Roll-AE: A Spatiotemporal Invariant Autoencoder for Uncovering Neuronal Electrophysiological Patterns

Tommaso Dreossi, Rounak Dey, Emily Fox, Daphne Koller

Track: Full Paper Track

Keywords: autoencoder, electrophysiology, self-supervised learning

TL;DR: Roll-AE, a new autoencoder, extracts spatiotemporally invariant features from MEA recordings and outperforms standard autoencoders in classification tasks, while characterizing electro-physiological traits in iPSC-derived neuronal cultures.

Abstract: Micro-electrode array (MEA) assays enable high-throughput recording of the electrophysiological activity in biological tissues, both in vivo and in vitro. While various classical and deep learning models have been developed for MEA signal analysis, the majority focus on in vivo experiments or specific downstream applications in vitro. Consequently, extracting relevant features from in vitro MEA recordings has remained largely dependent on particular curated features known as neural metrics. In this work, we introduce Roll-AE, a novel autoencoder designed to extract meaninful spatiotemporally invariant features from in vitro MEA recordings. We demonstrate that 1) Roll-AE’s embeddings outperform those from standard autoencoders across various classification tasks, and 2) Roll-AE’s embeddings effectively characterize electrophysiological phenotypic traits in induced Pluripotent Stem Cells (iPSC)-derived neuronal cultures.

Attendance: Tommaso Dreossi

Submission Number: 34