Early reading skills and the ventral occipito-temporal cortex organization

Learning to read impacts the way the ventral occipitotemporal cortex (VOT) reorganizes. The postulated underlying mechanism of neuronal recycling was recently revisited. Neuroimaging data showed that voxels weakly specialized for visual processing keep their initial category selectivity (i.e., object or face processing) while acquiring an additional and stronger responsivity to written words. Here, we examined a large and diverse group of six-year-olds prior to formal literacy training (N = 72) using various data analysis techniques (univariate, multivariate, rapid adaptation) and types of stimuli (print, false fonts, houses, faces) to further explore how VOT changes and adapts to the novel skill of reading. We found that among several visual stimuli categories only print activated a wide network of language related areas outside of the bilateral visual cortex, and the level of reading skill was related to the strength of this activation, showing the development of the reading circuit. Rapid adaptation was not directly related to the level of reading skill in the young children studied here, but it clearly revealed the emergence of the reading network in readers. Most importantly, we found that the reorganization of the VOT is not in fact an "invasion" by reading acquisition-voxels previously activated for faces started to respond more for print, while at the same time keeping their previous function. We can thus conclude that the revised hypothesis of neuronal recycling is supported by our data.


Introduction
Learning to read impacts neural reorganization, especially in the language-related regions. This reorganization includes a significant tuning for print in the left ventral occipitotemporal cortex (VOT) and other cortical regions processing language (Brem et al., 2010;Chyl et al., 2018;Monzalvo & Dehaene-Lambertz, 2013;Pleisch et al., 2019). Importantly, the localization of regions sensitive to visual words is to a large extent culturally invariant in both adults (Rueckl et al., 2015), and children, despite small differences related to the orthographic transparency (Chyl et al., 2021;Feng et al., 2020). Functional segregation of the so-called "visual word form area" was recently proposed, with the posterioreanterior axis of the VOT involved in two different aspects of visual word recognition: a posterior part, responsible for visual feature extraction (perceptual), and an anterior part involved in integrating information from and to the language network (lexical; Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018). Another functionally distinct region of the VOT plays a key role in face perception and is known as a "fusiform face area" (FFA; Kanwisher & Yovel, 2006). However, the exact neural mechanism of the emergence of print-specificity is still a topic of ongoing debate. According to the neuronal recycling hypothesis (Dehaene & Cohen, 2007), groups of neurons that prior to reading acquisition were responsible for processing other visual categories, such as faces, start to respond selectively to visual letters and words after reading acquisition. Consistent with this hypothesis, among 5-and 6-year-olds with varying reading performance, the spatial extent of a letter-selective VOT region was positively correlated with reading skill and negatively correlated with the extent of an adjacent face-selective region (Centanni et al., 2018). A study on illiterate and literate adults further showed that reading ability is associated with reduced activation to faces in the left VOT and increased right-hemispheric activation for faces (Dehaene et al., 2010). However, the extent and pace at which the left hemisphere takes over print processing is still a subject of debate.
Other evidence does not support the idea that reading "invades" the cortical region that originally showed selectivity for faces. A behavioral study investigating whether this developmental co-dependency has a cognitive cost found that learning to read does not affect face recognition ability in young children (Kü hn et al., 2021). The hypothesis that face and word recognition share the same cortical circuits was challenged by a study that examined longitudinally ten 6year-olds during the first critical year of reading acquisition. It was found that voxels in the VOT that started to respond to words along with the acquisition of reading also kept their initial category selectivity for objects (tool) processing (Dehaene-Lambertz et al., 2018). Based on this, a revised cortical recycling model for reading was proposed, suggesting that reading "invades" weakly specialized voxels in the VOT cortex, placed in close proximity to the faces selective voxels, leaving previously stabilized cortical responses unchanged. In line with this idea, a recent longitudinal study found that both word-selective regions expand and become more selective from childhood to adolescence, but it was also the case for the face-selective regions (Nordt et al., 2021). So far, only one study examined VOT selectivity in pre-readers and found that face-selective responses in the left VOT positively correlate with pre-readers' rapid automatized naming speed (RAN; an important predictor of reading, Norton & Wolf., 2012) measured at the same time point, as well as with their reading fluency measured two years later (Liebig et al., 2021).
To further examine early reading-induced VOT changes, the present study investigates pre-school children's VOT responses to faces and words using the fMRI rapid adaptation (fMRI-RA) design and both univariate and multivariate analyses. FMRI-RA paradigms are capable of probing neuronal tuning in fMRI (Grill-Spector et al., 2006) different from the average blood oxygenation level-dependent (BOLD)-contrast response to stimuli (e.g., faces versus objects). Repeated presentation of the same stimuli leads to a smaller BOLD-contrast response reflecting the similarity of the neuronal activation patterns, and this effect is known as the repetition suppression effect. As such fMRI-RA provides a tool for characterizing functional properties of neural populations (Grill-Spector & Malach, 2001) and has been applied to examine the specialization of VOT regions for face (Jiang et al., 2006) and word processing (Glezer et al., 2016;Glezer et al., 2009) in skilled adult readers. The fMRI-RA was also applied when testing readers with dyslexia, both adults and children. In this case, the authors observed a significant diminished cortical adaptation to spoken words, written words, visual objects, and faces, compared to controls (Perrachione et al., 2016). Specifically, adaptation for speech was found to be reduced in the frontotemporal language network, while for written words it was reduced in the left VOT and frontotemporal areas. Recent EEG findings support the idea of diminished adaptation to speech in dyslexia, showing that implicit phonological learning in the frontotemporal language network builds up at a slower pace compared to typical readers (Zhang et al., 2021). Interestingly, reduced VOT adaptation in individuals with dyslexia was also found for the repetition of nonlinguistic objects and faces (Perrachione et al., 2016). In the same study, better reading skills in adults and children with dyslexia were associated with greater repetition-induced neural adaptation to speech in the left planum temporale, though no such relations were observed for the visual stimuli in the VOT (Perrachione et al., 2016). The dysfunction of rapid neural adaptation was hypothesized to underlie the impaired reading development in dyslexia. These results suggest that the development of reading skills might be related to neural adaptation, but this phenomenon was never studied in emerging young readers and prereaders. By employing this paradigm, we aimed to probe the specialization of the VOT and language-related areas, and see how the adaptation effects are related to reading level.
Even though neural adaptation analysis provides important information on the neural response for specific stimuli, a direct comparison of fMRI adaptation effects and multivariate pattern analysis (MVPA) in the visual cortex led to the conclusion that MVPA might be more sensitive to subtle differences in stimulus representation, even though the two methodological approaches provide highly correlated results (Sapountzis et al., 2010). MVPA also looks at the contributions of multiple voxels, not by averaging them as in the case of univariate approach, but rather using the patternclassification approach to analyze the pattern of activation as a whole (Norman et al., 2006). Activation patterns in the VOT in response to written words (but not nonlinguistic stimuli such as faces) examined using MVPA were recently shown to be less reproducible in poor readers than in typical readers, regardless of the language they speakeFrench or Chinese (Feng et al., 2020).
Here we extended this line of work by studying both adaptation effects and multivariate patterns in young children's VOT responses to different visual categories. We analyzed data from 72 kindergartens, who were never formally taught to read, and tested visual category selectivity in relation to their emerging reading skills. Because 24 (out of the 72) children from the studied group already acquired basic reading skills, we matched them to 24 (out of the 72) prereading peers and performed group comparisons. We investigated patterns of neural engagement and information coding across the VOT during the processing of real words, letterlike symbols, and the nonlinguistic categories: faces and houses. We expected developmental changes in the reading network also outside of the VOT, e.g., in dorsal brain regions such as the left precentral gyrus (PrCG), inferior frontal gyrus (IFG), and superior temporal gyrus (STG; Chyl et al., 2018;Brem et al., 2010;Monzalvo & Dehaene-Lambertz, 2013). Better reading skills were predicted to be related to stronger activation for word processing in the whole reading network, as shown in previous literature Turkeltaub et al., 2003). Neural adaptation to stimulus repetition was, until now, studied only in older children, where diminished adaptation in dyslexia might be explained by lower reading level or reduced experience. Here, studying a younger group of children with varying reading levels, we aimed to answer the question whether individual differences in reading skills are related to the differences in the adaptation effect for both linguistic and non-linguistic stimuli (Perrachione et al., 2016). We expected to see the adaptation effects for words in the language regions only in Readers, as for Prereaders print is not meaningful yet. We expected to observe a possible right hemispheric shift in face processing in readers, and reduced face processing in the left VOT, proportional to an expected increase in word processing in the latter region (Dehaene et al., 2010;Dehaene-Lambertz et al., 2018;Feng et al., 2020;Monzalvo et al., 2012). Finally, employing MVPA, we aimed to look at the reproducibility of the brain response for words and faces in the perceptual and lexical areas of the VOT (Lerma-Usabiaga et al., 2018), and FFA (Feng et al., 2020). In particular, we were interested to see how reproducibility of the brain response is related to reading skills, and what is the relationship between words and faces processing. MVPA, in contrast to the adaptation analysis, would provide us with information not only on the magnitude of the difference between the groups but also on the similarity of more spatially dispersed activation patterns within the ROIs. MVPA ROI-wise analysis will disentangle whether the left VOT takes over the word processing via destabilization of neural response for other categories (neuronal recycling hypothesis), or, alternatively, without any loss to previously established processing patterns (revised neuronal recycling hypothesis). If the former is true we predict higher stability for words in readers than prereaders in the VOT and at the same time higher stability for faces in prereaders than readers in the FFA. Finding higher stability for words in readers than prereaders without any differences for faces would support the revised hypothesis.

