UC San Diego

Machine learning (ML) is a powerful but data intensive tool. Its application is limited in the natural sciences due to the cost of acquiring data and the complexity of the input-output relationships. Various methods have attempted to use prior domain knowledge to constrain the ML search space, minimize the amount of data required, improve ML interpretability, and better replicate natural systems. Here, we demonstrate a novel workflow that fully incorporates ML into a mechanistic equation, effectively simplifying the ML challenge from finding both the structure and the parameters of the input-output mappings to just the parameters. We show that this framework generates latent variable models that match biological expectations when predicting gene expression from promoters. We then explore how this method can be applied to a kinetic model where some of the latent parameters have been measured before. This framework generalizes previous methods which aimed to parameterize differential equation models, and we anticipate that it will have wide applications wherever a process can be modeled with an equation and ML can be used to estimate the parameters.