Developmental dyslexia consists of a specific and persistent failure to acquire efficient reading skills despite conventional instruction, adequate intelligence, and socio-cultural opportunity [1]. Many competing neuro-cognitive hypotheses aim to explain dyslexia. The phonological hypothesis, which is the most influential account for reading problems, postulates deficits related to the access or the manipulation of phonemic information, or both, preventing efficient learning of graphemes/phonemes correspondences that are crucial to reading; e.g. [2, 3]. By contrast, the auditory processing deficit theory proposes that phonological deficits are secondary to a more basic impairment in (rapid) auditory processing [4, 5]. The visual magnocellular hypothesis posits the existence of low level visual disorders related to abnormal thalamic magno-cells [6] that are involved in the processing of moving stimuli and would thus be important for reading activities related to saccadic eye movements [7]. Alternatively, the visuo-attentional hypothesis situates the impairment in the encoding of letter sequences, and this latter deficit would be dissociated from phonological deficits [8]. The cerebellar hypothesis relates dyslexia to a general learning disorder that includes a failure to automatize reading and writing skills, i.e. dyslexia is regarded as an impaired automatization of high-order sensory-motor procedures essential in reading [9, 10]. Finally, the general magnocellular theory encompasses the latter four theories (basic auditory, basic visual, attentional and cerebellar) by discussing each deficit as a consequence of a general magnocellular defect [11]. According to this last theory, phonological deficits are secondary in comparison to other deficits.

This multiplicity of theories aiming to explain dyslexia reflects the heterogeneity of behavioural deficits. It is indeed becoming accepted that dyslexia is not a unique entity but might reflect different neuro-cognitive pathologies [12]. As a matter of fact, dissimilar behavioural types have been proposed for a long time. One of the first classification was proposed by Boder in 1973 [13]. Dyslexic children were distinguished on the basis of their 'sensory' deficit, i.e. they would either be classified as dysphonetic (having phonological problems), or dyseidetic (having visual problems). More recently, based on the classification of patients with acquired dyslexia, Castles and Coltheard [14] proposed to distinguish phonological from surface developmental dyslexic children. In this case, dyslexics are split into patients with assembling problems, i.e. grapheme/phoneme association, vs. patients with addressing problem, i.e. lexicon access. However, it is also recognized that developmental dyslexic children can have both assembling and addressing problems. Recent data and modelling also suggest that different subgroups can be distinguished within each subtype and that signs can be shared between these subgroups [15]. For instance, whereas patients with acquired surface dyslexia have preserved phonological abilities [14], developmental surface dyslexics present, in addition to a lexicon addressing deficit, mild phonological disorders sometimes in conjunction with a letter decoding deficit. In addition, this latter deficit (letter decoding) could also be observed in some developmental phonological dyslexics [15]. In another study [16] on adult dyslexics, Ramus et al. also found that the phonological/surface distinction does not hold for developmental dyslexia as all subjects presented phonological problems. In addition, for 70% of the subjects, phonological problems were associated with cerebellar, visual and/or auditory deficits. Other studies also pointed out the existence of different subgroups that often do not strictly follow the above mentioned theories. For instance, Heim and colleagues [17] identified, using a combined cluster/discriminant analysis, three behavioural subgroups. One subgroup had phonological deficits only (33.3% of the dyslexic subjects), another group had phonological, basic auditory and visual magnocellular deficits (35.6%), whereas the remaining group had attentional problems only (31.1%). Reid et al. [18] also observed different clusters, but using a deviance analysis: 6.6% of dyslexics had a magnocellular deficit only, another 6.6% had a cerebellar deficit only, 26.6% had a naming (fluency) deficit only, and another 20% had a phonological awareness deficit only. Other subjects (40%) presented a combination of the mentioned deficits. These and other studies [19–23] therefore suggest that one theory cannot explain all of the behavioural deficits associated with dyslexia. It is therefore not surprising that brain studies that aim to find the biological counterparts of cognitive deficits do not always agree one with the other, depending on the sample at hand and the cognitive tests used.

