Drug Sensitivity Prediction Based on Multi-stage Multi-modal Drug Representation Learning.

Area of Science:

• Computational biology
• Pharmacogenomics
• Artificial intelligence in medicine

Background:

• Personalized cancer treatment requires accurate prediction of anticancer drug responses.
• Current models face challenges in comprehensively representing drug properties and cell-drug interactions.
• Improving prediction accuracy can enhance patient survival and reduce healthcare costs.

Purpose of the Study:

• To develop a drug sensitivity prediction model, ModDRDSP, that leverages multi-stage, multi-modal drug representations.
• To comprehensively capture drug properties and model complex cell-drug interactions for improved prediction accuracy.
• To establish a superior model for predicting anticancer drug responses in personalized medicine.

Main Methods:

• Drug representation learning using deep hierarchical bi-directional GRU (DSBiGRU) for SMILES and deep message-crossing network (DMCN) for molecular graphs.
• Integration of cell line multi-omics data using a convolutional neural network (CNN).
• Ensemble deep forest algorithm for final drug sensitivity prediction.

Main Results:

• ModDRDSP demonstrated superior performance compared to four leading industry models.
• Ablation experiments validated the contribution of each module within the ModDRDSP model.
• Case studies confirmed ModDRDSP's effectiveness in predicting drug sensitivity.

Conclusions:

• ModDRDSP offers a powerful and validated approach for predicting anticancer drug sensitivity.
• The model's multi-modal representation and integration of multi-omics data contribute to its enhanced performance.
• This advancement supports the development of more effective personalized cancer therapies.