Disruption of functional networks in dyslexia: A whole-brain, data-driven analysis of connectivity

Abstract

Background

Functional connectivity analyses of fMRI data are a powerful tool for characterizing brain networks and how they are disrupted in neural disorders. However, many such analyses examine only one or a small number of a priori seed regions. Studies that consider the whole brain frequently rely on anatomic atlases to define network nodes, which may result in mixing distinct activation timecourses within a single node. Here, we improve upon previous methods by using a data-driven brain parcellation to compare connectivity profiles of dyslexic (DYS) versus non-impaired (NI) readers in the first whole-brain functional connectivity analysis of dyslexia.

Methods

Whole-brain connectivity was assessed in children (n = 75; 43 NI, 32 DYS) and adult (n = 104; 64 NI, 40 DYS) readers.

Results

Compared to NI readers, DYS readers showed divergent connectivity within the visual pathway and between visual association areas and prefrontal attention areas; increased right-hemisphere connectivity; reduced connectivity in the visual word-form area (part of the left fusiform gyrus specialized for printed words); and persistent connectivity to anterior language regions around the inferior frontal gyrus.

Conclusions

Together, findings suggest that NI readers are better able to integrate visual information and modulate their attention to visual stimuli, allowing them to recognize words based on their visual properties, while DYS readers recruit altered reading circuits and rely on laborious phonology-based “sounding out” strategies into adulthood. These results deepen our understanding of the neural basis of dyslexia and highlight the importance of synchrony between diverse brain regions for successful reading.

Introduction

Developmental dyslexia, or specific reading disability, is defined as an unexpected difficulty in accuracy or fluency of reading for an individual’s age, intelligence, level of education, or professional status (1). It is the most commonly diagnosed learning disability in the United States, with prevalence estimates ranging from 5 to 17.5 percent (2–4).

Functional magnetic resonance imaging (fMRI) studies have largely converged on three left-hemisphere brain areas whose activity is altered in dyslexia. These are: (1) a parietotemporal region, (2) the inferior frontal gyrus (5–7), and (3) an occipitotemporal region, including the visual word-form area (8–12), part of the fusiform gyrus that is seemingly responsible for rapid word recognition. These areas and their right-hemisphere homologs show over- or underactivation to reading tasks in dyslexic relative to non-impaired readers, in both children and adults (5, 13–16).

However, activation-based analyses of fMRI data do not address how brain regions interact and therefore may give an incomplete picture of the neural underpinnings of dyslexia. Functional connectivity analyses can detect differences in areas with similar magnitudes of activation, but whose activity is differentially synchronized with other areas across subject groups and/or types of stimuli. This synchrony between anatomically distinct regions may be equally or more important for cognitive performance than the magnitude of activation in any single region (17).

While there have been some functional connectivity studies of dyslexia (18–22), all of these have examined connections between specific regions chosen a priori. These regions follow the literature of task-based results and represent only a small fraction of the brain. Previous studies, then, potentially fail to provide a complete picture of the connectivity profiles of dyslexic versus non-impaired readers.

In this paper we present, to our knowledge, the first whole-brain functional connectivity study of dyslexia. Connectivity was examined in two large datasets of children and young adults. We used a novel method for parcellating voxels into functional subunits to use as nodes (23, 24) and contrasted the connection strengths—defined as correlations between signal timecourses—between dyslexic and non-impaired readers for all possible pairs of nodes to determine networks that are differentially connected across reading groups.

We hypothesized that these data-driven methods would reveal differences in previously characterized areas, but also in new, less well-characterized networks, thereby revealing the functional brain networks that crucially support the complex cognitive task of reading.

