Altered reward system reactivity for personalized circumscribed interests in autism

Sixty-seven youth, ages 8–17 years, were enrolled in this study, including 45 with ASD (without intellectual disability) and 22 TDC. Five participants with ASD did not attempt the scan, and one child with ASD was excluded for extreme hydrocephalus. All participants stayed within our head motion thresholds during the scan (i.e., root mean square < 1.75 mm of maximum displacement and .0175 rad of translation). The final imaging sample comprised 39 youth with ASD and 22 TDCs. All participants had a General Conceptual Ability (GCA) ≥ 75, equivalent to full-scale IQ, as measured by the Differential Ability Scales—Second Edition [19].

Youth with ASD received an expert clinical diagnosis based on Diagnostic and Statistical Manual of Mental Disorders–Fourth Edition–Text Revision criteria (DSM-IV-TR) [20]; the Autism Diagnostic Interview-Revised [21], and the Autism Diagnostic Observation Schedule [22] were used by experienced clinicians to inform diagnostic decisions. Youth with ASD were excluded if parents reported any known genetic, current mood or psychotic disorder, neurological disorder, premature birth (gestational age ≤ 37 weeks), or other significant medical conditions that affects brain functioning. Youth on atypical antipsychotics were excluded. Youth on psychostimulants were asked to withhold on the day of the study (n = 5), and youth taking other psychoactive medication were included (selective serotonin reuptake inhibitors: n = 10, selective norepinephrine reuptake inhibitors: n = 3, alpha2a-agonist: n = 3).

TDCs were excluded if parents reported any known genetic, language, learning, neurological, or psychiatric disorder, premature birth, or first- or second-degree relative with ASD. Youth were also excluded if parents reported elevated symptoms on any scale from the Child and Adolescent Symptom Inventory (CASI-4R; [23]. Groups did not differ in age and sex ratio, but the ASD group had a marginally lower GCA than TDCs (Table 1).

For this study, we developed the Interest Preference Assessment (IPA), a short interview adapted from the Interests Scale [24, 25], to directly evaluate each child’s most favorite interest/hobby to be used in the fMRI task (see Additional file 1). The IPA is comprised of two sections: The first section consists of a list of 25 categories of interest (e.g., machines or figuring out how things work, animals, people), and children are asked to rate each interest on a scale from 1 to 5, with 5 indicating “I could do this activity or talk about this topic all the time.” The second part of the IPA asks children to identify their primary interest and answer questions assessing the interference and intensity of that interest (e.g., “How much time do you spend doing or thinking about this interest or hobby? Does this get in the way of other responsibilities? Do you get annoyed or upset when you are asked to stop talking or doing this interest or hobby?”). Most children identified a singular primary interest, but a few children identified more than one interest as their “favorite.” In these cases, we asked them to pinpoint the interest they would most enjoy having in the upcoming MRI session. Examples of interests include videogames, professional sport teams, musicians/actors, toys, elevators, and movies (Additional file 1: Table S1).

Additionally, all parents were asked to complete the Social Responsiveness Scale-2nd Edition (SRS-2; [26]), the Repetitive Behavior Scale-Revised (RBS-R; [27, 28]), and the Interests Scale (IS; [24, 25]) to dimensionally assess behaviors characteristic of ASD.

Participants were compensated for their participation in this study. Written informed consent was obtained from all participants and their parents. This study was approved by the local Institutional Review Board.

We used an incentive delay task (IDT) in a blocked fMRI design that is an adaptation of the “classic” IDT [29] and aims to examine participants’ motivation to receive either a social or a personalized type of reward based on individual interests (vs. neutral outcome). To maximize ecological validity, we utilized short movie clips of actors expressing facial expressions along with other nonverbal gestures in the social reward condition as well as short video clips depicting personalized interests in the interest reward condition (Fig. 1 and Additional file 1 for details).

