Background
Humans are sensitive to the contingent relationships between cues and outcomes in social or non-social contexts, such as the chances of having someone interact with you after making eye contact or seeing rain soon after hearing thunder. These associations rely on mechanisms of inference [
1,
2], whereby the brain tries to predict the most likely outcome by computing probabilities based on past experiences. Cast in the Bayesian framework, making optimal inferences requires adjusting our prior knowledge when there is a relevant discrepancy (i.e. prediction error) between our prediction and the actual outcome [
1]. The weight of prediction errors depends on the relative precisions of top-down predictions and bottom-up sensory inputs. The precision of prediction errors should be high when mismatches represent informative changes, but low when signalling noise. Minimizing prediction errors, and therefore surprise, can be achieved through optimal predictive mechanisms [
3] that track different forms of uncertainty. Not only does the expected uncertainty (e.g. probabilistic cue-outcome association) need to be tracked, but also the unexpected uncertainty (e.g. unsignalled reversal of the associations) [
4].
In daily life, dealing with uncertainty can be challenging, especially for individuals with Autism Spectrum Disorders (ASD) who report a high intolerance of uncertainty [
5‐
7]. Indeed, in addition to the two core symptoms of ASD that are impairments in social communication and interaction, and restricted and repetitive behaviours [
8], autistic individuals often report difficulties in tolerating unexpected changes. More broadly, autistic individuals could perceive the world as being more unpredictable than neurotypicals (NT) [
9]. Accounts of ASD formulated from a Bayesian perspective suggest an atypical functioning of the predictive brain in ASD [
10‐
15]. Their increased inflexibility or their needs for a routine could be a way to restore some predictability (i.e. to compensate for feelings of constant unpredictability). More precisely, the first Bayesian accounts of ASD suggested that perception might not be biased by priors so much, either because prior precision is low [
10] or because sensory precision is high [
11]. Another hypothesis suggested a high and inflexible precision of prediction errors in ASD (HIPPEA,
12). If prediction errors are not flexibly modulated and are always given a high weight in ASD, it would lead to suboptimal prior updating and to (almost) constant sensations of surprise.
These hypotheses of ASD were elaborated after re-interpreting the existing literature on learning and perception in ASD within the Bayesian framework, but did not come from tasks that were specifically designed to test these hypotheses. Since the formulations of these hypotheses, several studies attempted to test them more directly. A recent systematic review evidenced some differences between ASD and NT individuals in predictive learning and predictive responses [
16]. They found that most of the studies investigating the predictability of repeated stimuli showed reduced habitation in ASD (e.g.
17,
18). Most studies on structural priors (i.e. priors learned over long time scales) reported no differences between NT and ASD (e.g.
19,
20), whereas studies on contextual priors (i.e. priors learned over shorter time scales) gave more heterogeneous results in ASD. For instance, low-level perceptual tasks in autistic adults showed slower prior learning in ASD [
21] or more inflexible priors [
22]. Results of associative learning tasks in ASD often depend on context. In uncertain contexts, the ability to learn and update associations was typical in autistic children (colour-reward association task,
23), but atypical in autistic adults (tone-visual outcome association,
24). In their study [
24], a tone (high or low) was probabilistically associated with an outcome (face or a house) and participants simply had to report what the outcome was. By modelling the response times (modulated by the expectedness of the outcome), the authors showed that autistic adults tended to overestimate the volatility of the environment [
24]. Yet, a recent study suggested that this might be restricted to subgroups of autistic individuals [
25]. In another associative learning study where adults had to learn an association between a tone and a rotation direction in a very uncertain context, autistic adults managed to learn a prior, but failed to update it after an unexpected change in contingency [
26]. Another study relying on a large cohort of children and adults performing a probabilistic reversal learning task also showed poorer performance in ASD and reduced flexible behaviours [
27]. Finally, another probabilistic reward learning task in a volatile environment showed that having more autistic traits was associated with worse performance [
28]. Overall, these behavioural studies tend to indicate that priors can be learned in ASD, but maybe with a different dynamic, more inflexibility or with decreased abilities in people with ASD or high autistic traits.