Materials and Methods

1. Participants

All data was acquired with participants’ written consent and in accordance with the Institutional Review Board of Yale University. Data are from two previously unpublished sets of dyslexic readers (DYS) and non-impaired controls (NI). Participants were classified as dyslexic based on sub-25^th percentile performance on one of the following batteries: (1) the Gray Oral Reading Test-IV (GORT; (25)) Fluency score (the composite of rate and accuracy), (2) either GORT subtest score (rate or accuracy), or (3) the Test of Word Reading Efficiency (TOWRE; (26)) total score, combined with a sub-25^th percentile score on the Woodcock-Johnson battery (27). NI readers scored above the 25^th percentile on all GORT, WJ, and TOWRE measures.

The first data set contained 80 children (“young readers,” mean age 8.9 years); the second contained 104 young adults (“older readers,” 64 NI, 40 DYS; mean age 20.8 years). After applying exclusion criteria to ensure there were no significant differences in head motion between groups (see Supplemental Information), 75 young readers (43 NI, 32 DYS) were entered into the final analysis.

Participant demographic and behavioral data are shown in Table 1. All participants were right-handed. Dyslexia and attention disorders are highly comorbid (28); while data on attention measures and ADD/ADHD status of participants were not explicitly collected, none were on psychoactive medication at the time of scanning. NI and DYS participants differed significantly on all behavioral reading measures (p < 0.001). Groups did not differ significantly in age (p > 0.05). There was a significant relationship between sex and group in young readers, with the proportion of males higher in the DYS group (Cochran-Mantel-Haenszel chi-squared test: χ² = 6.27, p = 0.01). There was no difference in sex breakdown in older participants (χ² = 0.71, p = 0.4). While IQ scores (29) were higher in the NI group, all DYS participants had normal-range IQ scores; it is difficult to balance groups along this dimension as reading ability and cognition are strongly linked (30). Still, studies have found that the presence or absence of reading disability is a better predictor of performance than IQ scores on tasks involving reading and related behaviors (31). To control for the observation that IQ was lower in the DYS group, we performed ancillary analyses in a subset of the cohort matched for performance IQ; see Supplemental Information for a description of these analyses and their results. Results from these IQ-matched subsets were highly consistent with the original analyses, thus data from the full cohort are presented here.

Table 1.

Participant demographic and behavioral data

	Young readers							Older readers
		DYS, N = 32			NI, N = 43			DYS, N = 40			NI, N = 64
		M	(z)	SD	M	(z)	SD	M	(z)	SD	M	(z)	SD
Age		9.13		1.03	8.69		1.16	20.83		2.99	19.72		3.06
WAISI IQ: Full scale		98.7	−0.09	13. 6	122. 5	1.50	13. 9	102. 6	0.17	13. 0	114. 9	0.99	11. 2
WAISI IQ: Verbal		102. 4	0.16	12. 4	125. 7	1.71	11. 6	102. 0	0.13	14. 5	115. 5	1.03	12. 6
WAISI IQ: Performance		95.5	−0.30	13. 7	115. 0	1.00	17. 2	101. 7	0.12	14. 7	110. 5	0.70	11. 7
WJ: Word ID		86.4	−0.90	5.9	121. 8	1.45	10. 8	83.7	−1.09	8.6	109. 3	0.62	9.5
WJ: Word attack1		88.1	−0.79	5.7	114. 2	0.95	9.9	78.4	−1.44	7.4	101. 2	0.08	6.9
TOWRE: Word reading		84.9	−1.01	11. 0	118. 9	1.26	9.9	78.4	−1.44	8.0	99.1	−0.06	12. 2
TOWRE: Phonologic decoding1		81.1	−1.26	5.7	114. 1	0.94	10. 0	68.8	−2.08	8.1	98.8	−0.08	8.3
GORT: Rate		5.50	−1.50	2.05	14.02	1.34	2.05	5.23	−1.59	1.66	11.98	0.66	1.39
GORT: Accuracy		4.28	−1.91	1.89	12.14	0.71	2.82	3.73	−2.09	2.16	12.91	0.97	1.98
GORT: Fluency		4.28	−1.91	1.76	13.28	1.09	2.54	2.40	−2.53	1.53	14.08	1.36	2.16
GORT: Comprehension		8.31	−0.56	3.27	13.33	1.11	3.11	7.78	−0.74	2.33	10.45	0.15	1.97
Gender (n)	M	22			17			24			33
	F	10			26			16			31

