Demystifying Multitask Deep Neural Networks for Quantitative Structure-Activity Relationships.

Area of Science:

• Computational chemistry
• Artificial intelligence
• Machine learning

Background:

• Deep neural networks (DNNs) excel in AI tasks like computer vision and NLP.
• DNNs show promise for quantitative structure-activity relationship (QSAR) predictions, often outperforming traditional methods.
• Multitask DNNs, predicting multiple QSAR properties simultaneously, can outperform single-task models but lack a clear explanation for their success.

Purpose of the Study:

• To investigate the mechanism behind multitask DNNs' predictive performance in QSAR.
• To understand how information is shared between unrelated QSAR tasks within a multitask DNN.
• To develop a strategy for effectively utilizing multitask DNNs in QSAR modeling.

Main Methods:

• Exploration of information borrowing in multitask DNNs during prediction.
• Analysis of the relationship between activity correlation and predictive performance.
• Development of a strategy using prior domain knowledge to select correlated training sets for multitask DNNs.

Main Results:

• Multitask DNNs borrow "signal" from molecules with similar structures across different QSAR tasks.
• The impact of this information borrowing (positive or negative) is contingent on the correlation of molecular activities.
• A novel strategy for selecting training sets with correlated activities for multitask DNNs was developed and validated.

Conclusions:

• The predictive advantage of multitask DNNs is explained by structure-based signal borrowing, moderated by activity correlation.
• A new strategy enhances multitask DNN performance by incorporating domain knowledge to select correlated training sets.
• This work provides a framework for more effective application of multitask DNNs in QSAR research.