Training deep neural density estimators to identify mechanistic models of neural dynamics.

Pedro J Gonçalves,Michael Deistler,Giacomo Bassetto,Chaitanya Chintaluri,David S Greenberg,Sara A Haddad,Kaan Öcal,William F Podlaski,Tim P Vogels,Jakob H Macke,Jan-Matthis Lueckmann,Marcel Nonnenmacher

doi:10.7554/elife.56261

Abstract

Mechanistic modeling in neuroscience aims to explain observed phenomena in terms of underlying causes. However, determining which model parameters agree with complex and stochastic neural data presents a significant challenge. We address this challenge with a machine learning tool which uses deep neural density estimators-trained using model simulations-to carry out Bayesian inference and retrieve the full space of parameters compatible with raw data or selected data features. Our method is scalable in parameters and data features and can rapidly analyze new data after initial training. We demonstrate the power and flexibility of our approach on receptive fields, ion channels, and Hodgkin-Huxley models. We also characterize the space of circuit configurations giving rise to rhythmic activity in the crustacean stomatogastric ganglion, and use these results to derive hypotheses for underlying compensation mechanisms. Our approach will help close the gap between data-driven and theory-driven models of neural dynamics.

Highlights

New experimental technologies allow us to observe neurons, networks, brain regions, and entire systems at unprecedented scale and resolution, but using these data to understand how behavior arises from neural processes remains a challenge
When comparing a state-of-the-art genetic algorithm (Indicator Based Evolutionary Algorithm, IBEA, [Bleuler et al, 2003; Zitzler and Kunzli, 2004; Van Geit et al, 2016]) to SNPE, we found that the parameter-settings favored by IBEA produced simulations whose summary features were as similar to the observed data as those obtained by SNPE high-probability samples (Appendix 1—figure 10)
We demonstrate how SNPE can be used to identify the posterior distribution over both membrane and synaptic conductances of the stomatogastric ganglion (STG) (31 parameters total) and how the full posterior distribution can be used to study the above questions at the circuit level

Summary

Introduction

New experimental technologies allow us to observe neurons, networks, brain regions, and entire systems at unprecedented scale and resolution, but using these data to understand how behavior arises from neural processes remains a challenge. To test our understanding of a phenomenon, we often take to rebuilding it in the form of a computational model that incorporates the mechanisms we believe to be at play, based on scientific knowledge, intuition, and hypotheses about the components of a system and the laws governing their relationships. The goal of such mechanistic models is to investigate whether a proposed mechanism can explain experimental data, uncover details that may have been missed, inspire new experiments, and eventually provide insights into the inner workings of an observed neural or behavioral phenomenon (Herz et al, 2006; Gerstner et al, 2012; O’Leary et al, 2015; Baker et al, 2018).

Objectives

Methods

Results

Conclusion