Towards robust and replicable sex differences in the intrinsic brain function of autism

For discovery analyses, we examined the R-fMRI dataset with one of the largest number of females and males in both the autism and the NT groups available to date, selected from the Autism Brain Imaging Data Exchange (ABIDE) repositories ABIDE I and II [22, 23]. The final ABIDE I and II dataset of N = 1019 included N = 82 females with autism, N = 362 males with autism, N = 166 neurotypical females (NT F), and N = 409 neurotypical males (NT M), aggregated across 13 sites. Specific selection criteria are described in Supplementary Methods in the Additional file 1 and depicted as a figure in the Additional file 1: Figure S1. Briefly, we selected cases between 7 and 18 years of age (the ages most represented across ABIDE sites), with MRI data successfully completing brain image co-registration and transformation to standard space, with FIQ between 70 and 148 and with mean framewise displacement (mFD) [54] within three times the interquartile range (IQR) + the third quartile (Q3) of the sample (i.e., mFD 0.39 mm). Further steps included matching for mean age across groups, as well as for mFD and IQ within diagnostic groups. This latter step limited the number of exclusions while keeping average group motion low (mFD < 0.2 mm) and sampling biases that may result when matching neurodevelopmental conditions to NT around intrinsic features such as IQ [55, 56]. At each step, any sites with less than three individual datasets per diagnostic/sex groups were excluded. Demographics and characteristics of this sample are summarized in Table 1 and in Supplementary Methods in the Additional file 1.

We examined five whole-brain R-fMRI metrics previously reported to reflect typical sex differences and found to be atypical in autism, including (1) PCC-iFC, (2) VMHC, (3) ReHo, (4) DC and (5) fALFF (see Additional file 1: Supplementary Methods). R-fMRI image pre-processing steps included: slice time correction, 24 motion parameters regression [57], component-based noise reduction (CompCor) [58], removal of linear and quadratic trends, and band-pass filtering (0.01–0.1 Hz, for all metrics but fALFF). Functional-to-anatomical co-registration was achieved by Boundary Based Registration (BBR) using FSL FLIRT [59]. Linear and nonlinear spatial normalization of functional echo planar images (EPIs) to Montreal Neurological Institute 152 (MNI152) stereotactic space (2 mm³ isotropic) was done using ANTS registration (Advanced Neuroimaging Tools) [60]. Computation of voxel-mirrored homotopic connectivity (VMHC) followed registration to a symmetric template. All R-fMRI derivatives were smoothed by a 6 mm FWHM Gaussian kernel. To account for site and collection time variability across each of the data collections in ABIDE I and II data repositories, site effects were removed using the ComBat function available in python [61] (https://​github.​com/​brentp/​combat.​py). This approach has been shown to effectively account for scanner-related variance in multi-site R-fMRI data [61]. For further details see Additional file 1: Supplementary Methods.

Statistical Z-maps were generated within study-specific functional volume masks including all voxels in MNI space present across all subjects. Main effects of diagnosis and sex along with their interaction were explored by fitting a general linear model (GLM) including diagnosis or/and sex as the regressors of interest respectively, and age and mean framewise displacement (mFD) [54] as nuisance covariates. In primary analyses, we did not include FIQ as a covariate as this is thought to be suboptimal when comparing groups selected from populations carrying intrinsic IQ differences such as autism and NT [56]. Nevertheless, to provide an indication as to whether IQ may affect primary findings, in supplementary analyses FIQ was also included as an additional nuisance regressor. We applied gaussian random field theory correction based on strict voxel-level threshold of Z > 3.1 as recommended by [62] and cluster level P < 0.01, given the assessment of five R-fMRI metrics in the same study (i.e., P 0.05/5 R-MRI metrics = 0.01).

