Brain functional connectivity correlates of autism diagnosis and familial liability in 24-month-olds

Abstract

Background: fcMRI correlates of autism spectrum disorder (ASD) diagnosis and familial liability were studied in 24-month-olds at high (older affected sibling) and low familial likelihood for ASD.

Methods: fcMRI comparisons of high-familial-likelihood (HL) ASD-positive (HLP, N = 23) and ASD-negative (HLN, N = 91), and low-likelihood ASD-negative (LLN, N = 27) 24-month-olds from the Infant Brain Imaging Study (IBIS) Network were conducted, employing object oriented data analysis (OODA), support vector machine (SVM) classification, and network-level fcMRI enrichment analyses.

Results: OODA (alpha = 0.0167, 3 comparisons) revealed differences in HLP and LLN fcMRI matrices (p = 0.012), but none for HLP versus HLN (p = 0.047) nor HLN versus LLN (p = 0.225). SVM distinguished HLP from HLN (accuracy = 99%, PPV = 96%, NPV = 100%), based on connectivity involving many networks. SVM accurately classified (non-training) LLN subjects with 100% accuracy. Enrichment analyses identified a cross-group fcMRI difference in the posterior cingulate default mode network 1 (pcDMN1)- temporal default mode network (tDMN) pair (p = 0.0070). Functional connectivity for implicated connections in these networks was consistently lower in HLP and HLN than in LLN (p = 0.0461 and 0.0004). HLP did not differ from HLN (p = 0.2254). Secondary testing showed HL children with low ASD behaviors still differed from LLN (p = 0.0036).

Conclusions: 24-month-old high-familial-likelihood infants show reduced intra-DMN connectivity, a potential neural finding related to familial liability, while widely distributed functional connections correlate with ASD diagnosis.

Keywords: Default mode network; Familial; Functional connectivity; Infant; MRI.

Fig. 1

Functional network architecture in 24-month-old infants. A Average correlations for 230 spherical regions of interest (ROIs). The ROIs are sorted by network. B ROIs brain map (dorsal, lateral, and medial views). Colors correspond to network assignments. C Cerebellar ROIs on a flattened cerebellar surface [62]. aDMN1: anterior default mode network 1; aDMN2: anterior default mode network 2; aFP: anterior frontoparietal network; Aud: auditory network; CO-SubC: cingulo-opercular-subcortical network; DAN: dorsal attention network; MotM: motor-mouth network; pcDMN1: posterior cingulate default mode network 1; pcDMN2: posterior cingulate default mode network 2; pFP: posterior frontoparietal network; Sal: salience network; SM1: somatomotor network 1; SM2: somatomotor network 2; SM3: somatomotor network 3; tDMN: temporal default mode network; US: unspecified (not assigned); Vis1 and Vis2: visual 1 and 2

Fig. 2

Multi-dimensional scaling (Euclidean distance) indicates (A) some separation (p = 0.012) between fcMRI correlation matrix cluster means (colored asterisks) for high-likelihood positive (HLP) and low-likelihood negative (LLN); but (B) not for HLP versus high-likelihood negative subjects (HLN; p = 0.047); nor for (C) HLN versus LLN subjects (p = 0.225). D Overall, the HL and LL groups did not differ significantly from one another (p = 0.112). Group label colors in figure inset

Fig. 3

Scores on the first two principal components, computed on the set of features retained by the crystallized support vector machine classifier. The classifier, trained on high-likelihood positive and high-likelihood negative subjects only, was highly accurate (overall = 99.1%, males = 98.5%, females = 100%). LLN subjects were excluded from the training and were all correctly classified as unaffected. There is clear separation of HLP and HLN/LLN subjects, with only one incorrect positive classification (circled). Negative predictive values: overall = 100%, male = 100%, female = 100%. Positive predictive values: overall = 95.6%, male = 94.4%, female = 100%

Fig. 4

Connections important for accurate support vector machine (SVM) classification. The brain-wide support vector classifier was trained with high-likelihood positive and high-likelihood negative subjects. Consensus features (across leave-out folds) important for accurate classification are depicted on an adjacency matrix, and the 100% consensus features are shown on lateral, posterior, and dorsal views of the brain. Colors indicate network membership (see Fig. 1)

Fig. 5

Enrichment analyses and secondary validation. A 3-group (HLP, HLN, LLN) ANOVA screen for functional connection differences suggests a clustering of nominally significant group differences for ROI pairs in the posterior cingulate default mode network 1 (pcDMN1)– temporal default mode network (tDMN) network pair (p = 0.007). A Lower triangle: ANOVA F-statistics. Upper triangle: 5% threshold applied to the F-statistics. The 230 ROIs are sorted by network (defined in Fig. 1). Black border: pcDMN1- tDMN. B Functional connections in the pcDMN1-tDMN network pair were used to predict ADOS CSS (five-fold cross-validated linear regression) with a Mean Squared Error of 7.01 (empirical p = 0.0396, on 25,000 permutations). C Locations of implicated connections within pcDMN1 - tDMN visualized on posterior, dorsal, and lateral views of the brain

Fig. 6

Enrichment analyses for pairwise Tukey-corrected t-tests characterize the pcDMN1-tDMN network pairs when comparing (A) the smaller LLN (n = 27) and HLP (n = 23) groups (p = 0.0461) and (B) the LLN and the larger HLN (n = 91) groups (p = 0.0004). Locations of implicated connections within the enriched network pair are visualized on dorsal, posterior, and lateral views of the brain. C, D The group mean functional connectivity profiles are remarkably consistent for pcDMN1 - tDMN ROI-pairs that show group differences: LLN > HLP (in C) and LLN > HLN (in D). The ROI-pairs are ordered by increasing LLN mean functional connectivity for clarity

Fig. 7

Enrichment analyses with student t-tests for the pcDMN1-tDMN network pair. A High-likelihood (HL) subjects compared with LLN (p = 0.0036). All subjects had a typical ADOS CSS = 1. Locations of implicated connections within the enriched network pair are visualized on dorsal, posterior, and lateral views of the brain. B Group mean functional connectivity profiles for pcDMN1 - tDMN ROI-pairs that show group differences: LLN > HL. The ROI-pairs are ordered by increasing LLN mean connectivity for clarity