Powerful, transferable representations for molecules through intelligent task selection in deep multitask networks.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Drug discovery and materials science

Background:

• Deep learning chemical representations offer significant potential for drug discovery and materials innovation.
• Current methods face limitations including high computational cost, potential for inherited bias, and substantial data requirements.

Purpose of the Study:

• To address the limitations of deep learning chemical representations by proposing a novel methodology.
• To reduce the cost, mitigate bias, and overcome data scarcity issues in generating effective chemical representations.

Main Methods:

• Utilizing multi-task learning (MTL) in conjunction with transfer learning (TL).
• Employing pairwise task affinity to calculate task similarity for programmatic task selection, thereby avoiding bias introduction.
• Testing the proposed methodology on diverse real-world datasets.

Main Results:

• Demonstrated the methodology's effectiveness in complex and low-data environments.
• Showcased the deep representation's ability to capture more expressive task-based information compared to traditional cheminformatics fingerprints.
• Validated the potential for reduced cost and inherited bias in representation generation.

Conclusions:

• The proposed MTL and TL approach effectively overcomes key limitations in deep learning for chemical representations.
• This method enhances the expressiveness and utility of learned representations for applications in drug discovery and materials innovation.
• Task similarity analysis provides a robust way to guide representation learning and assess its quality.