Detecting pairwise correlations in high-density recordings: open science in action

Detecting pairwise correlations in
high density recordings: open science in action

Stephen J Eglen                  Cambridge Computational Biology Institute
https://sje30.github.io          University of Cambridge
sje30@cam.ac.uk                  @StephenEglen

Slides: http://bit.ly/eglen-comp

Code/data http://github.com/sje30

CNS 2017 Workshop Neuronal population recordings
These slides are available under a creative common CC-BY license.

Acknowledgements

Paul Charlseworth, Ellese Cotterill, Catherine Cutts, Tom Edinburgh, Gerrit Hilgen, Ole Paulsen, Evelyne Sernagor.

BBSRC, EPSRC, Wellcome Trust, Software Sustainibility Institute.

What are retinal waves?

Correlating spike trains

Retinal waves: Wong et al (1993) (fig 8, 9)

How to interpret figures?

1. three regions: [0,1], 1, [1, ∞]
2. lower firing rates -> higher CI

What method should we use?

Phase 1: finding a short list

34 measures in literature + 1 from us => 35.

Six necessary properties:

1. Symmetric
2. Robust to variations in firing rate
3. Robust to amount of data
4. Bounded [-1, +1]
5. Robust to variations in bin width (Δt)
6. Anticorrelation should be clear

Dependence on firing rate

Phase 2: Desirable properties

Desirable properties:

1. D1: Ignore periods when both neurons are inactive.
2. D2: minimal assumptions on structure.
3. D3: aside from Δt, minimise number of parameters

Four methods:

1. Kerschensteiner and Wong correlation (D1, D2)
2. Spike count correlation (D2, D3)
3. Kruskal et al. binless correlation measure (D1?, D2, D3)
4. Tiling coefficient (D1, D2, D3)

Tiling measure

Evidence

Blankenship et al (2011)

Correlations decay with age

Inference methods

Correlograms: needles in haystacks

Desirable (binary) properties
of burst detectors

1. Deterministic
2. No assumptions on ISI distribution
3. Minimal parameters
4. Computational efficiency

Desirable (qualitative) properties
of burst detectors

5. Non-bursting trains
6. Non-stationary trains
7. Regular short bursts
8. Nonstationary bursts
9. Regular long bursts
10. High frequency bursts
11. Noisy trains

Evaluation

Ranking of burst analysis methods

Burst analysis: mouse retinal neurons

Example trains of cells derived from human stem cells

Burst analysis:
networks derived from human stem cells

Reproducible research

To do this work we needed access to data. We have released these sets.

Then and now

Access to code

English is a poor way of unambigiously describing algorithms.

Claerbout and Donoho: the scholarship is not the article; the “scholarship is the complete software […]”

“Talk is cheap. Show me the code.” (Torvalds, 2000)

2/8 burst detectors were available (as R packages).

Specific recommendations

1. Include enough code to reproduce key figure/result from your paper (“modeldb”).
2. Provide toy examples if your project is too intensive to expect others to run in a few hours.
3. Version control (github)
4. Licence (MIT)
5. Provide data (“good enough” formats)
6. Provide tests
7. Use standards (where you can)
8. Use permanent URLs (Zenodo/figshare)