The Bayesian polyvertex score (PVS-B): a whole-brain phenotypic prediction framework for neuroimaging studies

Abstract

The traditional brain mapping approach has greatly advanced our understanding of the localized effect of the brain on behavior. However, the statistically significant brain regions identified by the standard mass univariate models only explain minimal variance in behavior despite increased sample sizes and statistical power, highlighting the nonsparseness of the explanatory signal in the brain. We introduced the Bayesian polyvertex score (PVS-B), a wholebrain prediction framework that aggregates the effect sizes across all vertices to predict individual variability in behavior. The PVS-B estimates the posterior mean effect size at each vertex with the summary statistics from the brain mapping approach and the correlation structure of the imaging phenotype. Empirical data showed that the PVS-B was able to double the variance explained of the total composite cognition score by an nBack fMRI contrast when compared to prediction models based on the mass univariate parameter estimates as well as models based on p-value thresholds. A fivefold improvement in variance explained by the PVS-B was observed using the stop signal fMRI contrast to predict individual variability in the stop signal reaction time. We believe that the PVS-B can shed light on the multivariate investigation of brain-behavioral associations and will empower small scale neuroimaging studies with more reliable and accurate effect size estimates. Data used in the preparation of this article were obtained from the Adolescent Brain Cognitive Development (ABCD) Study (https://abcdstudy.org), held in the NIMH Data Archive (NDA). The ABCD Study is a multisite, longitudinal study designed to recruit more than 10,000 children age 9-10 and follow them over 10 years into early adulthood.

Towards Reproducible Brain-Wide Association Studies

“…Currently, brain-wide association studies are trapped in the initial discovery phase. Larger samples will inevitably allow multivariate models, both those currently in use and future improvements currently being developed [38][39][40][41] , to be employed by individual research groups. Smaller, investigator-initiated neuroimaging studies will continue to be just as important for human neuroscience as they were in the days when samples of N>10,000 for brainwidewide association studies (BWAS) were still an impossibility.…”

Distinct regionalization patterns of cortical morphology are associated with cognitive performance across different domains

“…We recently demonstrated that there were significant associations between the regionalisation of CSA and CTH (when controlling for global structural measures) and the fluid and crystallised composite scores from the NIH Toolbox within the ABCD baseline sample using the Multivariate Omnibus Statistical Test: the MOSTest 24 . This test aggregates vertex-wise associations across the cortex thereby improving our power for detecting widely distributed effects compared to the standard neuroimaging approach [24][25][26] . Given the graded nature of the biology underlying the patterning of the cortex this is a more appropriate statistical approach compared to standard neuroimaging methods.…”

“…This allowed us to determine the unique association between relative cortical morphology and cognition and compare this to the predictive power of brain structure without controlling for global measures. The method used here is outlined in detail by Zhao and colleagues 25 . The association between each imaging phenotype and each cognitive task was modelled using the mass univariate approach such that the behaviour of interest was predicted independently at each vertex using a general linear model (GLM).…”

Generalization of Cortical MOSTest Genome-Wide Associations Within and Across Samples

“…Having discovered genomic loci in training folds, we perform replication of these loci in the test sets. To perform replication for each SNP we calculate a PolyVertex Score (PVS) (similar to (6,7)), to project the test imaging phenotype data to a single scalar value for each individual using projection weights learned in the training data. This approach is similar to the widely used method of Polygenic Risk Scores (PRS) in genetics(8), where instead of predicting a phenotype we are predicting a single genomic variant and instead of using distributed effects across the genome as predictors we use the distributed effects across the cortex, estimated in the training set.…”

“…Having discovered genomic loci in training folds, we perform replication of these loci in the test sets. To perform replication for each SNP we calculate a PolyVertex Score (PVS) (similar to (6,7)) from imaging data in the test set for each MOSTest discovered locus. This PVS aggregates the distributed effects across the cortex by taking a weighted sum across all vertices using mass univariate z statistics as weights from the training set.…”