Gaussian Process Modeling of Bioorthogonal Cycloaddition Reactivity from Sparse Data

Bioorthogonal tetrazine ligations are widely used due to their exceptionally fast kinetics, yet quantitative prediction of their reactivity grounded in experiments remains challenging because experimentally measured rate constants are scarce and unevenly distributed across chemical space.In this work, we demonstrate that Gaussian process (GP) regression models can capture meaningful reactivity trends from sparse kinetic datasets of tetrazine ligations.Using a curated collection of experimental second-order rate constants, we show that GP models achieve competitive predictive performance despite limited data availability.We compare Gaussian process models to a neural network baseline and find that, while the deep learning approach can reach similar point-prediction accuracy, Gaussian processes exhibit more reliable performance in the low-data regime.Importantly, the flexibility of GP models enables data-efficient learning of structure-reactivity relationships without requiring large training datasets.Together, these results establish Gaussian processes as practical surrogate models for modeling reactivity trends in tetrazine ligations and related bioorthogonal cycloadditions, providing a foundation for data-efficient exploration and optimization of bioorthogonal reaction space.