Variegation of autism related traits across seven neurogenetic disorders

Abstract

Gene dosage disorders (GDDs) constitute a major class of genetic risks for psychopathology, but there is considerable debate regarding the extent to which different GDDs induce different psychopathology profiles. The current research speaks to this debate by compiling and analyzing dimensional measures of several autism-related traits (ARTs) across seven diverse GDDs. The sample included 350 individuals with one of 7 GDDs, as well as reference idiopathic autism spectrum disorder (ASD; n = 74) and typically developing control (TD; n = 171) groups. The GDDs were: Down, Williams-Beuren, and Smith-Magenis (DS, WS, SMS) syndromes, and varying sex chromosome aneuploidies ("plusX", "plusXX", "plusY", "plusXY"). The Social Responsiveness Scale (SRS-2) was used to measure ARTs at different levels of granularity-item, subscale, and total. General linear models were used to examine ART profiles in GDDs, and machine learning was used to predict genotype from SRS-2 subscales and items. These analyses were completed with and without covariation for cognitive impairment. Twelve of all possible 21 pairwise GDD group contrasts showed significantly different ART profiles (7/21 when co-varying for IQ, all Bonferroni-corrected). Prominent GDD-ART associations in post hoc analyses included relatively preserved social motivation in WS and relatively low levels of repetitive behaviors in plusX. Machine learning revealed that GDD group could be predicted with plausible accuracy (~60-80%) even after controlling for IQ. GDD effects on ARTs are influenced by GDD subtype and ART dimension. This observation has consequences for mechanistic, clinical, and translational aspects of psychiatric neurogenetics.

Trial registration: ClinicalTrials.gov NCT00001246 NCT00013559.

Conflict of interest statement

The authors declare no competing interests.

© 2022. This is a U.S. government work and not under copyright protection in the U.S.; foreign copyright protection may apply.

Figures

Fig. 1. Synopsis of results from pairwise mixed model ANOVAs (above diagonal) and ANCOVAs with cognitive impairment covaried (below diagonal).

To test for differential ART subscale profiles between each unique pair of GDD groups, we ran 21 (number of unique GDD group pairings) 2 (GDD group) × 5 (SRS-2 subscale) mixed-model ANOVAs. For these analyses, group effects (e.g., GDD 1 is more or less impaired overall than GDD 2 on the SRS-2 subscales) and group*subscale interaction effects (e.g., there is a difference in the SRS-2 profile for GDD 1 vs. GDD 2) were evaluated and results are presented above the diagonal. Parallel analyses were also completed using ANCOVA including cognitive impairment as a covariate and results are presented below the diagonal. When interpreting the figure, note the following. Main effects of group were denoted with a ‘G’; group*scale interactions were denoted with an ‘I’. Instances in which there was a main effect of group or group*scale interaction that did not survive Bonferroni correction are denoted with a single asterisk (*p < 0.05); those that survived Bonferroni correction are denoted with a double asterisk (**p < 0.05—Bonferroni corrected). Finally, to aid interpretation, color coding was implemented as follows. When a main effect of group was identified that survived Bonferroni correction, the cell in the matrix was color coded blue. When there was a Bonferroni-corrected group*subscale interaction, the cell was color coded yellow. Instances in which there was both a main effect of group (magnitude of impairment) and a group* subscale interaction (SRS-2 profile difference) following Bonferroni correction were color coded green.

Fig. 2. Gene dosage disorder (GDD) scores for different autism-related traits (ARTs).

Top panel: Point-line graph showing score profiles for each GDD across ARTs. Color encodes group. GDDs are in solid lines, and the benchmark autism spectrum disorder (ASD) and healthy volunteer (HV) groups are in dashed lines. Middle panel: Boxplots for each ART showing GDD group score distributions. Bottom panel: Heatmaps for each ART showing Cohen’s d effect sizes for all pairwise GDD group comparisons (column group vs. row group). Asterisks denote statistically significant comparisons (*nominal p < 0.05, **surviving Bonferroni correction for multiple comparisons).

Fig. 3. Autism-related trait (ART) scores for different gene dosage disorders (GDDs).

Top panel: Point-line graph showing score profiles for each ART across GDDs. Colored solid lines encode ARTs. Cognitive impairment scores are shown as a reference (dashed gray). Middle panel: Boxplots for each GDD showing score distributions for each ART. Bottom panel: Heatmaps for each GDD showing Cohen’s d effect sizes for all pairwise ART comparisons (column group vs. row group). Asterisks denote statistically significant comparisons (*nominal p < 0.05, **surviving Bonferroni correction for multiple comparisons).

Fig. 4. Relationships between autism-related trait (ART) scores and cognitive impairment within each gene dosage disorder (GDD).

Scatterplots and linear fit lines showing the relationship between increasing cognitive impairment (x-axis: “IQ_as_Tscore”) and ART score value (y-axis) faceted by GDD (rows) and ART subscale (columns). Robust correlation coefficients are provided for each cell. Color encodes GDD. Dashed lines show population norm values (50, black) and 2 standard deviations above this norm (70, gray). Note that IQ is inverted and transformed to a distribution with mean = 50, sd = 10 to form “IQ_as_Tscore” (i.e., IQ_as_Tscore > 70 is equivalent to IQ < 70).

Fig. 5. Average prediction accuracy for each GDD group achieved by each of the four machine learning (ML) models.

All the models yield plausible (above chance) prediction accuracy. Without IQ, group LASSO with 65 items performs better than LASSO with 5 subscales for most of the GDD groups except WS where their prediction accuracies are similar. With IQ information, the comparison of ML model performance with 65 items vs. 5 subscales varies across the GDD groups. The error bar indicates 95% confidence interval of the 1000 bootstrap samples across the 5-fold cross-validation.

Fig. 6. Feature importance for the LASSO model with IQ included.

The coloring represents both the feature importance and the direction of feature effect. Positive values indicate an increased likelihood of a certain GDD group while negative values suggest a reduced probability of a certain GDD group. Higher absolute values (i.e., feature importance values) indicate greater consistency of a feature being selected as an important predictor to improve prediction accuracy across the 5-fold cross-validation. Awareness is consistently selected as an important predictor for SMS and PLUSX. Cognition is consistently selected for predicting SMS and WS. Motivation is consistently selected for predicting WS, PLUSX, and PLUSY. RIRB is consistently selected for predicting SMS and DS.