The IDT is a simple speeded button press task to examine neural responses to rewards that can be gained dependent upon a person’s ability to quickly and accurately respond to a target symbol in each trial. An online tracking algorithm was implemented to continuously monitor and adjust the target duration according to individual performance to achieve an accuracy rate of ≥ 50% (see Table 2). In social reward (SR) trials, target hits resulted in a short video clip of a person showing happy facial expressions, including nodding with a smile and a praising nonverbal gesture—the “thumbs up” sign. The outcome for misses was a video clip of a person showing neutral expressions, with slight natural motion (e.g., eye blinks). In personalized interest reward (IR) trials, target hits resulted in a short video clip of a person’s favorite interest (e.g., Minecraft©), whereas the outcome for misses was a video clip of a tree with slight natural movement (i.e., breeze gently moves leaves). The outcome in the neutral (NR) condition was always a video clip of “TV static.” The SR video clips portrayed six different adult actors (half female, half male, of various races and ethnicities representative of the local region) and six unique clips of personalized interests were presented in the IR condition; each of which was randomly repeated eight times throughout the experiment. Actor’s videos were chosen from a larger pool based on ratings of the actor’s likability as well as based on her authenticity of depicting approval; only actors who had the highest ratings on both scales were included (see [18]). Interest clips were chosen from various youtube© videos to depict the participants’ individual interests. All video clips fulfilled the following criteria: HD quality, multicolored content, disabled audio, .avi format, resolution 720 × 400 pixels, 25 frames/s, and 1500 ms duration; and the actor or interest/hobby is depicted in the center of the screen.

There were two 6 min 44 s runs that presented the three experimental conditions (i.e., SR, NR, and IR) block-wise, with a fixation period interspersed after each block. Altogether, 6 blocks per condition and 18 fixation periods were presented across the two runs. Within each run, trial types were designated using intuitive cues signaling the reward type that could be obtained in the ongoing trial for correct performance: a plus sign for SR trials, the numeral zero for NR trials, and a double-plus sign for IR trials. Each block consisted of a total of 7 trials across a total of 3 presentations per run (21 trials per run and 42 trials per condition across both runs). Each trial started with a condition cue for 250 ms, followed by a variable anticipation phase (three arrows were displayed in the middle of the screen; 1000–1669 ms) and the target appearance (white solid square; individually tailored duration of 100–590 ms based on the online tracking algorithm). Feedback was presented immediately after target disappearance for a duration of 1500 ms, followed by a jittered inter-trial interval (660–1340 ms). Jittering was optimized so that each trial lasted exactly two TRs (4.68 s). Block duration was 32.76 s. The fixation period between each block flashed a crosshair in each corner of the screen and then the center of the screen for the duration of a single TR (2.34 s × 5 crosshairs = 11.7 s).

To ensure that all participants fully understood the different cue-outcome relations, i.e., to avoid a learning component during the experiment, each participant received training right before the scan. This was followed by a performance-based test of their understanding as well as a practice session of the task. Only participants who fully understood the task instructions (i.e., remembered each cue-outcome relation in all three reward conditions for hit and miss trials) moved on to the practice session of the task, followed by the real fMRI scan. All participants passed the test with 100% accuracy.

All imaging data were collected using a Siemens Verio 3T scanner (Erlangen, Germany) and a 32 multichannel headcoil. Functional data consisted of two 6-min 44 s runs of whole-brain T2*-weighted BOLD echo planar images with 173 volumes acquired per run, including two “dummy” scans at run onset allowing for T1 magnetic saturation (40 oblique axial slices, isotropic voxel size = 3.5 mm, TR/TE = 2340 ms/25 ms, flip angle = 60°). Two high-resolution structural MR images were acquired for the registration of fMRI data to MNI space: A T1-weighted MPRAGE sequence of the entire brain (176 sagittal slices, isotropic voxel size = 1 mm, TR/TE = 1900 ms/2.54 ms, flip angle = 9°) and a FLASH sequence collected in the same plane as the fMRI data (number of slices = 40, slice thickness = 3.5 mm, TR/TE = 300 ms/2.46 ms, flip angle = 60°).