Participants
354 children took part in a screening session prior to the study. All participants attended the last grade of kindergarten, which is obligatory for children aged 6 years, before elementary school in Poland and does not include reading instruction. Children who were not able to read a single word or a pseudoword in a reading task, and met other inclusion criteria: being born at term (>37 weeks), right-handedness, no history of neurological illness or brain damage, and no ADHD symptoms, were invited to participate in the project (N ¼ 98). For the fMRI session, we invited as many subjects as possible, given time and money restrictions. Nonetheless, our sample size is larger than the average N from the similar studies. Eightyeight children (M age ¼ 6.74, SD ¼ .37, min ¼ 5.5, max ¼ 7.38, 36 girls, 52 boys) completed the fMRI task. Of these children, 13 moved too much during the fMRI procedure as defined by Art Repair and 3 additional children had poor quality single subject masks as detected by Mask Explorer (see fMRI data acquisition and analyses), leaving 72 children (M age ¼ 6.74, SD ¼ .39; min ¼ 5.5, max ¼ 7.38, 33 girls, 39 boys) for the analyses. Even though all children were recruited as pre-readers, about onethird developed initial reading skills before the fMRI session. That is why for group analyses we selected two groups of children that differed in reading level: Readers, who were able to read 17 or more words or pseudowords in the administered reading tasks (N ¼ 24), and Prereaders (N ¼ 24), who read less than 4 words or pseudowords (2 letter-long words and pseudowords, that were provided first). The study was approved by the Research Ethics Committee at the SWPS University of Social Sciences and Humanities in Warsaw and was carried out following the provisions of the World Medical Association Declaration of Helsinki. Parents of children who participated signed an informed consent form and children agreed verbally.

