Shared anomalies in cortical reading networks in Chinese and French dyslexic children

To determine whether the neural anomalies underlying developmental dyslexia are universal across languages or influenced by the writing system, we tested 10-year-old Chinese and French children, with or without dyslexia, in a cross-cultural fMRI paradigm. We compared their brain responses to words written in their known script, faces and houses while they were asked to detect a rarely presented star. We observed that impaired reading scores were correlated with a decreased activation to words in several key regions of the reading circuit, including left fusiform gyrus, superior temporal gyrus/sulcus, precentral and middle frontal gyrus. In ROIs previously reported as sensitive to dyslexia, we observed main effects of dyslexia common to Chinese and French readers, without interaction with the children’s native language, suggesting a cross-cultural invariance in the neural anomalies underlying dyslexia. Multivariate pattern analyses further confirmed that dyslexics exhibit a reduced activation to written words in the left fusiform gyrus and left posterior superior temporal gyrus, and not merely a greater inter-individual variability. The impairments in these regions may reflect the causes as well as the consequences of orthographic and phonological deficits in dyslexia in different languages. The current study highlights the existence of common brain mechanisms for dyslexia even in highly different writing systems.


9
(Words vs. Faces vs. Houses) as a within-subject factor. The ANOVA analysis was intended to 232 test the main effect of Category (One category > mean of the other two categories) across the 233 whole group (N = 96). We report effects at a threshold of p < 0.001 at the voxel level and p < 234 0.05 family wise error (FWE) corrected for multiple comparisons at the cluster level (denoted 235 p FWE_corr ). We also separately report these category effects in each group (N = 24) at a threshold 236 of p < 0.001 at the voxel-level, non-corrected at the cluster-level to provide the reader with the 237 full information on activation patterns in each group. 238 To deepen our analyses of the differences between groups in their response to words (vs. 239 fixation), we focused our analyses on a mask of voxels corresponding to the word-specific 240 activation across all children determined as by the Words > [Faces, Houses] contrast (p < 0.001 241 voxel level and p FWE < 0.05 cluster level). We performed an ANOVA with Language (French 242 vs. Chinese) and Dyslexia (control vs. dyslexic) as between-subject factors. To provide readers 243 with full information, we report these results at the threshold of voxel-level p < 0.001, non-244 corrected at the cluster-level. 245 246

Brain-behavioral correlation analysis 247
The above analyses were based on a binary distinction between dyslexics and controls. To 248 better understand the effect of reading on brain activations, we replaced the categorical 249 distinction (controls vs dyslexics) by a continuous variable of reading performance (standard 250 scores in dyslexia screening measures) across both Chinese and French children (N = 96) and 251 studied the correlation between this variable and word activation across the entire brain. We 252 report effects at a threshold of p < 0.001 at the voxel level and p < 0.05 FWE corrected for 253 multiple comparisons at the cluster level.  Table S4), and averaged the reported coordinates (x, y, z respectively) to 262 create a 6-mm-radius sphere of the averaged locus as a ROI (Fig 3). 263 Due to the limited number of published neuroimaging studies of Chinese dyslexia, no 264 meta-analysis was available to summarize the available evidence into a pooled estimate. 265 However, atypical activation in a lateral prefrontal region within BA 9 has been reported in 266 Chinese dyslexics [30,31] and this region was repeatedly found to be more involved in reading 267 Chinese than alphabetic languages [10,45,46]. Besides, previous studies also often reported 268 that Chinese reading networks are more symmetrical in the ventral visual system. We thus 269 included the foci in both left middle frontal gyrus and right occipital cortex that were reported 270 in several meta-analyses on Chinese typical reading [45][46][47][48] and created representative ROIs as 271 above (see Table S5). In total, we obtained 10 ROIs related to dyslexia in alphabetic languages 272 and 3 additional ROIs potentially related to Chinese typical reading and dyslexia (Fig 3). 273 We extracted the mean contrast-weighted beta values for the words vs. fixation contrast in 274 each ROI in each child and entered these values in an ANOVA with Language (Chinese vs. 275 French) and Dyslexia (controls vs. dyslexics) as between-subject factors. The false discovery 276 rate (FDR) multiple-comparison method was implemented to take into account the multiple 277 ROIs. We did the same analyses in the same ROIs for the contrasts of faces vs fixation and 278 houses vs fixation. The FDR corrected p value is denoted as p FDR_corr . 279 280