WASI, Wechsler Abbreviated Scale of Intelligence (26); WJ, Woodcock-Johnson (25); TOWRE, Test of Word Reading Efficiency (24); GORT, Gray Oral Reading Test IV (23). IQ, WJ and TOWRE measures are given in standard scores normalized to mean 100, standard deviation 15; GORT measures are standard scores normalized to mean 10, standard deviation 3. Groups differed significantly at p < 0.001 on all listed behavioral measures (IQ, WJ, TOWRE and GORT).

¹Test uses non-words (pseudowords)

Each subject underwent four functional runs while performing a word- and non-word-rhyming task. The effect of task was regressed out for purposes of the connectivity analysis (see Supplemental Information). Regressing out activation associated with task leaves residual fluctuations that more closely represent intrinsic, spontaneous neural activity (32). In using residuals, we avoid our results being dominated by activation coupled to the processing of each stimulus and instead examine fluctuations that occur independently of individual task stimuli. While overall brain state during task performance could subtly affect these fluctuations, in the case of dyslexia—a disorder with the specific task of reading—these effects may serve to enhance the contrast in functional organization between groups. See Supplemental Information for MRI acquisition parameters, information about task regression and other preprocessing steps.

2. Connectivity analysis pipeline

Analysis consisted of the following steps: (a) defining network nodes using a novel whole-brain parcellation; (b) calculating connectivity matrices for each subject; (c) comparing matrices to identify significant group differences in correlation strength; and (d) correcting for multiple comparisons using the network-based statistic (32). See Fig. 1 for a schematic of this analysis.

Fig. 1. Schematic of connectivity analysis.

Steps in the present approach to connectivity analysis: (A) Use groupwise parcellation algorithm (24) to divide whole brain into N (~200) nodes by grouping voxels with maximally similar timecourses. (B) Calculate NxN connectivity matrices of normalized correlation values for each possible pair of nodes for each subject. (C) Perform two-tailed t-test on each cell of the matrix. (D) Threshold based on chosen t-score, calculate largest connected component of suprathreshold edges in each direction (NI > DYS and DYS > NI). (E) Correct for multiple comparisons by randomly permuting group assignment and recalculating largest connected component size to create a null distribution for expected component size due to chance. (F) Visualize any components surviving correction.

a. Defining network nodes

A major consideration in functional connectivity analyses is how to delineate brain regions to use as nodes. Especially in the cortex, where structural boundaries are often unclear, choices such as how to draw node boundaries and how many nodes to include can greatly impact the properties — and neurobiological validity — of the resulting network (34).

Previous studies have commonly defined nodes in one of four ways: (1) anatomically, using the Brodmann-based automatic anatomic labeling atlas (AAL; (35)) (36–39); (2) functionally, according to task data from the same or previous fMRI studies (40, 41) — but note this approach provides limited brain coverage depending on which areas activated to the task(s); (3) by treating each individual voxel as a node (42, 43); or (4) using algorithms independent of anatomy or function that divide the brain into arbitrarily defined regions (44).

With the exception of the voxel-based approach, the problem with these parcellation schemes is the potential to mix activation timecourses within a single node, resulting in an average timecourse that does not accurately represent any of the individual voxels within. Such flawed averaging could render meaningless results. In the present study, we used a novel groupwise parcellation method to divide the brain into regions optimized to contain voxels with similar temporal signal fluctuations (23, 24). This purely data-driven approach yields nodes with more coherent timecourses than nodes defined using the AAL, even at the same whole-brain resolution (24). See Fig. 1A for an example of a whole-brain functional parcellation.