2.2.
Word and pseudoword reading tasks The task was part of a standardized test battery for the early detection of reading problems, specifically designed to diagnose children starting school education (Szczerbi nski & Pelc-Pę kala, 2013). The child's task was to read aloud as many words as possible in 30 s. Two parallel sheets of increasing word length were provided, starting with 2-letter words. The instruction was as follows: "Read as many words from this list as possible. Read from top to bottom and do not skip any word. Read as quickly and as correctly as possible. You will have only half a minute for that.". After 30 s the experimenter said: "Stop", and repeated the procedure for the second list. The total number of correctly read words in a minute (the sum of the two lists) was calculated for each child. The pseudoword reading task had an analogous procedure, c o r t e x 1 6 0 ( 2 0 2 3 ) 1 3 4 e1 5 1 with two parallel pseudoword lists of increasing length. For group selection, we used a sum of word and pseudoword reading for a cut-off between Readers and Prereaders (see above); for correlations, we used the word reading score. For a detailed description of other applied measures, see Supplementary Materials 1.

Behavioral measures
During the initial screening session, children performed a letter knowledge task, word reading and pseudoword reading tasks, phoneme analysis and phoneme deletion tasks (the Decoding Test; Szczerbi nski & Pelc-Pę kala, 2013). After the screening, prereaders were invited for the second behavioral session. During this measurement point, orthographic awareness was tested with the orthographic awareness task (Awramiuk et al., 2015). Vocabulary knowledge was assessed with the Picture Vocabulary Test: Comprehension (Haman et al., 2012). Rapid automatized naming was tested with object and color naming subtests (Szustrowa & Jaworowska, 2003;Jaworowska & Anna, 2013). Children also performed tests of digit and syllable repetition. Non-verbal IQ was assessed with the Cattell Culture Fair Intelligence Test (CIFT; Ko c-Januchta et al., 2013). Visual attention was tested with a subtest of the IDS Intelligence Scale measuring visual selective attention (Jaworowska et al., 2012) and Corsi Blocks task measuring visuospatial memory span (Fischer, 2001). Grammatical development was measured with a sentence repetition task (Smoczy nska et al., 2015). Numerical processing skills were tested with a magnitude comparison task (Vanbinst et al., 2018). Additional in-house tests of phonological awareness were also provided (See Supplementary Materials). On the day of the fMRI session, participants again performed the Decoding Test with the subtests of letter knowledge, word reading, pseudoword reading, and parallel versions of the phoneme analysis task (Szczerbi nski & Pelc-Pę kala, 2013). In addition, several questionnaires were administered to the participant's parents. We asked about their education, reading history (ARHQ; Lefly and Pennington 2000), and reading habits (parents' beliefs and personal habits) as well as about their children's pre-reading habits (e.g., shared reading activities, going to the library together with parents). For all the in-house questionnaires and descriptions of the behavioral tests, see Supplementary Materials 2.
The behavioral data were analyzed in Python. The results reported in the tables are corrected with Holm-Bonferroni sequential correction procedure (Holm, 1979). For further details, i.e. demographic and behavioral data of the whole group and selected subgroups of readers and prereaders, see Table 1.

Behavioral measures analysis
Pearson correlations were used to investigate the relationship between word reading and demographic information across the whole group. We also performed Pearson correlations between word reading and letter knowledge and other reading related skills. Holm-Bonferroni correction was used to correct for multiple comparisonseeight demographic measures in total were considered as well as 14 different behavioral tasks.
Next, we compared behavioral performance between readers and pre-readers with the use of two sample t-tests. Again, Holm-Bonferroni sequential correction was used to correct for the number of tests/

fMRI task and procedure
Before the neuroimaging procedure, children were familiarized with the task and MRI environment in a mock scanner. During the fMRI task, children were instructed to lay still, carefully observe images appearing on the screen, and press the left button on the response pad after a red dot appeared on the image. The task was composed of four runs (5.13 min each, 20.5 min in total), with at least one break offered to the child after the first two runs. Stimuli were organized in the mini blocks in which 6 items were presented, with a visual stimulus presentation time of 700 ms and a 500 ms of interstimulus interval. There were five different categories used: (1) 3e5 letter, highly frequent Polish nouns written in courier font (2) the same words written in letter-like BACS false font (Vidal et al., 2017), (3) neutral children faces (Meuwissen et al., 2017), (4) houses dissimilar to faces (Filliter et al., 2016), and (5) objects known from everyday life (Brodeur et al., 2014). Shine toolbox was used to equalize luminance and contrast of the visual stimuli (Willenbockel et al., 2010). Our univariate analysis showed that the object category was an outlier; especially the condition Objects_mix evoked much stronger activation than any of the other categories. Most likely, this relates to the fact that Objects_mix stimuli had a much larger intra-category variability, e.g., in shape and size than the other four categories. That is why in this paper, the Objects condition was excluded from further analysis.
Each of the four runs consisted of 30 blocks. In each block, 6 items were presented. After each block, a blank screen was presented for 1e3 sec. In the Repetition (rep) blocks, the same image was presented six times in a row. In the Mixed (Mix) blocks, six different items from one of five categories were presented. The number of blocks for each condition was balanced in each run (i.e., 3 Adapt and 3 Mix blocks from each category were shown). In each run, in 10 out of 30 blocks, a targetered doteappeared on the screen, simultaneously with one of the six stimuli (in total, 5.5% of the stimuli in the run were accompanied by the target dot). The target could appear in any position in the block besides the first. The block type and number of each category in which the target appeared were balanced in each run. The experiment was programmed in Presentation Software (Neurobehavioral Systems, Albany, CA). See Fig. 1 for block types and categories used in the experiment.