Anterior-to-Posterior ROI Analysis in the visual cortex 281
To visualize whether the anterior-to-posterior and lateral-to-medial organization of 282 activation to words (or to other categories) in the ventral visual cortex was similar between 283 dyslexics and controls among both Chinese and French participants, a sliding-window ROI 284 analysis was also conducted. We moved the ROI along the y-axis in the left hemisphere with 285 constant x and y coordinates (respectively x = -48 and z = -16). Along the y-axis, six 6-mm-286 radius spheres were regularly spaced along the anterior-posterior axis, with the center 287 positioned at y = -73, -64, -55, -46, -37, -28 respectively. The x-axis and z-axis were set based 288 on the peak of the word-specific activation (Words > [Faces, Houses]) in all participants. To 289 better characterize bilateral activation of ventral visual system in reading, counterparts of these 290 ROIs in the right hemisphere were also included in the analysis. 291 11 First, we examined the responses to words vs fixation along this anterior posterior axis in 292 an ANOVA with Language and Dyslexia as between-subject factors, ROI (6 y-axis position = -293   73, -64, -55, -46, -37, -28) and Hemisphere (left and right) as within-subject factors. Second,  294 we investigated the responses to the other two visual categories following the same logic. We 295 determined two x-axis sites (x = ± 39, x = ± 30) based on the location of the peak of Faces > 296 other categories and Houses > other categories in our participants (see Table 2) and performed 297 two separate ANOVAS with the same factors than above (Language, Dyslexia, 6 y-axis 298 positions and Hemisphere). In each ANOVA, we corrected for multiple comparisons using the 299 FDR method. 300 301

Individual peak analyses 302
To investigate whether the dyslexic children had a greater inter-individual variability in 303 brain localization, we conducted individually defined ROI analyses. We focused on those 304 regions showing significant main effects of dyslexia or language in the group activation analysis 305 (i.e. left FFG, MFG, precentral, STS, pSTG, and SPL). We searched for active voxels (Words > 306 fixation) in a sphere (radius = 12 mm) centered on the peak coordinates identified by the whole 307 group activations (Words > [Faces, Houses]). We eliminated voxels with z-value inferior to 1 308 and selected the 10 strongest activated voxels within the search area. We first derived the 309 individual center of mass of these voxels by averaging their x, y, z coordinates. We calculated 310 the distance between this center of mass and the group peak coordinates in each child. Second, 311 we averaged the beta values measured in these voxels to obtain the maximal activation in each 312 child. We then entered those measures into Language × Dyslexia ANOVAs to investigate 313 whether dyslexic children differed in peak location and activation intensity compared to control In the MVPA analysis, we also focused on the regions showing significant main effects of 320 dyslexia or language in the univariate activation analysis. We drew a sphere with 9-mm radius 321 12 centered on the averaged coordinates of foci reported in meta-analyses, and then intersected 322 each sphere with the whole group activations (Words > [Faces, Houses]) to obtain a group mask 323 (~3052 mm 3 ). All the voxels within the mask were included for MVPA analysis. 324 Secondly, within each ROI, we used the correlation-based multi-voxel pattern analysis to 325 quantify the within-subject reproducibility of activation patterns. Thus, we calculated the 326 correlation coefficients between the pattern of response evoked by words relative to fixation 327 during the first run and the pattern of response evoked by each category (words, faces and 328 houses) relative to fixation during the second run in each subject. The correlation coefficients 329 were further converted into Z-scores. For each ROI, we then entered these correlation 330 coefficients into an ANOVA with Language (Chinese vs French), Dyslexia (control vs dyslexia), 331 and Condition (within-category correlation, e.g. words with words, vs between category 332 correlation, e.g. words with faces, words with houses, and faces with houses) as factors. 333 We performed a similar MVPA analysis in the bilateral face fusiform areas to investigate 334 whether dyslexic children showed reproducible activation patterns to faces. Bilateral face ROIs 335 were spheres with a 9-mm radius centered on the reported peak coordinates in the face-selective 336 activation in previous studies (left [-39, -45, -18], right [39, -45, -18]) [49]. We intersected each 337 sphere with the whole group activations (Face > [Words, Houses]) to obtain a mask (~3052 338 mm3). All the voxels within the mask were included for MVPA analysis. We then calculated 339 the correlation coefficients between the pattern of response evoked by faces relative to fixation 340 during the first run and the pattern of response evoked by each category (faces, houses and 341 words) relative to fixation during the second run in each subject. 342 Note that six French controls and three French dyslexics finished only one run of the visual 343 task, so that they were not included in this MVPA analysis. For those children who had 4 runs, 344 we used their first two runs to calculate the correlation coefficients between runs. The FDR 345 multiple-comparison method was again used as a correction for the multiple ROIs tested. 346 347 Results 348

