synthetic lethal effects

Abstract

Synthetic lethal (SL) pairs are pairs of genes whose simultaneous loss-of-function results in cell death, while a damaging mutation of either gene alone does not affect the cell's survival. This makes SL pairs attractive targets for precision cancer therapies, as targeting the unimpaired gene of the SL pair can selectively kill cancer cells that already harbor the impaired gene. Limited by the difficulty of finding true SL pairs, especially on specific cell types, current computational approaches provide only limited insights because of overlooking the crucial aspects of cellular context dependency and mechanistic understanding of SL pairs. As a result, the identification of SL targets still relies on expensive, time-consuming experimental approaches. In this work, we applied cell-line specific multi-omics data to a specially designed deep learning model to predict cell-line specific SL pairs. Through incorporating multiple types of cell-specific omics data with a self-attention module, we represent gene relationships as graphs. Our approach achieves the prediction of SL pairs in a cell-specific manner and demonstrates the potential to facilitate the discovery of cell-specific SL targets for cancer therapeutics, providing a tool to unearth mechanisms underlying the origin of SL in cancer biology. The code and data of our approach can be found at https://github.com/promethiume/SLwise.

Keywords: Cell specific target identification; Deep learning; GNN; Multi-omics; Synthetic lethality.

Fig. 1

The framework of our method. The SLwise model incorporates a multiple graph learning approach to integrate diverse cell-specific omics data. Individual graph neural networks are first applied to each omics data (e.g. ES, paralogs). Two layers of GraphSAGE networks are utilized to generate node embeddings by aggregating neighbor features. An attention mechanism then combines the omics embeddings into an integrated representation. These node embeddings are fed into a deep neural network to make SL predictions.

Fig. 2

The performance of evaluation in three different cell lines under different split test set.

Fig. 3

The performance of transferable evaluation in three different cell lines.

Fig. 4

Representative predicted gene pair (BCL2L2-WEE1) and its SL mechanism analysis in A375 cell line. (A). The prediction score and rank of all candidate SL gene pairs. The blue dots are SL pairs with positive labels. Predictions falling on the left side of the dotted orange line represent the top 5% ranking. (B). Among the downregulated genes of BCL2L2 and WEE1, RHOA and CNOT9 are the common significant driver gene with a CERES score below − 0.5 in A375. (C). The Reactome pathway analysis demonstrates the presence of cellular damage. (D). The Hallmark analysis demonstrates the presence of cellular damage.

Fig. 5

Representative predicted gene pair (UBC-UBE2L6) and its SL mechanism analysis in HT29 cell line. (A). The prediction score and rank of all candidate SL gene pairs. The blue dots are SL pairs with positive labels. Predictions falling on the left side of the dotted orange line represent the top 5% ranking. (B). Among the downregulated genes of UBC and UBE2L6, STIL is the only common significant driver gene with a CERES score below − 0.5 in A375. (C). The Reactome pathway analysis demonstrates the presence of cellular damage. (D). The Hallmark analysis demonstrates the presence of cellular damage.