Artificial intelligence-driven phenotyping of zebrafish psychoactive drug responses

MPTP-Treated Zebrafish Recapitulate ‘Late-Stage’ Parkinson’s-like Cognitive Decline

“…A convolution neural network (CNN) sequentially applies convolution operators to an input and, thus, extracts the higher-level features, from basic lines and gradients to shapes [32]. Here, we used the ResNet34 CNN architecture, which offers the best balance between training time, complexity and the prediction quality [33], and has already been successfully applied to analyzing zebrafish behavioral data [34]. In the present study, we trained the computational model to predict MPTP concentrations using video recordings of SAB in fish in the Y-maze test.…”

“…In the present study, we trained the computational model to predict MPTP concentrations using video recordings of SAB in fish in the Y-maze test. We extracted a fish position from the videos during the test using the EthoVision XT-10 software, cut each locomotor track into short 30-s frames, converted each track to an image, and used those images to train a neural network, similar to [34].…”

“…This network was analysed using the Louvain algorithm, which assigns a community to each node based on the optimal data modularity. The method was chosen for its high-quality detection of distinct data clusters and fast data processing speed (see [34] for details).…”

“…Prediction results are usually described as a confusion matrix-a square matrix where each row represents actual classes, and columns reflect the predicted classes. To assess significance of the model predictions, a permutation test shuffled all class labels and computed a baseline accuracy-a minimum level to consider prediction statistically 'significant', as in [34]. We then constructed a similarity graph using the confusion matrix as a matrix of edge weights and detected communities (small subsets of the graph nodes that share similar properties).…”

“…We then constructed a similarity graph using the confusion matrix as a matrix of edge weights and detected communities (small subsets of the graph nodes that share similar properties). Finally, we used the Louvain community detection method, as it yields optimal partitions [36], and has already been successfully used for zebrafish behavioral analyses [34].…”