Out-of-distribution evaluation of active learning pipelines for molecular property prediction

Area of Science:

• Computational chemistry
• Machine learning
• Data science

Background:

• Active learning (AL) minimizes data needs for machine learning (ML) models.
• Molecular property prediction is data-intensive, making AL valuable.
• Evaluating AL on out-of-distribution (OOD) data is crucial for real-world applications but understudied.

Purpose of the Study:

• To evaluate active learning (AL) for molecular property prediction, specifically focusing on out-of-distribution (OOD) performance.
• To develop and assess an AL framework using Evidential Deep Learning (EDL) for predicting solvation energy.
• To analyze the impact of data diversity and dataset similarity on AL performance.

Main Methods:

• Developed an AL framework utilizing prediction uncertainties from Evidential Deep Learning (EDL).
• Trained models on in-distribution data and augmented with OOD data sampled from PubChem.
• Investigated AL performance using random sampling versus EDL-guided selection.
• Examined generalization by starting with a less diverse in-distribution dataset.

Main Results:

• Evidential Deep Learning (EDL)-based active learning (AL) demonstrated superior performance over random sampling for OOD molecular property prediction.
• Analysis revealed how training-test dataset similarity influences AL effectiveness.
• Investigated distributional differences in molecule selection between AL and random sampling.

Conclusions:

• Active learning (AL) with Evidential Deep Learning (EDL) is a promising strategy for enhancing molecular property prediction, especially for OOD data.
• The study highlights the importance of considering dataset similarity and distributional characteristics in AL.
• EDL-based AL offers a more effective approach than random sampling for navigating OOD molecular data.