While the investigation of prediction learning in ASD has made good progress at the behavioural level (for a review:
15), little is known about its underlying neurobiological mechanisms in ASD. As mentioned above, ASD might be characterized by an atypical precision of priors [
10], sensory inputs [
11] or prediction errors (i.e. sensory/prior balance,
12). Precision is thought to be encoded by the post-synaptic gain of superficial neurons signalling prediction errors [
29,
30]. This gain quantifies the effect of a presynaptic input on a post-synaptic output, and mostly relies on glutamatergic NMDA receptors [
29,
30]. According to this framework, an atypical encoding of precision in ASD should be associated with an abnormal (glutamatergic) neuromodulation. Top-down predictions would be mediated by slow NMDA glutamatergic receptors [
31], in particular in the prefrontal cortex [
32]. Bottom-up prediction errors would be mediated by fast AMPA glutamatergic receptors and GABA
A receptors [
31]. Another hypothesis suggests that rapid glutamatergic and GABAergic neurotransmission would represent prediction errors, whereas slower neuromodulators (such as acetylcholine) would encode the precision of prediction errors [
33,
34]. All these hypotheses point towards the key role of glutamate (excitatory) and GABA (inhibitory) in predictive mechanisms. Abnormal concentrations or signalling of these neurotransmitters would alter the ability to encode the precision of predictions or prediction errors, and so, to optimally learn predictions. Consistently, computational simulations suggested that an increased glutamate/GABA ratio would be consistent with a decreased influence of contextual priors [
35]. Given the role of glutamate and GABA in predictive mechanisms and the hypotheses of altered predictive abilities in ASD, the question arises whether abnormal glutamate/GABA signalling could be related to the Bayesian accounts of ASD.
Interestingly, abnormalities in the glutamatergic and GABAergic neurotransmission were found in ASD, including atypical neurotransmitter concentration, receptor density or enzyme dysregulation (e.g. [
36,
37]). On the one hand, some studies support the
hyperglutamate theory in ASD and are based on increased glutamate concentrations in the serum and plasma, on the upregulation of NMDA receptors in animal models induced with valproic acid, or on the behavioural ameliorations triggered by antagonists of glutamatergic receptors [
38]. On the other hand, some other studies support the
hypoglutamate theory in ASD and rely on dysfunction of certain glutamatergic receptors and on beneficial effects of glutamatergic agonists in ASD [
38]. Studies using the in vivo and non-invasive method of magnetic resonance spectroscopy (MRS) in autistic adults showed that the level of Glx (glutamate and glutamine) depended on the brain region: Glx was either increased (amygdala–hippocampal complex [
39], auditory cortex [
40], sensorimotor cortex [
41], anterior cingulate cortex (ACC) [
42], cerebellum [
43]), decreased (ACC [
44,
45], central nuclei [
46,
47]) or not different from NT (frontal regions [
46,
48‐
51], parietal regions [
39,
44,
46], occipital regions [
49,
52], cingulate cortex [
50,
53], thalamus [
44]). Regarding GABA, most of the evidence is in favour of a decreased action of GABA in ASD [
37,
54]. In autistic adults, MRS studies mostly found decreased GABA levels (sensorimotor cortex [
55], supplementary motor area [
56], ACC [
43], cerebellum [
43]) or no group differences (frontal regions [
47,
49,
51,
52,
56,
57], occipital regions [
49,
52,
55,
56], auditory cortex [
52,
58], ACC [
59], striatum [
47]).
It is also important to note that glutamate is converted into GABA by the glutamic acid decarboxylase (GAD, either the 65 or 67 isoforms), whose concentration was found to be reduced in ASD [
60]. This would lead to an increase in glutamate and decrease in GABA in ASD, which is consistent with the results reported above. Consistently, high Glx/GABA ratios measured with MRS were associated with more autistic traits (e.g. superior temporal cortex [
61,
62]). Glutamate is also a precursor of glutathione (GSH), which can be simultaneously measured with edited MRS [
63]. GSH is a key component regulating oxidative stress and several other cellular and genetic pathways that contribute to protect cells. GSH therefore plays a neuroprotective role in the brain, and could, for instance, be released to protect neurons when there is an excess of glutamate [
64]. An imbalance in the GSH redox system could contribute to the neuropathology of ASD (e.g. [
64]).
To sum up, several findings indicate an imbalance in glutamate and/or GABA in autistic adults. As these neurotransmitters are hypothesized to play a key role in encoding predictions and prediction errors, an alteration of the glutamate/GABA signalling could underlie the difficulties in making predictions, as observed at the behavioural level in ASD. Glutamate is a precursor of GABA but also of GSH, a neuroprotective molecule. Although there is no specific hypothesis directly relating GSH to prediction learning, it contributes to the same metabolic cycle as glutamate and GABA. An imbalance in GSH could alter neuronal functions and indicate abnormal Glutamate metabolism. Our exploratory approach is therefore to determine the existence of a deficit/excess in GSH, which could contribute to shedding light on the functioning of brain regions hypothesized to be involved in encoding priors.
