4.1 Participants.
Participants were recruited from healthcare clinics in Los Angeles, through advertising in the local community and social media, and by word-of-mouth. Inclusion criteria for all participants included: (a) aged 8-17 years old; (b) IQ of at least 75 on either Full-Scale Intelligence Quotient (FSIQ), or Verbal Comprehension Index (VCI) of the Wechsler Abbreviated Scale of Intelligence 2nd edition (WASI-II)56 c) right-handed as assessed by a questionnaire adapted from Crovitz and Zener.57 Exclusion criteria for all participants included: (a) history of head injury with loss of consciousness greater than 5 min; (b) not sufficiently fluent in English or parent who did not have English proficiency (as not all assessments have been validated in other languages); (c) born before 36 weeks of gestation; (d) contraindications to participating in MRI; (e) on probiotics/prebiotics for the past two weeks; and (f) on antibiotics in the past month.
Additional inclusion criteria for the NT group were: (a) no first-degree relatives diagnosed with ASD; (b) a t- score<65 on the Conners-3AI parent58 indicating no attention deficit hyperactivity disorder; (c) a t-score<60 on the Social Responsiveness Scale 2nd edition (SRS-2)59 indicating low likelihood of ASD; and (d) no psychological or neurological disorders. For the ASD group, the Autism Diagnostic Observation Schedule (ADOS-2)60 and the Autism Diagnostic Interview-Revised (ADI-R)61 were administered by a research-certified assessor to confirm ASD diagnosis. Of the ASD participants, 9 were taking antidepressants, 1 was taking anticonvulsants, 11 were on stimulants, and 3 were taking antipsychotics at the time of participation (metabolites were adjusted for medication usage, see Methods section 4.4). No NT participants were taking medication at time of participation. All participants were instructed to abstain from antibiotic usage for 30 days and probiotics for 14 days prior to participation. All participants and parents were evaluated for their capacity to give informed consent and provided their written child assent and parental consent in accordance with the study protocols approved by the University's Institutional Review Board.
4.2 Behavioral Measures. The study took place over two days. On the first day, behavioral measures and assessments were completed, and fMRI task training and desensitization took place. On the second day, participants brought their stool sample and participated in the fMRI sessions.
In addition to the screening measures, parents completed the Sensory Experiences Questionnaire (SEQ-3)62 to assess sensory processing and the Screen for Child Anxiety Related Emotional Disorders (SCARED-P)63 to measure anxiety symptoms. Participants completed the Disgust Propensity and Sensitivity Scale- Revised (DPSS-R)64 to assess frequency of disgusting experiences and the emotional impact of disgusting stimuli, the Alexithymia Questionnaire for Children to measure alexithymia (AQC)65, and the Body Perception Questionnaire-Short Form (BQP-SF)66 to measure interoception.
We also collected data on variables that could impact the gut microbiome. The following variables were collected from the parent or the child: birth delivery method, prenatal antibiotic usage, antibiotic and probiotic usage during infancy, gastrointestinal symptoms (GSRS), stool consistency (Bristol Stool Form Scale)67, sleep (Adolescent Sleep Wake Scale [ASWS]68; and Family Inventory of Sleep Habits [FISH]69, and current medication usage. For diet, parents were asked to select a diet that best reflects what their child consumes on a regular basis. We then grouped the diets into the American (high consumption of whole grains, some processed foods such as frozen and packaged foods as well as whole grain pasta and breads, limited quantities of poultry, fish, eggs and dairy, and vegetables and fruits are consumed in moderate to large quantities) or other (Mediterranean, Paleo, Vegetarian, Gluten-Free, Dairy-Free, low FODMAP, or other).
