Published chemical literature has accumulated a wealth of reaction data, but this data comes from diverse sources, is unevenly distributed, and is clearly biased towards highly selective "successful experiments." Can such literature data be used for reliable machine learning modeling and further guide new catalytic reactions? Professor Luo Sanzhong's team and collaborators at Tsinghua University recently reported OrgAIcat, a machine learning framework for asymmetric organocatalysis. The research team manually compiled over 22,000 data points on Aldol and Michael addition reactions from 844 papers to construct the iSynth dataset. Based on this, the team developed R-SPOC, a reaction descriptor that requires no quantum chemical calculations, for predicting enantioselectivity and screening catalysts. When OrgAIcat was combined with Bayesian optimization, a challenging Aldol reaction was improved from 41% ee to 96% ee, with the entire optimization process requiring only 12 experiments. These results demonstrate that after rigorous compilation, proper characterization, and sufficient validation, historical literature data can still be used to guide experiments, despite its heterogeneity and publication bias. The article was published as an open access Research Article in CCS Chemistry, the flagship journal of the Chinese Chemical Society.
Background information:
Asymmetric catalysis is an important means of obtaining chiral drugs, functional materials, and fine chemicals. Over the past two decades, organocatalysis has developed various activation modes, including enamines, imine ions, and hydrogen bonding. However, the discovery of new catalysts and reaction conditions still relies heavily on chemists' experience and trial-and-error processes. Even considering only the Aldol and Michael addition reactions studied in this paper, the combination of variables such as catalyst, substrate, solvent, additives, and temperature creates a vast reaction space.
Machine learning can help researchers quickly screen reaction conditions, but modeling first faces the problem of data sources. High-throughput experimental data are usually of consistent quality but are concentrated in a narrow chemical space; patent and large reaction databases have broader coverage but often lack the fine stereochemical information needed for enantioselectivity modeling. Academic literature preserves a large number of asymmetric catalysis results, but also suffers from problems such as inconsistent experimental conditions, missing negative results, and repeated reports of commonly used catalysts. This research raises the question: Do these data still contain structure-selectivity rules that can be used to predict new reactions?
Highlights of this article:
1. Construct the iSynth dataset from 844 articles: The research team manually compiled organocatalytic Aldol and Michael addition reactions published between 2000 and 2020. After structural normalization, reaction condition verification, and rigorous data cleaning, they ultimately obtained 13,360 Aldol reactions from 478 publications and 9,194 Michael addition reactions from 366 publications. Each data point included information on catalyst, reactants, products, solvent, additives, temperature, and enantioselectivity. The research team converted the reported ee (efficiency energy) into the corresponding free energy difference ΔΔG, and the subsequent regression model used ΔΔG as the prediction target.
The dataset contains a variety of catalysts and substrates, but their distribution is highly uneven: high-ee results are prevalent, and a few commonly used substrates, catalysts, and conditions are repeatedly reported. Based on orders of magnitude estimation using combinations of variables already present in the dataset, the potential combinations for Aldol and Michael addition reactions are approximately 1016and 1017, respectively, while existing literature records are only on the order of 104. This demonstrates that the existing data covers only a small portion of the potential reaction space and exhibits significant selection bias.
2. R-SPOC simultaneously describes the catalyst, structural changes, and reaction conditions: To characterize complex organocatalytic reactions, the team developed the reaction-level descriptor R-SPOC based on their previous SPOC descriptor. This representation consists of three complementary parts: (1) the structure and physicochemical characteristics of the catalyst; (2) the structural changes represented by the difference between the product and reactant descriptors; and (3) reaction conditions such as solvent, additives, and temperature. R-SPOC does not require explicit transition state searches or quantum chemical calculations, thus enabling it to handle tens of thousands of literature data sets at a relatively low computational cost. Ablation experiments show that the catalyst structure provides the most predictive information, while reaction transformation and condition information also contribute independently. The model performs best when all three parts are used together.
In the unified 10-fold cross-validation, theExtraTreesregression model based on R-SPOC achieved R² values of 0.716 and 0.774 fortheAldolandMichaeladditionreactions, respectively, with mean absolute errors of ΔΔG of 0.282 and 0.265 kcal/mol, and corresponding mean absolute errors of ee of approximately 7.30% and 6.78%. A further developed four-category model categorizes reactions into low, moderate, good, and excellent selectivity intervals, with overall AUC values of 0.908 and 0.922, respectively, providing a more intuitive candidate classification for experimental screening.
