Optimizing Model Learning Performance on a Challenging Heck Reaction Yield Data Set.

Area of Science:

• Organic Chemistry
• Machine Learning
• Computational Chemistry

Background:

• Machine learning (ML) development in organic synthesis is hindered by limited data availability.
• Existing literature-based datasets often suffer from sparse distribution and high-yield bias, restricting ML model performance.
• The HeckLit dataset, comprising 10,002 cases from the Heck reaction, offers a broad chemical space for ML applications.

Purpose of the Study:

• To establish a comprehensive, ML-compatible dataset for Heck reaction yields from literature.
• To address the challenges of data sparsity and high-yield preference in literature-derived datasets.
• To improve the predictive accuracy of ML models for organic synthesis reactions.

Main Methods:

• Development of the HeckLit dataset, a literature-mined collection of 10,002 Heck reaction yield cases.
• Application of feature distribution smoothing (FDS) to address data sparsity.
• Implementation of a subset splitting training strategy (SSTS) to optimize model learning.
• Evaluation of model performance using R-squared (R²) metric.

Main Results:

• The HeckLit dataset covers an extensive chemical space, significantly larger than high-throughput experimentation datasets.
• Initial ML model performance on HeckLit yielded R² = 0.318, indicating limited learning ability.
• Feature distribution smoothing (FDS) did not improve model performance.
• The subset splitting training strategy (SSTS) significantly boosted model performance, achieving R² = 0.380.

Conclusions:

• The HeckLit dataset provides a valuable resource for advancing ML in organic synthesis.
• The subset splitting training strategy (SSTS) is an effective method for improving ML model performance on sparse, literature-derived datasets.
• The study proposes a criterion for subset division, offering a new approach to learning from large-scale chemical data.