Intelligent predicting reaction performance in multi-dimensional chemical space using quantile regression forest

Jing Dong, Xuechun Mu, Zelin Zhang, Yuqing Zhang, Xiaohui Yang, Puyu ZhangMatch 2021 · DOI

Abstract: Buchwald-Hartwig amination reaction is widely applied in synthetic organic chemistry, which faces tedious and complex experimental process. In 2018, an interesting yield prediction technique is proposed via machine learning (random forest) in Science. However, the method is based on point prediction with many feature descriptors. For tackling these problems, complements and improvements have been made from the perspectives of machine learning and statistics, including feature dimensionality reduction, distributed prediction and visualization, so as to provide accurate and reliable decision information.

Introduction

Predicting chemical reaction yields is a fundamental challenge in synthetic organic chemistry. Traditionally, this process relies on extensive trial-and-error experimentation. A significant milestone was reached in 2018 when researchers demonstrated that machine learning, specifically Random Forests, could predict yields for the Buchwald-Hartwig amination. However, standard point-prediction models lack a measure of uncertainty and often rely on an excessive number of descriptors, which can lead to overfitting and reduced interpretability. This work addresses these limitations by introducing a framework based on Quantile Regression Forests (QRF) and streamlined feature selection.

Methodology

The proposed approach moves beyond simple point predictions to provide full probability distributions of reaction yields. The pipeline consists of two primary stages: feature selection and distributed prediction.

Overview of the proposed methodology, including feature selection, QRF prediction, and statistical density fitting.

Feature Selection

To reduce dimensionality and eliminate noise, a two-step screening process was implemented. First, Between-category to within-category sums of squares (BW) was used to rank features. Subsequently, the Least Absolute Shrinkage and Selection Operator (LASSO) was applied to remove redundant correlations. This process successfully reduced the initial 120 descriptors to a core set of 21 essential features.

Quantile Regression Forest and Density Estimation

Unlike standard Random Forests that average tree outputs, QRF retains all observations in leaf nodes to estimate the conditional distribution of the yield. To provide an intuitive visualization for chemists, Kernel Density Estimation (KDE) with an Epanechnikov kernel was used to transform the QRF output into a smooth probability density curve, representing the likelihood of various yield outcomes.

Experimental Dataset

The study utilized a comprehensive dataset of 4,608 reaction data points for the Buchwald-Hartwig amination. The chemical space includes 23 isoxazole additives, 15 aryl/heteroaromatic halides, 4 palladium catalysts (ligands), and 3 bases.

Chemical components of the Buchwald-Hartwig amination dataset used for model training and validation.

Results and Discussion

The refined model achieved high predictive accuracy while significantly reducing computational complexity. Key results include:

  • Dimensionality Reduction: An 82.5% reduction in input features (from 120 to 21) without sacrificing performance.
  • Predictive Performance: The model maintained an $R^2$ of 0.92 and an RMSE of 7.46.
  • Reliability: Prediction intervals (90%, 95%, and 99%) covered actual experimental results with high fidelity (96.21% to 100%).
  • Outlier Analysis: The framework identified 17 statistical outliers; their removal further improved the $R^2$ to 0.933, demonstrating the model's utility as a diagnostic tool for identifying anomalous chemical behavior or experimental errors.

Conclusion

By integrating uncertainty quantification into yield prediction, this framework provides chemists with actionable confidence intervals rather than just single-point estimates. Future work will focus on extending this methodology to other complex reaction types and integrating it into active learning loops to optimize chemical space exploration.

Summary updated

Abstract

Buchwald-Hartwig amination reaction is widely applied in synthetic organic chemistry, which faces tedious and complex experimental process. In 2018, an interesting yield prediction technique is proposed via machine learning (random forest) in Science. However, the method is based on point prediction with many feature descriptors. For tackling these problems, complements and improvements have been made from the perspectives of machine learning and statistics, including feature dimensionality reduction, distributed prediction and visualization, so as to provide accurate and reliable decision information.
BibTeX
@inproceedings{dong2021intelligent,
  author = {Jing Dong and Xuechun Mu and Zelin Zhang and Yuqing Zhang and Xiaohui Yang and Puyu Zhang},
  title = {{Intelligent predicting reaction performance in multi-dimensional chemical space using quantile regression forest}},
  booktitle = {Match},
  year = {2021},
  doi = {10.46793/match.87-2.299p},
}
Figure