3. Validate the model using multiple data partitioning methods and external data: If the same type of high-frequency catalyst or similar reaction appears in both the training and test sets, random partitioning may overestimate model performance. To address this, the team added several validations:(1)In data partitioning based on catalyst skeleton organization, the model achieved R² = 0.705 and MAE = 0.277 kcal/mol, which is close to the results of random partitioning; (2) In the low similarity test, the model performance dropped to R² = 0.632 and MAE = 0.318 kcal/mol; (3) When predicting based solely on the frequency of each reaction component in the literature, the performance of the baseline model was significantly lower than thatof OrgAIcat; (4) After downsampling high-frequency catalysts, substrates, solvents, and additives, reducing the total data volume by about40%, the model performance only decreased slightly; (5) In the external test consisting of6new papers from 2021-2024, the predictions of Aldol and Michael addition reactions also achieved relatively good performance.
4. Validate the model using unreported experimental systems: In addition to external testing based on literature, the team conducted independent experimental validation. First, with substrates and conditions remaining constant, they screened 24 newly synthesized and previously unreported chiral primary amine catalysts for Aldol reactions. OrgAIcat predicted selectivity differences caused by catalyst structural changes well, with a Pearson correlation coefficient of 0.736. Subsequently, the validation was extended to the Aldol reaction of hydroxyacetone with different aldehydes, while simultaneously changing the catalyst, additives, and substrates. The model's mean absolute error of ΔΔG in this multivariate system was 0.227 kcal/mol, with a Pearson correlation coefficient of 0.776. For reactions related to the training data but with specific combinations not found in the database, the model can still provide a reference for catalyst and condition screening.
5. Combining OrgAIcat with Bayesian optimization, 12 experiments increased ee from 41% to 96%: The team selected the Aldol reaction, with an initial selectivity of only 41% ee , to test the model's effectiveness in practical optimization. OrgAIcat first virtually screened 1,354 organic catalysts in iSynth, then combined dimensionality reduction, clustering, commercial availability, and expert judgment to select 10 model-recommended catalysts, adding 13 laboratory-produced catalysts to form the final candidate pool. Combining these 23 catalysts with catalyst dosage, solvent, additives, additive dosage, and temperature, a total of 8,280 candidate conditions were obtained. The team then used Bayesian optimization to arrange the experiments. The first round of experiments improved the selectivity from 41% ee to 86% ee; after 4 rounds and 12 experiments, a final selectivity of 96% ee was achieved. The optimal catalyst came from the candidates recommended in the OrgAIcat virtual screening stage. The entire optimization process simultaneously utilized model screening, expert judgment, and Bayesian optimization. OrgAIcat first screens for a large catalyst space; researchers then add their own laboratory catalysts based on existing experience, supplementing structural types that are less common in literature data; Bayesian optimization is used to select the experimental conditions for the next round.
Summary and Outlook:
In summary, this study compiled over 22,000 data points on Aldol and Michael addition reactions from 844 articles and established the OrgAIcat model. This model employs the R-SPOC descriptor, which requires no quantum chemical calculations, and can be used for enantioselectivity prediction and catalyst screening. Combined with Bayesian optimization, the selectivity of the target Aldol reaction was improved from 41% ee to 96% ee in just 12 experiments. Tests with recent literature and unreported laboratory reactions further validated the model's predictive ability for relevant organocatalytic systems.
This work demonstrates that existing literature data, after being standardized and organized, can provide direct reference for catalyst screening and condition optimization for new reactions. In the future, with the addition of data from low-selectivity and failed experiments, as well as the inclusion of more reaction types and catalytic modes, OrgAIcat is expected to be used in a wider range of asymmetric catalytic systems, helping researchers narrow down the range of candidate catalysts and reaction conditions before experiments, and guiding the efficient optimization of new reaction systems.
---
About the journal: CCS Chemistry is the Chinese Chemical Society's flagship publication, established to serve as the preeminent international chemistry journal published in China. It is an English language journal that covers all areas of chemistry and the chemical sciences, including groundbreaking concepts, mechanisms, methods, materials, reactions, and applications. All articles are diamond open access, with no fees for authors or readers. More information can be found at https://www.chinesechemsoc.org/journal/ccschem .
About the Chinese Chemical Society: The Chinese Chemical Society (CCS) is an academic organization formed by Chinese chemists of their own accord with the purpose of uniting Chinese chemists at home and abroad to promote the development of chemistry in China. The CCS was founded during a meeting of preeminent chemists in Nanjing on August 4, 1932. It currently has more than 120,000 individual members and 184 organizational members. There are 7 Divisions covering the major areas of chemistry: physical, inorganic, organic, polymer, analytical, applied and chemical education, as well as 31 Commissions, including catalysis, computational chemistry, photochemistry, electrochemistry, organic solid chemistry, environmental chemistry, and many other sub-fields of the chemical sciences. The CCS also has 10 committees, including the Woman's Chemists Committee and Young Chemists Committee. More information can be found at https://www.chinesechemsoc.org/ .
