Abstract
Cognitive models of reading assume that speech production occurs after visual and phonological processing of written words. This traditional view is at odds with more recent magnetoencephalography studies showing that the left posterior inferior frontal cortex (pIFC) classically associated with spoken production responds to print at 100–150 ms after word-onset, almost simultaneously with posterior brain regions for visual and phonological processing. Yet the theoretical significance of this fast neural response remains open to date. We used transcranial magnetic stimulation (TMS) to investigate how the left pIFC contributes to the early stage of reading. In Experiment 1, 23 adult participants (14 females) performed three different tasks about written words (oral reading, semantic judgment, and perceptual judgment) while single-pulse TMS was delivered to the left pIFC, fusiform gyrus or supramarginal gyrus at different time points (50–200 ms after word-onset). A robust double dissociation was found between tasks and stimulation sites—oral reading, but not other control tasks, was disrupted only when TMS was delivered to pIFC at 100 ms. This task-specific impact of pIFC stimulation was further corroborated in Experiment 2, which revealed another double dissociation between oral reading and picture naming. These results demonstrate that the left pIFC specifically and causally mediates rapid computation of speech motor codes at the earliest stage of reading and suggest that this fast sublexical neural pathway for pronunciation, although seemingly dormant, is fully functioning in literate adults. Our results further suggest that these left-hemisphere systems for reading overall act faster than known previously.
Significance Statement
Recent neuroimaging data suggest that left posterior inferior frontal cortex, classically associated with spoken production, responds to print simultaneously with left fusiform and supramarginal gyri, each responsible for visual and phonological processing, contrary to traditional serial cascade models of reading. While the region is now known to mediate different aspects of cognitive processing, the functional significance of this fast neural response remains unclear. Using transcranial magnetic stimulation, we show that early inferior frontal activation plays a specific and causal role in speeded oral reading at 100 ms after word-onset. This fast sublexical neural pathway for pronunciation, although seemingly dormant, is fully functioning in literate adults. We propose that the left-hemisphere reading systems act differently and faster than known previously.
Introduction
Literate adults produce spoken sounds from written text in only a fraction of a second (Brysbaert, 2019). Traditional cognitive models illustrate that reading expertise is accomplished by cascading activations from letter/word forms to phonological, semantic, and motor representations (Rumelhart and McClelland, 1982; Coltheart et al., 2001). This rapid sequential activation relies on a distributed left-hemisphere network (Fig. 1), including the fusiform gyrus (FG) for visual word-form processing, the supramarginal gyrus (SMG) for print-to-sound translation, and the posterior inferior frontal cortex (pIFC) for speech production (see for review, Pugh et al., 2001; McCandliss and Noble, 2003; Turker and Hartwigsen, 2021). Event-related potential (ERP) and magnetoencephalographic (MEG) studies show that the FG, SMG, and pIFC respond to print at 150–200, 250–400, and 300–500 ms after stimulus onset, respectively (Marinkovic et al., 2003; Dien, 2009; Vartiainen et al., 2011; Thesen et al., 2012), largely consistent with the cascade models of reading.
Neurocognitive components of reading in the left hemisphere (adapted from Jobard et al., 2003; Kearns et al., 2019). Following early visual and orthographic processing in lateral occipital cortex (LOC) and fusiform gyrus (FG), written words are transformed into spoken outputs either through the lexico-semantic route associated with middle temporal gyrus (MTG)/anterior temporal lobe (ATL; thick lines) or through the phonological assembly route associated with supramarginal gyrus (SMG; dashed lines) prior to the motor execution in the primary motor cortex (M1).
Interestingly, however, more recent work shows that the left pIFC, known as “Broca's region,” responds at 100–200 ms simultaneously with the left FG (Pammer et al., 2004; Cornelissen et al., 2009; Wheat et al., 2010; Klein et al., 2015), suggesting that these neurocognitive components may not act sequentially as predicted by cognitive models of reading. Clearly, this leaves open whether pIFC is activated by input signals bypassing FG/SMG or by fast interregional interactions via FG/SMG. This is important because early pIFC activity may reflect top-down modulatory signals over posterior regions to promote early visual processing, rather than specific language processing (Woodhead et al., 2014). Indeed, such fast prefrontal predictive signals are known to facilitate rapid perceptual processing in FG (Bar et al., 2006). Since literacy enhances functional connectivity between occipitotemporal and frontal cortices (Hannagan et al., 2015), early pIFC activation in reading may be generated by a similar neural mechanism for visuoperceptual processing.
Several MEG studies attributed early pIFC response to the rapid activation of articulatory codes since the region is classically associated with speech production (Wheat et al., 2010; Klein et al., 2015). Yet this account should be revisited because the left pIFC is now known to mediate different aspects of linguistic/nonlinguistic processing (Amunts et al., 2010; Hagoort, 2014). Indeed, the same region responds to print very fast even when no spoken response is required (Cornelissen et al., 2009). The temporal dynamics of the region needs further investigation because bottom-up neural pathways for reading change with behavioral task demands (Nakamura et al., 2006).
