Shorter cortical adaptation in dyslexia is broadly distributed in the superior temporal lobe and includes the primary auditory cortex.

Studies of the performance of individuals with dyslexia in perceptual tasks suggest that their implicit inference of sound statistics is impaired. Previously, using two-tone frequency discrimination, we found that the effect of previous trials' frequencies on the judgments of individuals with dyslexia decays faster than the effect on controls' judgments, and that the adaptation (decrease of neural response to repeated stimuli) of their ERP responses to tones is shorter (<xref ref-type="bibr" rid="bib22">Jaffe-Dax et al., 2017</xref>). Here, we show the cortical distribution of these abnormal dynamics of adaptation using fast-acquisition fMRI. We find that faster decay of adaptation in dyslexia is widespread, although the most significant effects are found in the left superior temporal lobe, including the auditory cortex. This broad distribution suggests that the faster decay of implicit memory of individuals with dyslexia is a general characteristic of their cortical dynamics, which also affects sensory cortices.

Introduction

Dyslexia, a specific and significant impairment in the development of reading skills that is not accounted for by mental age, visual acuity problems, or inadequate schooling (World Health Organization, 2016), affects ~5% of the world’s population (Lindgren et al., 1985). Though individuals with dyslexia are diagnosed for their reading impairments, they also often have difficulties in simple non-linguistic perceptual tasks (Mcanally and Stein, 1996; Ahissar et al., 2000; Sperling et al., 2005; Giraud and Ramus, 2013). These can be largely explained as resulting from inefficient use of stimulus statistics that characterize the experiment (the ‘Anchoring Deficit hypothesis’; Ahissar et al., 2006; Oganian and Ahissar, 2012; Jaffe-Dax et al., 2015). In these tasks, participants are not aware of the effect of previous stimuli. But their perception — in particular when retention is required, as for example in serial discrimination tasks — tends to contract to the estimated mean of previous stimuli (contraction bias; Raviv et al., 2012, 2014). This contraction to the mean merges the (implicit) predicted stimulus (based on previous exposures) with the current sensory estimate, forming a coherent percept.

The neural mechanism that may underlie the implicit learning of experimental statistics is adaptation; that is, an automatic, implicit, and stimulus-specific decrease of the response to repeated stimuli. Importantly, the rate of decay of the behavioral effect of previous trials in serial discrimination is similar to the rate of decay of neural adaptation, as measured by magnetoencephalography (MEG) (Lu et al., 1992). Inspired by this finding, we recently compared both the behavioral dynamics and the rate of adaptation (event-related potential [ERP] responses) of good readers (i.e., the control group) and dyslexic participants (Jaffe-Dax et al., 2017). All participants performed serial two-tone frequency discrimination in four blocks with different Trial Onset Asynchronies (TOAs, i.e., temporal interval from the onset of a trial to the onset of the next trial). Both the magnitude of perceptual contraction to the mean frequency of previous trials and the magnitude of neural adaptation (P2 and N1 components that are automatically produced by the auditory cortex [Mayhew et al., 2010]) decayed faster in participants with dyslexia (ERP [Jaffe-Dax et al., 2017]).

As ERP responses cannot be used to localize the cortical source of this group difference, we now recruited the participants from the ERP study (Jaffe-Dax et al., 2017) to take part in an fMRI study with a similar protocol, which allowed us to characterize which brain areas show shorter adaptation in dyslexia. Using the ERP-based protocol in the scanner, we measured the BOLD response (s) to tones for each TOA, and calculated the time constant of adaptation (fitting an exponential decay function). To identify the areas of significant group difference in time constants of adaptation, we used two methods: data-driven whole-brain analysis, and hypothesis-driven regions of interest (ROI) based on areas suggested to produce N1 and P2 (e.g. Mayhew et al., 2010). All cortical regions that responded to tone discrimination showed a tendency to decay faster in participants with dyslexia. Whole-brain analysis revealed significant group differences in the left superior temporal lobe and in the right insular cortex, and ROI analysis revealed significant differences in the left primary auditory cortex.

