Fast and effective molecular property prediction with transferability map

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Transfer learning is crucial for molecular property prediction with limited data.
• Existing methods risk negative transfer or require extensive target task training.

Purpose of the Study:

• To develop a method for quantifying transferability between molecular property prediction tasks.
• To enable informed source dataset selection for improved transfer learning.

Main Methods:

• Proposed Principal Gradient-based Measurement (PGM) using an optimization-free scheme.
• Calculated principal gradients to approximate model optimization directions.
• Measured transferability as the distance between source and target principal gradients.

Main Results:

• Developed a quantitative transferability map for source dataset selection.
• PGM effectively guided transfer learning across 12 benchmark datasets.
• Demonstrated improved target task performance with PGM guidance.

Conclusions:

• PGM provides fast and effective guidance for transfer learning in molecular property prediction.
• This approach enhances efficiency in discovering drugs, materials, and catalysts.
• Offers a quantitative understanding of task relatedness prior to transfer learning.