Summary

Brain Predictability toolbox (BPt) represents a unified framework of machine learning (ML) tools designed to work with both tabulated data (e.g. brain derived, psychiatric, behavioral and physiological variables) and neuroimaging specific data (e.g. brain volumes and surfaces). This package is suitable for investigating a wide range of different neuroimaging-based ML questions, in particular, those queried from large human datasets.

Availability and implementation

BPt has been developed as an open-source Python 3.6+ package hosted at https://github.com/sahahn/BPt under MIT License, with documentation provided at https://bpt.readthedocs.io/en/latest/, and continues to be actively developed. The project can be downloaded through the github link provided. A web GUI interface based on the same code is currently under development and can be set up through docker with instructions at https://github.com/sahahn/BPt_app.

Objective

Independent of weight status, rapid weight gain has been associated with underlying brain structure variation in regions associated with food intake and impulsivity among pre-adolescents. Yet, we lack clarity on how developmental maturation coincides with rapid weight gain and weight stability.

Methods

We identified brain predictors of 2-year rapid weight gain and its longitudinal effects on brain structure and impulsivity in the Adolescent Brain Cognitive Development^SM Study®. Youth were categorized as Healthy Weight/Weight Stable (WS_HW , n = 527) or Weight Gainers (WG, n = 221, >38lbs); 63% of the WG group were healthy weight at 9-to-10-years-old.

Results

A fivefold cross-validated logistic elastic-net regression revealed that rapid weight gain was associated with structural variation amongst 39 brain features at 9-to-10-years-old in regions involved with executive functioning, appetitive control and reward sensitivity. Two years later, WG youth showed differences in change over time in several of these regions and performed worse on measures of impulsivity.

Conclusions

These findings suggest that brain structure in pre-adolescence may predispose some to rapid weight gain and that weight gain itself may alter maturational brain change in regions important for food intake and impulsivity. Behavioural interventions that target inhibitory control may improve trajectories of brain maturation and facilitate healthier behaviours.

Correlating transcriptional profiles with imaging-derived phenotypes has the potential to reveal possible molecular architectures associated with cognitive functions, brain development and disorders. Competitive null models built by resampling genes and self-contained null models built by spinning brain regions, along with varying test statistics, have been used to determine the significance of transcriptional associations. However, there has been no systematic evaluation of their performance in imaging transcriptomics analyses. Here, we evaluated the performance of eight different test statistics (mean, mean absolute value, mean squared value, max mean, median, Kolmogorov-Smirnov (KS), Weighted KS and the number of significant correlations) in both competitive null models and self-contained null models. Simulated brain maps (n = 1,000) and gene sets (n = 500) were used to calculate the probability of significance (Psig) for each statistical test. Our results suggested that competitive null models may result in false positive results driven by co-expression within gene sets. Furthermore, we demonstrated that the self-contained null models may fail to account for distribution characteristics (e.g., bimodality) of correlations between all available genes and brain phenotypes, leading to false positives. These two confounding factors interacted differently with test statistics, resulting in varying outcomes. Specifically, the sign-sensitive test statistics (i.e., mean, median, KS, Weighted KS) were influenced by co-expression bias in the competitive null models, while median and sign-insensitive test statistics were sensitive to the bimodality bias in the self-contained null models. Additionally, KS-based statistics produced conservative results in the self-contained null models, which increased the risk of false negatives. Comprehensive supplementary analyses with various configurations, including realistic scenarios, supported the results. These findings suggest utilizing sign-insensitive test statistics such as mean absolute value, max mean in the competitive null models and the mean as the test statistic for the self-contained null models. Additionally, adopting the confounder-matched (e.g., coexpression-matched) null models as an alternative to standard null models can be a viable strategy. Overall, the present study offers insights into the selection of statistical tests for imaging transcriptomics studies, highlighting areas for further investigation and refinement in the evaluation of novel and commonly used tests.

Effect sizes are commonly interpreted using heuristics established by Cohen (e.g., small: r = .1, medium r = .3, large r = .5), despite mounting evidence that these guidelines are mis-calibrated to the effects typically found in psychological research. This study's aims were to 1) describe the distribution of effect sizes across multiple instruments, 2) consider factors qualifying the effect size distribution, and 3) identify examples as benchmarks for various effect sizes. For aim one, effect size distributions were illustrated from a large, diverse sample of 9/10-year-old children. This was done by conducting Pearson's correlations among 161 variables representing constructs from all questionnaires and tasks from the Adolescent Brain and Cognitive Development Study® baseline data. To achieve aim two, factors qualifying this distribution were tested by comparing the distributions of effect size among various modifications of the aim one analyses. These modified analytic strategies included comparisons of effect size distributions for different types of variables, for analyses using statistical thresholds, and for analyses using several covariate strategies. In aim one analyses, the median in-sample effect size was .03, and values at the first and third quartiles were .01 and .07. In aim two analyses, effects were smaller for associations across instruments, content domains, and reporters, as well as when covarying for sociodemographic factors. Effect sizes were larger when thresholding for statistical significance. In analyses intended to mimic conditions used in "real-world" analysis of ABCD data, the median in-sample effect size was .05, and values at the first and third quartiles were .03 and .09. To achieve aim three, examples for varying effect sizes are reported from the ABCD dataset as benchmarks for future work in the dataset. In summary, this report finds that empirically determined effect sizes from a notably large dataset are smaller than would be expected based on existing heuristics.

