Hybrid interpretable biophysical modeling of fluorescent protein biosensor function

Fluorescent protein biosensors are designed to sense physiological signals and report changes through altered fluorescence emission. These chimeric proteins combine ligand-binding and fluorescent protein domains such that binding allosterically regulates fluorescence. For optimal function, the reversible ligand-dependent transitions between conformational states must be tailored to the conditions of the intracellular environment. Biosensor function is not easily characterized by a single number, but instead depends on the kinetics of these transitions, which in turn are controlled by protein sequence and structure. However these parameters cannot be directly measured making biosensor optimization difficult. To address this challenge, we developed a hybrid model incorporating neural network layers and interpretable biophysics that infers the underlying dynamics. To assay biosensor function in the cellular context, we generated hundreds of sequence variants of the calcium biosensor GCaMP and expressed each in cultured primary neurons for testing. Given the amino acid sequence and the observed fluorescence, the network predicts the parameters for mechanistic models of biosensor function and cellular calcium handling and generates a reconstruction of the fluorescence trace. By modeling the response of biosensors in a physiologically relevant context, we identify how mutations alter the underlying biophysical properties. Similarly, by monitoring calcium transients using many biosensors with variable kinetic properties, the sources of heterogeneous dynamics across cells can be disentangled.