ethoscopy

A data-analysis toolbox utilising Pandas, Seaborn, and Plotly to curate, clean, analyse, and visualise behavioural time series data. Whilst the toolbix was created around the data produced from an Ethoscope (a Drosophila monitoring system), if the users data follows the same structure for time series data all methods can be utilised.

Head to the tutorial for an in-depth walk through. The Jupyter notebooks in tutorial_notebook/ are the authoritative, runnable examples.

For more information on the Ethoscope system, click here - If using in conjenction with Ethoscope data this software contains functions for loading the Ethoscope data into ethocopy from .db files both locally and in remote ftp servers.

At its core ethoscopy is a subclass of the data manipulation tool Pandas. The dataframe object has been altered to contain a linked metadata dataframe which contains experimental information. This secondary dataframe can be used to filter the data containing dataframe, as well as a store of information per specimen during analysis.

Ethoscopy contains methods to perform common analytical techniques per specimen in the data table, such as removing dead specimens, interpolating missing values, or calculating sleep from movement. Addtionally, specialist anlysing tools have been implemented for analysing circadian rhythm, such as periodograms, and for generating hidden Markov models (HMM) to understand latent behavioural states. HMMs are trained utilising hmmlearn in the background and come accompanied with a range of visualisation tools to understand the generated model.

For lifespan experiments, survival_table() and km_survival_plot() estimate Kaplan-Meier survival curves with Greenwood confidence bands. Death is detected from sustained immobility using the same rule as curate_dead_animals(), and specimens still moving when their recording ends are censored rather than counted as deaths. Pass subject_cols when one animal was recorded across several sessions so its recordings are treated as one individual.

-- Update to 2.0 --

This new update sees a whole refactoring of the code base to make everything more streamline and keep the package up to date with the new versions of pandas and numpy. Gone are seperate classes for periodograms and HMM based analysis, all are under one class behavpy(). Addtioanlly, now the user can choose between plotter packages, Seaborn and Plotly, and choose a desired colour pallete. The previous used package Plotly can balloon the size of jupyter notebooks, putting a strain on storage, despite being great for data exploration. If you just want static plots, use Seaborn. But be wary of comparison, the backend for Plotly plots is all calculated in ethoscopy applying z-score and bootstrapping to quantification plots, whereas Seaborn based plots will use the Seaborn internal tools for errors and averaging.

The latest update is backwards compatible with all previously saved behavpy dataframes. However, post loading they should be re-initiated as the new behavpy class. See in the getting started a demonstration of what to do.

Addtionally, the concat method ( behavpy_object.concat() ) for combining dataframes has been shifted to a function that is imported automatically. Call etho.concat(df1, df2) or etho.concat(*[df1, df2]) instead. There are other minor changes to method and argument names, which are reflected in their docstrings and in the tutorial.

Getting Started

Ethoscopy can be installed via pip from PyPi

We recommned installing ethoscopy into a virtual environment due to specific package versions.

python pip install ethoscopy

Example of use

Ethoscopy is primarily made to work in a Jupyter notebook environment and should be imported in as so:

import ethoscopy as etho

Generate a behavpy dataframe object as so:

data = pandas_dataframe
metadata = pandas_dataframe

df = etho.behavpy(data, metadata, check = True, canvas = 'plotly', palette = 'Set2')

# select only the data from specimens in experimental group 2
filtered_df = df.xmv('experimental_column', 'group_2')

Loading and re-initialising old data (saved pre 2.0)

import ethoscopy as etho
import pandas as pd

df = pd.read_pickle('path/to/your/file.pkl')
df = etho.behavpy(df, df.meta, check = True, canvas = 'plotly', palette = 'Set2')

Choosing a sleep rule

Since ethoscopy 3.0, sleep_annotation needs a sleep rule. To reproduce earlier results, add rule="classic", or declare it once at the top of a notebook:

import ethoscopy as etho
etho.set_sleep_rule("classic")   # every later sleep_annotation call uses it

Old notebooks can also be re-run without editing them by setting the environment variable ETHOSCOPY_SLEEP_RULE=classic. An explicit rule= argument always wins over set_sleep_rule(), which wins over the environment variable, as with matplotlib's rcParams. Without any of them, sleep_annotation raises an error explaining the choice.