Critically, it remains unknown whether the early pIFC activity, although associated with articulatory codes (Wheat et al., 2010) or top-down modulatory signals (Woodhead et al., 2014), plays any causal role in reading. Since ERP/MEG allows correlational, rather than causal, inferences about structures and functions (Pascual-Leone et al., 2000), it is essential to identify behavioral effects arising when this neural activity is suppressed during reading. In fact, Wheat et al. (2013) tackled the question using transcranial magnetic stimulation (TMS) but found no behavioral changes during 100–200 ms after word onset, leaving the issue unresolved to date.
We addressed these questions using chronometric TMS over the key neural components of reading. In Experiment 1, we applied single-pulse TMS to pIFC, FG, and SMG at different time points (50–200 ms after word onset) while participants performed three different tasks, i.e., oral reading, semantic judgment, and perceptual judgment about written words (Fig. 2). While the left pIFC encompasses structurally and functionally heterogeneous subregions (Amunts et al., 2010; Hagoort, 2014), we targeted the inferior frontal/precentral junction associated with speech motor codes (Hickok and Poeppel, 2007; Hickok, 2012). If early pIFC activity mediates top-down modulation over occipitotemporal regions, TMS applied to this region will impede behavioral responses equally in the three tasks. Conversely, if this early activity reflects fast activation of articulatory codes, TMS will likewise disrupt oral reading, sparing semantic and perceptual judgment. Moreover, if early pIFC activity causally contributes to reading, TMS will yield measurable changes in behavior ∼100–200 ms after word onset.
Behavioral task and target cortical regions. a, Single-pulse TMS was delivered to the left ventral premotor cortex (pIFC: x = −56, y = −4, z = 18), fusiform gyrus (FG: x = −46, y = −56, z = −12), and supramarginal gyrus (SMG: x = −45, y = −39, z = 45). b, Sequence of visual stimuli. Target words were colored randomly either black or gray on a white background. In Experiment 1, participants performed oral reading, semantic judgments, and perceptual judgments while receiving TMS over one of the three cortical targets. In each task, single-pulse TMS was applied at four different time points, i.e., 50, 100, 150, and 200 ms relative to word onset. In Experiment 2, participants performed two different word production tasks, i.e., oral reading and object naming, under the same TMS procedure (see Materials and Methods).
Materials and Methods
Participants
A total of 50 healthy right-handed volunteers participated in the present study (27 participants [17 females; age range, 19–38 years, mean (SD) = 25.89 (5.91) years] in Experiment 1 and 23 participants [15 females; age range, 19–52 years, mean (SD) = 28.17 (9.82) years] in Experiment 2). All of them were native Japanese speakers with normal (or corrected-to-normal) vision and had received >12 years of school education [mean (SD) = 16.81 (2.42) years in Experiment 1 and 16.61 (2.63) years in Experiment 2]. None of them had known neurological or psychiatric disorders and reported any reading, writing, or learning disabilities. It is safe to assume that all participants could fluently read stimulus words in the syllabic kana script (see below), for which fluent reading is known to be achieved in the very early stage of school education (Wydell and Butterworth, 1999) such that reading impairments in kana are known to be rare (∼1%) even in the elementary-school age population (Uno et al., 2008). Four participants were excluded from analyses because of technical errors during data acquisition in Experiment 1. We performed post hoc power analysis (alpha level = 0.05) for each of the significant differences reported (see Results) and confirmed that statistical power (1-β) exceeded 0.91 in Experiment 1 and 0.97 in Experiment 2, respectively. All participants provided written informed consent prior to the experiment. The protocol of this study was approved by the ethical review committee of the National Rehabilitation Center for Persons with Disabilities.
Behavioral tasks
In Experiment 1, visual stimuli consisted of 100 Japanese nouns written with 2–3 characters in the syllabic kana script [mean (SD) word length = 2.5 (0.5) characters, 3.6° (1.1°) in visual angle]. Half of them represented animal names (e.g., cat, duck, snake), while the other half plant names (e.g., rose, mint, plum). Mean lexical frequency per million words was 10.19 (SE = 2.06) for animal names and 8.79 (SE = 1.14) for plant names according to the Balanced Corpus of Contemporary Written Japanese (Maekawa et al., 2013). The two categories were matched with each other in terms of lexical frequency (t(98) = 0.60; p > 0.55). It is of note that kana is a highly regular writing system with 46 basic characters, each of which has almost one-to-one correspondence with sound (e.g., “サ” is always pronounced as /sa/).
