Machine Learning for Inference in Biophysical Neuroscience Simulations

Abstract

A central challenge in neuroscience is that many properties of neural systems cannot be measured exactly. This limits our understanding of these systems and our ability to build simulations that match experimental recordings or predict neural responses to unseen stimuli. Inference allows scientists to identify parameters—the properties that cannot be measured exactly—such that biophysical simulations are consistent with experimental measurements of neural activity. However, previous inference methods struggle with complex biophysical neuroscience simulations or can infer only a limited number of parameters.

This thesis presents new inference methods for biophysical simulations in neuroscience. To overcome limitations of previous methods, we leverage recent advances in machine learning, particularly in neural density estimation and automatic differentiation. By inferring biophysical properties, we open up possibilities to build accurate biophysical simulations of neural activity, to study parameter degeneracy, and to gain insight into properties of neural systems. While this thesis focuses on biophysical neuroscience simulations, many of the developed methods are broadly applicable and are already being used across various scientific disciplines.

The thesis consists of two main parts. In the first part, we develop and apply methods for neural simulation-based Bayesian inference (SBI). SBI uses neural networks to invert mechanistic computer simulations, thereby providing Bayesian estimates of parameters given experimental measurements. We use these methods to study how a natural constraint on neural circuits—low metabolic cost—constrains circuit properties. We then develop new methods for improving the flexibility, robustness and efficiency of SBI. Finally, we present sbi, a Python toolbox for simulation-based Bayesian inference which implements many popular SBI methods and allows domain scientists to apply these methods to their simulators and measurements.

In the second part, we show that differentiable simulation enables the identification of parameters that align biophysical simulations with experimental recordings or computational tasks. We develop Jaxley, the first differentiable biophysics simulator for neuroscience and use it to fit parameters of biophysical simulations with gradient descent. This approach scales parameter inference to large-scale biophysical models, including morphologically detailed single-cell and network models, and demonstrates that differentiable simulation can overcome previous limits on the number of parameters.

Overall, the presented methods and results demonstrate that machine learning unlocks new possibilities for constructing biophysical simulations in neuroscience. By overcoming a central challenge—inferring parameters from measurements of neural activity—we hope that our methods will enable new insights into the cellular and synaptic contributions to biological intelligence.