fMRI data acquisition and analyses
The fMRI data were acquired on a 3-T Siemens Trio scanner (Siemens, Berlin, Germany) using a whole-brain echo-planar imaging sequence with a 32-channel head coil, 32 transverse slices providing whole-brain coverage, 3.5 mm slice thickness, 2 sec repetition time, 30 ms echo time, 90 flip angle, 192 mm 2 field of view, voxel resolution ¼ 3.5 mm 3 , 64 Â 64 matrix. Anatomical data were acquired using the same 32-channel c o r t e x 1 6 0 ( 2 0 2 3 ) 1 3 4 e1 5 1 coil with a T1-weighted sequence with 176 slices, 1 mm slice thickness, 2.53 sec repetition time, 3.34 ms echo time, 7 flip angle, 256 Â 256 matrix, 1 Â 1x1 mm 3 voxel size. fMRI data preprocessing and analyses were performed using Statistical Parametric Mapping (SPM12, Welcome Trust Center for Neuroimaging, London, UK), run on MATLAB R2016b (MathWorks Inc). Images from the 4 runs were realigned to the mean functional image. Then structural images from single subjects were coregistered to their mean functional image. Coregistered anatomical images were segmented using pediatric tissue probability maps (Wilke et al., 2008), and normalized. Functional images were normalized using the composition of flow fields and a group-specific template and spatially smoothed with an 8 mm isotropic Gaussian kernel (univariate analysis only). The data were modeled for each run, using the canonical hemodynamic response function convolved with the experimental conditions and fixation periods. Besides adding movement regressors to the design matrix of each subject, the Art Repair toolbox (Mazaika et al., 2007) was used to interpolate motion-affected volumes (percent threshold ¼ 4 (% variation in the global intensity), scan-to-scan movement threshold ¼ 1.5 mm, total movement threshold ¼ 5 mm), and the ART toolbox was used to reject motion-affected volumes surpassing the movement threshold of 3 mm and a rotation threshold of .05 radians. If more than 20% of volumes were motion-affected, a child was excluded from the analysis. In the final group of 72 participants, in each child, on average 95% of volumes were of good quality. Mask Explorer (Gajdos et al., 2016)   c o r t e x 1 6 0 ( 2 0 2 3 ) 1 3 4 e1 5 1 was used to additionally inspect the quality of the individual masks computed at the 1st level of analysis, with a cut-off point of 90%, i.e., at least 90% of the template was required to be covered by the mask. Of the larger original sample, 11 children were excluded based on excessive motion, 3 because of bad masks, and 2 children met both exclusion criteria. The exclusion criteria used in this study were established prior to the analysis and used in our previous studies.
In the univariate analysis, the general linear model approach was used to analyze the data acquired in the four runs for each participant. Contrasts of the category_mix against baseline (overall activation masks) were computed to show the activity for the given category. Moreover, to search for the category-specific activation for faces and words, we contrasted the categories with each other: words_-mixebacs_mix, faces_mixehouses_mix, and, additionally, words_mixehouses_mix and faces_mix. To examine neural adaptation effects for each category, contrasts of categor-y_mixecategory_adapt were computed. On the whole-brain level, one-sample t-tests were used to test the main effects in the whole group of children. Multiple regression analyses with the number of words read per minute were used to investigate individual differences in brain activity related to the level of reading skill in the whole sample. Two-sample ttests models were built to directly compare two matched groups of Readers and Prereaders. All univariate analyses are reported at a significance level of p < .005, corrected for multiple comparisons using False Discovery Rate (FDR) cluster correction p < .05.
In the multivariate analysis, similar general linear models on the first level of analysis were created, but separately for each run, and using non-smoothed data, similarly to previous studies (e.g., Feng et al., 2020). Motion-affected volumes were not interpolated, but otherwise we applied the same procedure and criteria as those of the univariate analysis, including using the ART toolbox to reject motion-affected volumes surpassing the movement threshold of 3 mm and a rotation threshold of .05 radians. Neither spatial smoothing nor interpolation were used in this analysis, as we aimed to perform only minimal preprocessing of the MVPA-analyzed data. To investigate the fine-grained representation of words and faces, multivariate analysis was performed in 6 VOT regions of interest (ROIs) localized based on previous functional MRI studies using MNI space (x, y, z). The first two ROIs were based on the averaged stereotaxic coordinates of 'lexical' [LEX: À41.54, À57.67, À10.18] and 'perceptual' [PER: À39.44, À71.52, À7.98] left VOT clusters in a previous study investigating functional and structural segregation of the left VOT (Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018). The coordinates of the third ROI were based on the face-selective activation [FFA: À39, À45, À18] found in previous studies (Downing et al., 2006;Feng et al., 2020). Similarly to other studies, we also defined their right-hemispheric homologues for all three ROIs. ROIs were built as 7 mm spheres around the given coordinates using the Marsbar toolbox (Brett et al., 2002) implemented in Matlab. The size of the ROIs was selected to provide maximum coverage of the VOT and avoid overlaps. Then, we intersected each ROI with the overall activation mask [words_mix þ faces_mix þ Houses_mix þ Bacs_mix] to obtain a fair representation of responsive voxels. PER and LEX ROIs were fully included in the mask (181 voxels each), but after the intersection with the mask, the FFA ROI was reduced to 122 voxels in the left hemisphere and 109 voxels in the right hemisphere.
In the following step of the MVPA analysis, within each ROI we quantified the within-subject reproducibility of the patterns of activation, by calculating in each subject, the correlation coefficients between the pattern of responses evoked within the category (e.g. words_mix) vs. between all the visual categories (words_mix, bacs_mix, houses_mix and faces_mix). We computed correlations between the conditions of interest (words_mix and faces_mix) relative to the baseline across all runs (1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, 3 and 4), and averaged the coefficients, following the procedure described in the previous paper (Feng et al., 2020). Only words and faces were used for computing withincategory correlations, as the other two conditions were considered as control conditions, used for the betweencategory correlations (i.e., words_mix with faces_mix, face-s_mix with houses_mix, houses_mix with bacs_mix etc.). The correlation coefficients were converted into Z-scores and averaged. Scores of within_mix for words and faces, as well as the between_mix score, were entered into ANOVA repeated measures models and used for Pearson correlations with the number of words read per minute. Holm-Bonferroni correction was implemented to correct for the number of ROIs (Holm, 1979). Significant clusters were labeled using the Automated Anatomical Labeling (AAL) atlas implemented in the Python package AtlasReader (Notter et al., 2019). All glass brain figures were prepared with the Python package Nilearn (Abraham et al., 2014).