Results

We recruited 20 participants with dyslexia and 19 good readers from our previous study (Table 1; Jaffe-Dax et al., 2017) and asked them to perform two-tone frequency discrimination in separate blocks with four trial-onset intervals (TOAs) of 3, 6, 9, and 15 s. Before entering the scanner, all participants performed a short four-block training session with simulated scanner noise to familiarize them with in-scan conditions (Sperling et al., 2005; Chait et al., 2007). The two groups exhibited similar accuracy (72.4 ± 6% vs. 73 ± 4.6%, z = 0.5, p=0.57). In scans, good readers (controls) performed better than dyslexic participants (Mean ± SEM: 82.5 ± 1.6% vs. 76.3 ± 2.2%, z = 2.6, p<0.01 in Mann-Whitney U-tests), suggesting that they gained more from the short pre-scan practice (in line with the faster learning reported by Jaffe-Dax et al., 2017).

Table 1.

General characteristics of the participants in this study (mean and standard deviation).

The assessments used in this study were the same as in our previous study (Jaffe-Dax et al., 2017).

	Control group	Dyslexic group	Mann-Whitney z value
	N = 19	N = 20
Age (years)	25.9 (2.6)	24.5 (2.6)	1.7 n.s.
General cognitive (scaled)
Block design	13.1 (3.2)	12.4 (2.9)	0.6 n.s.
Digit span	11.1 (2.9)	7.5 (1.9)	4.0****
Phonological speed [items/minute]
Pseudo-word reading rate	64.0 (25.2)	31.9 (9.7)	3.9****
Single-word reading rate	101.6 (35.2)	69.2 (21.3)	3.0***
Word pattern recognition rate	69.8 (15.6)	41.7 (11.8)	4.5****
Passage reading rate	142.2 (23.9)	100.7 (17.4)	4.5****
Spoonerism rate	9.9 (3.0)	5.7 (3.1)	3.8****
Phonological accuracy [% correct]
Pseudo-word reading accuracy	90.6 (11.9)	63.5 (18.4)	4.0****
Single-word reading accuracy	97.2 (4.3)	89.0 (6.5)	3.7****
Word pattern recognition accuracy	100.0 (0.0)	98.27 (3.1)	2.5**
Passage reading accuracy	98.7 (1.2)	95.4 (2.3)	4.1****
Spoonerism accuracy	90.8 (6.9)	77.8 (17.2)	2.5**

*p < 0.05; **p<0.01; ***p<0.005; ****p<0.0005.

To evaluate the dynamics of cortical adaptation in each group, we used the following procedure. First, we determined which Talairach voxels responded to the task (standard generalized linear model [GLM], p<0.001, false discovery rate [FDR] corrected) when all participants were considered. For each of these voxels, we calculated the dynamics of adaptation, among control participants and among participants with dyslexia, as follows. We estimated over the mean blood oxygenation level dependent signal (BOLD) response of each group in each of the four TOA conditions (block design; see 'Materials and methods'). Using these s, we fitted an exponential decay model (Lu et al., 1992; Jaffe-Dax et al., 2017): to each voxel. In this model, denotes the time constant of adaptation, is the asymptotic level of BOLD and is the magnitude of adaptation. Figures 1A and 1B show the distribution of the mean fitted s for the control and dyslexic groups, respectively. Their comparison illustrates the broadly distributed trend of faster decay in the dyslexic group.

Figure 1.

Cortical distribution of the groups’ mean estimated time constants () of adaptation, calculated separately for each of the responding voxels.

(A) Control participants. (B) Participants with dyslexia. The estimated s for participants with dyslexia were consistently shorter than those estimated for the control group. Significant group differences in the whole-brain analysis (Monte-Carlo cluster-level corrected: cluster threshold of 44 voxels; see 'Materials and methods') are outlined in magenta. The left and right primary auditory cortices, which were estimated as a source of P2 (ERP) component, are outlined in orange. An ROI analysis (see text) revealed a significant group difference in the left primary auditory cortex (Figure 2).

Figure 2.

BOLD response as a function of TOA in the primary auditory cortex of each hemisphere.

Blue: control. Red: dyslexic. AC: the 3 subregions that comprise the primary auditory cortex, outlined in orange in Figure 1.

