Efficient Exploration of Chemical Reaction Space via Surrogate-Driven Chemical Representations

T. Stuyver, G. Chen, J. E. Alfonso-Ramos

The efficient exploration and screening of chemical reaction space requires machine-learning (ML) models capable of making accurate predictions within and across reaction families.[1] Yet, when only small, high-quality, task-specific datasets are available, conventional ML architectures often suffer from limited accuracy and poor generalizability. A common strategy to mitigate this challenge is to augment models with physically motivated quantum-mechanical (QM) descriptors, which has been shown to yield robust, data-efficient reactivity predictions—even with only a few hundred training data points.[2,3] Unfortunately, generating QM descriptors is computationally demanding, limiting the practicality of this approach for truly large-scale or high-throughput screening. Surrogate models that predict descriptors on the fly from molecular graphs offer a promising workaround.[2,3] Especially when combined with large pre-computed literature QM datasets,[4-6] such surrogates enable generalizable screening pipelines that require minimal task-specific training data and incur only marginal computational cost. In this talk, I will highlight our recent efforts in this area,[6,7] focusing on results showing that downstream models perform even better when using the surrogate’s internal hidden representations rather than the predicted descriptors — except when the descriptor set is tightly aligned with the mechanistic physics of the target reaction class.[7] These latent embeddings from surrogate models encode rich and transferable chemical information, making them a powerful foundation for data-efficient predictive chemistry. Finally, I will briefly discuss ongoing work assessing the benefits of using these pre-trained hidden representations within Bayesian Optimization workflows.

1. N. Casetti et al., Chem. Eur. J. 2023, 29, e202301957.
2. Y. Guan et al., Chem. Sci. 2021, 12, 2198-2208.
3. T. Stuyver; C. W. Coley, J. Chem. Phys. 2022, 156, 084104.
4. R. Ramkrishnan et al., Sci. Data 2014, 1, 1-7.
5. P. C. St. John et al., Sci. Data 2020, 7, 244.
6. J. Alfonso-Ramos; R. Neeser; T. Stuyver, Digit. Discov. 2024, 3, 919-931.
7. G. Chen; T. Stuyver, Digit. Discov. 2025, 4, 3227-3237.