Machine learning model turns decades of literature data into a fast route to highly selective organocatalysis

Decades of published chemistry have produced an enormous archive of reaction data, yet that archive has rarely been trusted to guide new experiments. The results come from scattered sources, cover an uneven slice of chemical space, and skew heavily toward the high-performing reactions that journals prefer to publish. A team led by Professor Luo Sanzhong at Tsinghua University, working with collaborators, has now shown that this imperfect historical record can nevertheless be turned into a working predictive tool for asymmetric organocatalysis, one of the most important methods for making chiral drugs, functional materials, and fine chemicals.

The researchers report OrgAIcat, a machine learning framework for asymmetric organocatalysis published as an open access Research Article in CCS Chemistry, the flagship journal of the Chinese Chemical Society. The foundation of the work is a manually compiled dataset the team calls iSynth. Drawing on 844 papers published between 2000 and 2020, the researchers extracted more than 22,000 data points on two workhorse reaction classes, Aldol and Michael additions.

After structural normalization, verification of reaction conditions, and rigorous data cleaning, the final dataset contained 13,360 Aldol reactions from 478 publications and 9,194 Michael addition reactions from 366 publications. Each entry recorded the catalyst, reactants, products, solvent, additives, temperature, and the measured enantioselectivity. Because enantioselectivity is reported in the literature as enantiomeric excess, or ee, the team converted those values into the corresponding free energy difference, ΔΔG, and used that thermodynamic quantity as the prediction target for their regression models.

The unevenness of the compiled data was itself a central finding. High-ee results dominate the literature, and a small number of popular substrates, catalysts, and conditions appear again and again. By estimating the number of possible combinations of the variables already present in the dataset, the team calculated that the potential reaction space contains roughly 10 to the 16 combinations for Aldol reactions and 10 to the 17 for Michael additions, while the literature records only about 10 to the 4.

In other words, published data covers a vanishingly small fraction of the accessible space, and it does so with strong selection bias toward successful, selective experiments. To represent such complex reactions for machine learning, the team developed a reaction-level descriptor called R-SPOC, built on their earlier SPOC descriptor. R-SPOC combines three complementary pieces of information: the structure and physicochemical characteristics of the catalyst, the structural change captured as the difference between product and reactant descriptors, and the reaction conditions, including solvent, additives, and temperature.

Crucially, the descriptor requires no explicit transition state searches or quantum chemical calculations, which allows it to be applied to tens of thousands of literature entries at relatively low computational cost. Ablation experiments showed that catalyst structure carries the most predictive information, while the reaction transformation and the condition descriptors each contribute independently, and the model performs best when all three are used together. In a unified 10-fold cross-validation, an ExtraTrees regression model built on R-SPOC achieved R² values of 0.716 for Aldol reactions and 0.774 for Michael additions, with mean absolute errors in ΔΔG of 0.282 and 0.265 kcal/mol respectively.

Expressed in the more familiar selectivity metric, those errors correspond to mean absolute errors in ee of approximately 7.30% and 6.78%. The team also built a four-category classification model that sorts reactions into low, moderate, good, and excellent selectivity intervals, reaching overall AUC values of 0.908 and 0.922 for the two reaction classes, a format that offers experimenters a more intuitive ranking of candidate reactions. Because random train-test splits can flatter a model when the same frequent catalysts or near-identical reactions appear on both sides, the researchers ran a battery of tougher validations.

When data were partitioned according to catalyst skeleton organization, performance barely changed, yielding R² of 0.705 and a mean absolute error of 0.277 kcal/mol. In a low-similarity test designed to probe genuinely unfamiliar chemistry, performance dropped to R² of 0.632 and an error of 0.318 kcal/mol, a decline the authors report transparently. A baseline model that predicted purely from how often each reaction component appeared in the literature performed significantly worse than OrgAIcat, showing the model learns structure-selectivity relationships rather than memorizing popularity.

When high-frequency catalysts, substrates, solvents, and additives were downsampled, cutting the dataset by roughly 40%, performance fell only slightly. Finally, an external test on six new papers published between 2021 and 2024 showed the model retained relatively good predictive performance on reactions reported after the training data were collected. The most demanding test came from the laboratory itself.

Holding substrates and conditions constant, the team screened 24 newly synthesized chiral primary amine catalysts that had never been reported, in Aldol reactions. OrgAIcat tracked the selectivity differences caused by changes in catalyst structure with a Pearson correlation coefficient of 0.736. The validation was then extended to the Aldol reaction of hydroxyacetone with different aldehydes, a harder problem in which the catalyst, additives, and substrates all varied simultaneously.

In that multivariate setting the model achieved a mean absolute error in ΔΔG of 0.227 kcal/mol and a Pearson correlation coefficient of 0.776. For reactions related to the training data but featuring specific combinations absent from the database, the model still offered useful guidance for choosing catalysts and conditions. The framework then faced a practical optimization challenge.

The team selected an Aldol reaction with a starting selectivity of only 41% ee. OrgAIcat first virtually screened 1,354 organic catalysts in the iSynth dataset; combining dimensionality reduction, clustering, commercial availability, and expert judgment, the researchers selected 10 model-recommended catalysts and added 13 laboratory-produced catalysts to form a candidate pool of 23. Crossing those catalysts with catalyst loading, solvent, additives, additive loading, and temperature produced 8,280 candidate conditions.

Bayesian optimization then scheduled the experiments. The very first round lifted selectivity from 41% ee to 86% ee, and after four rounds and just 12 experiments the reaction reached 96% ee. Notably, the optimal catalyst came from the candidates recommended during the OrgAIcat virtual screening stage.

The workflow illustrates a division of labor: the model narrows a vast catalyst space, chemists supplement it with laboratory catalysts representing structural types underrepresented in the literature, and Bayesian optimization chooses the next experimental conditions. The study demonstrates that historical literature data, once standardized and organized, can directly inform catalyst screening and condition optimization for new reactions, despite its heterogeneity and well-known publication bias. The authors note that the picture will improve further as data from low-selectivity and failed experiments are added and as more reaction types and catalytic modes are incorporated.

With such extensions, OrgAIcat is expected to apply across a broader range of asymmetric catalytic systems, helping researchers shrink the field of candidate catalysts and conditions before the first experiment is run and guiding efficient optimization of new reaction systems. For a discipline whose discovery process has long depended on experience and trial and error, the result suggests that the accumulated record of past experiments, carefully curated, can become an active instrument for designing the next ones.

Subject of Research: Machine learning prediction of enantioselectivity in asymmetric organocatalysis using curated literature data

Article Title: OrgAIcat predicts enantioselectivity in organocatalysis using literature data and rapidly optimizes reactions

Article References: OrgAIcat predicts enantioselectivity in organocatalysis using literature data and rapidly optimizes reactions. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: machine learning, organocatalysis, enantioselectivity, asymmetric catalysis, Aldol reaction, Michael addition, Bayesian optimization, iSynth dataset, R-SPOC descriptor, catalyst screening, CCS Chemistry, data-driven chemistry