# Machine Learning for Separating Dopamine and Octopamine Electrochemical Signals in Drosophila

## TL;DR

This paper introduces a machine learning method to separate dopamine and octopamine signals in fruit flies, enabling more accurate study of their roles in behavior.

## Contribution

A deep learning-based regression approach using a modified U-Net to deconvolve overlapping electrochemical signals of dopamine and octopamine.

## Key findings

- The method achieved less than 10% estimation errors for dopamine and octopamine.
- Predicted concentrations showed strong agreement with experimental data (r = 0.93, CCC = 0.93).
- Machine learning enables reliable separation of overlapping signals in Drosophila.

## Abstract

Drosophila melanogaster, the fruit
fly, uses the neurotransmitters dopamine and octopamine to mediate
learning, enabling adaptive behaviors such as reward seeking and punishment
avoidance. Their colocalization in the mushroom bodies makes it challenging
to study their individual contributions. Fast-scan cyclic voltammetry
allows subsecond monitoring of neurotransmitter dynamics, but simultaneous
detection of dopamine and octopamine remains difficult due to overlapping
oxidation and reduction peaks in their voltammograms. Traditional
signal separation methods, such as principal component regression,
assume fixed voltammogram shapes across time. However, this assumption
fails for octopamine, which exhibits time-varying voltammograms due
to secondary oxidation processes at the same potential as dopamine
oxidation. In this study, we use a deep learning-based regression
approach that analyzes color plots to separate dopamine and octopamine
signals collected in Drosophila. Using
the distinct primary oxidation peak of octopamine as input, a modified U-Net architecture was trained as a regression model to
predict the secondary oxidation peak and subtract it from the dopamine-octopamine
mixture to isolate dopamine contributions. The method achieved normalized
root-mean-square errors of 0.06 for dopamine and 0.08 for octopamine,
calculated against ground truth components from computationally generated
mixtures. Thus, estimation errors are under 10% and there is reliable
signal separation. Applications to experimentally measured mixtures
demonstrate accurate signal decomposition, with the predicted dopamine
and octopamine concentrations showing strong agreement (r = 0.93, CCC = 0.93) in scatter plot analysis. Thus, machine learning
provides a robust framework to deconvolute overlapping electrochemical
signals from octopamine and dopamine, facilitating simultaneous neurochemical
detection.