3. Road Map
★ Introduce & Define: Brain Basis Set (BBS),
a method for quantifying and interpreting
inter-individual differences in fMRI patterns,
and for predicting phenotypes.
★ Cover several examples of BBS modeling using
publicly available datasets (Human
Connectome Project Young Adult (HCP-YA) &
Adolescent Brain Cognitive Development
(ABCD) data.
4. The Big Picture
★ Resting state functional connectomes are massive and complex.
or
Dense, high dimensional and redundant
information?
Sparse, low dimensional and unique
features?
★ Open question: Do connectomes differ across individuals in a
correspondingly massive number of ways, or do most differences take a
small number of characteristic forms?
5. The Method: Brain Basis Set (BBS)
Basis Set = Chosen # of top components from PCA decomposition of subjects x features matrix
Input can be rest or
task connectivity
matrices, task
activation maps,
structural images.
Each subject
represents a single
row. Their
connectivity matrix
is vectorized, so
columns represent
ROIs
6. Why PCA is Ideal
★ The world is complex and nonlinear,
and so is the human brain.
★ However, building linear models aids
in interpretability.
★ Dimension reduction compresses a large
set of features (edges/voxels) onto a new
feature subspace of lower dimension
without losing the important information.
7. Assessing low-rank structure in fMRI data: Converging Evidence
1. Estimate Intrinsic Dimensionality
2. Test out-of-sample reconstruction using
components from basis set
3. Use basis set in phenotype prediction
model
8. Examples of BBS working
★ Human Connectome Project Young Adult (HCP-YA): 1200 healthy adults,
age 22-35
○ Resting state fMRI connectomes (60 min of data per subject)
○ Task activation maps from 7 different tasks (19 contrasts)
★ Adolescent Brain Cognitive Development (ABCD) study: 2,206 9-10
year olds
○ Resting state fMRI connectomes (8 min of data per subject)
9. Phenotype BBS
General Executive 0.44
Processing Speed 0.39
Penn Progressive
Matrices
0.30
ASR Externalizing 0.24
ASR Internalizing 0.20
ASR Attention 0.21
NEO-Openness 0.18
NEO-Conscientiousness 0.19
NEO-Extroversion 0.13
NEO-Agreeableness 0.19
NEO-Neuroticism 0.00
HCP Resting State Prediction
Number of Components Used to Predict
MeanCorrelationbetweenPredicted&
ObservedPhenotype
Sripada et al. Scientific Reports (2019)
10. HCP Task Activation Prediction
Predicting g,
general intelligence
factor
Towards a “Treadmill Test” for Cognition: Reliable Prediction of Intelligence From Whole-Brain Task Activation Patterns.
Sripada C, Angstadt M, Rutherford S (2018)
Dubois et al. (2018) PHILOSOPHICAL
TRANSACTIONS OF THE ROYAL
SOCIETY B: BIOLOGICAL SCIENCES
11. ABCD Resting State Prediction
r = 0.31
r = 0.06
r = 0.15
Leave one SITE out cross validation = Generalizability
Thompson et al. (2018) Developmental Cognitive Neuroscience
12. Examples of BBS not working
★ Small sized studies, the curse of dimensionality problem
features (connections/voxels) >> observations (subjects)
★ Prediction of certain phenotypes (NEO Neuroticism in HCP,
Speed/Flexibility in ABCD).
★ Psychiatric phenotypes are difficult to predict
13. Summary
★ Introduction to Brain Basis Set for quantifying meaningful across
subject variability which can be leveraged to build predictive models.
★ Convinced you why we should all be discovering low rank structure
in our data.
★ Take home examples of successful and unsuccessful phenotype
prediction using resting state and task fMRI data.
★ To Do: More data! More models! More out-of-sample predictions!
Thank organizers for the opportunity to discuss my work, and ski the swiss and french alps. It has been a bucket list experience for me.
I would like to begin by acknowledging the wonderful lab I work in.
Chandra Sripada is the principal investigator. He has a unique education with an MD in Psychiatry & a PhD in Philosophy. Most of our collaborators are in Statistics and Computer Science. Our lab is very interested in data science and efficient computing practices applied to neuroscience research questions. The method I will introduce today is described fully in a paper just accepted a few days ago in Scientific Reports. Open access preprints are available on Bioarxiv for all of the results I will present today.