Post-hoc analyses were conducted to functionally characterize the sex-by-diagnosis interaction result(s). First, to explore the cognitive domains implicated in the cluster(s), we quantified the percentage of its overlap with 12 cognitive ontology maps [63] thresholded at P = 1e−5. We labelled these components based on the top five tasks each component recruits [2]. Second, we used the Neurosynth Image Decoder (http://​neurosynth.​org/​decode/​) [64] to visualize the terms most strongly associated with the significant cluster. After excluding anatomical (e.g., occipital) and redundant terms (synonyms [e.g., saccades and eye movements], plurals [e.g., object and objects] or noun/adjective/adverb equivalents [e.g., vision and visual]), we visualized the top 27 terms showing correlations with the cluster map between r = 0.64 and r = 0.10. Third, to explore potential clinical relevance of the significant cluster, we explored brain-behavior relationships as a function of sex within the autism group. Specifically, we ran a GLM examining the interaction between biological sex and available ADOS calibrated severity total score (CSS) [65], as well as social-affect (SA) and restricted, repetitive behavior (RRB) subscores (see Additional file 1: Supplementary Methods) with the dependent variable(s) being the R-fMRI metric(s) extracted from the cluster mask(s) showing a statistically significant sex-by-diagnosis effect(s).

Post-hoc analyses to functionally characterize this VMHC sex-by-diagnosis interaction indicated that the VMHC cluster in superior lateral occipital cortex overlapped with cognitive maps involved in higher-order visual, oculomotor, cognitive flexibility and language-related processes (Fig. 3a). Further, as shown in Fig. 3b, the most common terms were primarily related to lower-order visual processing and higher-order visual cognition, such as ‘visuospatial’ and ‘spatial attention.’ To explore potential clinical relevance of the VMHC dorsolateral occipital cluster, we explored brain-behavior relationships as a function of sex, within the autism group using three available ADOS scores (calibrated severity total score, and non-calibrated social affect and RRB subscores; see Additional file 1: Supplementary Methods). Although not surviving a strict Bonferroni correction for multiple testing (i.e., 0.05/3 = 0.02), an interaction effect was observed for ADOS social affect scores. It revealed that more severe social affect deficits (F_(1,311) = 4.44, p = 0.036) were associated with decreased VMHC in females with autism (r = − 0.29), but not in males with autism (r = 0.03). Given that ABIDE data were aggregated and released when calibrated social affect scores [69] were not available to assess potential differences in language abilities and age, analyses were repeated after including ADOS module (ADOS Module 2 to 4) as a nuisance covariate: results remained unchanged (F_(1,306) = 5.0, p = 0.026) as they did also after removing the few data with the less represented ADOS module 2 (see Additional file 1: Supplementary Methods). There were no significant findings, with regard to the CSS total score and non-calibrated RRB sub-score (Fig. 3c), even at an exploratory statistical threshold of P < 0.05.

The same pattern of results identified in discovery analyses was observed when the datasets were pre-processed using GSR or ICA-AROMA, across the three R-fMRI metrics in all the clusters identified in the primary analyses across main effects of diagnosis, sex and their interaction; effect size ranges from small to moderate as in discovery analyses (η_p² range = 0.01–0.07; Fig. 2b, Fig. 4, Additional file 1: Table S6 and Additional file 3: Figure S6).

VMHC reflects inter-hemispheric homotopic relations. The strength has been suggested to index coordinated cross-hemispheric processing: stronger VMHC indexes weaker hemispheric specialization and vice versa [33, 70]. Several lines of evidence support the notion that the neurobiology of autism is related to atypical hemispheric interactions, including homotopic connectivity and hemispheric lateralization [35, 71‐80]. VMHC and functional hemispheric lateralization have also been shown to be sex-differential in NT [33, 81, 82]. The dorsolateral occipital association cortex identified in our discovery analyses is known to serve hemispherically specialized processes, such as visuospatial coordination [83]. Thus, our findings of NT males’ VMHC in dorsolateral occipital cortex being lower than that of NT females are consistent with the notion of increased hemispheric lateralization in this cortical region in NT males relative to NT females. In our data, females with autism instead showed even lower VMHC than NT males, while males with autism showed slightly higher VMHC than NT males. This pattern is indicative of ‘gender-incoherence’ [20] as males and females with autism display the opposite pattern expected in NT per their biological sex. Findings of ‘gender incoherence’ have been reported in earlier neuroimaging studies of autism using different modalities [3, 84, 85]. Among them, several R-fMRI studies explicitly focusing on detecting sex-by-diagnosis interactions (i.e., the regression model included a sex-by-diagnosis interaction term) [3, 11] yielded a pattern of results consistent with ‘gender incoherence.’ In contrast, other studies [12‐14] reported a pattern consistent with the ‘extreme male brain’ model [19]—i.e., a shift towards maleness in both females and males with autism. While the seemingly diverging conclusions of these two sets of studies may be attributed to methodological differences, such as the extent of brain networks explored and the statistical modelling employed, findings from our prior work suggest that both shifts towards either maleness or femaleness co-occur in the intrinsic brain of males with autism, in a network-specific manner [2]. However, such prior work did not include female data. Thus, by not directly assessing sex-by-diagnosis interactions, unlike the present study, results could not point to patterns affecting diagnostic differences between the sexes versus those that are common to autism across sexes [4]. This is relevant for efforts focusing on identifying underlying mechanisms. Findings resulting from sex-by-diagnosis interactions may shed light on sex-differential mechanisms that are atypical in autism and may reflect sex-specific susceptibility mechanisms. On the other hand, atypicalities common for both sexes may reflect factors central to the emergence of autism, regardless of whether they overlap with patterns known to be differential between sexes [86]. Interestingly, a recent study based on a sample selected from GENDAAR [16] revealed that the iFC between the nucleus accumbens (selected a priori) and a region of the dorsolateral occipital cortex partially overlapping with that identified by our VMHC analyses, was differentially modulated by the aggregate number of oxytocin receptor risk alleles in females with autism versus NT females and versus males with autism. Although VMHC was not directly tested in the said study [16], its result in dorsolateral occipital cortex is consistent with our observation of atypical sexual differentiation of this visual network region and, together, suggest the need for future whole-brain studies of oxytocin effects in autism.