To locate regions in which the fitted differs significantly between the groups, we conducted a whole-brain analysis, in which we fitted to each voxel, for each participant separately. To reduce the impact of outliers resulting from the noisy estimation of (due to this single subject and single voxel analysis), we assessed group difference with a non-parametric test (Mann-Whitney U test), in which extreme values are not over-weighted. We corrected for multiple comparison bias by requiring a cluster of contingent voxels with a significant group difference (cluster corrected for p<0.05, dictated 44 spatially contingent voxels, based on Monte-Carlo cluster-level correction). Significant regions were found in the left superior temporal cortex (TAL: −54,–18, 10) and in the right insular cortex (TAL: 39,–2, −8), outlined in magenta in Figure 1A and 1B. The superior temporal cortex is known to be involved in a broad range of auditory tasks, including simple tone discrimination (Daikhin and Ahissar, 2015), language (Fedorenko et al., 2010) and music (Fedorenko et al., 2011), and even social tasks (e.g. Deen et al., 2015). Thus, the group difference for this area was expected given the behavioral results. The right insular cortex is multi-modal (Bushara et al., 2003), and is also involved in introspection (Craig et al., 2000). A comparison of  Figures 1A and 1B suggests that other regions might have mean group differences (e.g., frontal cortices), but due to large inter-subject variability in these regions, the group differences were not significant. This large variability might account for the spurious dots of large values scattered throughout the cortical map (Figures 1A and 1B).

In addition to the whole-brain, data-driven analysis, we conducted an ROI, hypothesis-driven analysis. Given our previous ERP findings that P2 (and to a lesser extent N1) of individuals with dyslexia shows shorter adaptation (Jaffe-Dax et al., 2017), we conducted an ROI for the estimated cortical source of this ERP component. In a seminal MEG study (Lütkenhöner and Steinsträter, 1998), the source of N1 was attributed to Planum Temporale, and that of P2 to Heschl's gyrus. A more recent study (Mayhew et al., 2010) combined ERP and fMRI, and found significant correlations between variability in ERP-measured N1–P2 complex and BOLD responses in several regions, including the primary (Heschl's gyrus) and secondary (Planum Temporale and STG areas) auditory areas. Given these estimates, we conducted an ROI analysis on both primary and secondary auditory areas. We used a combined cytoarchitectonic (Morosan et al., 2001) and myeloarchitectonic (Dick et al., 2012) definition of these areas (the primary auditory cortex being composed of three sub-regions and two secondary auditory areas [Planum Temporale and Planum Polare]). We fitted the exponential decay model to the s averaged over the right and the left primary auditory cortices (composed of 99 voxels each, denoted by the orange outlines in Figure 1A–1B), and over the two right and left secondary cortices. We found significant differences between the groups' s in the left primary auditory cortex (z = 2.6, p<0.01, effect size r = 0.42; Mann-Whitney U-tests). In the right primary auditory cortex, the group difference showed the same trend, but did not reach significance (z = 1.5, p=0.15, effect size r = 0.23; Mann-Whitney U-tests). Figure 2 shows the s estimated for the left and right primary auditory cortices of the control (blue) and dyslexic (red) participants on each of the four TOA blocks. None of the other sub-regions of auditory cortex yielded significant group differences.

Taken together, the whole-brain and ROI analyses revealed a significant group difference in the timescales of adaptation in the left superior temporal cortex, left primary auditory cortex, and the right insular cortex. BOLD activity in both auditory regions was previously shown to correlate with the magnitude of N1–P2 responses (Mayhew et al., 2010). The right insular cortex is probably not associated with the P2 response that we measured with ERP and was therefore not predicted by our previous study. In addition, the general (though not significant) trend of shorter adaptation in participants with dyslexia was consistent across all responding voxels.

Discussion

We characterized the cortical distribution of the decay of BOLD adaptation for participants with dyslexia and control participants, thus extending our previous behavioral and ERP study (Jaffe-Dax et al., 2017). We found a broadly distributed tendency for shorter adaptation in dyslexia. We further assessed group difference in the primary and secondary auditory cortices, associated with the production of P2 (Mayhew et al., 2010). Reports of previous studies asking whether these areas manifest neuro-typical structure and function in dyslexia are mixed. For example, Clark et al. (2014) reported early anatomical abnormalities, whereas Boets et al. (2013) reported adequate stimulus resolution. We now found a significant group difference in the left primary auditory cortex, and a similar tendency, which did not reach significance, in the right primary auditory cortex.

The broad distribution of dyslexics' faster decay of adaptation is in line with recent observations of a domain general abnormally small adaptation in dyslexia (Perrachione et al., 2016). The researchers compared BOLD responses to stimulus repetitions (in blocks of ~10 s) with BOLD response to non-repeated stimuli (auditory and visual), and found reduced stimulus-specific adaptation in the auditory (superior temporal), visual (fusiform and lateral occipital [LO]), and associative (insular and inferior frontal) cortices. Importantly, they found a three-way interaction of group (control and dyslexic) x condition (repeated and non-repeated) x time (within the block), where controls' increase of adaptation along the block was larger than that of participants with dyslexia. This observation, which results from controls' accumulative adaptation with stimulus repetition, is fully consistent with our observation of dyslexics' faster decay of adaptation.