4.3 Stool Sample Collection
After completion of their day 1 visit, participants were given a stool collection kit (specimen cup, wooden spatula, plastic bag, Fisherbrand Scientific Commode Specimen Collection System, gloves, ice packs, and an insulated transportation container). They were instructed to collect a stool sample within 72 hours of their MRI, freeze the sample at home, and transport it in the insulated transportation container with ice packs to the lab. Once in the lab, the sample was placed in a -80° C degree freezer for storage (first at USC, and then at UCLA where they were aliquoted and stored in a -80° C degree freezer). Aliquoted samples were shipped on dry ice with a stool collection log to Metabolon for further processing and analysis on their global metabolomics and bioinformatics platform (Metabolon, 617 Davis Drive, Durham, NC).
4.4. Preprocessing of Metabolomic data
Peak area values were log transformed and KNN imputation was applied for missing data (Do et al., 2018). Next, data was Z score normalized and adjusted for use of antidepressants, vitamins, supplements, laxatives, antihistamines, stimulants, cognition enhancers, and antipsychotics. Specifically we regressed out significant medication/supplement effects identified using a backward selection approach (function “MASS::stepAIC” in R using the BIC, i.e., log(n) degrees of freedom). The adjusted features were then used in downstream analyses. A priori metabolite targets of interest included 26 named metabolites in the tryptophan pathway (see Figure 1).
4.5 Brain Imaging
4.5.1 Scanning Parameters:MRI data were acquired on a 3 Tesla MAGNETOM Prisma (Siemens, Erlangen, Germany) with a 20-channel head coil. A 5-min structural T1-weighted MPRAGE was acquired for each participant (TR = 1950 ms, TE = 3.09 ms, flip angle = 10°, 256 × 256 matrix, 176 sagittal slices, 1 mm isotropic resolution). Each functional scan consisted of an echo-planar imaging (EPI; 150 whole-brain volumes) acquired with the following parameters: TR = 2 s, TE = 30 ms, flip angle = 90°, 64 × 64 matrix, in-plane resolution 2.5 × 2.5 mm, and 41 transverse slices, each 2.5 mm thick, covering the whole-brain with a multiband factor of three. Spin Echo EPI field mapping data was also acquired in AP and PA directions with identical geometry to the EPI data for EPI off-resonance distortion correction (TR = 1,020 ms, TE1 = 10 ms, TE2 = 12.46 ms, flip angle = 90°, FOV = 224 × 224 × 191 mm3, voxel size = 2.5 mm isotropic).
4.5.2 Scanning Procedure: All participants completed a practice MRI session in a mock MRI scanner prior to the fMRI tasks to become familiarized with the task and the MRI environment and to increase comfortability and minimize head motion. Functional MRI procedure, task stimuli, fMRI acquisition, and data preprocessing were completed following the protocol previously published in Kilroy et al., 2021.38 We utilized a head-motion cut-off of absolute FD>1.5 mm. Five participants (4 ASD, 1 NT) were excluded for head motion in the watching facial expressions and hand actions task, 3 (2 ASD, 1 NT) for disgust processing, and 2 (1 ASD, 1 NT) from watching others being touched. There were no significant differences in relative head motion between the two groups for the disgust processing (t=0.981, p=0.33) and watching others being touched (t=-1.029, p=0.307) tasks, but significant differences were present for the observation of facial expressions and hand actions task (t=-2.572, p=0.015). Please see fMRI processing in Section 4.6.2 for details regarding motion correction.
4.5.3 fMRI Tasks: The task-based fMRI paradigms were selected based on existing literature showing significant ASD vs. NT differences during these tasks, their relevance to key ASD symptomatology (socio-emotional processing and sensory sensitivities38,49,70,71, and/or their relevance to vagally mediated emotional processes (disgust and emotion processing).42 The fMRI tasks included: watching videos of facial expressions/body actions, physical and social disgust processing tasks, and watching videos of others being touched (Supplemental Figure 1). Stimuli were presented using the Psychophysics Toolbox72 on MATLAB. During all tasks, participants were instructed to simply watch all videos and remain as still as possible. fMRI tasks are described in A-C below.
