Deep Forest-Based Intelligent Yield Predicting of Buchwald-Hartwig Coupling Reaction

Xuechun Mu, Jing Dong, Lichao Peng, Xiaohui YangMatch 2022 · DOI

This academic paper, published in MATCH Communications in Mathematical and in Computer Chemistry, presents a machine learning approach to a classic problem in organic synthesis: predicting the success of a chemical reaction.

TL;DR

Researchers have applied a "Deep Forest" algorithm to predict the yield of Buchwald-Hartwig coupling reactions, achieving a state-of-the-art $R^2$ of 0.94. By combining the representation learning of deep learning with the efficiency of ensemble tree models, the authors outperformed traditional Random Forests and even Convolutional Neural Networks (CNNs) while requiring significantly less computational power.

What Problem Does This Paper Address?

The Buchwald-Hartwig coupling reaction is a cornerstone of modern organic chemistry, essential for creating Carbon-Nitrogen (C-N) bonds found in many pharmaceuticals. However, predicting the "yield" (how much product you actually get) is notoriously difficult because it depends on a complex interplay between aryl halides, ligands, bases, and additives.

In 2018, a landmark paper in Science demonstrated that Random Forest (RF) models could predict these yields with high accuracy. However, RF models have a major limitation: they lack "feature learning." They take the data as-is without discovering deeper, hierarchical patterns within the chemical descriptors. While deep learning (like Neural Networks) can learn these patterns, they often require massive datasets and grueling hyperparameter tuning—luxuries chemists don't always have.

Methods

The authors propose using Deep Forest (also known as gcForest), a non-neural network deep learning model.

How it works:

  1. Cascade Structure: Unlike a standard forest that makes a single pass, Deep Forest uses a multi-layered "cascade." Each layer consists of multiple random forests.
  2. Representation Learning: Each layer processes the input and passes its "class probability" (or predicted yield) to the next layer. This output is concatenated with the original features, allowing the model to "learn" increasingly complex representations of the chemical space.
  3. Automatic Complexity: The model automatically determines the number of layers needed, stopping when performance plateaus.

The researchers used three types of chemical descriptors calculated via Spartan software: molecular, atomic, and vibration descriptors. They tested the model on two tasks: Regression (predicting the exact percentage yield) and Classification (categorizing yield as Low, Medium, or High).

Datasets

The study utilized the high-throughput screening database originally published by Doyle et al. in Science (2018). This dataset includes:

  • 15 Aryl halides
  • 4 Ligands
  • 3 Bases
  • 23 Additives (isoxazoles)

The data was split into a 70% training set and a 30% test set. The authors also performed "out-of-sample" testing by training on some additives and testing on entirely new ones to see how well the model generalizes to unknown chemicals.

Key Insights and Ideas

  • Deep Learning without the "Black Box" Headache: Deep Forest provides the benefits of deep architecture (hierarchical learning) but is much easier to train than a CNN. It has fewer hyperparameters and is robust to the "small" sample sizes often found in experimental chemistry.
  • The "Aryl Halide" Dominance: Through feature importance analysis (Gini importance), the authors found that aryl halides are the most significant factors affecting yield. Specifically, the Carbon-3 NMR shift and electrostatic charges on the aryl halide were the strongest predictors.
  • Efficiency Matters: The Deep Forest model achieved its results in just 8 seconds on a standard CPU, whereas a CNN took nearly two minutes to achieve lower accuracy. This makes it a highly practical tool for lab-bench chemists.

Key Results

The Deep Forest (DF) model consistently outperformed all other tested algorithms, including Random Forest (RF), Support Vector Machines (SVM), and Neural Networks (NN).

  • Regression Performance:
    • Deep Forest: $R^2 = 0.94$, RMSE = 6.86
    • Random Forest: $R^2 = 0.92$, RMSE = 7.7
    • CNN: $R^2 = 0.93$, RMSE = 7.3
  • Classification Performance:
    • The DF model achieved an accuracy of 93.72% and a Kappa statistic of 0.906, indicating "almost identical" agreement between predicted and actual yield categories.
  • Generalization: In out-of-sample additive testing, the model maintained high accuracy (up to 94.44% for specific additives), proving it can predict the behavior of new reactants it hasn't seen before.

Limitations and Future Work

While the paper demonstrates excellent results, a few areas remain for future exploration:

  • Descriptor Dependency: The model relies on descriptors calculated via Spartan. The accuracy of the prediction is inherently tied to the quality and relevance of these physical-chemical features.
  • Scope of Reactions: The study is limited to the Buchwald-Hartwig reaction. While the method is generalizable, the model itself would need to be retrained for different reaction types (e.g., Suzuki or Heck couplings).
  • Data Requirements: Although Deep Forest is better with small data than traditional deep learning, it still requires high-throughput experimental data to be effective, which may not be available for more niche or novel reaction classes.

Summary updated

Abstract

Buchwald-Hartwig coupling reaction is widely used in organic chemical synthesis, yield prediction is particularly important. In 2018, Science reported a yield prediction method based on random forest, but this method lacks feature learning. Therefore, an intelligent prediction and analysis method of coupling reaction yield based on deep forest is proposed. Combined with the advantages of deep learning and ensemble learning, the new deep model in the form of non-neural network is explored, which has good characterization learning ability and low difficulty in adjusting parameters, realizes the efficient prediction of chemical reaction, and analyzes the factors that have a significant impact on the prediction of reaction yield.
BibTeX
@inproceedings{mu2022deep,
  author = {Xuechun Mu and Jing Dong and Lichao Peng and Xiaohui Yang},
  title = {{Deep Forest-Based Intelligent Yield Predicting of Buchwald-Hartwig Coupling Reaction}},
  booktitle = {Match},
  year = {2022},
  doi = {10.46793/match.88-1.005m},
}
Figure