We should note that the timescale of impaired retention in dyslexia is slower than that of sensory memory (iconic or echoic memory is <<1 s). Still, it is expected to affect the perception of individuals with dyslexia owing to their reduced cross-trial retention. It is therefore expected to impede their performance in a broad range of perceptual tasks, such as entrainment to rapid stimuli (e.g., Witton et al., 1998; Goswami, 2011; Lehongre et al., 2013), but only in protocols that contain stimulus repetitions. Indeed, it was shown that listeners entrain better to familiar compared with unfamiliar stimuli (Doelling and Poeppel, 2015; Kumagai et al., 2017). The advantage of familiarity is expected to be smaller in dyslexia. The faster decay of implicit perceptual memory of individuals with dyslexia is expected to also impede their long-term accumulation of stimulus statistics, and consequently to reduce the complexity and richness of their long-term accumulated categorical representations (e.g., Perrachione et al., 2011; Banai and Ahissar, 2017).

In summary, the data collected in this study point to the specific neural structures that underlie the ‘anchoring deficit’ in dyslexia, namely a reduced use of stimulus statistics (Ahissar et al., 2006; Ahissar, 2007; Oganian and Ahissar, 2012). These data suggest a broad multimodal (e.g., Jaffe-Dax et al., 2016) cortical distribution that includes, but is not limited to, sensory areas.

Materials and methods

In the two-tone frequency discrimination task, subjects were asked to indicate which of two sequentially presented tones had a higher pitch. The tones were 50 ms long, presented at comfortable intensity, and were drawn from a uniform distribution between 800 Hz and 1250 Hz. The frequency difference within each pair was randomly drawn at between 1% and 20% (following the protocol in Jaffe-Dax et al., 2017). In the pre-training session (8 min), each participant performed one block of each of the four Trial Onset Asynchronies (TOAs) of 3, 6, 9, or 15 s, administered in four separate blocks in random order (each block consisted of 16 trials). These TOAs are longer than those in our previous ERP experiment (1.5, 3, 6, and 9 s, Jaffe-Dax et al., 2017), because the controls' ERP (N1 and P2) response at 9 s was still larger than that at 6 s. Each block had a constant TOA of 3, 6, 9, or 15 s. In the scanner, each participant performed three runs of four blocks (of 16 trials). The block design allowed us to measure the dynamics of adaptation in timescales that are independent of the sluggishness of typical hemodynamic response function (HRF), as we analyzed each block as a whole, and not on a trial-by-trial basis. Specifically, we modelled the magnitude of the BOLD signal in each block as a function of its TOA. This block design was used to estimate on the basis of the magnitude of the BOLD response. However, the number of trials was too small for robust estimation of behavioral context effects, which are based on the difference in success rate (binary scores for each trial) between trials that gain and those that are hampered by the context (Jaffe-Dax et al., 2017). Stimuli were digitally constructed using Matlab 2015b (The Mathworks Inc., Natwick, MA, USA) and administered through inserted sound-attenuating MR compatible S14 earphones (Sensimetrics Corporation, Malden, MA, USA). The demographic, cognitive and reading assessments of this cohort are described in Jaffe-Dax et al. (2017).

Before the functional scan, high-resolution (1 × 1 × 1 mm resolution) T1-weighted magnetization-prepared rapid acquisition gradient-echo (MPRAGE) images were acquired using a 3T Magnetom Skyra Siemens scanner and a 32-channel head coil at the ELSC Neuroimaging Unit (ENU). The cortical surface was reconstructed from the high-resolution anatomical images using standard procedures implemented by the BrainVoyager QX software package (version 2.84; Brain Innovation, The Netherlands). The functional T2*-weighted MRI protocols were based on a multislice gradient echo-planar imaging and obtained under the following parameters: TR = 1 s, TE = 30 ms, flip angle = 90°, imaging matrix = 64 × 64, field-of-view=192 mm; 42 slices with 3 mm slice thickness and no gap were oriented in AC-PC plane, covering the whole brain, with functional voxels of 3 × 3 × 3 mm and multiband parallel imaging with an acceleration factor of 3 (Moeller et al., 2010).