A. Watching videos of facial expressions and hand actions (n=78; 38 NT [19 female, 19 male], 40 ASD [11 female, 29 male]):One 9-minute run with five 15-sec blocks of video-stimuli were shown. As Supplemental Figure 1A shows, blocks consisted of one of three categories of stimuli: emotional facial expressions (e.g., happy expression), non-emotional facial expressions (e.g., tongue to lip), or bimanual hand actions (e.g., playing the xylophone). Each video was presented for 3.75 sec followed by a 1.25 sec black screen between each stimulus, there were 3 videos per block, and both male and female actors were included in each block. For further details on stimuli, please see Kilroy et al., 2021.38
B. Disgust Processing (n=46; 22 NT [12 female, 10 male], 24 ASD [6 female, 18 male]): There were four categories of stimuli, disgusting foods, disgusted facial expressions, neutral foods, and neutral facial expressions (Supplemental Figure 1B). The neutral and the disgusted facial expressions were chosen from an online repository (NimStim)73 and from previous research (EmStim)38 then edited and counterbalanced so that each participant saw the same actor displaying a neutral and disgusted facial expression. To ensure that the neutral food images were indeed items the participant truly had no preferential or disgusting feelings for, all participants were administered a questionnaire prior to participating in the study, to assess their preferences for each food stimuli. For each participant, 18 images were used from each stimulus category. One fMRI run was presented to all participants, including six blocks per stimulus category. Within each 15-sec block, three different images from the same category were presented with a 250-ms fixation crosshair between each stimulus (e.g., three different disgusting food images). Thus, the fMRI task consisted of 24 blocks (5 per stimulus category), lasting for a single 10-min run. For further details on stimuli, please see Jayashankar et al., in review.42
C. Watching others being touched (n=37; 19 TD [11 female, 8 male], 18 ASD [5 female, 13 male]): Participants watched four different videos where a person strokes the arm of another person in the MRI scanner with: 1) their hand with glove on (social touch), 2) a dry sponge (object touch), 3) their hand with glove on hovering next to the person’s arm (social touch control), and 4) a dry sponge hovering next to the person’s arm (object touch control) following asimilar to the protocol used in Green et al., 2015 (Supplemental Figure 1C).74 Each video was 15 sec long followed by a 15 sec rest block. During the rest blocks, participants were shown a black crosshair in the middle of a white screen. Excluding an initial junk block, five blocks of each stimulus condition were alternated with rest in a pseudo-random sequence. Stimuli conditions were counterbalanced across participants. We note that our original intention was to physically touch participants while in the scanner, but as this run was largely conducted during the initial period of the COVID-19 pandemic, we were obliged to remain at a 6-ft distance from our participants, and thus used videos of touch instead, as this has previously been shown to show strong effects.75,76
4.6 Analysis
4.6.1 Behavioral Group Differences. Independent samples t-tests and Fisher’s exact tests were conducted to determine ASD-NT differences in demographic and behavioral variables. Significance was set to p<.05.
4.6.2 fMRI
fMRI Preprocessing:All analyses followed best practices in fMRI analysis, as detailed in our prior studies.38 The data analytic approach used to address each of our research questions utilized FMRIB’s Software Library 6.0 (FSL).77-81 Standard preprocessing pipeline was performed involving: (a) structural T1 brain extraction and non-brain tissue removal; (b) smoothing with 5 mm FWHM Gaussian kernel; (c) B0 unwarping along y-axis; (d) high pass filter with 100 sec cutoff; (e) realignment using MCFLIRT to obtain motion estimates; and (f) Independent component analysis (ICA). Preprocessed data was fed into the ICA AROMA algorithm82, which filtered out noise and motion components from the whole brain signal. Registration to the MNI-152 standard atlas using 12 degrees-of-freedom affine transformation and FNIRT nonlinear registration were performed.78,79
Within-group analyses: Individual participants' statistical images were subjected to higher-level mixed-effects analyses using FSL's FLAME Stage 1 algorithm, modeling the stimulus conditions for each participant as separate regressors. For watching facial expressions and hand actions, regressors included: emotional faces, non-emotional faces, and bimanual hand actions. For disgust processing, regressors included: disgusting foods, neutral foods, disgusted facial expressions, neutral facial expressions. Subject-specific head motion parameters were used as nuisance regressors. For observation of others being touched, regressors included social touch and object touch.
