Unsupervised cell line embedding using pairwise drug response correlation.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• Human cell line models are crucial for disease research and drug discovery.
• Data heterogeneity and fragmentation in cell line characterization hinder optimal utilization.
• Existing methods struggle to integrate diverse drug screening and omics data.

Purpose of the Study:

• To develop an unsupervised deep learning model for integrating heterogeneous cell line data.
• To create a unified cell line embedding for enhanced drug discovery applications.
• To improve machine learning performance in predicting drug response and synergy.

Main Methods:

• An unsupervised deep learning model utilizing contrastive learning was developed.
• The model integrated heterogeneous drug response screening data into a unified cell line embedding.
• SHapley additive explanations (SHAP) were used to identify key genes influencing the embedding.

Main Results:

• The unified cell line embedding significantly enhanced downstream machine learning tasks.
• Model performance improved in predicting drug synergy and cell line drug response.
• Identified genes contributing to the embedding are linked to cancer drug resistance.

Conclusions:

• The developed deep learning embedding effectively integrates heterogeneous cell line data.
• This approach enhances the predictive power of machine learning models in drug discovery.
• The findings provide insights into genetic factors influencing cancer drug resistance.