Preprocessing of functional scans in BrainVoyager included 3D motion correction, slice scan time correction, and removal of low frequencies up to three cycles per scan (linear trend removal and high-pass filtering). The anatomical and functional images were transformed to the Talairach coordinate system using trilinear interpolation. Each voxel’s time course was z-score normalized and smoothed using a 3D Gaussian filter (full width at half maximum [FWHM] of 4 mm). A standard (two gamma) hemodynamic response function (Friston et al., 1998) was convolved with the trial timings of each TOA block to build four predictors for the subsequent GLM analysis. For all task-responsive voxels (p<0.001, FDR corrected; Benjamini and Yekutieli, 2001), each TOA condition was modeled separately to account for its contribution to the measured BOLD signal in each voxel. Specifically, a single value was obtained for each TOA condition. An exponential decay model (see 'Results') was fitted to these values, and its parameters were estimated for each voxel in each subject using a least-square method. For ROI analysis, the MNI coordinates of auditory cortex subdivision were obtained from Morosan et al., 2001 and translated into Talairach coordinates using Yale BioImage Suite Package (sprout022.sprout.yale.edu/mni2tal/mni2tal.html; Lacadie et al., 2008). The BOLD signal was averaged for each ROI and then the values of the four TOA blocks were fitted to the exponential decay.

Whole-brain significance results were corrected for multiple comparison false-positive biases by a Monte-Carlo cluster correction (Forman et al., 1995), implemented using a plug-in of BrainVoyager (Goebel et al., 2006). The Monte-Carlo procedure was given an a-priori chosen probability for type I error of 0.05 and yielded a cluster threshold of 44 voxels. Non-parametric tests (Mann-Whitney’s U-test) were used for group comparisons, as we did not assume a normal distribution, as in our previous study (Jaffe-Dax et al., 2017).

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Widespread shorter cortical adaptation in dyslexia" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Andrew King as the Senior Editor. The following individual involved in review of your submission has agreed to reveal his identity: John Stein (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

This study builds on the authors' previous work that demonstrated using behavioral measurements and EEG recordings that dyslexics show a faster decay of implicit memory than controls, which may account for their longer reading times. Here they used fMRI to investigate which cortical areas show these shorter neural adaptation effects. The authors calculate time constants of adaptation and find these to be significantly different in the dyslexic subjects in a region of left non-primary cortex and in the right anterior insula close to frontal operculum. The reviewers agreed that this study has a solid theoretical and empirical foundation in the previous work carried out by the authors and by other groups, and that the results are presented clearly and potentially add important new details to the framework for understanding dyslexia. The success of the authors in demonstrating differences in the time constant between the groups using the sluggish BOLD response is notable. Nevertheless, the reviewers had some difficulty reconciling these data with the previous neurophysiological data.

Essential revisions:

1) The study does not show widespread significant shorter cortical adaptation as suggested in the title. The regions in which significant group differences are small and include high-level multimodal cortex that is not predicted by the previous work. The N1 and P2 responses are generally regarded as arising from non-primary auditory cortex in the posterior superior plane and the previous neurophysiological data fitted with the idea that there might be a less stable representation of sensory information over time in dyslexia promoted by this group, which is a compelling idea. A specific test of this idea would require a comparison of the time constants in the generator of the N1, conventionally regarded as in planum temporale (there are probably two generators in PT as suggested by the work of Lutkenhoner). The reviewers were unanimous that a region of interest analysis is needed, though this might be focused on this region for which there is a strong prior based on the previous work.

2) Examination of the areas in which there is a decrease in tau without correction for multiple comparisons shows a widespread swathe of differences that I find difficult to reconcile with the hypothesised deficit in the sensory trace. For example, there is shorter tau in primary somatosensory cortex and motor cortex. Do the authors think this is consistent with the anchoring deficit model?

3) The authors suggest that poorer behavioral performance by the dyslexia group in the scanner could be related to their impairment in learning the task after training compared to the control group. An alternative possibility that needs to be discussed is that these results are due to the added acoustic noise of scanning, since continuous fMRI sampling as opposed to sparse sampling was employed. It is well known that individuals with dyslexia tend to have trouble with noise exclusion, and a tone-discrimination task with relatively sensitive pitch differences will be disproportionately difficulty for dyslexic individuals in the presence of noise. There is a paper by Chait et al., 2007 to that effect (tones in noise) specifically, if a citation is needed, as well as the larger noise-exclusion literature (e.g. Sperling et al., 2005 and others).