Because the brain undergoes significant changes between childhood and young adulthood, we parcellated young and older readers separately in order to delineate nodes that were maximally functionally coherent within each age group. For younger readers, 20 parcellations were generated from randomly selected groups of 30 NI participants, and the parcellation with the best whole-brain reproducibility score was selected and applied to all participants. For older readers, the best parcellation was chosen from 20 parcellations generated from randomly selected groups of 45 NI participants.

This process resulted in parcellations of 205 nodes for young readers and 207 nodes for older readers. The parcellations, defined in standard space, were warped to individual subjects’ space using the transformation from each subject’s 3D anatomical image to the MNI brain (45). While parcellations were defined using data from NI participants, there was no difference in whole-brain inhomogeneity scores between the NI and DYS groups (p = 0.30 in young readers, p = 0.26 in older readers). Inhomogeneity is defined as the Euclidean distance between the timecourse of an individual voxel to the mean timecourse of the node to which it belongs; the whole-brain homogeneity index for an individual subject is calculated by summing Euclidian distances for all voxels (see (24)).

b. Computing connectivity matrices

Using the respective group parcellation (younger or older), for each subject we computed the pairwise correlation coefficient between the timecourses of each possible pair of nodes using data from all runs from that subject. R-values were normalized to z-scores using the Fisher transformation, resulting in a 205×205 symmetric connectivity matrix for each young participant, and a 207×207 matrix for each older participant. Each cell of the matrix represents the strength of the connection, or “edge,” between two particular nodes.

c. Computing edge-wise differences

Using the individual-subject matrices as inputs, we performed a two-tailed t-test on the values of each cell to create a “difference matrix” of t-scores representing the magnitude of group difference at each edge (see Fig. 1C). We then thresholded this matrix for statistical significance at p < 0.01 (see Fig. 1D). Calculations described in the next section were performed on the left and right tails separately such that direction of differences could be identified (i.e., NI > DYS or DYS > NI).

d. Multiple-comparison correction

To correct for multiple comparisons, we used the network-based statistic (NBS; (33)). This method, analogous to cluster correction for graphs, provides a solution to the statistical problem of massively multiple comparisons in a whole-brain connectivity analysis. In the NBS, using the difference matrix as the input (subsection c above), the largest fully connected network of suprathreshold edges, or “component,” is identified and its extent defined as the number of edges it comprises. Finally, these calculations are repeated for a large number of trials in which subjects’ group assignments are randomly permuted to create a null distribution for the expected component size due to chance (see Fig 1E).

3. Group data analyses

We performed the analysis described above to compare connectivity between the NI and DYS groups in both data sets (alpha = 0.01; young readers: t threshold = ±2.645 (df = 73); older readers, t threshold = ±2.625 (df = 102); both groups, K = 5000 randomizations in the NBS step).

Results

1. Young readers

In young readers, we identified two differentially connected networks: one more strongly connected in the NI group (component size = 337 edges; p < 0.01, corrected) and one more strongly connected in the DYS group (component size = 415; p < 0.01, corrected). The full NI > DYS (red) and DYS > NI (blue) networks are shown in Supplement: Fig. S1A. We refer to the NI > DYS network as simply the “NI” network, and the DYS > NI network as the “DYS” network.

Although the components that survived correction at p < 0.01 represented only approximately 3 percent of the total possible edges in the brain, this number is sufficiently large to make visualization of all edges challenging. To focus our analysis on regions where connectivity was maximally different between groups, we further reduced these figures to include only those nodes with a sum of edge differences of at least 15 as well as their functional partners (Fig. 2A; see Table 2 for a list of these six nodes and their coordinates; see Supplemental Table S1 for a list of coordinates of all functional partners for each selected node). A node’s sum of edge differences was defined as the sum of its edges in both the NI and DYS networks—in other words, its total number of differential connections. The advantage of this method is the ability to detect nodes with the similar numbers of edges but different functional partners in each network. For example, node D in younger readers had eight edges in both the NI and DYS networks, so in a standard degree measurement its difference would be zero (see Table 2). However, preserving information about the location of connections and calculating the sum of edge differences—16 in this case—reveals this node as a locus of significant divergence between groups.

