Using Active Learning to Develop Machine Learning Models for Reaction Yield Prediction.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Computer-aided synthesis planning is crucial for drug discovery.
• Machine learning models require large datasets, often limited by experimental budgets.
• Active learning strategies can optimize data collection for model training.

Purpose of the Study:

• To investigate the impact of active learning on reaction yield prediction models.
• To evaluate the influence of different machine learning algorithms and initial data sizes.
• To assess the robustness of active learning in reducing experimental data needs.

Main Methods:

• Utilized two public high-throughput experimentation datasets.
• Implemented active learning strategies, specifically output margin.
• Compared active learning to random sampling for model training.
• Analyzed feature importance in trained machine learning models.

Main Results:

• Active learning achieved a predefined Area Under the Receiver Operating Characteristic Curve (AUROC) faster than random sampling on both datasets.
• Model accuracy improved with active learning, particularly when few features were important.
• The choice of machine learning algorithm and initial data points influenced prediction performance.

Conclusions:

• Active learning is an effective strategy for optimizing data collection in reaction yield prediction.
• This approach can significantly reduce the experimental data required for training accurate models.
• Further research can explore active learning's broader applicability in computational chemistry.