Chemical reaction networks (CRNs) are a common framework for modeling biochemical systems, but inferring their behavior from data is computationally hard. A new arXiv preprint proposes a joint approach that performs model selection and parameter estimation at the same time, rather than treating them as separate steps.
The method builds on simulation-based Bayesian inference, using neural posterior estimation to learn from simulated data. This lets the model discriminate between overlapping reaction pathways while also estimating reaction rates, which the authors say are two demanding problems that CRN inference must solve together.
Because the source is a single preprint, the claims are not independently verified, and the abstract does not report quantitative results. Still, the joint framing suggests a possible path toward more efficient inference for complex reaction networks.