Molecular property prediction in the ultra-low data regime

Area of Science:

• Molecular property prediction
• Machine learning in chemistry
• Materials science

Background:

• Data scarcity is a significant challenge for machine learning (ML) in molecular property prediction and design.
• Multi-task learning (MTL) can improve performance by leveraging correlations between properties, but imbalanced datasets can lead to negative transfer, reducing its effectiveness.

Purpose of the Study:

• To introduce Adaptive Checkpointing with Specialization (ACS), a novel training scheme for multi-task graph neural networks.
• To mitigate inter-task interference in MTL while retaining its benefits, particularly in data-scarce scenarios.

Main Methods:

• Developed ACS, a training scheme for multi-task graph neural networks.
• Validated ACS on diverse molecular property benchmarks.
• Applied ACS to predict properties of sustainable aviation fuels.

Main Results:

• ACS consistently matched or surpassed the performance of recent supervised methods across multiple benchmarks.
• Demonstrated the ability of ACS to build accurate predictive models using as few as 29 labeled samples for sustainable aviation fuel properties.
• Showcased the practical utility of ACS in real-world, low-data applications.

Conclusions:

• ACS effectively addresses data scarcity challenges in molecular property prediction and design.
• The proposed method enables reliable predictions in low-data regimes, accelerating AI-driven materials discovery.
• ACS broadens the applicability of multi-task learning in scientific domains with limited experimental data.