Functional image processing and statistical analyses were then carried out using FEAT (FMRIB’s Expert Analysis Tool), part of FMRIB’s Software Library (FSL) package [30, 31]. Prior to image analysis, the first two images of the functional data set were discarded because of the non-equilibrium state of magnetization. Each time series was despiked using AFNI’s 3ddespike program, motion-corrected, temporally filtered (nonlinear high-pass filter with a 1/90 HZ cutoff calculated with FSL’s cutoffcalc), and a 3D Gaussian filter (FWHM = 5 mm) was applied to account for individual differences in morphology and local variations in noise. Voxel-wise regression analyses were performed on each of the participant’s runs using FILM (FMRIB’s Improved Linear Model). Motion parameters (i.e., six parameters corresponding to three directions of translation and three axes of rotation) were entered as nuisance regressors (absolute mean displacement: ASD = 0.13 ± 0.05 mm, TDC = 0.14 ± 0.04 mm, p = 0.38; relative mean displacement: ASD = 0.09 ± 0.04 mm, TDC = 0.10 ± 0.04 mm, p = 0.79). Each task condition (SR, IR, and NR) was coded as an explanatory variable (EV) and convolved with a double gamma function, along with its temporal derivative. Each EV yielded a per-voxel parameter estimate (β map) that represented the activation magnitude associated with that regressor. In order to create comparisons of interest, β maps were contrasted. Functional data were registered to MNI stereotactic space using affine transformations. Next, within-subject analyses across runs employed a fixed-effects model, whereas group-level inferential statistical analyses were carried out on each contrast of interest (e.g., IR vs. SR, IR vs. NR, and SR vs. NR) using FMIRB’s linear analysis of mixed effects (FLAME1+2), followed by two-sample t tests (TDC vs. ASD). Whole-brain Z-statistic (Gaussianized T) maps were thresholded using clusters determined by a voxel level of Z ≥ 3.1 (i.e., p ≤ 0.001) and an FWE-corrected cluster-significance threshold of p ≤ 0.05 to strictly control type I errors [32]. In addition, to test our a priori hypothesis of greater reward system activation for individual CI reward than social reward in ASD vs. TDC, we examined group differences in six regions of interest (ROI), including ventral striatum/nucleus accumbens (Nacc), dorsal striatum/caudate, ACC, vmPFC, OFC, and insula, which were anatomically defined areas from the Harvard-Oxford structural probabilistic atlases [33]. We applied a FWE-corrected threshold of p ≤ 0.05 across each region using the Randomise v2.1 program as part of FSL [34].

Reaction times (RTs) for hits (in ms) and task accuracy (correct response rate in %) on the IDT were analyzed using a three (reward: NR vs. SR vs. IR) by two (group: ASD vs. TDC) repeated-measures MANOVA model, followed by planned contrasts. This analysis revealed a main effect of reward [F(4,56) = 143.65, p < 0.001, Cohen’s d > 1.4], but no main effect of group [F(2,58) = 0.53, ns] or group-by-reward interaction effect [F(4,56) = 1.13, ns]. This suggests that the different reward conditions similarly affected behavioral performance across groups. The following univariate ANOVAs showed that the significant reward effect was related to both speed (p < 0.001, Cohen’s d > 1.4) and accuracy (p < 0.001, Cohen’s d > 1.4). Regarding RT, post hoc contrasts revealed fastest RTs for SR, slowest RTs for NR and with the IR condition intermediate (all ps < 0.001, all Cohen’s ds > 1.06). This indicates that incentive manipulations within the experimental task were successful. Post hoc contrasts for accuracy revealed the highest the correct response rate for SR, the lowest the correct response rate for IR and the NR condition intermediate (all ps < 0.004, all Cohen’s ds > 0.77) (see Table 2). Although we intended to maintain an equal average accuracy across conditions and participants by adjusting the target duration according to individual trial-by-trial performance, our finding of different accuracy rates for the three incentive conditions is in line with prior studies [11].

Considering reward system responsiveness across both groups for the two high-level contrasts, i.e., SR > NR and IR > NR, the whole-brain analysis revealed robust brain activation (i.e., k ≥ 10) in ventral striatum (including Nacc), dorsal striatum (including caudate nucleus, and putamen), thalamus, amygdala, ACC, vmPFC, insula, and OFC (Fig. 2). Considering differential reward system activation for the two reward types, the IR > SR comparison showed greater BOLD responses in a cluster, comprising Nacc, caudate, thalamus, ACC, vmPFC, insula and OFC (MNI_peak = − 2, 42, 0; Z_max = 9.16; k = 8047) as well as a cluster within posterior cingulate cortex (MNI_peak = 0, − 30, 28; Z_max = 9.00; k = 166). The reversed comparison (i.e., SR > IR) yielded significant activation differences within right and left insula (MNI_right-peak = 36, − 18, 20; Z_max = 5.18; k = 802, and MNI_left-peak = − 42, − 18, 6; Z_max = 4.87; k = 500).

Following up our hypothesis of reward “imbalance” in ASD (i.e., enhanced reward system activation for CI reward, but diminished responsiveness for social reward), we specifically investigated group activation differences in response to IR versus SR. Using whole-brain cluster thresholding that strictly controls type I errors [32], neither the IR > SR contrast nor the SR > IR contrast revealed significant group activation differences. The additional ROI analyses, however, demonstrated that the right caudate (MNI_peak = 12, 14, 14; t_max = 3.14; Cohen’s d = 0.84; k = 26) as well as the left caudate (MNI_peak = − 12, 6, 12; t_max = 3.14; Cohen’s d = 0.89; k = 96) were more active for IR than SR in youth with ASD relative to TDC; or put differently, the caudate was less active to SR than IR in ASD versus TDC (Fig. 3). None of the other a priori ROIs (i.e., Nacc, vmPFC, ACC, OFC, and insula) revealed significant group differences for IR > SR and SR > IR.