Behavioral Results 349
As shown in Table 1, the four groups were matched in age and sex. Because the reading 350 tests are not directly comparable due to the writing differences, we did not compare French and 351 13 Chinese scores directly. However, both Chinese and French dyslexia performed worse than their 352 respective controls in the dyslexia screening measures (CCRT for Chinese children, t = 18.03, 353 p < 0.001 and "L'alouette" for French children, t = 14.66, p < 0.001). More information on the 354 behavioral profile of Chinese and French children is presented in the supplementary material 355 (see Table S1

Category-specific activations 363
We first examined the brain activations to each category (i.e. Words, Faces, and Houses) 364 relative to the other two categories among all participants (see Fig 1A and Table 2 Table 3). 384 When a categorical distinction was made between normal readers and dyslexics, no 385 significant cluster differed between these two groups in either direction (dyslexics > normal 386 readers and normal readers > dyslexics), when analyzing either the words > fixation or the 387 words > others contrast. However, a few voxels reached the voxel threshold (p = 0.001) in 388 regions corresponding to the above, more sensitive correlation analysis, including the left 389 fusiform gyrus, left precentral and left superior temporal sulcus (see figure S7A). reduced activation in dyslexics relative to controls. Importantly, all of these effects were 417 significant within each language group (see Fig 4). Keeping constant x = ± 39 and z = -16, we also studied the activation to Faces vs fixation 444 along the y-axis (ranging from -79 to -22). We observed a significant Hemisphere × ROI 445 interaction. This effect was due to greater right than left face activation at several y coordinates. 446 Besides, the main effect of dyslexia reached significance, due to a lower activation to faces in 447 dyslexics compared with controls bilaterally and in both languages (see Fig S9C). 448 Along the medial house specific activation at x = ± 30 and z = -16, we similarly studied 449 the activation to Houses vs fixation. We found a significant triple interaction of Language × 450 Hemisphere × ROI. French children had greater right than left activation at each of the six 451 anterior-posterior y coordinates (all p FDR_corr < 0.005) while Chinese children had the same 452 pattern only at four sites (y = -73, -46, 37, 28). We also observed a significant Dyslexia × ROI 453 interaction, with decreased activation to Houses in dyslexics in several sites (see Fig S9D). We computed the similarity separately for within-category patterns (words in run 1 and words 484 in run 2) versus between-category patterns (average of words-faces, words-houses and faces-485 houses, each in run 1 versus run 2). If the representation of words is more stable than that of 486 other categories in these regions, then we should observe a significant main effect of condition 487 (a greater correlation within than between-category). If the activation pattern is less 488 reproducible in dyslexics than in controls, a significant interaction of condition × dyslexia 489 should be found. In all these regions, when pooling over all subjects, there was an overall 490 replicable pattern of activation evoked by words, as indicated by a significant main effect of 491 condition, with a greater correlation coefficient within than between categories (all p FDR_corr < 492 0.001) (see Fig 5 and Fig S10A). Crucially, we also observed a significant interaction of 493 Condition × Dyslexia (control vs dyslexia) in the left FFG (F (1, 83) = 10.14, p = 0.002, p FDR_corr 494 = 0.006) and in the left pSTG (F (1, 83) = 15.75, p < 0.001, p FDR_corr < 0.001). Post-hoc analysis 495 found that normal readers, but not dyslexics, exhibited a significantly similar pattern of 496 activation from one run to the next. Those results show that the above differences between 497 normal readers and dyslexics were not due to an artifact of group averaging, and that individual 498 dyslexics exhibited a genuinely less reproducible activation patterns in these regions (Fig 5). In the present study, we examined whether similar impairments in reading circuits were 509 observed in Chinese and French dyslexics. Our goal was to study whether the same 510 neurobiological mechanisms were involved in reading disorders, independently of the size of 511 the speech units mapped to characters, and of the complexity of the characters. We investigated 512 this question in 10-year-old children using a similar paradigm in both countries with a 513 minimally demanding task which was equally easy for everyone (i.e. detecting a star), in four 514 matched groups (French and Chinese × dyslexics and normal readers). 