In-Silico Molecular Binding Prediction for Human Drug Targets Using Deep Neural Multi-Task Learning.

Area of Science:

• Computational biology
• Drug discovery
• Genomics

Background:

• Deep neural multi-task learning (MTL) shows promise for in-silico prediction of molecular binding in human genomes, excelling with imbalanced data and avoiding overfitting.
• The interrelation between tasks in MTL is crucial but can also negatively impact performance, an effect often underestimated.
• Understanding task relatedness is key to optimizing MTL for complex biological predictions.

Purpose of the Study:

• To investigate the impact of target similarity on the effectiveness of multi-task learning for human drug target binding prediction.
• To compare the performance of multi-task learning (MTL) versus single-task learning (STL) across different target set compositions.
• To develop an improved MTL approach for predicting human drug target binding.

Main Methods:

• Utilized molecular interaction data from the ChEMBL database for human targets.
• Trained and evaluated various deep neural multi-task and single-task networks.
• Clustered human targets based on sequence similarity in their binding domains and selected diverse target sets for analysis.

Main Results:

• The performance of MTL was highly dependent on the sequence similarity within the target set.
• MTL underperformed compared to STL for diverse target sets and overall human targets.
• MTL significantly outperformed STL for target sets composed of highly similar targets.

Conclusions:

• Target similarity is a critical factor for successful multi-task learning in molecular binding prediction.
• A novel approach, Multiple Partial Multi-Task learning, was developed, enhancing MTL for human drug target binding prediction.
• This study provides insights into optimizing MTL strategies for specific biological prediction tasks.