To further examine the differential caudate response, we extracted individual β values from this bilateral cluster, separately for group and reward type (see lower right of Fig. 3). The repeated-measures ANOVA performed on the β values as the dependent variable confirmed the significant group-by-reward interaction effect [F(1,59) = 10.46, p = 0.001, Cohen’s d = 0.84]; main effects of reward and group were non-significant (ps > 0.51). Follow-up within-group t tests yielded a significantly greater caudate response for interest reward than social reward in the ASD group [t(38) = − 3.05, p = 0.004, Cohen’s d = 0.53], while TDC showed marginally stronger caudate activation for social reward compared to interest reward [t(21) = 1.83, p = 0.081, Cohen’s d = 0.51]. Between-group t tests yielded significantly stronger caudate activation for IR in the ASD group compared to TDC [t(59) = 2.79, p = 0.007, Cohen’s d = 0.74], but no significant activation differences between groups in response to SR within this cluster [t(59) = − 1.28, ns, Cohen’s d = 0.34]. The group-by-reward interaction effect remained significant when age and GCA (i.e., FSIQ) were controlled for (p < 0.001, Cohen’s d = 0.93).

Given the nonselective nature of this analysis [35], we additionally investigated unbiased individual β values averaged across an anatomically defined bilateral caudate mask (based on the Harvard-Oxford probabilistic atlas). We found that the group-by-reward interaction effect was still significant, but with a lower, though still medium strong, effect size [F(1,59) = 4.34, p = 0.04, Cohen’s d = 0.55]. Please also note that there were no significant between-group differences in brain activation for the neutral condition using whole-brain or ROI analyses (NR > fixation for structurally defined bilateral caudate: t(59) = − 1.04, ns).

Because our study sample comprised of substantially more youth with ASD than TDC, we repeated our structural ROI analyses for right and left caudate in an age- and IQ-matched sub-sample of 22 individuals with ASD (14 males, 8 females; age 12.3 ± 2.4 years; IQ 110.2 ± 17.2) and the initial group of 22 TDC (17 males, 5 females; age 12.9 ± 2.1 years; IQ 111.9 ± 18.0). The results remained virtually the same, with significantly greater bilateral caudate activation for IR than SR in the 22 youth with ASD relative to the 22 TDC (right MNI_peak = 10, 14, 14; t_max = 3.55, left MNI_peak = − 12, 6, 12; t_max = 3.24). Taken together, these results suggest that the caudate nucleus responded differently in ASD versus TDC dependent on the reward type. This effect was driven by particularly greater dorsal striatum responsiveness for personalized interest reward than social reward in ASD, while TDCs did not show significant differences—although a reverse direction—in their response pattern to both reward types.

To specifically test predictions derived from the social motivation hypothesis, we correlated the individual β values from the cluster that distinguished the two groups (i.e., magnitude of caudate response for IR > SR) with the degree of social dysfunction in ASD as assessed by the autism diagnostic observation scale (ADOS) severity score [36] and the SRS-2 social communication and interaction (SCI) subscale’s T-score [26]. This analysis yielded a significant positive correlation between (right) caudate responses and SRS-2 SCI scores (r(39) = 0.31, p = 0.05; ADOS: Spearman’s ρ(39) = − 0.02, ns), such that stronger caudate activation for interest reward than social reward was related to greater social impairment in the ASD group (Fig. 4). Because of our a priori hypothesis, we did not apply corrections for multiple comparisons.

We further explored associations between bilateral caudate responses for IR > SR and RRBI symptoms in ASD as assessed with the RBS-R (total score), IS (total score), ADI-R (RRB total score), ADOS-2 (RRB total score), and SRS-2 (restricted interests and repetitive behavior subscale T-score). This analysis revealed a significant positive correlation with ADOS-2 RRB symptom expression (Spearman’s ρ(39) = 0.41, p = 0.011; Additional file 1: Table S2), such that stronger caudate responsiveness for CI reward than social reward was associated with greater (clinically observed) RRBIs in our ASD sample. This correlation, however, would not survive correction for multiple comparisons.

We additionally found a significant negative correlation between (left) caudate activation differences to IR > NR and ΔRT for IR vs.NR in the ASD group (r(39) = − 0.32, p = 0.05). This suggests that stronger caudate responsiveness to CIs versus neutral outcome was associated with faster behavioral responses by the participants with ASD to attain this individual reward. Please note, though, that this correlation would not survive correction for multiple comparisons.