In this work, we propose a simple yet effective solution to the problem of connectome inference in calcium imaging data. The proposed algorithm consists of two steps. First, processing the raw signals to detect neural peak activities. Second, inferring the degree of association between neurons from partial correlation statistics. This paper summarises the methodology that led us to win the Connectomics Challenge, proposes a simplified version of our method, and finally compares our results with respect to other inference methods.

1. 1/11
Simple connectome inference from partial
correlation statistics in calcium imaging
Antonio Sutera, Arnaud Joly, Vincent Fran¸cois-Lavet, Zixiao Aaron Qiu
Gilles Louppe, Damien Ernst and Pierre Geurts (aka The AAAGV Team).

2. 2/11
Introduction
From time-series of the neuron activity, the goal is to infer the
directed connections between neurons.
i j

3. 3/11
Introduction
But this problem is difficult because of...
masking effects
low sampling rate
⇒
slow decay of fluorescence.
Time [s]0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8

4. 4/11
Framework of our solution Table of contents
How-to :
1. Deal with calcium fluorescence signals : signal processing.
2. Find the network : the inference method.
3. Improve the method : averaging and tuning.
One step further :
Improvement of each stage of our solution.
Comparison with other methods.
Even further :
The “full method” in a nutshell.

5. 5/11
Signal processing
1
2 3
4
Signal processing
0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 10
0.5
0.0
0.5
1.0
1.5
2.0
2.5
0.2
0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10
1.0
1.5
2.0
2.5
3.0
1. Applying a low-pass filter
2. Applying a high-pass filer
3. Applying a hard-threshold
filter
4. Magnify the importance of
spikes that occur in case of
low global activity

6. 6/11
Connectome inference from partial correlation statistics
pi,j
partial correlation
=
degree of association
between neurons i and j
known
known
Partial correlation coefficient pi,j
between neurons i and j defined by :
pi,j = −
Σ−1
ij
Σ−1
ii Σ−1
jj
,
Advantages and
drawbacks :
Only direct associations
Filter out spurious
indirect effects
Easy to compute
(with lots of data)
Edge orientation
Sensitive to the value of
parameters in the
filtering process

7. 7/11
Improvements
Approximation for partial correlation statistics using only the
800 first principal components.
Averaged partial correlation statistics over various values of
the parameters
Improve robustness (over all networks)
Reduce variance of its prediction
Decrease the sensitivity to the filtering process

8. 8/11
Each stage of our solution brings some improvement
Average scores on normal-1,2,3,4 datasets.

9. 9/11
Comparison with other methods
Average scores on normal-1,2,3,4 datasets.

10. 10/11
An optimized version to win the challenge : “Full method”
A tuned signal processing,
Weighted average of partial correlation statistics,
Prediction of edge orientation.

11. 11/11
For every details of our method...
The description of the method :
Sutera, A., Joly, A., Fran¸cois-Lavet, V., Qiu, Z. A., Louppe,
G., Ernst, D., Geurts, P. (2014). Simple connectome
inference from partial correlation statistics in calcium imaging.
Code available at :
https://github.com/asutera/kaggle-connectomics