Background and aims

Recently, we demonstrated that a distinct pattern of structural covariance networks (SCN) from magnetic resonance imaging (MRI)-derived measurements of brain cortical thickness characterized young adults with alcohol use disorder (AUD) and predicted current and future problematic drinking in adolescents relative to controls. Here, we establish the robustness and value of SCN for identifying heavy alcohol users in three additional independent studies.

Design and setting

Cross-sectional and longitudinal studies using data from the Pediatric Imaging, Neurocognition and Genetics (PING) study (n = 400, age range = 14-22 years), the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) (n = 272, age range = 17-22 years) and the Human Connectome Project (HCP) (n = 375, age range = 22-37 years).

Cases

Cases were defined based on heavy alcohol use patterns or former alcohol use disorder (AUD) diagnoses: 50, 68 and 61 cases were identified. Controls had none or low alcohol use or absence of AUD: 350, 204 and 314 controls were selected.

Measurements

Graph theory metrics of segregation and integration were used to summarize SCN.

Findings

Mirroring our prior findings, and across the three data sets, cases had a lower clustering coefficient [area under the curve (AUC) = -0.029, P = 0.002], lower modularity (AUC = -0.14, P = 0.004), lower average shortest path length (AUC = -0.078, P = 0.017) and higher global efficiency (AUC = 0.007, P = 0.010). Local efficiency differences were marginal (AUC = -0.017, P = 0.052). That is, cases exhibited lower network segregation and higher integration, suggesting that adjacent nodes (i.e. brain regions) were less similar in thickness whereas spatially distant nodes were more similar.

Conclusion

Structural covariance network (SCN) differences in the brain appear to constitute an early marker of heavy alcohol use in three new data sets and, more generally, demonstrate the utility of SCN-derived metrics to detect brain-related psychopathology.

Background

Nicotine and illicit stimulants are very addictive substances. Although associations between grey matter and dependence on stimulants have been frequently reported, white matter correlates have received less attention.

Methods

Eleven international sites ascribed to the ENIGMA-Addiction consortium contributed data from individuals with dependence on cocaine (n = 147), methamphetamine (n = 132) and nicotine (n = 189), as well as non-dependent controls (n = 333). We compared the fractional anisotropy (FA), axial diffusivity (AD), radial diffusivity (RD) and mean diffusivity (MD) of 20 bilateral tracts. Also, we compared the performance of various machine learning algorithms in deriving brain-based classifications on stimulant dependence.

Results

The cocaine and methamphetamine groups had lower regional FA and higher RD in several association, commissural, and projection white matter tracts. The methamphetamine dependent group additionally showed lower regional AD. The nicotine group had lower FA and higher RD limited to the anterior limb of the internal capsule. The best performing machine learning algorithm was the support vector machine (SVM). The SVM successfully classified individuals with dependence on cocaine (AUC = 0.70, p < 0.001) and methamphetamine (AUC = 0.71, p < 0.001) relative to non-dependent controls. Classifications related to nicotine dependence proved modest (AUC = 0.62, p = 0.014).

Conclusions

Stimulant dependence was related to FA disturbances within tracts consistent with a role in addiction. The multivariate pattern of white matter differences proved sufficient to identify individuals with stimulant dependence, particularly for cocaine and methamphetamine.

To identify neuroimaging biomarkers of alcohol dependence (AD) from structural magnetic resonance imaging, it may be useful to develop classification models that are explicitly generalizable to unseen sites and populations. This problem was explored in a mega-analysis of previously published datasets from 2,034 AD and comparison participants spanning 27 sites curated by the ENIGMA Addiction Working Group. Data were grouped into a training set used for internal validation including 1,652 participants (692 AD, 24 sites), and a test set used for external validation with 382 participants (146 AD, 3 sites). An exploratory data analysis was first conducted, followed by an evolutionary search based feature selection to site generalizable and high performing subsets of brain measurements. Exploratory data analysis revealed that inclusion of case- and control-only sites led to the inadvertent learning of site-effects. Cross validation methods that do not properly account for site can drastically overestimate results. Evolutionary-based feature selection leveraging leave-one-site-out cross-validation, to combat unintentional learning, identified cortical thickness in the left superior frontal gyrus and right lateral orbitofrontal cortex, cortical surface area in the right transverse temporal gyrus, and left putamen volume as final features. Ridge regression restricted to these features yielded a test-set area under the receiver operating characteristic curve of 0.768. These findings evaluate strategies for handling multi-site data with varied underlying class distributions and identify potential biomarkers for individuals with current AD.