Multiscale Probabilistic Modeling - A Bayesian Approach to Augment Mechanistic Models of Cell Signaling with Machine-Learning Predictions of Binding Affinity

HollyHuber

University of Southern California

"Multiscale Probabilistic Modeling - A Bayesian Approach to Augment Mechanistic Models of Cell Signaling with Machine-Learning Predictions of Binding Affinity"

Recently, the Nobel Prize winning machine learning (ML) model, AlphaFold, expanded its protein structure prediction capabilities from monomers to multimers with AlphaFold3. Here, we investigate this expanded utility in the novel context of mechanistic models of cell signaling. These models describe cell signaling events, such as binding, amongst a network of molecules, mostly proteins, and have been applied to answer both clinical and fundamental biology questions. For example, cell signaling models have been used to propose improvements to CAR-T cell therapies and to elucidate cellular ‘decision making’. Use of these models is oftentimes limited by a sparsity of data for parameterization. Thus, in this work, we introduce a Bayesian framework that incorporates information about protein structure to guide parameter inference for mechanistic models. Rather than searching all plausible parameter values, we can refine our search by considering information that is specific to the proteins involved in the signaling event. Excitingly, we find augmenting mechanistic models of signaling to be uniquely compatible with established ML models. We test our approach on two signaling models. In both cases, our approach improves parameter estimates—however, these improvements do not significantly change prediction performance on test data for either model. We find that this is due to a lack of sensitivity between the informed parameters and the test outputs. In contrast, when we examine an output that is sensitive to the changed parameters, we see a clear change in the predicted dynamics. We note that our proposed approach is limited to parameters of reversible, bimolecular binding reactions. Yet, excitingly, mechanistic models of cell signaling are often comprised of such reactions, ensuring the relatively wide applicability of our inference approach in this context.
Additional authors:

---