Along with the sex-dependent autism patterns, our analyses found statistically significant main effects of diagnosis in inter-hemispheric interactions indexed by VMHC in distinct cortical circuits. These were localized along the midline of the DN (paracingulate/frontal cortex consistently and PCC/precuneus) where main effects of PCC-iFC and ReHo also converged. Our results are consistent with prior reports of atypical intrinsic organization of the DN in autism [12, 23, 26, 87‐89]. Together they support a common, sex-independent role of DN in autism. This is also highlighted by a recent autism neurosubtyping study that identified three latent iFC factors, all sharing DN atypicalities along with their neurosubtype-specific patterns [90]. Building on this evidence to disentangle the specific role of each of the factors affecting autism in sex-independent and sex-dependent ways, a necessary next step is to engage in novel large-scale data collection efforts including more female data.

The growing awareness of the replication crisis in neuroscience [91‐93] motivated our analyses examining robustness and replicability of findings. While a comprehensive and systematic reproducibility assessment is beyond the scope of the present study, here we focused on examining whether the findings observed in the discovery analyses were also seen after using different preprocessing pipelines—robustness—as well as in fully independent, albeit of convenience samples (i.e., not harmonized a priori with each other)—replicability. To this end, given the lack of consensus on quantitative metrics of replicability, we opted to use measures of effect size. These are considered complementary to null hypothesis significance testing [94]. In the context of this study, given the use of convenience samples of different sizes, their selection was considered an advantageous and practical means to provide information on the magnitude of group differences in diagnosis and sex, as well as their interaction. Here, we considered findings to be robust and/or replicable for any non-negligible effects (i.e., η_p² ≥ 0.01 [95]). We reasoned that given their distributed and heterogenous nature [6], atypicalities in the autism connectome can stem from a combination of differently sized non-negligible effects, as shown for autism in other biological domains such as genetics [96, 97].

With this in mind, across the two preprocessing methods examined here, the patterns of findings were consistent with those observed in discovery analyses across all R-fMRI metrics and effects. These robust results are consistent with a prior study by He and colleagues [46] reporting that differences in a wider range of pre-processing pipelines have marginal effects on variation in diagnostic group average comparisons. Our study confirms and builds on this earlier report by extending findings of robustness to sex group mean differences and their interactions with diagnosis.

A more nuanced picture emerged from the inter-sample analyses as replicability varied by sample, across the effects and R-fMRI metrics examined. Specifically, while inter-sample main effects of sex were moderately to largely replicable across R-fMRI metrics on both independent samples (~ 50 to 80% of the clusters in GENDAAR and EU-AIMS-LEAP, respectively), replicability of diagnostic effects significantly varied by sample (86 to 29%) across R-fMRI metrics. This is at least in part consistent with findings by King et al. [50] who showed that, depending on the R-fMRI feature examined, diagnostic group differences varied across samples. Even in this scenario, King et al. [50] also reported that findings of decreased homotopic connectivity in autism were relatively more stable than other R-fMRI metrics. This observation, combined with the replicability of our VMHC sex-by-diagnosis interaction findings in the larger of the two independent samples (EU-AIMS LEAP), suggests that measures of homotopic connectivity may have specific biological relevance for autism. It is also possible that given the moderate to high test–retest reliability, VMHC is more suitable in efforts assessing replicability [98, 99].