A sequence of events used for the TMS experiment is illustrated in Figure 2B. Each trial comprised central fixation (500 ms), a target word (100 ms), and a backward mask (100 ms) followed by a blank response screen (2,300 ms). The mask stimuli consisted of a string of squares matched in length with preceding targets. This backward masking was used in line with the previous fMRI study by Pammer et al. described above (Pammer et al., 2004), which allowed us to tap early neurocognitive processes while minimizing later top-down influence on visual word processing (Dehaene et al., 2001). Target words were colored randomly either black or gray on a white background (Fig. 2). To approximately match the level of difficulty between the two judgment tasks (perceptual and semantic), the black-versus-gray contrast in target stimuli was set to be sufficiently low such that overall accuracy during perceptual judgment could fall around ∼90%. As described below, TMS was applied to cortical regions of interest at four different time points (50–200 ms after word onset) relative to word onset.
The TMS experiment consisted of three blocks in each of which participants received TMS over one of the three cortical targets (pIFC, SMG, and FG; see below). Each block included three subblocks each of which was assigned to one of the three behavioral tasks described below. In the oral reading task, participants read aloud target words as quickly as possible. For each participant, oral responses were recorded with a high-resolution condenser microphone and analyzed offline by the offline visual inspection of the digitized speech waves. That is, reading latencies from word onset were defined on trial-by-trial basis as the first detectable amplitude in the speech waveforms (Schuhmann et al., 2009, 2012; Zhang et al., 2018). In the semantic judgment task, participants decided as accurately and as quickly as possible whether target words represented animals or plants. Perceptual judgment was used as a visual control task in which participants decided whether targets were colored black or gray. Each task consisted of 100 trials and lasted 6 min. In the semantic and perceptual judgment tasks, participants responded by button press with their right index and middle fingers.
Participants were seated at a viewing distance of 57 cm with a head and chin rest to reduce head movements. Participants received three blocks (∼20 min) in each of which TMS was delivered to one of the three target sites. The three blocks were separated by short breaks (∼15 min), which was required to determine the optimal position and orientation of the TMS coil for each target site. Given the short longevity of neural effects induced by single-pulse TMS (see below), this temporal spacing between blocks was sufficient to eliminate possible carry-over effects from preceding blocks. In each block, participants received single-pulse TMS while they were engaged in one of the three tasks according to prespecified task instructions. The order of stimulation sites (i.e., pIFC, SMG, and FG) and tasks (i.e., semantic judgment, oral reading, and perceptual judgment) was randomly counterbalanced across participants such that tasks and stimulation sites are not correlated with each other to eliminate the possible effects of learning from behavioral data analysis. Before the start of the experiment, participants performed a short practice session for each task. Stimulus presentation and response collection were controlled using the E-prime software (version 3.0; Psychology Software Tools).
In Experiment 2, we used the same set of 100 nouns used in Experiment 1 for the oral reading task. For the object-naming task, we additionally collected a set of 100 easily identifiable object drawings (animals, commodities, fruits, plants, etc., ∼4.1° in visual angle) from freely available resources on the Internet. Object names were selected such that their word length [mean (SD) = 2.5 (0.5) moras] and lexical frequency [mean (SE) = 12.28 (1.17)] were matched with those of the word stimuli described above (p > 0.09 for all). The event sequence for the object-naming task comprised central fixation (500 ms), a target object (150 ms), and a backward mask (100 ms) followed by a blank response screen (2,300 ms). The mask stimuli consisted of a random Gaussian-noise image roughly matched in size with preceding targets. Otherwise, the design, materials, and procedures were all identical to Experiment 1.
TMS procedure
For each participant, a structural T1-weighted MRI scan was obtained prior to the TMS experiment. Single-pulse monophasic TMS was generated using a 70 mm figure-of-eight coil with Magstim Rapid (Magstim) in Experiments 1 and 2. The coil positions were guided by the Brainsight neuronavigation system (Rogue Research). Prior to the experiments, we measured the resting motor threshold in the first dorsal interosseous muscles for each participant, which was defined as the lowest stimulus intensity required to elicit motor-evoked potentials in at least five consecutive trials. In both experiments, the intensity of TMS was set to 120% of the motor threshold which corresponded to ∼70% of the maximum stimulator output [mean (SD) = 71.43 (7.49) %]. Low-frequency single-pulse TMS (∼0.3 Hz) at this stimulation intensity can induce an estimated electric field strength of ∼60 V/m around cortical targets (Numssen et al., 2024) and transiently suppress variously different types of cognitive processing (Chambers et al., 2004a; Lou et al., 2004; Beauchamp et al., 2010; Ronconi et al., 2014; Savoie et al., 2020; Xiao et al., 2020; Osada et al., 2021; Luber et al., 2022; see also Silvanto and Cattaneo, 2017 for review), including reading and naming (Devlin et al., 2003; Skarratt and Lavidor, 2006).
Cortical target sites were determined using the known spatial coordinates of the left pIFC, FG, and SMG (see below) in the standardized brain of the Montreal Neurological Institute (MNI). This is because it is generally difficult to pinpoint the optimal target locations by the visual inspection of MRI scans due to the structural variability of the individual brain and the structural/functional heterogeneity within target structures. We therefore relied on the a priori spatial coordinates data since reported locations of the target structures are well consistent across different functional neuroimaging studies (see below). Moreover, since TMS is known to have a spatial resolution of ∼5–10 mm (O'Shea and Walsh, 2007; Sliwinska et al., 2014), the spatial extent of the induced electric field roughly corresponds to that of the spherical volume typically used for region-of-interest (ROI) analysis (∼5 mm in radius) in functional brain imaging (Nakamura et al., 2007; Brem et al., 2009, 2010; Eichert et al., 2020), thus covering each of the three cortical targets described below.