In order to better understand prediction learning in autistic adults and to investigate its neural correlates, we conducted a behavioural study combined with MRS measurements of Glx, GABA and GSH. NT and autistic adults performed an associative learning task where a high or low pitch tone was predictive of a clockwise or counterclockwise rotation of a pair of dots (same paradigm as in [
66]). There was some expected uncertainty about the cue-outcome association (i.e. 75% of the trials presenting the main association), as well as a part of unexpected uncertainty as the association could suddenly reverse. Participants had to make explicit predictions about the outcome and to report what they perceived. In a subset of ambiguous trials, the dots did not rotate, but we expected participants to be biased by their predictions (i.e. to report the expected rotation). Therefore, there was an explicit measure of prediction learning, as well as a measure of prior bias. Participants also filled in questionnaires assessing their autistic traits or symptoms and their intolerance of uncertainty. Finally, Glx, GABA and GSH were simultaneously quantified using edited MRS in a low-level perceptual region (i.e. the medial occipital cortex) and in a higher-level region hypothesized to play a role in learning predictions [
67,
68], (i.e. the right inferior frontal gyrus, IFG).
The current study had several objectives. First, we aimed at better characterizing prediction learning in ASD and relating it to the autistic symptomatology. For this purpose, we assessed whether autistic adults managed to learn and update their predictions and whether they were biased by their expectations. The Bayesian hypotheses suggesting hypo-prior [
10] or high sensory precision [
11] in ASD would predict a decreased prior bias and a decreased performance to make predictions. However, in light of the recent empirical findings cited above, we predicted that autistic individuals would learn a prior but may be more inflexible to update it. Within the Bayesian literature of ASD, defaults in making accurate predictions are supposed to underlie the autistic symptomatology and to account for their higher intolerance of uncertainty. Therefore, we predicted that the percentage of correct predictions would be negatively correlated with the questionnaire scores assessing social difficulties, atypical sensory sensitivity and intolerance of uncertainty. We also measured whether this predictive ability varied along the autism spectrum using the Autism-spectrum Quotient, as autistic traits are continuously distributed in the population [
69] and were correlated with performance in a probabilistic learning task [
28]. Second, our other main goal was to investigate the neuromolecular correlates of prediction learning in ASD using MRS. Precisely, we expected the Glx/GABA ratio to be correlated with the ability to make predictions in lower- and higher-level regions. We were especially interested in assessing whether there would be such a correlation in the IFG, as it would be involved in encoding predictions and prediction errors [
67,
68], and that atypicalities were reported in this region in ASD. To better characterize the functioning of these two regions in ASD, we also compared the concentrations in Glx, GABA and GSH between groups. Given the MRS literature, we hypothesized that these concentrations would not differ in ASD versus NT in the occipital cortex. There are no published MRS results in the right IFG in autistic adults; yet, given its potential role in making predictions and the altered abilities observed in ASD, we hypothesized that Glx levels might be increased and/or GABA levels decreased, which may be associated with an increase in GSH.
MRS data analysis
HERMES MRS spectra were analysed using Gannet 3.1 [
78], implemented in MATLAB 2020b. Frequency domain spectra were frequency- and phase-corrected using spectral registration (specRegHERMES [
79]). Preprocessing of the MRS data also included a 3 Hz line broadening filter (exponential apodization function), fast Fourier transformation, time averaging, frequency and phase corrections based upon fitting of the Choline and Creatine signals, and pairwise rejection of data for which fitting parameters were greater than three standard deviations from the mean. GABA and Glx were fitted with a three Gaussian model using nonlinear least square fitting between 2.79 and 4.10 ppm. The unsuppressed water spectrum was modelled between 3.8 and 5.6 ppm using a single Gauss-Lorentzian mode with phase and linear baseline parameters, also using nonlinear least-squares fitting. The GSH spectrum was estimated from 2.25 and 3.5 ppm using a single Gaussian to model the 2.95 ppm GSH signal with four Gaussians to model co-edited unwanted signals and a nonlinear baseline.
Upon a careful visual inspection of all the MR spectra, and the criterium that the fit error for Glx and GABA + should not exceed 15%, some participants were excluded from the analyses (see Table
2). The mean fit error, the full-width at half-maximum (FWHM), the water drift and signal-to-noise ratios (SNR) of the remaining participants are given in Table
2. None of these quality metrics differed between groups except for the GABA + concentration in the OCC VOI and the GSH concentration in the IFG VOI. The tissue fractions did not differ between groups in any VOI (Table
2). Individual spectra are shown in Fig.
