Uncertainty-aware synthetic lethality prediction with pretrained foundation models.

Area of Science:

• Genomics
• Computational Biology
• Cancer Therapeutics

Background:

• Synthetic lethality (SL) is a promising cancer therapy strategy.
• Experimental identification of SL gene pairs is costly and limited to well-studied genes.
• Current computational methods struggle to generalize to novel genes due to reliance on curated networks.

Purpose of the Study:

• To develop a novel computational framework, Cilantro-sl, for predicting synthetic lethality gene pairs.
• To leverage pretrained biological foundation models for enhanced prediction accuracy and generalization.
• To provide uncertainty-aware predictions for reliable experimental prioritization of therapeutic targets.

Main Methods:

• A two-stage, graph-free framework utilizing pretrained biological foundation models.
• Stage 1: Context-aware embeddings from single-cell models, in silico gene knockouts, and CRISPR viability data.
• Stage 2: Pairwise feature derivation and classification, incorporating conformal prediction for uncertainty quantification.

Main Results:

• Cilantro-sl accurately predicts SL gene pairs, outperforming existing methods.
• Demonstrated zero-shot generalization to unseen gene pairs and genes.
• Ablation studies confirmed the importance of viability pretraining and gene priors, independent of PPI and GO features.

Conclusions:

• Cilantro-sl offers a scalable and robust method for discovering novel synthetic lethality targets.
• The framework transforms biological representations into practical, uncertainty-aware hypotheses for cancer therapy.
• Enables efficient identification of high-confidence SL candidates for experimental validation.