Using the MARSBAR toolbox (Brett et al., 2002), we constructed three 5-mm-radius spherical ROIs in the left pIFC, FG, and SMG in the standardized MNI brain space provided by Harvard-Oxford brain atlases (https://neurovault.org/collections/262/). While the left pIFC is known to comprise several subdivisions associated with different aspects of language processing (Amunts et al., 2010; Hagoort, 2014), we selected its ventral premotor sector (Brodmann area 6) responsible for speech production. That is, we created a target ROI for pIFC centered at x = −56, y = −4, z = 18 by averaging the spatial coordinates data provided by previous fMRI studies that isolated the inferior frontal/precentral junction associated with speech motor codes (Kemeny et al., 2006; Brown et al., 2009; Zheng et al., 2010; Eichert et al., 2020). The center coordinates of the ROIs were located at x = −46, y = −56, z = −12 for FG (Bolger et al., 2005) and x = −45, y = −39, z = 45 for SMG (Hartwigsen et al., 2010). For the left FG, the middle part of the left fusiform cortex responsible for visual word-form analysis is shown to fall at the same location with a standard deviation of ∼5 mm, irrespective of tasks and languages (Cohen et al., 2002; Jobard et al., 2003; Bolger et al., 2005). The spatial coordinates for the left SMG were used in a previous TMS study by Hartwigsen et al. (2010), which relied on several neuroimaging studies of phonological processing during reading (Price et al., 1997; Devlin et al., 2003; McDermott et al., 2003). The three ROIs were then mapped onto each participant's native brain space using the spatial normalization procedure implemented in Statistical Parametric Mapping 12 (https://www.fil.ion.ucl.ac.uk).
On each trial, single-pulse TMS was delivered at four different time points, i.e., 50, 100, 150, or 200 ms after word onset. In addition, we created a fifth type of trials in which TMS was delivered 100 ms prior to word onset. These trials served as a baseline condition for each task for each site, since the effects of single-pulse TMS are known to disappear in ∼10–20 ms (Ilmoniemi et al., 1997; Walsh and Cowey, 2000), that is, well before the word onset in each trial. Note that this baseline condition is thus equivalent to sham stimulation because single-pulse TMS delivered 100 ms prior to word stimuli cannot interfere with the execution of tasks due to the short longevity of neural effects induced by TMS (Stoeckel et al., 2009; Sliwinska et al., 2012; Jackson et al., 2015; Alexander et al., 2017). On the other hand, while single-pulse TMS before word onset may affect some more general levels of cognitive processing, e.g., response preparation and attentional modulation, such global and non-specific effects of TMS should occur systematically across tasks and sites and thus can be eliminated in the between task/site comparisons described below.
Consequently, there were 20 trials for each timing per task per site per participant, comparable with the number of trials used in previous chronometric TMS studies (typically 10–20 trials per condition; Duncan et al., 2010; Schuhmann et al., 2012; Jackson et al., 2015). The four time points for testing (50, 100, 150, and 200 ms) were spaced so that the effects of TMS could be separated between two consecutive time windows given the known duration of single-pulse TMS. The earliest and latest of the time points were determined because the fastest pIFC response to print is shown to appear in 100–150 ms after word onset and no earlier than 50 ms (Pammer et al., 2004; Wheat et al., 2010) and (2) the other two neural components of reading (i.e., FG and SMG) start to respond in 150–250 ms after word onset (Marinkovic et al., 2003; Pammer et al., 2004; Dien, 2009; McDonald et al., 2010; Vartiainen et al., 2011). In addition, the lexical frequency of target words was evenly distributed across all the five time points described above (i.e., no significant difference between any two time points, p > 0.96 for both experiments).
Statistical analysis
Reaction time (RT) data for incorrect responses or outliers (±3 SD from the mean) in each participant were excluded from analysis for each experiment. A preliminary inspection of the accuracy and RT data revealed that participants made more errors but responded faster in perceptual judgments than in oral reading and semantic judgments in Experiment 1 (see Results). Given the between-task differences in speed–accuracy trade-off, we chose to use composite RT–accuracy scores (Townsend and Ashby, 1983) as a primary performance measure to assess the behavioral effects of TMS. For each participant, these composite scores were calculated by dividing RT by accuracy for each timing for each task for each site. Since fast decisions tend to be less accurate (Heitz, 2014), this composite index allowed us to control for the speed–accuracy trade-off in behavioral performance across different tasks by weighing fast performance against poor accuracy (Townsend and Ashby, 1983). Specifically, these composite scores are equal to plain RTs in the absence of errors and increase in proportion with the number of errors such that higher scores overall represent poorer performance. This behavioral index is typically used for controlling for between-group differences in speed–accuracy trade-offs (Akhtar and Enns, 1989; Simon et al., 2008; Choi et al., 2021) and commonly employed in previous TMS studies (Chambers et al., 2004b; Cattaneo et al., 2009; Jackson et al., 2015).
