Classification of Odor-Derived Electroantennograms with Machine Learning

TL;DR

This paper shows how machine learning can classify odors based on electrical signals from insect antennae, opening new possibilities for chemical sensing.

Contribution

The study introduces a novel method for odor classification using machine learning on antennal LFP time-series data.

Findings

Machine learning models accurately predicted VOC concentration and identity from LFP responses.

The method successfully classified complex VOC mixtures using LFP waveform features.

Results suggest potential applications in detecting agricultural pests and human diseases.

Abstract

Insects have a keen ability to detect numerous odors in their environment. These odors, known as volatile organic compounds (VOCs), provide the insect with information about food, predators, and mates that may be in the area and are detected by olfactory receptors expressed by sensory neurons on the antennae. When VOCs are transduced by the olfactory sensory neurons, the antennal electrical potential dynamically changes, causing a local field potential (LFP) response to occur. Research has used the LFP amplitude for determining VOC concentration, but only recently have antennal LFPs been posited to be able to be used for VOC discrimination and identification. To close this gap, we use the time-series response of the antenna to odors as well as principal components of these responses to capture the characteristics of the LFP response, including waveform dynamics, intensity, slope, and duration. We use antennae of the Manduca sexta moth to record LFPs generated in response to Floral and disease associated VOC’s. Using machine learning approaches (support vector machines and random forests) trained on the LFP responses, we were able to predict and classify individual VOCs across a range of concentrations, as well as complex mixtures that elicited a given LFP waveform from an excised antenna. These results demonstrate that antennal olfactory responses can be used for the classification of differing VOC features, including concentration, identity, and duration, and have implications for diverse chemical sensing applications, such as search-and-rescue, the presence of agricultural pests, and the presence of human disease.

Figures