Neocortical cell type classification from electrophysiology recordings using deep neural networks

Raymond Wang, Sang Min Han, Marta Agnieszka Gajowa, Chunlei Liu

Keywords: neuron type classification, convolutional neural network, electrophysiology

Abstract: Understanding the neural code requires identifying different functional units involved in the neural circuits. One way to identify these functional units is to solve a neuron type classification problem. For decades, current clamp electrophysiology recordings have provided the means to classify the neurons based on subtle differences in action potential shapes and spiking patterns. However, significant variations in neuronal type definitions, classification pipelines, and variability in the neuronal activities make unambiguous determination of neuron type challenging. Previous solutions to this electrophysiology-based cell type classification problem consisted of dimensionality reduction juxtaposed with clustering using hand-crafted action potential features. Recent discoveries have allowed genetic-based cell-type classifications, which have fewer ambiguities, but they are less practical in vivo and have even lower throughput. Leveraging the unprecedented ground truth data published in the Allen Institute Cell Types Database, which contains anatomical, genetic, and electrophysiology characterizations of neurons in the mouse neocortex, we construct a robust and efficient convolutional neural network (CNN) that successfully classifies neurons according to their genetic label or broad type (excitatory or inhibitory) solely using current-clamp electrophysiology recordings. The CNN is configured as a multiple-input single-output network consisting of three subnetworks that take in the raw time series electrophysiology recording as well as the real and imaginary components of its Fourier coefficients. Our single pipeline method is fast and streamlined while simultaneously outperforming previous methods and achieving more classification classes using only single current-clamp trace as the input. This end-to-end convolutional neural network-based classification method removes the need for hand-crafted features, specific knowledge, or human intervention for quick identification of the cell type with high accuracy, enabling interpretation of the experimental data in a bias-free manner and a much broader scientific context.

One-sentence Summary: A robust and efficient convolutional neural network successfully classifies neurons according to their genetic label and broad type using only current-clamp electrophysiology recordings.