The striking clinical and biological heterogeneity in autism should be considered as a major contributor to discrepancies in findings of studies focusing on the main effects of diagnostic group means contrasts/interactions [100‐103]. Against this background, we interpret our replicability findings on diagnostic effects and, in turn, diagnosis-by-sex interactions. Inter-sample differences may have contributed to the more variable results of replicability on the diagnosis main effects. These may include autism symptom level, age, and IQ, albeit secondary analyses suggested that the examined IQ range did not substantially affect the pattern of discovery results. For example, the EU-AIMS LEAP sample was on-average older, had lower VIQ and most notably, lower symptom severity across all subscales of the ADOS and ADI-R than the ABIDE sample. On the other hand, the GENDAAR sample (which has greater number of replicable diagnostic mean group patterns) did not differ from ABIDE in these variables, except for mean age. Furthermore, a fact that is often neglected, is that the NT groups may also present with considerable sample heterogeneity between studies [100, 104]. For instance, our NT controls in the EU-AIMS LEAP sample had lower VIQ than both ABIDE and GENDAAR NT controls. This has potentially influenced the low replicability of diagnosis main effects in EU-AIMS LEAP.

In contrast, sex-by-diagnosis interaction effect on VMHC in the dorsolateral occipital cortex was replicable in the larger sample, the EU-AIMS LEAP, but not in GENDAAR. Small samples introduce larger epistemic variability (i.e., greater variation related to known and unknown confounds) [105]. Increasing the number of subjects/data allows mitigating epistemic variability and, thus, capturing the underlying variability of interest. Thus, although the rate of replicability for the main effect of diagnosis in EU-AIMS LEAP was limited, accounting for biological sex, a known key source variability in autism, may have substantiated a replicable sex-by-diagnosis pattern in this larger sample. In line with sample size concerns, using four datasets sized between 36 and 44 individuals selected from the ABIDE repository, He et al. [46] found low similarity rates of diagnostic group-level differences on the strength of iFC edges in contrast with the largely similar pattern of results across pipelines. Of note, unlike prior efforts [46, 50], we controlled for site effects within each of the samples (i.e., ABIDE, GENDAAR and EU-AIMS LEAP), using ComBat. Future large-scale harmonized data collections are needed to control and assess the impact of inter-sample variability. Taken together, these findings highlight that sample differences can impact replicability.

Beyond clinical and biological sources of variation, samples may differ in MRI acquisition methods, as well as in approaches used to mitigate head motion during data collection and its impact on findings [106]. Adequately controlling for head motion remains a key challenge for future studies assessing inter-sample replicability. For the present study, we excluded individuals with high motion, retained relatively large samples with group average low motion (mean ± standard deviation of mFD range = 0.09–0.16 ± 0.06–0.10 mm), as well as included mFD at the second-level analyses as a nuisance covariate. Overall, the extent to which each sample-related factor affects replicability needs to be systematically examined in future well-powered studies. Only this type of studies will allow for emerging subtyping approaches to dissect heterogeneity by brain imaging features using a range of data-driven methods [107, 108], including normative modelling [72, 109, 110].

Inter-sample differences and methodological differences, beyond nuisance regression, may have contributed to some differences in findings between the present and earlier studies, conducted with independent or partially overlapping samples [11, 23, 41]. For example, Alaerts et al. [11] also examined sex-by-diagnosis interaction in PCC-iFC in a dataset selected from ABIDE I only. Although their pattern of results was consistent with the ‘gender incoherence’ model, the resulting circuit(s) did not involve the dorsolateral occipital cortex as identified with VMHC in the present study. Along with differences in samples selected from the same data repositories, other methodological choices may also affect results. For example, prior studies differed with the present one in the inclusion of sex-by-diagnosis interaction [17], the extent of the whole-brain voxel-based analyses [15, 16], or the statistical threshold utilized [23]. Nevertheless, it is remarkable that even in light of these differences, consistent results have emerged including the overarching atypical inter-hemispheric interactions in autism, and sex-dependent and sex-independent atypical intrinsic brain function across distinct functional networks.