Sylmarie Davila-Montero, Michigan State University
Conference: 2020 Society of Toxicology Annual Meeting, Anaheim, CA
Abstract Title: Improving identification of neuroactive compounds using temporal information from microelectrode array recordings of cortical neural networks and a semi-supervised classification algorithm
Abstract: Human exposure to environmental chemicals can result in acute neurotoxicity (NT), negatively impacting brain activity. In vitro microelectrode array recordings of neural network function following chemical exposure can be used to screen chemicals for NT hazard. These recordings capture temporal (from min to days) and spatial aspects of action potential activity as described by a set of network parameters (NPs). To determine if a compound is neuroactive, global NPs are extracted from 40 min neural recordings resulting in loss of temporal information (TI). The TI could improve identification of compound fingerprints and/or provide information on the mechanisms of action that mediate a neural network response after acute exposure of a compound. Here, data from 384 previously tested compounds were used to explore the properties of the TI to screen for acute neuroactive compounds using the response from a single high concentration (nominally 40 µM) and a window analysis technique. From recordings on day in vitro 12, a total of 19 NPs were extracted for each 1-min window of time with 50% overlaps, resulting in one time series (trajectory) per NP. Extracted NP trajectories were normalized per well and per window of time, and a moving median filter was applied to reduce outliers. A k-means trajectory clustering technique was used to find 10 clusters of trajectories for each NP and to assign cluster IDs to the trajectories of each compound. Then, for each compound, a vector with a total of 19 cluster IDs (one per NP) was assigned and used to classify the compounds as neuroactive or negative compounds using a Support Vector Machine (SVM) classifier. The entire classification model was trained with 73 compounds (42 neuroactives, 31 negatives) and yielded a classification accuracy of 93.2%. When using the model to classify the 384 compounds, 257 were identified as neuroactives and 127 as negatives. By comparison, when TI was excluded from the SVM classifier, classification accuracy of the same 73 neuroactive/negative compounds decreased to 86.3%. The higher classification accuracy of the SVM model that uses TI data demonstrates including TI is more effective for identifying acute neuroactive compounds when performing single-point screening. This abstract does not reflect policy of the US EPA.