Fig. 2. Whole-brain connectivity differences between groups.

Three-dimensional representation of the NI > DYS (“NI”) and DYS > NI (“DYS”) edge components (p < 0.01 after NBS correction) in young readers (A) and older readers (B). Red lines denote the NI network, blue lines the DYS network. These components were thresholded for sparsity to show nodes with a sum of edges differences ≥ 15 in younger readers and ≥ 10 in older readers along with all the nodes to which these suprathreshold nodes were connected, representing approximately 1 percent of the total number of possible edges in both age groups. Spheres are placed at the centroid of each node, and are scaled and colored according to their numbers of edges in the NI and DYS networks—i.e., large red nodes have many more NI edges than DYS edges, and vice versa for large blue nodes. Note that all figures are shown in neurological convention (subject-left is image-left): axial views are top-down and coronal views are from the posterior.

Table 2.

Node-level analysis in younger readers.

Nodes with a sum of edge differences of at least 15**:
Node	Region	Talairach	Pos. edges	Neg. edges	Sum of edges	Diff. of edges
A*	L sup cuneus (Ba19)	(−9, −84, 30)	12	6	18	6
B*	L/R pCC (Ba23)	(6, −41, 21)	4	13	17	−9
C*	L mFG (Ba8)	(−5, 44, 37)	9	7	16	2
D*	R calc. (Ba17/23)	(11, −71, 11)	8	8	16	0
G	R sTG/SMG (Ba39/40)	(58, −45, 23)	11	4	15	7
H	L/R Cing (Ba23)	(0, −12, 32)	1	14	15	−13
Nodes with all one-sided edge differences:
E*	R sup cuneus (Ba19)	(21, −81, 35)	14	0	14	14
F*	R SMG (Ba40)	(41, −27, 38)	0	14	14	−14

^*Shown in Fig. 3.

^**Shown in Fig. 2A.

L, left; R, right; Ba, Brodmann area; sup., superior; pCC, posterior cingulate cortex; mFG, medial frontal gyrus; calc., calcarine sulcus; sTG, sup. temporal gyrus; SMG, supramarginal gyrus; cing, cingulate cortex.

Individual nodes were selected for discussion in the following manner. We visualized the connectivity profiles of the top four nodes with the highest sum of edge differences (at least 16), along with their functional partners in both the NI and DYS networks, in Fig. 3(A, B, C, D). In addition, we reasoned that nodes with very strong directionality—i.e., those with all NI > DYS or all DYS > NI edges—were also important loci of network differences. Thus we also profiled the two nodes with the highest difference of edges (Fig. 3E, 14 NI/0 DYS edges; Fig. 3F, 0 NI/14 DYS edges). See Fig. S2/Table S4 (Supplemental Information) for node-level results from the IQ-matched subset.

Fig. 3. Visualization of connectivity profiles of selected nodes in young readers.

We display the connectivity profile of the six nodes with maximal group differences in connectivity (of a total of 205 nodes; see Results section 1 for how these nodes were selected). In each case the selected node is shown in green (see Table 2 for each selected node’s coordinates), while nodes more strongly connected to the selected node in the NI > DYS (“NI”) network are shown in red, and nodes more strongly connected to the selected node in the DYS > NI (“DYS”) network are shown in blue (see Supplemental Table S1 for the coordinates of all NI and DYS partner nodes). Red and blue lines simply schematize connectivity between the selected node and each of its functional partners. Approximate anatomical locations of selected nodes: (A) left superior cuneus (BA 19), (B) posterior cingulate cortex, (C) left medial frontal gyrus (BA 8), (D) right calcarine sulcus (BA 17/23), (E) right superior cuneus (BA 19), (F) right supramarginal gyrus (BA 40). Note that all figures are shown in neurological convention: subject-left is image-left, and axial views are top-down.

