Chemical reaction networks stop refitting per dataset: one training run, then model choice and rate constants in seconds

Pathway discrimination and rate-constant estimation for chemical reaction networks are merged into one joint inference task, and after training a new observation needs only a forward pass. The standard approach, sequential Monte Carlo approximate Bayesian computation, relies on handcrafted low-dimensional summary statistics and repeats simulation and sampling for every new dataset. On synthetic data from three candidate mechanisms, the network reached 94% global model-selection accuracy when substrate and product were both observed at 20 time points, and still 92% with product only at 10 time points.

A shared summary network compresses variable-length, multi-experiment time series into one fixed-length vector that drives both pathway classification and parameter density estimation. The summary network uses one-dimensional convolutions for temporal features and set attention to treat experiments under different initial conditions as an unordered set, so the learned representation does not depend on the order in which results are supplied. Each candidate mechanism gets its own conditional invertible neural network that maps rate constants to a latent variable for posterior density; evaluation used 3,000 independent test trajectories, with expected calibration error and Brier scores checking posterior-probability calibration.

On real catalytic protodeboronation data, the network assigned a posterior probability of 92.19% to the inhibitor-induced deactivation pathway. That experimental data contained only product concentration at 10 time points across three different catalyst loadings, with intermediate species unobserved. Synthetic trajectories drawn from these posteriors were consistent with the experimental observations, but posteriors for most rate constants were wide, indicating that measuring the product alone is insufficient to fully identify individual rate constants.

The work addresses computational chemistry and machine-learning readers studying reaction kinetics, in settings where the candidate mechanism set is known and observations are variable-length time series of a subset of species. The synthetic study covers three candidate mechanisms and scales to four; the real case discriminates three catalyst-deactivation hypotheses in catalytic protodeboronation. The authors propose scaling the architecture to a larger number of competing structures, such as the 20 commonly encountered reaction mechanisms, possibly with hierarchical classification heads for overlapping structures.

The framework currently relies on a closed-world assumption: it presumes the true data-generating process lies within the predefined candidate set. If the true mechanism is absent from the candidate space, the model network is forced to allocate probability mass across the available approximations. The authors state that moving toward open-world inference remains a critical challenge, and that flagging an experimental trajectory as an anomaly within the candidates considered is a direction for future research. In the real protodeboronation case, posteriors for most rate constants were wide, indicating that measuring the product alone is insufficient to fully identify all individual rate constants.