ASGCL: Adaptive Sparse Mapping-based graph contrastive learning network for cancer drug response prediction

Area of Science:

• Computational biology
• Genomics
• Drug discovery

Background:

• Personalized cancer treatment is crucial due to genomic variability among patients.
• Identifying effective drug treatments requires understanding complex interactions between cancer cells and drugs.
• Existing methods face challenges in accurately predicting drug responses.

Purpose of the Study:

• To develop an innovative computational approach for personalized cancer drug treatment.
• To unravel latent interactions in cancer cell lines and drug data.
• To improve the prediction of drug responses for clinical decision-making.

Main Methods:

• Introduced the Adaptive Sparse Graph Contrastive Learning Network (ASGCL).
• Utilized the GraphMorpher module for graph structure enhancement via node attribute masking and topological pruning.
• Employed dual-level contrastive learning (node and graph levels) and a combination of supervised and contrastive loss for end-to-end feature representation learning.

Main Results:

• ASGCL significantly outperforms existing methodologies in predicting drug responses.
• Ablation studies confirmed the efficacy and robustness of individual ASGCL components.
• The model demonstrated proficiency in identifying nuanced drug responses.

Conclusions:

• ASGCL offers a potent tool for guiding clinical decision-making in personalized cancer therapy.
• The approach effectively addresses the complexity of predicting drug responses based on genomic differences.
• Enhanced graph representation learning is key to improving predictive accuracy in precision oncology.