NSF4SL

NSF4SL predicts synthetic lethality (SL) interactions using a negative-sample-free self-supervised contrastive learning framework to prioritize gene pairs for cancer research.

Key Features:

• Negative-Sample-Free Model: Learns from positive SL data alone using a self-supervised approach without requiring labeled negative (non-SL) samples.
• Contrastive Learning Framework: Employs two interacting neural network branches that use contrastive learning to learn gene representations and distinguish potential SL relationships without negative data.
• Feature-Wise Data Augmentation: Applies feature-wise data augmentation to increase training diversity and mitigate sparsity in SL datasets.
• Gene Ranking Problem Formulation: Frames SL prediction as a gene ranking task rather than binary classification to prioritize candidate SL partners.

Scientific Applications:

• Cancer Research and Drug Target Identification: Predicts SL interactions to identify and prioritize candidate anti-cancer drug targets and gene pairs for therapeutic exploration.
• Reduction of Experimental Costs: Prioritizes computational candidates to reduce the extent of wet-lab screening and associated experimental validation costs.

Methodology:

NSF4SL uses self-supervised contrastive learning with two interacting neural network branches and feature-wise data augmentation to learn gene representations from positive SL examples without negative samples.