A central and exciting goal of the fMRI community has been to refine phenotypic prediction sufficiently so that functional connectomes can serve as reliable and objective “biomarkers” of clinically meaningful traits and dimensions. Individual differences in fMRI data has become a very important and popular topic, and has lead to successful prediction models. However, attempts to descriptively assess the nature and extent of population-wide across-individual variation are scarce. In other words, we have realized that individual differences contain relevant information, but we have yet to deeply explore the meaningful ways that people differ from one another.
Resting state connectomes are massive and complex. If there are 264 nodes in your parcellation, there are 34,000 connections. We hypothesized that the differences in a much smaller set of components can explain a sizable portion of how any two individuals meaningfully differ.
We begin with many subjects fMRI data. This can be resting state connectivity matrices or individual unthresholded task activation maps. We separate our subjects into training and test sets. We combine all of this data into a large matrix, where a subject’s data is vectorized and represents a single row and the columns or features represent our chosen ROIs. In our training set we run a matrix decomposition using PCA. Next we select a number of components to retain from the decomposition. Here, I show retaining 75 components, but in practice we have found this to be a range of 50-150.
Next in our training dataset, we calculate the expression scores for each of the n components for each subject by projecting their data onto the chosen subspace. We then fit a linear regression model with these expression scores as predictors and the phenotype of interest (i.e., intelligence) as the outcome. We save the Betas, the vector of fitted coefficients from this model. We then move to the test dataset and again calculate the expression scores for each of the n components for each test set subject. Our predicted phenotype for each test subject is the dot product of the Beta vector learned from the training dataset with the vector of component expression scores for that subject. So, the Basis set that I will refer to many more times in this talk is the chosen number of components from the PCA decomposition of the subject x features matrix.
I like to think of dimensionality reduction in terms of zipping a file. It compresses your data to a smaller, more manageable format but does not loose the important information and you can always go back into the high dimensional space if needed.
To test for the presence of low-rank structure in the (subjects x features) matrix we have several tricks. First, before the PCA decomposition, we estimate the intrinsic dimensionality of the matrix using a maximum likelihood estimation developed by Liza Levina. Across many large datasets like HCP and ABCD we find this estimation is always between 50-150. Next we test out of sample reconstruction. We take a subject who was not included in discovering the basis set, and try to reconstruct their connectome. In this graph shown here we observe the correlation between reconstructed and original connectomes as a function of components used, and we see a plateau around 100 components where adding in more components does not improve the reconstruction. Finally, we test how many components we need to predict phenotypes which I will show in detail.
In this paper we predict 11 phenotypes from the HCP study. The general executive and processing speed phenotypes are from a factor analysis of the NIH Toolbox, and we also predict Externalizing, Internalizing, and Attention from the ASR and the Neo 5-factor personality phenotypes. This table represents the held out test set of 100 unrelated subjects. And this figure represents 10-fold cross validation within our training set of approximately 900 subjects. I know this axes are a bit hard to read. On the x axis we show the number of components used in the prediction starting at 0 and going out to 500 and the y axis shows correlation between predicted and observed phenotypes. The point I want to make with this figure is we observed the same plateauing effect I showed a few slides ago in the out of sample reconstruction figure. That is if we continue to add more components and more information, we are not doing really any better at predicting our phenotypes meaning that these later components do not add meaningful information for predictive models.
In this study we again use HCP data but this time instead of resting state connectomes we use task activation maps from all 7 HCP tasks, and we are predicting G a general factor of intelligence that was introduced recently in a nice paper by Julian Dubois. This figure represents prediction of g in a held out test set of 100 unrelated subjects. The x axis shows the contrast used in the model. We found that tasks that tap executive processes were more predictive of intelligence. Our best prediction was using the N-back working memory task, specifically the 2bk-0bk contrast where we observed an average correlation of 0.68 between predicted and observed phenotype. fronto-parietal regions (executive function regions) as a major source of discriminative information for making subject-level predictions of intelligence.
In this study we use Brain Basis Set modeling on resting state data from the ABCD study to predict factors from a recent paper by Wesley Thomason and colleagues which used a Bayesian Probabilistic Principal Components Analysis on the behavioral data. Because this is a multi-site study, we chose to do a leave one site out cross validation and we are really excited about this because it represents that these results are generalizable across the entire population. Each dot in this figure represent the average prediction of a site, and on the X axis we show how many subjects were in this site. The red dot for each phenotype represents the mean prediction across all sites. As you can see we best predict the General Ability factor which is a little different than the g factor, but represent executive function.
Not necessarily due to the BBS method, as other types of predictive models do not work well for these phenotypes. Maybe we need different tasks to tap into these cognitive/clinical domains?