4) Some attempt should be made to explain why the decay is faster in dyslexics by fitting the results into what is already known about dyslexics' relative insensitivity to amplitude and frequency modulations. For example, lower frequency sensitivity may impair auditory entrainment to a tone; could this underlie the difference in adaptation time constant?

5) All the reviewers were struck by the capacity to extract adaptation and adaptation-recovery time course information from fMRI data, but it was agreed that a clearer explanation is warranted of how this can be studied using a method that has a time lag of several seconds.

6 The characteristics of the dyslexics should be given in brief for those who haven't read the previous paper. More details are also needed regarding the Monte Carlo simulations for cluster-level multiple comparison corrections to ensure that others could replicate this approach.

7) The discussion of the time course of adaptation with regards to the Perrachione et al.,2016 at the beginning of the Discussion section is confusing. In that paper, stimuli were presented for 8 repetitions in blocks of approximately 10 seconds, but each adapting or non-adapting stimulus was presented for 700 ms and at a rate of one per 1200 ms - so when the authors say "adaptation across a window of ~10 seconds was smaller than controls because it was largely recovered" it is unclear how they reconcile this with the stimulation procedure of the Perrachione study.

8) It should also be pointed out that Perrachione et al., 2016 reported a group x time x condition interaction in the adaptation response in their study (see Supplementary table S4) for all conditions, which they interpreted as evidence for increasing adaptation in the control group w/ repeated presentation of the same stimulus, whereas such increasing adaptation was attenuated or absent in the dyslexia group. The correspondence between this result and the present observations seems meaningful and is worth commenting on.

Essential revisions:

1) The study does not show widespread significant shorter cortical adaptation as suggested in the title. The regions in which significant group differences are small and include high-level multimodal cortex that is not predicted by the previous work. The N1 and P2 responses are generally regarded as arising from non-primary auditory cortex in the posterior superior plane and the previous neurophysiological data fitted with the idea that there might be a less stable representation of sensory information over time in dyslexia promoted by this group, which is a compelling idea. A specific test of this idea would require a comparison of the time constants in the generator of the N1, conventionally regarded as in planum temporale (there are probably two generators in PT as suggested by the work of Luktenhoner). The reviewers were unanimous that a region of interest analysis is needed, though this might be focused on this region for which there is a strong prior based on the previous work.

Thank you for this comment.

1) We changed the title of the paper to "Shorter cortical adaptation in dyslexia is broadly distributed in the superior temporal lobe and includes the primary auditory cortex".

2) The description of our ROI analysis now makes this point explicit. We now clarify that our ROI analysis was guided by the literature that mapped the cortical sources of P2 and N1 (Lütkenhöner and Steinsträter, 1998; Mayhew et al., 2010). The estimated source is somewhat broad, and includes all auditory areas, primary and secondary (PT and STG). We originally conducted an ROI mask for the primary regions (3 sub-regions, as in the original manuscript). We now added secondary auditory (PT and Planum Polaris) regions. Only one left primary sub-region showed significant group effect. Our whole brain STG observation is also broadly in line with Mayhew et al., (who found that STG is also correlated with variability in N1 and P2), yet specific coordinates are not provided in their paper.

2) Examination of the areas in which there is a decrease in tau without correction for multiple comparisons shows a widespread swathe of differences that I find difficult to reconcile with the hypothesised deficit in the sensory trace. For example, there is shorter tau in primary somatosensory cortex and motor cortex. Do the authors think this is consistent with the anchoring deficit model?

Yes. We added a short paragraph to the Discussion section to clarify this point. Behavioral observations (e.g., Jaffe-Dax et al., 2016) and imaging data (e.g., Perrachione et al., 2016) are consistent with a multimodal anchoring deficit that has specific temporal characteristics. This faster decay is "post sensory" (i.e. duration of echoic and iconic memories <1 second) and yet it affects perception. The study shows that poor cross trial retention is manifested in poor retention in primary and higher-level cortical areas.

3) The authors suggest that poorer behavioral performance by the dyslexia group in the scanner could be related to their impairment in learning the task after training compared to the control group. An alternative possibility that needs to be discussed is that these results are due to the added acoustic noise of scanning, since continuous fMRI sampling as opposed to sparse sampling was employed. It is well known that individuals with dyslexia tend to have trouble with noise exclusion, and a tone-discrimination task with relatively sensitive pitch differences will be disproportionately difficulty for dyslexic individuals in the presence of noise. There is a paper by Chaitet al., 2007 to that effect (tones in noise) specifically, if a citation is needed, as well as the larger noise-exclusion literature (e.g. Sperling et al., 2005 and others).