2. Older readers

In older readers, we identified an NI > DYS network (component size = 312 edges; p < 0.01, corrected) and a DYS > NI network (component size = 361 edges; p < 0.01, corrected), referred to as the “NI” and “DYS” networks, respectively. Full networks are shown in Supplemental Fig. S1B. As with results from younger readers, for visualization purposes, we further reduced these figures to show nodes with a sum of edge differences of at least 10 (Fig. 2B; see Table 3 for coordinates of each selected node; see Supplemental Table S2 for a list of coordinates of all functional partners for each selected node).

Table 3.

Node-level analysis in older readers.

Nodes with a sum of edge differences of at least 10**:
Node	Region	Talairach	Pos. edges	Neg. edges	Sum of edges	Diff. of edges
J*	L iPL (Ba7/39)	(−34, −56, 43)	4	14	18	−10
K	R calc./ling.	(13, −55, 12)	9	4	13	5
L	L precuneus	(−13, −62, 25)	4	9	13	−5
M	L calc./ling.	(−13, −50, 6)	12	0	12	12
N	L mTG (Ba19)	(−47, −68, 5)	3	9	12	−6
O	L pCC/calc.	(−10, −69, 11)	4	7	11	−3
P	L post iFG (Ba44/6)	(−43, 10, 31)	6	4	10	2
Q*	L ant iFG (Ba46/47)	(−42, 43, 8)	0	10	10	−10
Node corresponding to the VWFA:
R*	L Fusiform	(−42, −60, −12)	6	3	9	3

^*Shown in Fig. 4.

^**Shown in Fig. 2B.

L, left; R, right; Ba, Brodmann area; iPL, inferior parietal lobule; calc., calcarine sulcus; ling., lingual gyrus; mTG, medial temporal gyrus; pCC, posterior cingulate cortex; post, posterior; ant, anterior; iFG, inferior frontal gyrus.

To identify nodes for further discussion, we chose the node with the highest sum of edge differences (node J; Fig. 4A) as well as a node with entirely DYS > NI connections (node Q; Fig. 4B). Given the prominence of the visual word-form area (VWFA) in the dyslexia neuroimaging literature, we also show the connectivity profile of the node in the parcellation corresponding to the VWFA (node R, Fig. 4C; defined based on the Talairach coordinates for peak activation in the VWFA as reported in two recent seed-based connectivity studies of dyslexia [(20): (−40, −55, −10); (21): (−43, −62, −1); center of mass for parcellation node R, (−42, −60, −12)]). See Fig. S3/Table S5 (Supplemental Information) for node-level results from the IQ-matched subset.

Fig. 4. Visualization of connectivity profiles of selected nodes in older readers.

We display the connectivity profile of three nodes with maximal group differences in connectivity (of a total of 207 nodes; see Results section 2 for how these nodes were selected). In each case the selected node is shown in green (see Table 3 for coordinates of each selected node), while nodes more strongly connected to the selected node in the NI > DYS (“NI”) network are shown in red, and nodes more strongly connected to the seed in the DYS > NI (“DYS”) network are shown in blue (see Supplemental Table S2 for the coordinates of all NI and DYS partner nodes). Red and blue lines schematize connectivity between the selected node and each of its functional partners. Approximate anatomical locations of selected nodes: (A) left inferior parietal lobule/BA 7 (node J), (B) left anterior inferior frontal gyrus/BA 46 (node Q), (C) left fusiform gyrus/visual word-form area (node R). Note that all figures are shown in neurological convention: subject-left is image-left, axial views are top-down and coronal views are posterior.

Discussion

This data-driven approach revealed widespread differences in connectivity between dyslexic and non-impaired readers (Fig. 2). Many differences were in areas not previously detected by activation-based fMRI analyses, suggesting that dyslexia is a complex disorder that cannot be reduced to alterations in a small number of brain regions. Here, we discuss five themes that emerged from this whole-brain survey and subsequent node profiles. Relative to NI readers, DYS readers showed: (1) decreased connectivity along the visual pathway, as well as between visual and prefrontal regions; (2) decreased lateralization of language to the left hemisphere; (3) increased connectivity to limbic regions and the default mode; (4) reduced and altered connectivity to the visual word-form area; and (5) persistent connectivity to a left-hemisphere anterior language region.