Between-group Analysis:Between-group comparisons between the NT and ASD groups were performed using higher level mixed-effects analyses with FSL's FLAME 1 algorithm. Age, Sex, and IQ were used as covariates in all group-level analyses. For watching facial expressions and hand actions, groups were contrasted on: all stimuli>rest; emotional facial expressions>rest; non-emotional facial expressions>rest; all facial expressions>rest; hand actions>rest. For disgust processing, groups were contrasted on: disgusting foods>rest, disgusted faces>rest. For observation of others being touched, groups were contrasted on: social touch>rest and object touch>rest. For the facial expressions/hand action task and the disgust tasks, the resulting group-level images for all models were thresholded at voxel Z>3.1, with a cluster size probability correction threshold of p<.05. For observation of others being touched, a more lenient threshold (Z > 2.3 cluster size probability threshold of p<.05) was used to have more sensitivity to detect effects given the more subtle observation (rather than physical touch) task used, due to COVID-19 restrictions (see Methods 4.5.3.C). In addition, for disgust and facial expression/hand action tasks, for hypothesized regions of interest (ROIs), a small volume correction (SVC) analysis with a significance threshold of p<0.01 using predefined masks for disgust and observation tasks. For the facial expression/hand action task, we used structurally defined anterior insula parcellations from extant literature and the Harvard-Oxford atlas parcellations for the pACC and amygdala. For the IFGop, we used a hand-drawn anatomically derived ROI38 and previously published insula parcellations.83 For the disgust task, ROIs for SVC analysis were defined utilizing the Neurosynth database (which performs automated large-scale meta-analyses of fMRI data), using the search terms - “disgust”, “emotional faces”, and “food”, and we also included insula parcellations from extant literature.83 Functional ROIs were then masked with structural ROIs from the Harvard-Oxford atlas (thresholded at 30% probability) to ensure they captured non-overlapping regions.
4.4 General Linear Models (GLMs). GLMs were applied within the ASD group and across groups to test brain-behavior, brain-metabolite, and metabolite-behavior relationships.The GLMs included group as a factor, and sex, age, IQ and BMI were included as covariates. As a measure of effect size, we report the standardized beta (Std β). Std β between 0.10–0.29 is considered small, 0.30–0.49 medium, and greater than 0.50 large.84 Brain ROIs were chosen based on group differences in fMRI tasks as well as prior studies supporting atypicalities in brain activity in the chosen ROIs.38,42,49,70,71 The Benjamini-Hochberg method to correct for multiple comparisons was used; the false discovery reporting threshold set at 10% (FDR).85 We used FDR correction for the number of dependent variables in each analyses. Specifically for brain-metabolite and brain-behavior analyses, FDR correction was for the number of ROIs compared. For metabolite-behavior analyses, FDR correction was for the number of metabolites compared. To limit the number of comparisons, only metabolites that significantly correlated with brain activity were included in the metabolite-behavior analyses.
4.5 Mediation Models
Exploratory mediation analyses were conducted to determine if the brain was a mediator of the relationship between metabolite and behavior in the ASD group. The variables included in the mediation models were selected based on metabolites and behaviors that had significant associations with the same task-based ROIs. Mediation modeling was performed using lavaan in R. We estimated the bootstrapped 95% percentile confidence intervals for the indirect effects.86 Confidence intervals that do not contain zero are considered significant. Because age, sex, and BMI are collinear, we ran analyses using only age and FSIQ as covariates. Additionally, in terms of regressors for brain and behavior (as opposed to metabolites), it is less relevant to control for BMI.