For each participant, we then calculated the magnitude of TMS-induced behavioral interference by subtracting the composite scores of the baseline from those of the four different time points (50, 100, 150, and 200 ms) for each task for each site (Figs. 3, 4; see also Extended Data Figs. 3-1, 4-1 for raw composite scores at each time point). Since the resulting behavioral measures did not meet the assumptions of normality (p < 0.001; Shapiro–Wilk test) and homogeneity of variance (p < 0.001; Levene test), these values were square-root transformed to reduce variances and analyzed with nonparametric statistical testing. To determine when TMS started to interfere with behavior after stimulus onset, we first performed a one-sample Wilcoxon signed-rank test for each task for each time point for each site (two-tailed) to examine whether or not the negative impact of TMS on composite scores departed from zero. For each experiment, all p values derived from these one-sample tests (36 tests for Experiment 1 and 24 tests for Experiment 2) were corrected using Bonferroni's method for multiple comparisons. Because these initial tests revealed that the negative effects of TMS on behavior emerged at 100 ms after word onset (Tables 1, 2), all subsequent analyses described below were restricted to the three time points from 100 to 200 ms for both experiments.
Effects of TMS on behavioral performance in Experiment 1. For each site, behavioral effects of TMS are calculated using RT–accuracy composite scores relative to the baseline (see Materials and Methods) and plotted in the ordinate for each task for each time point (see Extended Data Fig. 3-1 for raw composite scores at each time point). In this and subsequent figures, light yellow bands indicate significant effects of interference induced by TMS (one-sample Wilcoxon signed-rank tests). Asterisks indicate significant site × task interaction between pIFC and other sites (see Results). Crucially, TMS of pIFC, but not that of FG and SMG, interfered with oral reading more greatly than semantic and perceptual judgments at 100 ms after word onset, suggesting a double dissociation in the very early stage of reading (see Results). Box plot elements are defined as follows: center line, median; box limits, upper and lower quartiles; whiskers, 1.5 × interquartile range.
Figure 3-1
Effects of TMS on behavioral performance in Experiment 1. For each site, RT-accuracy composite scores for each time point (see Materials and Methods) are plotted for each task for each time point. Download Figure 3-1, TIF file.
Effects of TMS on behavioral performance in Experiment 2. TMS of pIFC, as well as that of FG, disrupted oral reading but not object naming at 100 ms after word onset (see Extended Data Fig. 4-1 for raw composite scores at each time point). Critically, direct comparison between pIFC and other sites again revealed significant site × task interaction (asterisks), replicating the double dissociation at 100 ms observed in Experiment 1.
Figure 4-1
Effects of TMS on behavioral performance in Experiment 2. For each site, RT-accuracy composite scores for each time point are plotted for each task for each time point. Download Figure 4-1, TIF file.
Summary results of one-sample Wilcoxon statistics in Experiment 1
Summary results of one-sample Wilcoxon statistics in Experiment 2
Next, we further compared the behavioral effects of TMS across tasks, sites, and time points to examine the regional specific effects of TMS in the reading network more rigorously. This is most critical to determine the specific causal role of pIFC in the early stage of reading, because functional double dissociations can be specifically identified by comparing the behavioral effects of TMS between different sites in a task-relevant neurocognitive network (Nakamura et al., 2006; Ueki et al., 2006; Pobric et al., 2010; Schuhmann et al., 2012; Woollams et al., 2017; Osada et al., 2019; Pattamadilok et al., 2019; Osada et al., 2021; see also Sandrini et al., 2011 for review). We first examined residuals after regressing out the three factors of interest (tasks, sites, and time points) and found that they were also not normally distributed (p < 0.001 for both Experiments 1 and 2). Given the severe violations of normality assumption required for conventional statistical methods (e.g., linear mixed-effects models; see Lumley et al., 2002; Grilli and Rampichini, 2015 for review), we compared the effects of TMS across target sites, tasks, and time points by using nonparametric aligned rank transform ANOVA (Leys and Schumann, 2010; Wobbrock et al., 2011) implemented in the ARTool package for R (https://cran.r-project.org). To examine whether TMS effects on oral reading were specific to pIFC, we therefore performed an omnibus 3 × 3 × 3 ANOVA which included the effects of site (pIFC, FG, and SMG), task (reading, semantic, and perceptual), and timing (100, 150, 200 ms) as within-participant factors (see Results). All post hoc tests were corrected for multiple comparisons using Bonferroni's method.