1. Connectivity along the visual pathway, and between visual areas and prefrontal regions, is disrupted in dyslexia

Activity among nodes of the visual pathway was better synchronized in the NI group across ages. In young readers, the strength of NI network in occipital and occipitoparietal regions was notable on a whole-brain level (Fig. 2A). Specifically, young NI readers showed stronger connectivity between node A in the left superior cuneus and nearby nodes in bilateral visual association cortices (Fig. 3A); between node D in the right calcarine sulcus and bilateral visual association cortices (Fig. 3D); and between node E in the right superior cuneus and bilateral striate and extrastriate cortices (Fig. 3E). Taking functional coupling between primary visual cortex and its extrastriate partners to represent feed-forward and/or feedback activity, these results indicate that information transfer along the visual pathway follows better-formed connections in NI readers.

Furthermore, NI readers showed stronger functional connectivity between the visual pathway and prefrontal cortex (PFC). In young readers’ NI network, nodes C and E were more strongly connected to the medial PFC (Fig. 3C, E), and in older readers’ NI network, node J in the left inferior parietal lobule was more strongly connected to the left medial PFC and bilateral superior PFC (Fig. 4A). Evidence from several techniques—including fMRI (46), magnetoencephalography (47), and transcranial magnetic stimulation (48)—has shown that functional coupling between PFC and visual areas reflects top-down modulation of attention to visual stimuli. That this connectivity was stronger in NI readers suggests that they are better able to control activity in the visual pathway by focusing their attention on textual stimuli.

This role for attention systems working in concert with visual processing supports a growing body of evidence that dyslexia is, at least in part, a disorder of attention (49). Behavioral studies report impaired visual attention in individuals with dyslexia (50–52), including a reduced visual attention span (53), an impaired ability to shift attention (54), and deficits in serial search and spatial cueing (55). These deficits are general—that is, they affect dyslexic readers’ processing of not just text but all types of visual stimuli. However, previous fMRI studies of dyslexia have largely failed to find activation differences in primary visual and/or visual association cortices—areas where differences might be predicted given the general nature of the behavioral deficits.

The current study helps resolve the discrepancy between behavioral results and activation-based fMRI findings: While visual areas may not be differentially active across reading groups, they differ in the strength of their connections to one another and to other brain regions crucial for reading. In other words, the problem is not that dyslexic readers fail to use visual and attention regions altogether, but that they use them in different, and evidently less successful, ways.

2. DYS readers are slow to lateralize language to the left hemisphere

In most people, the left hemisphere is dominant for language (56) and language lateralization increases between 5 and 20 years of age (57). Present results show that NI readers have stronger left lateralization for language than DYS readers, who rely on bilateral systems. This extends previous findings, indicating that right-hemisphere homologs are not only more active in dyslexia, but also more strongly connected to the rest of the brain.

Young NI readers showed greater connectivity from the left IFG and anterior insula to node C in left medial PFC (Fig. 3C). The left IFG, which contains Broca’s area, is involved in language, including semantic, syntactic and phonological processing (58). In typically developing readers, reading initially recruits these anterior language regions as children learn to map visual word forms to their phonological and semantic content. Our results extend these findings, suggesting that successful reading in young people depends on the synchrony of activity in left-hemisphere language regions with the ipsilateral PFC, likely reflecting modulation of attention to language stimuli.