Thank you for pointing this out. We now clarify that the training period was conducted with noise simulating the noise in the scanner, which explains why performance of both groups was not reduced in the scanner compared to their training performance.

4) Some attempt should be made to explain why the decay is faster in dyslexics by fitting the results into what is already known about dyslexics' relative insensitivity to amplitude and frequency modulations. For example, lower frequency sensitivity may impair auditory entrainment to a tone; could this underlie the difference in adaptation time constant?

This is an important point. We predict that dyslexics' entrainment to unfamiliar stimuli will not be impaired. However, dyslexics' benefits from exposure are reduced compared with good reading controls. Since repetition lowers the measured thresholds, and since most protocols contain stimulus repetitions, we expect dyslexics to have poorer thresholds in a broad range of perceptual tasks (e.g., Witton et al., 1998; Goswami, 2011; Lehongre et al., 2013). For entrainment, it was specifically shown that tracking familiar stimuli is better than tracking similar yet unfamiliar stimuli. We now explain this in the Discussion section.

5) All the reviewers were struck by the capacity to extract adaptation and adaptation-recovery time course information from fMRI data, but it was agreed that a clearer explanation is warranted of how this can be studied using a method that has a time lag of several seconds.

The time lag of the fMRI signal is not detrimental to the estimation of the τ values more than to the standard β estimation. In this study we used a block design; each block was defined by its Trial Onset Asynchrony (TOA), and for each TOA there were 3 blocks. Then, a standard modeling with 4 conditions was applied to extract β values for each TOA condition (each calculated from a different block). Next, a model for estimating the decay time constant (τ) value was applied to each voxel: βTOA=a+bexp-TOA/τ, where τ denotes the timescale of adaptation, a is the asymptote level of the BOLD signal and b is the amplitude of adaptation. We added clarifications in the Materials and methods section regarding the nature and benefits of this block design.

6 The characteristics of the dyslexics should be given in brief for those who haven't read the previous paper.

Group characteristics were now added to the paper (new Table 1).

More details are also needed regarding the Monte Carlo simulations for cluster-level multiple comparison corrections to ensure that others could replicate this approach.

The cluster correction was performed using the cluster-size thresholding plugin of the BrainVoyager QX software (version 2.84; Brain Innovation, The Netherlands). The size of the cluster is determined by the significance level that we decide a-priori, as now better clarified in the text (a detailed description of this correction method is provided in Goebel et al., 2006, who implemented it in BrainVoyager).

7) The discussion of the time course of adaptation with regards to the Perrachione et al., 2016 at the beginning of the Discussion section is confusing. In that paper, stimuli were presented for 8 repetitions in blocks of approximately 10 seconds, but each adapting or non-adapting stimulus was presented for 700 ms and at a rate of one per 1200 ms - so when the authors say "adaptation across a window of ~10 seconds was smaller than controls because it was largely recovered" it is unclear how they reconcile this with the stimulation procedure of the Perrachione study.

Note that we refer to duration of adaptation as the time adaptation lasts rather than the duration of its induction. The duration of adaptation is characterized by the magnitude of the response within the time window of the repetition Block. Perrachione et al., 2016 tracked responses over 10 second (Block duration), and hence tracked the accumulative impact of adaptation as a function of repetition within this time window. The magnitude and duration of (the induced) adaptation both increase with repetition, but to a lesser extent in dyslexics. We now explain this more clearly in the Discussion section.

8) It should also be pointed out that Perrachione et al., 2016 reported a group x time x condition interaction in the adaptation response in their study (see Supplementary table S4) for all conditions, which they interpreted as evidence for increasing adaptation in the control group w/ repeated presentation of the same stimulus, whereas such increasing adaptation was attenuated or absent in the dyslexia group. The correspondence between this result and the present observations seems meaningful and is worth commenting on.

Thank you for pointing this out. We added this to the Discussion section – this is exactly our prediction – that controls' adaptation is accumulated across repetitions to a larger extent than dyslexics', due to the faster decay of dyslexics' implicit memory trace within this time window (~10 second, see comment 8). We now clarify it in the manuscript.