from functools import partial
data = etho.load_ethoscope(meta, reference_hour=9.0,
                           FUN=partial(etho.sleep_annotation, rule="k"))

rule="classic" is the 5-minute rule as before. A 10-s window is moving when any frame is faster than the velocity threshold (1.0), and sleep is 5 minutes or more without a moving window. On current ethoscope data, tracking noise and brief twitches break sleep into fragments under it: a 5-minute bout needs 30 clean windows in a row, so if a still fly crosses the threshold in 5% of windows only about a fifth of real rests survive.

rule="k" (k = 3 by default; "k2" is the stricter variant) counts movement only when it is sustained or walking. A window is awake if the fly walked (its median position moved more than 10 px from the previous window) or if at least k sustained movement events started in the 60 s around it. An event is a run of consecutive frames above the classic velocity test. Runs in which the position never leaves its pixel, and jumps of one or two frames that land back within a pixel, are tracking noise and are ignored. Frames the tracker inferred are dropped. Windows without frames follow untracked: with the default "immobile" they count as still, and the fly walked only if it is found more than 10 px from where it was last seen (background-subtraction tracking loses still flies, so this keeps their sleep); "break" never scores them as sleep. The output adds walking, sustained and micro_awake and keeps the classic columns, moving included. Positions from load_ethoscope are in pixels; for positions given as a fraction of the ROI width, pass pixel=1/roi_width.

Use rule="classic" to reproduce earlier analyses, and rule="k" for new ones, especially on DeepTubeTracker or exposure-first recordings. The evidence for the k-rule:

• Video ground truth at night. Against pixel-motion truth from video, the night-time error per fly was 0.03–0.07 with either tracker (classic: 0.11–0.37). A dead fly recorded with exposure-first acquisition scores 100% asleep.
• Arousal. In 136 air-puff runs (1,319 flies), puffs were delivered at random against sham firings. Flies the k-rule scores asleep respond like sleeping flies, by day as well as by night: the puff-evoked response (real minus sham) was +1.2 percentage points when asleep by every rule, +2.1 when only the k-rule called it sleep, +3.2 when only k = 3 did, and +5.2 when awake. So the k-rule's extra daytime sleep, small movements in place by a fly that does not walk, is sleep-like.
• Across the lab archive (219 recordings) it narrows the spread of sleep between recordings (interquartile range of per-recording median sleep 30 → 21 percentage points) and preserves the rebound after sleep deprivation and the fumin and Clk^Jrk (light phase) phenotypes.

Before scoring, check a recording's tracking noise:

qc = etho.motion_qc(raw_data)   # one row per fly and light phase

fp_rate_fixed is how often a still fly crosses the classic threshold and rest_survival_fixed the fraction of 5-minute rests that would survive it. Values above about 0.01 mean classic sleep is unreliable for that recording. A large untracked_fraction matters under either rule. Both rules are also in rethomics' sleepr, so both toolboxes score a recording alike.

Tutorial data

The six pickle files used by the tutorial notebooks (~36 MB total, dominated by overview_data.pkl at ~31 MB) are intentionally not shipped with the PyPI wheel to keep pip install ethoscopy lean. Fetch them once with:

import ethoscopy as etho
etho.download_tutorial_data()   # idempotent — skips files already on disk

By default this populates ~/.cache/ethoscopy/tutorial_data/ (user-writable on every platform, no root required). After that, get_tutorial('overview'), get_tutorial('circadian') and get_HMM('M'|'F') Just Work.

Lookup order. At load time, ethoscopy checks three locations in this order:

1. <site-packages>/ethoscopy/misc/tutorial_data/ — used by dev/editable installs and by the project's Docker image (which pre-populates it during build; see Docker/Dockerfile).
2. $ETHOSCOPY_TUTORIAL_DATA_DIR if set — useful for shared clusters where one admin-maintained copy serves many users.
3. ~/.cache/ethoscopy/tutorial_data/ — the default for download_tutorial_data().

Manual download. The canonical URLs live at https://github.com/gilestrolab/ethoscopy/tree/main/src/ethoscopy/misc/tutorial_data. Place the six *.pkl files in any of the three locations above.

Development and Testing

Ethoscopy includes a comprehensive test suite to ensure code reliability and prevent regressions.

Running Tests

Install development dependencies:

pip install -e ".[dev]"

Run all tests:

pytest

Run tests with coverage:

pytest --cov=ethoscopy --cov-report=term-missing

Or use the convenience script:

python run_tests.py

Test Categories

• Unit tests: Fast tests for individual functions (pytest -m unit)
• Integration tests: Workflow testing (pytest -m integration)
• Performance tests: Long-running tests (pytest -m slow)

For detailed testing information, see TESTING.md.

License

This project is licensed under the GNU-3 license