Young DYS readers, on the other hand, showed increased overall connectivity in the right hemisphere (see Fig. 2A, axial view). Specifically, two nodes in the visual pathway were more strongly connected to right-hemisphere language homologs in DYS readers: node D in the right calcarine sulcus was more connected to the right IFG (Fig. 3D); and node A in the left cuneus was more strongly connected to the right angular gyrus and right hippocampus/fusiform gyrus (Fig. 3A). (While activity and connectivity of the angular and fusiform gyri have been associated with successful reading (18, 59), this activity is also typically lateralized to the left hemisphere in NI readers.) In addition, node F in the right supramarginal gyrus showed increased connectivity to nodes both ipsilateral and contralateral in the DYS network (Fig. 3F).

Results from older readers indicate that by 20 years of age, differences in lateralization have diminished but not disappeared (Fig. 2B, axial view; Fig. 4A). It seems, then, that DYS readers do eventually attain left-weighted language connectivity, but at a slower rate—and to a lesser degree—than NI readers.

3. DYS readers show altered connectivity in the pCC/default-mode network

The posterior cingulate cortex (pCC) emerged as a locus of differential connectivity in young readers (Fig. 3B). Work in humans and non-human primates indicates that the pCC functions as a cortical hub, integrating information from distinct brain regions and networks (60). The pCC is also considered part of the default mode, or the network of regions that are active in the absence of any particular task (61). Our results show that in DYS readers, activity in the pCC is better synchronized with other areas of the default-mode network (dorsal and ventral cingulate cortices) and medial PFC, while in NI readers, pCC activity is better synchronized with visual association regions. Given the pCC’s “hub” status, stronger connectivity between the pCC and the dorsal visual pathway and left fusiform gyrus for the NI group (Fig. 3B) could reflect better integration and cognitive control of visual information by these readers.

Since, like visual areas, the pCC and other default-mode regions have largely failed to emerge in activation-based studies of dyslexia, it seems that these regions’ functional coupling, rather than their magnitude of activation, separates DYS from NI readers. Future studies using data obtained in the resting state as well as during task will help clarify how activity/connectivity in these regions relate to dyslexia.

4. Connectivity to the visual word-form area is reduced and altered in older DYS readers

The visual word-form area, while not differentially connected in young readers, emerged in older readers as more strongly connected in the NI group (Table 3). These results extend previous reports that in NI readers, VWFA activity increases with age (62, 63), and that its functional connectivity increases with age and reading skill into the young adult years (21). Our findings indicate that there is not yet a significant discrepancy in VWFA connectivity between dyslexic and younger good readers, but that the gap widens with age as good readers reach “expert” status.

In older readers, this node’s connections in the NI network were to bilateral extrastriate cortices, left IFG, and left medial PFC (Fig. 4C). These couplings could reflect successful information transfer between the VWFA and general visual association areas (extrastriate), mapping of word forms to language (IFG) and attentional modulation of this process (medial PFC), all preferentially recruiting the left hemisphere. DYS readers showed increased connectivity between the VWFA and other bilateral visual association regions, as well as to the right primary auditory cortex. Thus DYS readers do not completely fail to develop connectivity to the VWFA, but rather do so in altered, perhaps compensatory circuits.

5. DYS readers show persistent connectivity to a left anterior language region

A striking finding among older participants was increased connectivity to the left IFG in DYS readers (node Q; Fig. 4B). This node’s connections in the DYS network involved medial PFC, anterior cingulate and left caudate. While this area is multifunctional, many studies have attributed phonology in particular to this region (5). Its connectivity profile in older readers supports theories holding that dyslexic readers continue to rely on effortful mental articulation, or “sounding out,” strategies instead of transitioning to more rapid, visual-based systems.

Conclusion

We have presented the most comprehensive analysis of functional brain connectivity in dyslexic versus non-impaired readers to date. Of the approximately 20,000 connections investigated in each data set, a small but impressive fraction showed highly significant differences in functional connectivity between groups, suggesting that dyslexia is not simply attributable to dysfunction in a few specific language nodes. These data-driven, whole-brain results deepen and extend what was known from activation-based fMRI studies about the neural underpinnings of dyslexia.