Deep-learning-accelerated Bayesian inference for FRAP experiments

Abstract

Fluorescence recovery after photobleaching (FRAP) is an experimental method for determining properties such as the diffusion coefficient and binding rate of molecules in solutions, and is used extensively in areas such as food science and biology. By utilizing a high intensity laser the fluorescent molecules in a region of interest are bleached. The mean fluorescence intensity in the region, and its development over time, can be modeled as a function of the mobility parameters. Bayesian inference with regards to the model parameters, using the likelihood function, can then be performed. This likelihood function is very computationally heavy to evaluate. In this work a neural network has been implemented to approximate the likelihood function with regards to the four most central parameters in the model. The method is promising since three of the four marginal posterior distributions of the parameters were well approximated, with the result being comparable to traditional methods for posterior sampling. We demonstrated that the method is much more computationally efficient than traditional methods.