Overcoming selection bias in synthetic lethality prediction.

Area of Science:

• Genomics
• Computational Biology
• Cancer Therapeutics

Background:

• Synthetic lethality (SL) is a promising strategy for cancer therapy, targeting genes with simultaneous loss of function.
• Identifying novel SL pairs is challenging due to the vast number of potential gene combinations.
• Existing computational methods for SL prediction often overlook selection bias in experimental data, limiting their generalizability.

Purpose of the Study:

• To develop a computational method for predicting synthetic lethal gene pairs that is robust to selection bias.
• To improve the accuracy and generalizability of synthetic lethality prediction for potential anti-cancer drug development.

Main Methods:

• Developed selection bias-resilient synthetic lethality (SBSL) prediction models using regularized logistic regression and random forests.
• Utilized 27 molecular features derived from cancer cell lines, patient tissues, and healthy donors to describe gene pairs.
• Trained and validated models on approximately 8000 experimentally derived SL pairs across breast, colon, lung, and ovarian cancers.

Main Results:

• SBSL models demonstrated superior predictive performance, generalizability, and robustness to selection bias compared to existing methods.
• Gene dependency, a measure of a gene's essentiality for cell survival, was the most significant predictor in SBSL models.
• Random forests outperformed linear models when dependency features were excluded, emphasizing the importance of mutation mutual exclusivity, healthy co-expression, and tumor differential expression.

Conclusions:

• The developed SBSL method offers a more reliable approach to predicting synthetic lethal gene pairs.
• Addressing selection bias is crucial for building robust and generalizable computational models for synthetic lethality.
• The findings highlight the potential of SBSL for accelerating the discovery of targeted cancer therapeutics.