515 First, in a whole-brain analysis, we recovered the classical category-specific activations 516 for words, faces and houses in extra-striate visual areas across all participants but also in each 517 group ( Figure 1). Second, reading scores were correlated with the word activations in common 518 key-regions of the reading circuit (left VWFA, posterior superior temporal gyrus/sulcus, middle 519 frontal gyrus and precentral gyrus) but also in the right hemisphere (middle occipital and 520 fusiform gyri and precentral). Third, analyses based on ROIs from the literature confirmed these 521 results, and surprisingly, further identified a main effect of dyslexia in the left middle frontal 522 gyrus whose dysfunction was previously claimed to be specific to Chinese dyslexia [29,30]. In 523 all these analyses, no Language × Dyslexia interaction was significant, emphasizing common 524 neural anomalies in both languages. These results were replicated even when the best voxels in 525 these areas were chosen. However, we did observe some differences in activations depending 526 on the children's native language. Chinese reading tended to engage more symmetrical 527 activations in the visual system, with stronger activations in the right hemisphere than French 528 readers when we specifically tested the anterior-posterior organization of the fusiform region. 529 Chinese children also had stronger activations than French children in the left parietal region, 530 middle frontal region and posterior STG. 531 19 We concluded our analyses by examining the reproducibility of the activation patterns 532 between runs. The within-subject pattern of activity evoked by words was reproducible across 533 runs in normal readers in all key reading regions, underscoring that even in children, the reading 534 circuit is stable after 3 years of learning to read and can be reliably measured in a single fMRI 535 run. However, such was not the case for dyslexics, whose activity was significantly less reliable 536 in left fusiform and posterior superior temporal gyrus in both Chinese and French dyslexics. 537 We now discuss each of these results in turn. inter-cultural convergence should be expected. 558 In our study, although the main areas for reading were common to both groups, we also 559 observed modulations of the amplitude of brain activity within culturally universal brain 560 circuits. Chinese children had larger activations than French children in the (1)  Thanks to the individual peak location and intensity analyses, as well as the multivariate 606 pattern analyses, we could reject an alternative interpretation which, to the best of our 607 knowledge, was not explicitly tested in previous studies: the possibility that the hypoactivations 608 are an artifact of group averaging, solely due to greater inter-individual variability in the 609 localization of reading-related circuits in the dyslexic brain. Using individual peak, we observed 610 that the brain localization to words were not more dispersed among dyslexic participants than 611 among controls. Using MVPA, we showed that, within individual subjects, the activation 612 patterns in the VWFA in response to written words were less reproducible across runs in 613 dyslexics than in normal readers. This was solely the case for words, not for the other visual 614 categories. We did observe a slightly reduced activation to faces and houses in dyslexics relative 615 to controls, as previously reported in illiterate subjects [ Finally, we also observed an impairment in the left middle frontal gyrus as reported in 643 Chinese dyslexics by Siok et al. [29,30]. Activation at this location was modulated by both 644 reading score and language: overall, Chinese children had larger activation than French children 645 at this site, but in both languages, dyslexics also exhibited weaker activation than their controls: 646 effects of dyslexia and language on that region seemed to be additive. This result indicates that, 647 once again, an impairment in this region is not specific to Chinese but part of a universal 648 phenotype of dyslexia. 649 Unlike readers of alphabetic languages who can use grapheme-phoneme correspondences, 650 readers of Chinese must learn the phonology of characters as a whole and they may rely on 651 23 writing as a means for memorizing the large number of characters. This hypothesis is consistent 652 with the finding that writing skills could predict reading ability in Chinese children [77]. The 653 greater activation of the left MFG may reflect the greater reliance of Chinese children on writing 654 and the dysfunction of this region in Chinese dyslexia may reflect their impairment in linking 655 spelling with phonology through motor memories for writing [78]. Interestingly, novice readers 656 in alphabetic languages are also known to rely on a motor memory for hand gestures when 657 recognizing written words [79]. A recent study found that the motor representation can be 658 accessed automatically for subliminal words, in both Chinese and French adult readers [10]. 659 Our findings on children further suggest that the left MFG is likely to play a pivotal role in 660 successful reading acquisition that is independent of the writing system. 661

Conclusions 663
Thanks to several convergent analyses, we revealed that the neural anomalies underlying 664 developmental dyslexia are largely similar in French and Chinese readers. Across these very 665 different writing systems, the cultural invention of reading relies on similar brain resources. As 666 previously noted in an adult fMRI study [10], cultural variability is merely reflected in the 667 variable emphasis that different writing systems put on phonemes, syllables and whole words, 668 which in turn may modulate the severity of dyslexia and the degree of anomaly that can be 669 detected at different locations along the brain's reading circuitry.   Right superior temporal sulcus 57 -27 3 8.94e-10 6.02