Additionally, we examined whether behavioral effects of TMS differed across sessions since behavioral responses might have sped up while participants performed the same set of tasks three times in each experiment. For each task for each experiment, net effects of TMS per session were calculated for each participant by subtracting the composite scores of the baseline from those averaged over the four time points and submitted to one-way ANOVA treating the effect of session (three levels) as a within-participant factor. This supplemental analysis confirmed that the effect of session was nonsignificant irrespective of tasks in both Experiments 1 and 2 (ps > 0.3 for all).
E-field modelling
The spatial distribution of the TMS-induced electric field (E-field) was estimated using the SimNIBS software package (https://simnibs.github.io/simnibs/). For each participant, the high-resolution T1-weighted MRI volume was used to construct a tetrahedral surface mesh head model which included five types of head tissue (scalp, skull, cerebrospinal fluid, gray and white brain matter) with standard isotropic conductivity values (Wagner et al., 2004; Thielscher et al., 2011). For each cortical target site, the position and orientation of the TMS coil were exported from the Brainsight neuronavigation system and mapped into the SimNIBS coordinates to determine the magnetic vector potential at each mesh node. The strength of the E-field was then calculated using the finite element method for each target site for each participant (see https://osf.io/pa3jw/ for all individual E-field maps in Experiments 1 and 2). In group analysis, the individual-level E-field maps were normalized and rendered onto the surface-based MNI brain space using FNIRT (https://fsl.fmrib.ox.ac.uk/) and averaged across participants for each experiment (Fig. 5).
Computational modeling of the TMS-induced electric field (E-field). For each cortical target site, the strength of the E-field was estimated from the position and orientation of the TMS coil using the finite element method for each participant and mapped onto the surface-based MNI brain space across all participants. The resulting cortical E-field estimates show that single-pulse TMS pulses were successfully delivered to the three target sites in both experiments.
Mini meta-analysis of ERP studies of reading
We observed in Experiments 1–2 that the earliest effects of TMS emerged in FG at 100 ms after word onset (see Results). Given that neural effects induced by single-pulse TMS dissipate in 10–20 ms (Ilmoniemi et al., 1997; Walsh and Cowey, 2000), this unexpected finding suggests that the left FG already responds to print for visual word processing ∼100–120 ms after word onset, seemingly much faster than shown by most previous ERP/MEG studies. Thus, we additionally performed a simple meta-analysis to compare the timing of the earliest TMS effects (100–120 ms after word onset in Experiments 1 and 2) with the known time range of the N170 ERP component during visual word processing. To this end, we first examined the existing ERP literature to estimate the onsets of N170 recorded from the left posterior temporal region. Specifically, we selected 23 previous ERP studies of reading which met the following criteria (Table 3): (1) peer-reviewed research articles identified with PubMed and APA PsycNet, (2) ERP data were collected from healthy adults, (3) participants words or letter strings written in their native script, and (4) data analysis procedure is described in sufficient detail for isolating the N170 (N1) component.
ERP studies of reading included in the meta-analysis of N170 (N1)
Results
Experiment 1
Participants made few errors during the TMS experiment [mean (SD) error rate = 0.12 (0.76) % for oral reading, 4.12 (4.79) % for semantic judgments and 9.62 (10.83) % for perceptual judgments]. On the other hand, participants tended to respond much faster in perceptual judgments [mean (SD) reaction time (RT) = 455 (89) ms] than in oral reading [529 (100) ms] and semantic judgments [608 (87) ms]. To control for the between-task differences in speed–accuracy trade-off, we first computed composite RT–accuracy scores (see Materials and Methods) for each participant for each condition and then assessed the impact of TMS on behavior at the group-level of analysis (Fig. 3). TMS of the three cortical targets exerted an inhibitory influence on behavioral performance since these behavioral measures overall exceeded zero across different sites and tasks. To determine when TMS started to change behavioral measures, we initially used one-sample Wilcoxon signed-rank test to examine whether or not this negative impact of TMS departed significantly from zero for each site for each task for each time point (Table 1).
We found that oral reading was disrupted relative to the baseline when TMS was applied to the left pIFC at 100 ms (p = 0.031), 150 ms (p = 1.20 × 10−4), and 200 ms (p = 2.57 × 10−5) after word onset. TMS of the same site also interfered with semantic judgments only when delivered at 200 ms (p = 0.007) and did not affect perceptual judgments irrespective of timing (ps > 0.9 for all). In contrast, magnetic stimulation of the left FG interfered with both semantic judgments and oral reading. That is, semantic judgments were disrupted when TMS was delivered to FG at 100 ms (p = 0.008) and 200 ms (p = 0.046), whereas oral reading was impaired when TMS was applied at 100 ms (p = 0.005), 150 ms (p = 1.20 × 10−4), and 200 ms (p = 4.29 × 10−5) after word onset. However, TMS of FG did not interfere with perceptual judgments irrespective of timing (ps > 0.2 for all). Furthermore, oral reading was disrupted when TMS was applied to the left SMG at 150 ms (p = 0.015) and 200 ms (p = 1.20 × 10−4) after word onset. SMG stimulation also impaired semantic judgments when applied at 150 ms (p = 0.002) and 200 ms (p = 1.63 × 10−4) but not perceptual judgments irrespective of timing (ps > 0.9 for all). In summary, these initial tests revealed that the negative effects of TMS on behavior emerged at 100 ms or afterward. Subsequent analyses were therefore restricted to the three time points from 100 to 200 ms, as described below.