From a theoretical point of view, several arguments favour the idea that developmental disorders like dyslexia cannot be 'specific', i.e. cannot reflect an impairment in only one aspect of cognition like e.g. phonology [24], and this would explain why one cannot find one unique biological (area or network) counterpart of dyslexia. From the cognitive science perspective, developmental disorders should not be interpreted as an impairment in one cognitive process but rather as the endpoint of an abnormal developmental process, reflecting the interaction of deficient and compensatory processes [25]. Similarly, from a behavioural genetic perspective, genes that are involved in developmental disorders have both, specific and general effects [26] such that additional cognitive disorders or comorbidities, or both, should be observed. From an experimental point of view, recent reviews pointed out several brain areas with structural [27], or functional [12] abnormalities. Nevertheless, those reviews also agree on the heterogeneity of results across studies. In a recent paper, we [28] suggested that dyslexia has a multifocal origin in terms of brain morphology. Using a Voxel Based Morphometry (VBM) approach [29] we demonstrated that dyslexic subjects do not linearly differ from control subjects in term of local grey matter volumes. In other words, dyslexics (as a group) do not have lower or higher local grey matter volumes. By contrast, dyslexic subjects had significantly different patterns of volume variations from controls, mainly in the superior temporal suslcus, fusiform gyri and in the cerebellar declives. These patterns of volume variations correlated with pseudoword reading performances in both groups, such that dyslexics represented the lower tail of the distribution both in terms of volumes and in terms of performances. Significantly different correlations were also observed regarding phonological performances with stronger correlations in control than dyslexic subjects for the cerebellum and significant/present correlations for controls vs. absent for dyslexic subjects in the cerebrum. Finally, crisscrossed correlation patterns between dyslexic and control readers were also observed regarding the spelling performances. Our interpretation of the absence of net volume differences between groups is that, by testing a large sample of dyslexic subjects (N = 38), measures of brain volumes have too high variances (both in control and dyslexic readers), leading to accept the hypothesis of an absence of difference between groups (H0). Similarly, the possible heterogeneity of dyslexic subjects led to reject the hypothesis of a difference between groups (H1). In the present study, we further investigated this hypothesis of sample homogeneity by re-analysing the data from the same subjects tested in Pernet et al. [28]. We looked for brain areas where dyslexic subjects, as a group, were maximally different from controls. This was performed by classifying, on a single subject basis (as opposed to group comparison), each voxel of dyslexics' brains as within or outside the grey matter confidence interval observed in control subjects. Based on this brain classification, dyslexic subgroups were identified and we investigated if those subgroups showed behavioural differences. Our method contrasts with purely linear techniques (e.g. t-test) as it does not assume homogeneity of the patient population (the distribution can be e.g. bi-modal). This also contrasts with behavioural studies as subgroups are not defined by their task performances but rather defined by their brain distributions, i.e. their 'intermediate' or endophenotype [30].

Based on our previous results [28], we hypothesized that dyslexic subjects would differ maximally from controls over the left superior temporal gyrus (STG), the left and right fusiform gyri, and the left and right cerebellar declive (lobe VI). In addition, we expected only one group of dyslexic subjects over the left STG (as there was a tendency for group differences in [28]) but possibly several subgroups of dyslexics over the fusiform gyri and the cerebellum; leading to observe several brain phenotypes [30]. The new analyzes reported here revealed that dyslexics are best discriminated from controls (100% of dyslexics outside the confidence intervals) on the basis of cerebellar and lentiform nucleus volumes only. Furthermore, subgroups with higher or lower volumes in these areas differed behaviourally one from another, therefore comforting our hypothesis that dyslexia is an heterogeneous condition and, by extension, that it cannot be explained by a single 'specific' hypothesis.