2.7.
Data/code availability statement Behavioral data, ROI data, masked ROIs, parameters of the items used in the fMRI experiment, experimental Presentation protocol, SPSS dataset/script used in the ANOVA analysis, as well as fMRI statistical maps and SPM scripts are available online (https://osf.io/tp5uf/). The script used for the MVPA analysis was generously shared with us by dr Xiaoxia Feng and is not available on our osf. The ROIs can be downloaded and visualized as the Neurovault Collection (https:// neurovault.org/collections/12528/). The policy of the Laboratory of Brain Imaging of the Nencki Institute and informed consents signed by the Participants' parents permits external sharing of individual brain images only following completion of a data sharing and collaboration agreement. Accordingly, readers seeking access to the data should contact the corresponding author, and access will be granted after signing a data sharing and collaboration agreement. No part of the study procedures or analyses was pre-registered prior to the research being conducted. We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study in the Methods section.

Behavioral results
During the time that passed between screening and the MRI session (M in months ¼ 4.63, SD ¼ 2) children on average improved their reading and reading-related skills, as shown by the paired sample t tests: letter knowledge (t ¼ 9.03, p < .001), digit knowledge (t ¼ 3.47, p < .001), word reading (t ¼ 7.75, p < .001) and phoneme analysis (t ¼ 6.28, p < .001). In the following part of this section, in case the tests were administered both upon screening and the MRI session (letter knowledge, word reading, phoneme analysis), only tests administered during the MRI session are reported, and marked with "^" in Table 2. Among demographic measures, only the scores of the "child pre-reading habits questionnaire" (see supplementary materials) correlated positively with reading scores (r ¼ .43, p < .001). We found that letter knowledge was strongly correlated with word reading (r ¼ .68, p < .001), and pseudoword reading (r ¼ .70, p < .001), but also with phoneme analysis (r ¼ .60, p < .001), all measured on the same day as the MRI scan. A significant correlation with letter knowledge was found also for children's scores on the phonological awareness tasks based on alliterations (r ¼ .51, p < .001), and the sentence repetition task (r ¼ .42, p < .001), measured prior to the scanning. Word reading showed similar correlations with the other tests (sentence repetition: r ¼ .57, p < .001; phonological analysis: r ¼ .59, p < .001; alliteration task: r ¼ .48, p < .001). For the correlation matrix, see Supplementary Materials 3.
Even though demographically Readers and Prereaders groups were very similar (the only differences found between the groups was in children's pre-reading habits; see Table 1), they differed in many reading-related skills, such as letter knowledge, phonological awareness, vocabulary, and verbal working memory (Table 2).

Univariate analysis
We first examined the brain activations to each category, in the mixedebaseline, and the mixederepetition contrasts. Fig. 2 and Table 3 report the main condition effects (one-sample ttests), across the whole group of 72 children. Results show that for the non-linguistic stimulus categories (bacs, faces, and houses), the only region consistently activated in the c o r t e x 1 6 0 ( 2 0 2 3 ) 1 3 4 e1 5 1 mixed conditions, was the bilateral ventral occipitotemporal cortex. This was also the region that showed a significant repetition suppression effect for these categories, as demonstrated by the mixerepetition contrasts. For the words condition, areas activated by both mixed words and susceptible to the repetition suppression effect, also belonged to the broader language network, encompassing frontal, temporal, and parietal regions (see Table 4).
In the next step, we aimed to obtain the category-specific activations by computing words_mixebacs_mix, and face-s_mixehouses_mix contrasts (see Table 5). Across the whole group, including also prereaders, the contrast of words_-mixebacs_mix showed only a small difference in the bilateral Supplementary Motor Area (see Fig. 3). Hence we additionally performed a more lenient category comparison, words_mixe(houses_mix and faces_mix), to investigate word-related activation. This comparison showed word-related activity in predominantly left-lateralized temporal, frontal, and parietal regions of the language network. Additional comparisons of faces_mixehouses_mix showed activation of occipitotemporal regions and smaller clusters in the frontal areas. Finally, faces_mixe(words_mix and bacs_mix) followed a similar pattern but resulted in more robust results (see Fig. 3).
To explore the inter-individual variability of brain response to words and faces in relation to varying reading skills, independent of age, we performed a series of multiple regressions. We correlated the number of words read per minute with all contrasts of interest explored in the previous analyses, controlling additionally for age. The only significant positive correlations were increased activations for words_mix or word-specific contrasts (words-mixebacs_mix; words-mixefaces_mix and houses_mix) in the left VOT and precentral gyri in the better readers. Most of the clusters survived the correction for age. There were no significant correlations with reading skill for face processing. For adaptation effects (mixerep) we found a negative correlation between reading skill and the adaptation effect for faces in the left superior temporal gyrus; nonetheless, this effect disappeared after the correction for age (see Table 6 and Fig. 4).
In the next step, we compared directly readers to prereaders. Significant differences were revealed in overall activation to words and faces (words_mix and faces_mix vs. baseline), as well as category specific activation for words (words_mixe(faces_mix and houses_mix)), and faces (face-s_mixe(words_mix and bacs_mix)). Readers engaged the left VOT to a greater extent than Prereaders for the words_mix condition, with additional effects found in the left SMA and superior frontal regions for the words_mixe(faces_mix and houses_mix) contrast. At the same time, Prereaders activated more strongly left subcortical areas and the right FG and  Cerebellum for the faces_mix condition. Additionally, also for the contrast of faces_mixe(words_mix and bacs_mix), Prereaders activated a similar left SMA/Superior Frontal cluster more than Readers (see Table 7 and Fig. 5).
In the analysis of adaptation effects (mixerepetition), no significant differences were found between the groups.