We next compared the magnitude of behavioral interference across sites, tasks, and time points using aligned rank transform ANOVA (Leys and Schumann, 2010; Wobbrock et al., 2011). Specifically, we examined whether the observed inhibitory effects on oral reading were specific to pIFC using a 3 × 3 × 3 ANOVA which included the effects of site (pIFC, FG, and SMG), task (reading, semantic, and perceptual), and timing (100, 150, 200 ms) as within-participant factors. The main effects of site and timing were both significant (F(2,44) = 5.81, p = 0.006, ηp2 = 0.21 and F(2,44) = 27.93, p = 1.48 × 10−8, ηp2 = 0.56, respectively). The effect of task was neither significant (F(2,44) = 2.06, p = 0.140, ηp2 = 0.09) nor interacted with that of timing (F(4,88) =0.11, p = 0.980, ηp2 < 0.01). However, the effect of site interacted with those of task and timing (F(4,88) = 3.66, p = 0.008, ηp2 = 0.14 and F(4,88) = 5.41, p = 0.001, ηp2 = 0.20, respectively). Importantly, there was a significant three-way interaction between site, task, and timing (F(8,176) = 2.89, p = 0.005, ηp2 = 0.12). We therefore examined the effects of task and site separately for each of the three time points and found that the task × site interaction was significant for 100 and 150 ms (F(4,88) = 4.61, p = 0.006, ηp2 = 0.17; F(4,88) = 4.66, p = 0.006, ηp2 = 0.17) and not for 200 ms (F(4,88) = 0.30, p = 1.000, ηp2 = 0.01). These results are overall consistent with those of one-sample statistics described above and suggest that behavioral effects of TMS emerged differently across tasks and sites at 100 ms after word onset.
To disentangle the task × site interaction described above, we further compared the effects of TMS at 100 ms between pIFC and FG using a 2 × 3 ANOVA which included the site (pIFC and FG) and task (reading, semantic and perceptual). The main effect of site was again significant (F(1,22) = 12.75, p = 0.005, ηp2 = 0.37), whereas that of task was nonsignificant (F(2,44) = 3.30, p = 0.138, ηp2 = 0.13). Notably, there was significant interaction between task and site (F(2,44) = 4.78, p = 0.040, ηp2 = 0.18). We therefore performed three post hoc comparisons to examine whether task-specific effects of TMS differed in magnitude between pIFC and FG. Specifically, the inhibitory effects of pIFC stimulation as measured in composite score (Fig. 3) were +33 for reading, +7 for semantic judgments, and −27 for perceptual judgments, whereas those of FG stimulation were +33 for reading, +43 for semantic judgments, and +24 for perceptual judgments. In pairwise comparisons, we found that the between-task difference in reading versus semantic judgment was significantly greater for pIFC than for FG (i.e., +26 vs −10; p = 0.038). Likewise, the between-task difference in reading versus perceptual judgment was greater for pIFC than for FG (+60 vs +9; p = 0.012). However, the effect of task difference in semantic versus perceptual judgments did not differ in magnitude between the two target sites (+34 vs +19; p = 1.000). In sum, these findings show a double dissociation between pIFC and FG during visual word processing, where pIFC plays a causal and task-specific role in oral reading at 100 ms after word onset.
Next, we also performed a 2 × 3 ANOVA with site (pIFC and SMG) and task (reading, semantic, and perceptual) to compare the effects of TMS at the same time point between pIFC and SMG. The main effect of site was significant (F(1,22) = 7.96, p = 0.03, ηp2 = 0.27), whereas that of task did not approach significance (F(2,44) = 1.47, p = 0.721, ηp2 = 0.06). Again, there was significant interaction between site and task (F(2,44) = 6.30, p = 0.012, ηp2 = 0.22). While the inhibitory effects of SMG stimulation were +18 for reading, +33 for semantic judgments, and +22 for perceptual judgments (Fig. 3), the between-task differences in reading versus semantic judgment (+26 vs −15) and reading versus perceptual judgments (+60 vs −4) were both significantly greater for pIFC than for SMG (p = 0.031 and p = 0.002, respectively). In contrast, the effect of task difference in semantic versus perceptual judgments did not differ in magnitude between pIFC and SMG (+34 vs +11; p = 0.763). These results thus show a double dissociation between pIFC and SMG during visual word processing, again suggesting that pIFC causally and specifically contributes to oral reading at 100 ms after word onset.
