neurodata/mcc

Reproducible code for "Multiscale Comparative Connectomics"

Multiscale Comparative Connectomics

Reproducible code for the results shown in our manuscript Multiscale Comparative Connectomics (MCC).

Vivek Gopalakrishnan, Jaewon Chung, Eric Bridgeford, Benjamin D. Pedigo, Jesús Arroyo, Lucy Upchurch, G. Allan Johnson, Nian Wang, Youngser Park, Carey E. Priebe, and Joshua T. Vogelstein. “Multiscale Comparative Connectomics”. arXiv:2011.14990 (Nov. 2020).

Table of Figures and Tables

Click any link for individual instructions on how to generate that specific figure. Alternatively, execute the file code/run to generate all figures at once.

Figure 1: Overview of the statistical framework for multiscale comparative connectomics.

Figure 2: Average connectomes for each mouse strain with hierarchical structure labels.

Figure 3: Vertex embeddings of the corpus callosum obtained by the omnibus embedding.

Figure 4: Tractograms of the strongest signal edge, vertex, and community.

Figure 5: Whole-brain embeddings of all mouse connectomes in a low-dimensional space.

Figure 6: Our methods uncover more information about network topology than neuroanatomical metrics.

System Requirements

MCC uses both Python 3.8 and R 3.6.1. The analyses above depend on the following packages:

# Conda Python packages
jupyter==1.0.0
rpy2==3.3.6
statsmodels==0.12.1

# Pip install the latest version of graspologic
graspologic @ git+git://github.com/microsoft/graspologic@dev

# Conda R packages
r-base==3.6.1
bioconductor-complexheatmap==2.2.0
r-cairo==1.5_10
r-circlize==0.4.12
r-data.table==1.12.2
r-essentials==3.6.0
r-future==1.21.0
r-future.apply==1.7.0
r-ggplot2==3.1.1
r-globaloptions==0.1.2
r-igraph==1.2.4.1
r-mltools==0.3.5
r-tidyverse==1.2.1

# CRAN R packages
cdcsis==2.0.3

These analyses have been tested on macOS x64 and Linux x64.

Installation Instructions

We created a computational environment with these packages using a dedicated Miniconda kernel (should take about 5 mins to setup):

conda env create --file environment.yml --name mcc

To get rpy2 running on an M1 Mac, it's currently necessary to install it in API mode: RPY2_CFFI_MODE=API pip install rpy2. Also, it might be easier to install the R packages directly through the scripting interface instead of through conda since many compatibility conflicts haven't been resolved for new arm64 versions.

Reproducing Figures

Scripts to reproduce the figures in MCC are organized below.

Figure 1

• Run code/1_statistical_framework_graphs.ipynb (expected runtime: 5 seconds)
• This script uses igraph to generate the sample connectomes and graph models seen above

Figure 2

• Run code/2_plot_adjacency_matrices.ipynb (expected runtime: 5 seconds)
• This script uses ComplexHeatmap to generate average connectomes for each mouse strain

Figure 3

• Run code/3_cc_emedding.ipynb (expected runtime: 5 seconds)
• This script uses graspologic to embed the corpus callosum brain region of every mouse in a low-dimensional space

Figure 4

• Run code/4a_identifying_signal_components.ipynb (expected runtime: 2 min)
• This script uses graspologic and various k-sample hypothesis testing packages to identify the strongest signal edges, vertices, and communities

Tables 1-4

• Run code/4b_format_signal_components_tables.ipynb (expected runtime: 30 seconds)
• This script uses pandas to nicely format the results generated for Figure 4 into publication-ready tables
• Tables are found in the Supplement of the MCC manuscript

Figure 5

• Run code/5_whole_brain_emedding.ipynb (expected runtime: 5 seconds)
• This script uses classical multidimensional scaling (cMDS) to embed the results of the omnibus embedding in a low-dimensional space

Figure 6

• Run code/6_conditional_independence_anatomy.ipynb (expected runtime: 5 hours on a 48 core machine, probably much longer on a normal laptop)
• This script uses cdcsis to compute a bunch of conditional independence tests
• The purpose of this test is to determine if our methods recover information about network topology not encoded in neuroanatomy

mcc's Issues

Add effect sizes to p-values

• Are test statistics themselves effect sizes?
• Modify Tables 1-4

Update vertex-level simulation

The old simulation compared OMNI to {MASE, vertex-level network statistics}

Proposed changes

Modify the algorithms being compared:

• Add MDMR, DBS
• Rename vertex-level network statistics -> exponential random graph model
• Remove MASE

Performance metrics:

• Replace the three separate plots per row with an ROC curve

Update edge-level simulation

The old simulation compared DCorr to {T-test and Mann-Whitney}

Proposed changes

Modify the algorithms being compared:

• Add NBS, TF-NBS, UminP

Performance metrics:

• Replace the three separate plots per row with a covariate-adjusted ROC

Also include different p-value correction methods:

• Compare Bonferroni, Bonf-Holm, Benj-Hoch, Benj-Yeku