AI-Augmented Discovery of Nonproductive Parallel Catalytic Pathways

Kholodar et al. (2018) revealed that overstabilized covalent intermediates can dominate in parallel, nonproductive pathways—a finding missed by traditional models. Building on advances in ML-enhanced quantum chemistry (e.g., Chen et al., 2024; Abarbanel & Hutchison, 2024), this idea proposes to train ML models not just to find minimum energy pathways, but to actively search for and characterize these "dead-end" routes across diverse catalytic systems. By integrating reaction network enumeration with data-driven anomaly detection, we could uncover new targets for inhibitor design (e.g., for drug-resistant enzymes) or explain puzzling catalytic inefficiencies. This approach is distinct from the typical focus on productive pathways and could transform our ability to predict reaction outcomes, particularly in complex or crowded catalytic environments.

References:

1. Parallel reaction pathways and noncovalent intermediates in thymidylate synthase revealed by experimental and computational tools. S. A. Kholodar, Ananda K Ghosh, K. Świderek, V. Moliner, A. Kohen (2018). Proceedings of the National Academy of Sciences of the United States of America.
2. QupKake: Integrating Machine Learning and Quantum Chemistry for Micro-pKa Predictions. Omri D Abarbanel, Geoffrey R Hutchison (2024). Journal of Chemical Theory and Computation.
3. Constructing Accurate and Efficient General-Purpose Atomistic Machine Learning Model with Transferable Accuracy for Quantum Chemistry.. Yicheng Chen, Wenjie Yan, Zhanfeng Wang, Jianming Wu, Xin Xu (2024). Journal of Chemical Theory and Computation.