2c–j.
Table 2
MRS quality metrics, tissue fractions and concentrations
Number of MRS datasets | |
GABA+ | 24 | 25 | na | 22 | 24 | na |
Glx | 20 | 22 | na | 15 | 16 | na |
GSH | 26 | 26 | na | 20 | 23 | na |
Fit error (%) | |
GABA+ | 6.4 (± 2.8) | 8.2 (± 2.8) | * | 7.5 (± 2.0) | 6.8 (± 1.5) | ns |
Glx | 3.9 (± 0.8) | 4.4 (± 1.1) | ns | 5.3 (± 1.3) | 5.0 (± 1.1) | ns |
GSH | 7.3 (± 2.0) | 8.3 (± 2.6) | ns | 14.5 (± 6.3) | 10.9 (± 4.2) | * |
Signal to noise ratio |
GABA+ | 13.0 (± 3.8) | 11.5 (± 3.3) | ns | 12.9 (± 2.3) | 11.9 (± 2.1) | ns |
Glx | 19.9 (± 4.5) | 21.2 (± 6.1) | ns | 17.6 (± 4.7) | 16.6 (± 4.0) | ns |
GSH | 10.1 (± 1.9) | 9.8 (± 1.8) | ns | 9.0 (± 1.8) | 8.6 (± 1.3) | ns |
Water drift | 0.03 (± 0.01) | 0.03 (± 0.01) | ns | 0.02 (± 0.00) | 0.02 (± 0.00) | ns |
Full-width at half-maximum |
GABA+ | 21.9 (± 2.7) | 22.5 (± 2.5) | ns | 20.7 (± 2.1) | 21.0 (± 2.2) | ns |
Glx | 14.2 (± 2.6) | 13.6 (± 2.0) | ns | 15.8 (± 3.8) | 17.2 (± 5.2) | ns |
GSH | 8.7 (± 1.0) | 8.9 (± 0.9) | ns | 10.5 (± 2.7) | 11.1 (± 2.5) | ns |
Tissue fractions | |
Grey matter | 65% (± 4) | 66% (± 3) | ns | 54% (± 5) | 54% (± 4) | ns |
White matter | 25% (± 3) | 24% (± 3) | ns | 38% (± 6) | 38% (± 5) | ns |
Cerebrospinal fluid | 10% (± 2) | 10% (± 3) | ns | 8% (± 3) | 8% (± 3) | ns |
Concentrations |
GABA + (i.u.) | 2.94 (± 0.72) | 2.75 (± 0.61) | ns | 2.85 (± 0.53) | 2.83 (± 0.61) | ns |
Glx (i.u.) | 9.08 (± 2.90) | 9.35 (± 2.60) | ns | 8.12 (± 1.43) | 9.90 (± 3.14) | * |
GSH (i.u.) | 0.93 (± 0.17) | 0.92 (± 0.23) | ns | 1.13 (± 0.40) | 1.16 (± 0.36) | ns |
Glx/GABA + ratio | 3.75 (± 2.44) | 3.73 (± 1.52) | ns | 2.91 (± 0.68) | 3.59 (± 1.50) | ns |
The VOIs were co-registered to the anatomical T1-weighted image and segmented in Gannet through SPM12 in order to obtain the fractions of grey matter (
fGM), white matter (
fWM) and cerebrospinal fluid (
fCSF). GABA, Glx and GSH concentrations were quantified relative to the unsuppressed water signal, and corrected for tissue fractions using
α correction (e.g. [GABA]
αcor =
\(\frac{{\left[ {GABA} \right]}}{{f_{GM} + \alpha f_{WM} }}\), with
α = 0.5 as per [
80] which assumes a GM/WM ratio for GABA of 2:1 based on prior literature [
81]). Additional parameters included an assumed visible water concentration of 50,000 mM, an editing efficiency of 0.5 and T1- and T2-specific values for both GABA and water as described in [
80] where T1 and T2 values were estimated to be different between GM and WM. Gannet does not incorporate tissue compartment-specific values for GABA as these are currently not available, but for water tissue-specific values were used for GM, WM, and CSF, with different MR-visible concentrations as reported [
82]. Note that macromolecules at 3 ppm are co-edited with GABA, so GABA levels are reported as GABA+ [
83] and that an additional MM factor of 0.5 was used for the quantification of the estimated concentrations, as recommended by [
84]. The same approach was used for GSH due to lack of published information for GSH [
85]. Metabolite concentrations are given in institutional units [i.u.]. The tissue fractions are reported in Table
2.