Multivariate analyses
To further investigate more subtle group differences in distributed patterns of activity to words and faces, we built three separate ANOVA models to explore effects related to words (1), faces (2), and comparison of words and faces (3). We wanted to check whether the exact ROI, Group, Hemisphere and Specificity impact the stability of the activity patterns, and in all three models, MVPA was performed within predefined VOT ROIs. Since we were interested to see how reproducibility of the brain response is related to reading skills, we focused only on the main effect and interactions with the Group factor.
The first analysis (ANOVA 1) focused on the stability of the pattern of responses for words. In SPSS, we computed a 2x2x2x3 repeated measures ANOVA (Bonferroni corrected) with Specificity factor (within_words, between_categories), ROI (FFA, LEX, PER) and Hemisphere (Left, Right) as within-subjects factors, and Group (Readers, Prereaders) as a between-subject factor. We found a Group*Specificity*ROI interaction (F(2,92), F ¼ 3.863, p ¼ .024), which demonstrated that in both LEX and PER, the difference between the Specificity (within words > between categories) was significant only in Readers (LEXep ¼ .024; PERep ¼ .001), but not in Prereaders (see Fig. 6a). Next, we focused on the stability of the activation patterns for faces and computed a similar 2x2x2x3 model (ANOVA 2). No significant main effect or interaction with the Group was found in this analysis.
Finally, to answer the question if the stability of the brain response for words and faces differs, we computed 2x2x2x3 repeated measures model (ANOVA 3) with Category stability factor (within_facesebetween_categories, within_wordse between_categories), ROI (FFA, LEX, PER), and Hemisphere (Left, Right) as within-subject factors, and Group (Readers, Prereaders) as between-subject factor. We found a significant Category stability*ROI*Group (F(2,92) ¼ 3.541, p ¼ .033) interaction, which confirmed the different patterns of ROI-dependent stability of brain activity in Readers and Prereaders. While in Prereaders in all ROIs face processing was more stable than word processing (all p < .001), in Readers this was not the case in PER (LEX: p ¼ .014, FFA: p < .001), in which the stability of activity patterns to words and faces showed no significant difference In the last step, we examined how the stability of the patterns of brain activity across the 6 ROIs relates to children's individual differences in reading ability operationalized as a continuous measure of the number of words read per minute. We performed a series of correlations between the number of words read per minute and the correlations of activity patterns within_words and within_faces. The only significant correlation was between reading skills and within_words activation patterns in the left LEX ROI (r ¼ .299, p ¼ .011). The pattern of the results was the same after age was controlled for (left LEX, within_words: .299, p ¼ .011). These results, however, were no longer significant when corrected for the number of comparisons (six ROIs) with the Holm-Bonferroni sequential correction procedure.

Discussion
In this study, we aimed to investigate the patterns of neural activity across the ventral occipitotemporal (VOT) network while 72 6-year-old kindergarteners processed real words, letter-like symbols, and nonlinguistic categories of faces and houses. Given that the pre-existing organization of the VOT needs to largely adapt to the new, visual way of language processing, we focused on various aspects of variability related to children's reading status. We explored the sensitivity, specificity, and stability of cortical activity in response to the linguistic (words) and non-linguistic (faces) categories, as well as neural adaptation effects.

Behavioral and environmental effects
None of the studied children was formally schooled, but their level of reading and reading-related skills varied significantly. This variability to some extent can be explained by demographic factors. Thus, even though readers and prereaders were matched for SES (measured here with parental years of education), prereaders had less exposure to books and were not read to as often as children who already developed reading skills, as reflected by their lower score on the questionnaire measuring pre-reading habits (see Supplementary  Materials 1). This is also in line with previous studies showing that the home literacy environment is strongly correlated with children's language and literacy development (Puglisi et al., 2017;van Bergen et al., 2016).
Reading skills correlated significantly with many other reading-related skills which were superior in the reading group as shown by the direct group comparisons. These  reading-related skills, such as letter knowledge, phonological awareness, and sentence repetition (a proxy for verbal working memory and grammatical awareness) were previously reported as predictive for future reading skills and dyslexia, especially at a young age (e.g. Verwimp et al., 2020;Ozernov-Palchik et al., 2017). In semi-transparent Polish, early reading is predominantly supported by phonological abilities, as both these and our previous results showed (Chyl et al., , 2019. On the other hand, RAN was not found as significantly related to reading, as both correlations and group comparisons showed. This was somewhat surprising since RAN is considered a strong predictor for reading (Norton & Wolf., 2012). This absence of a relation with RAN scores might be attributed to the young age of the participants and the transparency of the Polish orthography (Furnes & Samuelsson, 2009, Dę bska, Łuniewska et al., 2021.  c o r t e x 1 6 0 ( 2 0 2 3 ) 1 3 4 e1 5 1 4.2.

