Improvement of multi-task learning by data enrichment: application for drug discovery

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Multi-task learning (MTL) is increasingly vital in drug discovery for predicting compound activities.
• Challenges remain in optimizing MTL models for enhanced prediction performance.
• Deep neural networks (DNNs) are powerful tools for complex predictive tasks in pharmacology.

Purpose of the Study:

• To investigate the impact of training data enrichment on multi-task deep neural network (MT-DNN) prediction quality in drug discovery.
• To evaluate how varying information capacity in training data affects model performance.
• To assess the generalizability of MT-DNNs for novel compound-target interactions.

Main Methods:

• Utilized three datasets: ViralChEMBL (classification), pQSAR(159), and pQSAR(4267) (regression).
• Developed feed-forward DNN-based multi-task models using the PyTorch framework.
• Evaluated four training data enrichment scenarios and two test data types to assess prediction accuracy.

Main Results:

• Training data enrichment is an effective strategy for enhancing MT-DNN prediction performance in drug discovery.
• The extent of performance improvement is contingent upon the quality and diversity (unique compounds and targets) of the training data.
• MT-DNNs struggle to predict interactions for compounds significantly dissimilar to those in the training set, even with enrichment.

Conclusions:

• Data enrichment strategies can significantly boost prediction accuracy in multi-task learning for drug discovery.
• Careful consideration of training data composition, including compound and target diversity, is crucial for maximizing benefits.
• Recommendations are provided for optimizing MT-DNNs to improve prediction accuracy and accelerate the identification of novel drug candidates.