This reposity cotains the code corresponding with:
@article{lts-behavior,
author = {Grace C. McKenzie-Smith* and Scott Wolf* and Julien Ayroles and Joshua Shaevitz},
title = {Capturing continuous, long timescale behavioral changes in \textit{Drosophila melanogaster}}
The data associated with this study can be found at URL and is described in detail within the paper. The data set encompasses over 2 billion pose instances, each paired with edge calls and metadata.
Brief examples of analysis are available here illustrating how to load and analyze tracks. This pipeline heavily depends on work from Gordon Berman and Kanishk Jain both from the MATLAB implementation of MotionMapper and the Python implementation MotionMapperPy.
The main refinements of this pipeline is dealing with memory constraints of the original MotionMapperPy implementation, separating steps to allow for drastic speed increases with parallelization, and the utilization of Lomb-Scargle periodograms to identify periodic behaviors in addition to continuous wavelet transforms (CWT).
The sequential run of the analysis pipeline is:
sbatch input-processing.slurmsbatch wavelets.slurmsbatch subsample_for_training.slurmpython lts-tsne-subsample-3.pysbatch embed.slurmpython lts-tsne-watershed-5.py
This script takes the raw data, interpolates missing data, and smooths data as needed. It also splits the data to the individual flies and saves the data in a format that is easy to load.
This script computes the wavelet transform of the data. It is a computationally expensive step and is parallelized over the flies as much as possible.
Because subsampling directly requires a lot of memory, this script subsamples the data for training the embedding for each file. It is parallelized over the flies. This also calculates
Generates training data and embeds the data using t-SNE. This training provides the model for the embedding of the full data set.
This script embeds the full data set using the model generated in the previous step. It's extremely slow and is parallelized over the fly hours but still takes multiple days to run on a cluster.
We used SLEAP for tracking the flies. The tracking code can be found in the tracking directory. The code is a slightly modified version that enables the input of MKV files.
The tracking process is parallelized over fly-hours to efficiently utilize GPU resources for variable-length data. Email gmckenziesmith@wesleyan.edu or swwolf@princeton.edu to obtain the trained SLEAP models.
We also included documentation on experimental methods, including the media recipe, in the documents directory.