Cortical processing of words
Even though the participants of the current study were never formally taught reading, and the highest reading performance was 35 words per minute, their brain processed printed words in a different manner than other visual stimuli. Whereas other visual categories only yielded consistent bilateral VOT activation, both standard univariate and fMRI-RA analysis showed that words engaged additional areas of the language network, encompassing frontal, temporal, and parietal regions. The engagement of the spoken language-related regions in word processing is a recognized marker of early reading, when children need to rely on phonology and slowly decode words letter-by-letter (Pugh et al., 2001), which is the most effective strategy in orthographies with transparent graphene-to-phoneme mappings (Chyl et al., 2021;Rueckl et al., 2015). Even though the examined group involved prereaders, most of the children were able to read at least a few words per minute (6.97 on average in the whole group), and the involvement of the broad linguistic network for processing  c o r t e x 1 6 0 ( 2 0 2 3 ) 1 3 4 e1 5 1 of simple single words is characteristic for this stage of reading development Dehaene-Lambertz et al., 2018;Monzalvo & Dehaene-Lambertz, 2013). When brain responses to the visual words were compared to the closely controlled artificial symbols (Vidal et al., 2017), the only significantly activated cluster was the bilateral supplementary motor area and superior frontal gyrus, showing that word-specific activations are still limited in the studied group. The involvement of these regions may be related to internal (subvocal) articulation, a strategy of reading used by younger (Brem et al., 2020;Chyl et al., 2018) or struggling (Chyl et al., 2019) readers. The adjacent regions, including the supplementary motor area, are known to be involved in the articulatory network (Brown et al., 2008;Hickok & Poeppel, 2007) and silent articulation (Richlan, 2014). The less stringent contrasts of the words_mixehouses_mix and faces_mix showed additional engagement of left parietal, temporal, and inferior frontal regions. This finding shows that at the earliest stages of reading development, artificial symbols mimicking characteristics of print are processed similarly to written words in the VOT. At the same time, the spoken languagerelated regions that are among the first to be involved during print processing in young readers  can be targeted with perceptually more dissimilar contrasts. This result points out the importance of careful consideration of the control condition; its specificity needs to be matched not only to the experimental condition and research question but also to the reading skill of the studied group. For example, in a study that inspired the current ROI selection, different control conditions evoked activity of different parts of the fusiform gyrus, but also showed differences in the gray matter structure (Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018) In this  previous work, the LEX ROI was sensitive for the contrast of words with pseudowords, consonants, and false fonts, while the PER ROI showed sensitivity for words compared to different types of checkerboards and scrambled words. In our ROI analysis, we did not find significant differences between the ROIs, which may be related both to still emerging specialization at this developmental stage, but also to the contrasts that were not designed to specifically target these areas. Furthermore, in another recent study young struggling readers had lower word selectivity in the left VOT not due to a lack of response to words, but due to high response to objects (Kubota et al., 2019; see also Dę bska et al., 2021). As we are currently continuing longitudinal measurements in the same children groups, in a future investigation it would be very interesting to characterize longitudinal changes in VOT and other brain regions for words, word-like symbol strings (Bacs), as well as each of the other visual categories. The responses evoked by words were to a large extent bilateral. Even though bilateral engagement for print can be observed also in adults, developmental findings show a decrease in right hemisphere activation for reading with increasing age and skill (Brown et al., 2005;Turkeltaub et al., 2003). Here, a significant right hemisphere engagement is in line with other studies in young children Monzalvo & Dehaene-Lambertz, 2013). Learning to read and increased reading performance is usually accompanied by increases in lateralization of occipitotemporal reading networks to the left hemisphere (Maurer et al., 2007;Pleisch et al., 2019;Turkeltaub et al., 2003). Some studies suggest that this lateralization is more dependent on reduced right hemispheric activation rather than increased left hemispheric activation (Seghier & Price, 2011). Perhaps in the young readers studied here, the next developmental step will be a leftward lateralization of the reading network expected in more proficient readers. For the other categories of word-like symbols, and non-linguistic faces and houses, both univariate activations and adaptation effects were restricted to bilateral VOT. This suggests that the processing of these categories was purely visual, and did not involve any linguistic processing.
Our results showed adaptation effects in the VOT for all visual categories, but we found no effects of reading status or correlations with reading skills for adaptation effects for any of the visual stimulus categories. In a previous study, older school-aged children and adults diagnosed with dyslexia) were shown to exhibit reduced rapid adaptation to stimulus repetition for different types of auditory and visual stimuli including written words (Perrachione et al., 2016). Because in Perrachione's study the deficit in rapid adaptation in children was studied with an auditory paradigm; perhaps a stable spoken language network could show the effects that are still too subtle for the developing visual language processing network. This might mean that at the early stage of literacy development, differences in reading are too small to modify visual stimulus selectivity. Alternatively, such differences may be specific to dyslexia, and not solely to the reading status, but this hypothesis needs to be verified in a longitudinal study (see also : Chyl et al., 2019). Such longitudinal studies will also be important to understand whether differences in neural adaptation in older children or adults with dyslexia are a consequence rather than a cause of reading failure. Fig. 6 e The results of multivariate analyses in Readers and Prereaders. a) interaction showing that only Readers present more stable processing of words in LEX and PER. b) interaction showing that in Prereaders face processing is more stable than word processing in all ROIs, whereas in Readers this difference is non-significant in PER.
The development of word-specific responses in the left VOT, IFG, and other regions of the language network in relation to the reading skill was found in the univariate analysis. Most of the effects found in the correlational analysis were independent of age. Similarly, group comparison with the groups matched for age showed that Readers had more activity in the left VOT and IFG, which complements the univariate correlational findings (for similar results, see Chyl et al., 2018). Correlational analysis performed on the MVPA results showed that only the stability of the activity patterns to words in the left LEX ROI were weakly related to the number of words per minute a child was able to read. This ROI, anatomically located in the middle part of the occipitotemporal sulcus, is associated with the long segment of the arcuate fasciculus (Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018), therefore might be the gateway connecting structurally the VOT circuit to other regions along the language network, i.e. angular, supramarginal and inferior frontal gyri, which we also found to be significantly activated for words. This result suggests that not only the left hemispheric LEX is a region where integration between the visual system and language network takes place (Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018), but also that this process takes place early in the reading development and prior to the formal reading training.

