�ENIGMA and COINSTAC��Theo G.M. van Erp, Ph.D.��Department of Psychiatry and Human Behavior�University of California Irvine, Irvine, CA, USA ��

ENIGMA Chairs Meeting 2020

ENIGMA

Enhancing NeuroImaging Genetics through Meta-Analysis

1400+ scientists

43 countries

50 working groups

Paul Thompson

​

Jason Stein

​

ENIGMA’s first consortium paper

​

Stein et al. VOLUME 44, NUMBER 5, MAY 2012, Nature Genetics

51 working groups (95 chairs), 31 disorders; 1,400 members, 400 institutions, 43 countries

​

​

​

3

33 clinical working groups

5

​

​

​

ENIGMA Disorder Working Group

Conduct prospective meta-/mega-analyses to:

​

1. Rank order case-control (disease-related) effect sizes of brain measures

​

• Examine factors that may moderate these effects sizes

​

• Compare brain effects across disorders

Schizophrenia

Jess & Theo

Bipolar Disorder

Chris, Derrek, & Ole

Major Depressive Disorder

Lianne & Dick

ENIGMA’s Analysis Approach

• Structural image analysis using FreeSurfer 5.1 or 5.3 (some 6.0dev) – earlier studies also used FSL
• Extract deep brain structure volumes, and mean cortical thickness and surface area measures within Desikan-Killiany Atlas ROIs (Desikan et al. 2006)
• used a common set of quality assurance procedures, which included review of distributions for outliers and visual inspection of deep brain structure and cortical surface labels
• Run statistical analyses in R (lm) to generate effect sizes (Cohen’s d and partial correlations)
• Meta-analyses using inverse variance-weighted random-effects model as implemented in the R package metafor (version 1.9-7) with FDR correction for multiple comparisons.

​

Cortical thickness and surface area are thought to be influenced by separate sets of genes

(Panizzon eta l. 2009; Winkler et al. 2010).

9 groups published their cortical papers

8

​

​

​

ENIGMA initial meta-analysis: �Dependent on Freesurfer common structure, and human agreement on covariates of interest

Standard CSV files of variables

R scripts

zip/tar file of results returned to central analysis team for meta-analysis

Automation

via

Linux Shell Scripts

Automating ENIGMA: COINSTAC

​

​

• Collaborative Informatics and Neuroimaging Suite

Toolkit for Anonymous Computation (COINSTAC)

​

- enables federated data analysis

A. D. Sarwate, S. M. Plis, J. A. Turner, M. R. Arbabshirani, and V. D. Calhoun, "Sharing privacy-sensitive access to neuroimaging and genetics data: a review and preliminary validation," Front Neuroinform, vol. 8, p. 35, 2014, 3985022.

​

S. M. Plis, A. D. Sarwate, D. Wood, C. Dieringer, D. Landis, C. Reed, S. R. Panta, J. A. Turner, J. M. Shoemaker, K. W. Carter, P. Thompson, K. Hutchison, and V. D. Calhoun, "COINSTAC: A Privacy Enabled Model and Prototype for Leveraging and Processing Decentralized Brain Imaging Data," Frontiers in Neuroscience, vol. 10, Aug 19 2016.

COINSTAC: Local analysis, globally aggregated

The process begins when the lead site defines the local and global analyses.

​

The local analysis is downloaded by each site (docker/singularity image within coinstac) to run a local data analysis that generates intermediate results that are combined in the global analysis that is run at the lead site.

1

Lead Site

From: Turner, MD., Gazula, H., et al., ENIGMA COINSTAC: Increasing neuroimaging data diversity with managed privacy, https://doi.org/10.13140/RG.2.2.18471.37285 

Participating sites --collaborating sites with similar data-- can see how their data is to be used (they can see all of the analyses proposed).

​

If they agree, they may join the consortium to participate in the meta-analysis, and map their data to the analysis.

2

Participating

Sites

3

Once the participating sites have joined the consortium and mapped their data, the lead site launches all of the local analyses.

​

Analyses at the participating run automatically and return the intermediate data to the lead site.

ENIGMA COINSTAC cross-diagnosis collaborations

ENIGMA SZ

rsfMRI

DTI

Structural

Regressions

Multivariate

Multimodal ML

ENIGMA BD

ENIGMA MDD

ENIGMA-COINSTAC: Advanced Worldwide Transdiagnostic Analysis Of Valence System Brain Circuits

Neural Circuitry Associated with Negative Symptom Domains

​

​

​

​

Strauss et al. 2019; JAMA Psychiatry

MPIs: Turner, Van Erp, & Calhoun

1R01MH121246

Overview of COINSTAC Analysis

• The consortium leader:
• names the consortium
• selects an analysis pipeline
• Members:
• join the consortium
• download a dockerized version of the consortium-specific analysis pipeline to their local COINSTAC instances
• map their data files for analysis (e.g., csv files with the relevant clinical and demographic data)
• The consortium leader:
• launches the pipeline.

COINSTAC ENIGMA Negative Symptom Domain Analysis Pipeline

• At each site:
• Computes SANS factor scores
• Runs regressions between SANS factor scores and deep brain structure volumes and cortical thickness and surface area measures
• At consortium leader site:
• Runs meta-analyses
• We will compare the COINSTAC meta-analysis output with the meta-analysis done via the method of distributing code and obtaining analysis output
• The first “real” ENIGMA COINSTAC test in early June involved seven sites from Australia to Germany
• This identified issues with the transfer of large files
• And some lack of robustness in the R scripts

​

• We ran a successful analysis with 7 sites on 7/30/2020.

Using COINSTAC

Where to download:

https://github.com/trendscenter/coinstac/releases

​

Currently v4.7.5 is the latest binary release:

​

After download and uncompressing, just double click on the “coinstac” icon

COINSTAC Login Screen

Computations

…

Download Docker/Synology Image

to local compute node

Remove Docker /Synology Image

from local compute node

Add a New Docker/Synology Image

Pipeline Menu

c

R code to run analyses at individual

sites + R code to run meta-analysis

Create a new pipeline

Consortia Menu

Start new consortium

Leave a consortium you joined

Join a consortium

Map data to a consortium you joined

Map Local Data

Launch the Pipeline

See the Remote and Local Pipeline Progress and get your results.

c

See that all sites have mapped

their data successfully

International multisite collaborations

Power - Statistical power to tackle new kinds of questions (‘cracking the brain’s genetic code’), impossible to solve before

​

Computational efficiency - Massive distributed computing, leverage infrastructure/data across 400 institutions across 40+ countries; beginning new cooperative machine learning, deep learning, new mathematics

​

Open Science & Reproducibility Studies are designed to ensure replication (as seen from Forest plots). Proposals = registered reports. Analyses run are public info.

​

Team Science experts from all fields working together makes for more expertise and a much stronger publication, and everyone learns and gains!

​

Advancing Science Rather than try to answer the same question independently, merge data already collected to drive new investigations

​

​

​

​

​

​

​

Acknowledgments

NSF: 1631819

NIH COINSTAC: R01DA040487

NIH COINSTAC ENIGMA: 1R01MH121246

NIH COINSTAC GIGA/ABCD: R01DA049238

Turner Lab