Improving anti-cancer drug response prediction using multi-task learning on graph convolutional networks.

Area of Science:

• Computational biology
• Bioinformatics
• Precision oncology

Background:

• Accurate prediction of anti-cancer drug response is crucial for personalized cancer treatment.
• Current computational methods often treat drug response prediction as isolated classification or regression tasks, neglecting inter-task relationships.

Purpose of the Study:

• To develop an advanced computational model for predicting anti-cancer drug efficacy by integrating multiple prediction tasks.
• To improve the accuracy and robustness of drug response prediction in precision oncology.

Main Methods:

• Proposed a Multi-task Interaction Graph Convolutional Network (MTIGCN) model.
• Employed graph convolutional networks for cell line and drug embedding generation.
• Utilized multi-task learning, combining drug sensitivity/resistance classification with regression and similarity network reconstruction.

Main Results:

• MTIGCN significantly outperformed seven state-of-the-art baseline methods on in vitro datasets.
• The model demonstrated strong predictive performance on independent in vivo datasets (PDX and TCGA).
• Case studies validated MTIGCN's capability to identify novel drug responses in cell lines.

Conclusions:

• MTIGCN offers a powerful and integrated approach for anti-cancer drug response prediction.
• The model's ability to leverage multi-task learning enhances feature representation and reduces overfitting.
• MTIGCN shows promise for advancing precision oncology by improving therapeutic strategy selection.