Cortical processing of faces
Face processing in the whole group of children took place in the bilateral VOT, which is in line with a previous study analyzing a similar group of kindergarteners. In this study, a non-linear hemispheric specialization for faces was shown, with more bilateral activation developing throughout childhood (Lochy et al., 2019). Here, the activations, in general, were largely bilateral, and we did not find any significant interaction of group and hemisphere in the processing of faces. Facespecific activations (faces_mixehouses_mix) were also present in the bilateral VOT, and also in the right IFG. As stated in a recent meta-analysis, right-hemispheric frontal involvement during face processing might be related to the mirror neuron response, automatic imitation, semantic aspects, or general control process (Muller et al., 2018). Regarding the adaptation effect, reduced activation in symmetric bilateral VOT was found for repeated vs. mixed faces in the whole group. This is partly in line with previous findings, which additionally reported adaptation in an "extended" face network (Haist & Anzures, 2016) including insula and amygdala in adults (Perrachione et al., 2016).
Similarly to the analyses focused on the relation between print processing and reading skills, a parallel set of analyses was performed for the faces_mix contrast, to check whether the newly learned reading skill may influence the processing of previously known categories. Prereaders displayed more activity for Faces_mix in left subcortical regions, the right cerebellum and the right fusiform gyrus than Readers. Moreover, the specific contrast with words and houses showed additionally the activity of the left frontal areas. Moreover, a negative correlation between the repetition suppression for faces and the number of words read per minute was found in the left middle and superior temporal gyrus. Even though this effect did not survive the age correction, the pattern of results suggest a link between reading skills and neural activity during face processing. Negative correlations between brain activity for faces and reading skills were found outside of the left VOT and in regions not specific for face processing (as reflected also by the one sample t tests), which suggests reorganization, but not the "invasion" of the regions occupied by face processing.

4.4.
Relationship between words and faces processing In our study we aimed to specifically test how reading changes brain activity for words and faces processing, and see whether the observed pattern supports the 1) original, or 2) revised version of the neural recycling hypothesis. Entering both categories of interest (words and faces) to one model enabled us to answer the question whether some or all of the ROIs in the VOT that are responsive for face processing at an early stage are "taken over" by the word processing (1). Alternatively (2), we could observe a "co-existence" of the development of faces and words related regions. We found that Readers started to show the stability in word processing to a level similar to the already stabilized pattern of face processing, which was comparable in Readers and Prereaders. It seems that the first region that shows the change in word stability is PER, previously described as responsible for processing of the perceptual characteristics of visual stimuli (Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018). Instead in LEX, the difference between word and face processing stability was smaller in Readers than Prereaders, perhaps reflecting the dynamics of word specific VOT development. Possibly, with higher reading expertise the stability of word processing in LEX would grow and follow the pattern observed in PER. The competition between the processing of words and faces in the same region of the VOT, suggested by the classical neuronal recycling hypothesis (Dehaene, 2005), does not seem to be supported by our data. In Readers, the stability of word processing grows, without the loss of earlier established processing patterns for, for example, faces.

Conclusions
To conclude, our findings seem to support the revised model of neuronal recycling rather than the standard model. The new visual categoryeprint starts to be processed in the cortex previously specialized for face processing, but leaves stabilized cortical responses unchanged (Dehaene-Lambertz et al., 2018). We show the emergence of the reading circuit, extending beyond the VOT to the language-related frontotemporal areas , that occurs in unschooled beginning readers. The extent of this network may vary depending on the contrast specificity (Lerma-Usabiaga, Carreiras, and Paz-Alonso 2018), with word specific activations largely dependent on children's reading skills. Rapid adaptation, found deficient in poor readers in a previous study (Perrachione et al., 2016), did not generalize to differences between Readers and Prereaders. This might mean that at the early stage of literacy development, differences in reading are too small to modify adaptation patterns or that adaptation deficits are indeed related to dyslexia, and not solely poorer reading.

Declaration of competing interest
None of the authors has a conflict of interest.

Open Practices
The study in this article earned Open Material badge for transparent practices. The materials for the study are available at: https://osf.io/tp5uf/.

Data availability
Data will be made available on request.