In addition, we further performed a 2 × 3 ANOVA with site (FG vs SMG) and task (reading, semantic, and perceptual) and confirmed that these main effects and their interaction were all nonsignificant (F(2,44) = 0.69, p = 1.000, ηp2 = 0.03 for task; F(1,22) = 0.32, p = 1.000, ηp2 = 0.01 for site; F(2,44) = 1.36, p = 0.799, ηp2 = 0.06 for interaction). Therefore, no further post hoc comparison was performed between FG and SMG.
Accordingly, the observed double dissociation between pIFC and other control sites shows that the fast pIFC activity plays a specific and causal role in generating the articulatory codes for reading, rather than spreading the top-down signals over the occipitotemporal regions for visual perceptual processing. Our results also suggest that the pIFC does not necessarily act downstream of FG and SMG and responds to print almost simultaneously with these posterior reading systems at 100 ms after word onset. Alternatively, however, it might be argued that the observed task-specific effects on reading arose from between-task differences in response modality because oral reading, unlike the two judgment tasks, required overt spoken production. In other words, it is still possible that the early effects of TMS are not specific to reading such that TMS of pIFC can similarly interfere with any other task involving spoken production. In Experiment 2, we therefore examined this issue by comparing the behavioral effects of TMS between oral reading and object naming (see Materials and Methods). We employed the two different word production tasks since object naming has been used as a standard control task in neuropsychological, neuroimaging, and brain stimulation studies of reading (see Ubellacker and Hillis, 2022 for review), including the most relevant TMS study by Wheat et al. (2013). More specifically, object naming served as an optimal control task for assessing the nature of the early effects of pIFC stimulation because reading and naming share most neurocognitive components of spoken production, except the critical process of fast phonological activation from visual inputs that occurs only in fluent reading (Moore and Price, 1999; Price and Mechelli, 2005). Consequently, TMS of pIFC will impede oral reading but not object naming at 100–200 ms after word onset if the early pIFC activity indeed mediates such rapid computation of speech codes from print.
Experiment 2
Participants again made few errors during the TMS experiment [mean (SD) error rate = 0.16 (0.89) % for oral reading and 1.20 (2.45) % for object naming]. As in Experiment 1, we calculated composite RT–accuracy scores relative to the baseline for each site for each task for each time point (Fig. 4) and examined whether the inhibitory effects of TMS significantly exceeded zero (Table 2). We again found that oral reading was disrupted when TMS was applied to left pIFC at 100 ms (p = 0.011), 150 ms (p = 0.010), and 200 ms (p = 0.001) after word onset. In contrast, TMS of the same site did not affect the behavioral performance in object naming at any time point (ps > 0.6 for all). We also found that TMS of FG disrupted oral reading at 100 ms (p = 0.001), 150 ms (p = 4.01 × 10−5), and 200 ms (p = 1.43 × 10−4) but did not affect object naming irrespective of timing (ps > 0.07 for all). These findings, consistent with those from Experiment 1, therefore suggest that the earliest effects of TMS emerge in FG at 100 ms after word onset and much faster than predictable from most previous ERP/MEG studies (see below for further analysis). On the other hand, magnetic stimulation of SMG impaired oral reading and object naming at 150 ms (p = 0.004) and 200 ms (p = 0.001), respectively.
As in Experiment 1, we further compared the effects of TMS between the three target sites with a 3 × 2 × 3 ANOVA which included the effects of site (pIFC, FG, and SMG), task (reading and naming), and timing (100, 150, 200 ms) as within-participant factors. The main effects of task and site were both nonsignificant (F(1,22) = 3.51, p = 0.074, ηp2 = 0.14 and F(2,44) = 1.68, p = 0.198, ηp2 = 0.07, respectively) but interacted with each other (F(2,44) = 6.08, p = 0.005, ηp2 = 0.22). On the other hand, the main effect of timing was significant (F(2,44) = 21.48, p = 3.10 × 10−7, ηp2 = 0.49) and interacted with those of task (F(2,44) = 5.46, p = 0.008, ηp2 = 0.20) and site (F(4,88) = 3.00, p = 0.023, ηp2 = 0.12). Importantly, there was significant three-way interaction between site, task, and timing (F(4,88) = 2.62, p = 0.040, ηp2 = 0.11). We therefore examined the effects of task and site separately for each time point. Again, we found that the task × site interaction was significant only at 100 ms (F(2,44) = 11.97, p = 2.11 × 10−4, ηp2 = 0.35) and not at 150 ms (F(2,44) = 1.33, p = 0.821, ηp2 = 0.06) and 200 ms (F(2,44) = 3.67, p = 0.101, ηp2 = 0.14).
Given the significant task × site interaction at 100 ms as described above, we next compared the magnitude of task-specific effects of TMS between pIFC and other control sites. First, the inhibitory effects of pIFC stimulation were +19 for reading and −23 for naming, whereas those of FG stimulation were +22 for reading and +10 for naming (Fig. 4). These effects of TMS were assessed using a 2 × 2 ANOVA with site (pIFC and FG) and task (reading and naming). The main effect of task was significant (F(1,22) = 9.46, p = 0.017, ηp2 = 0.30), whereas that of site was nonsignificant (F(1,22) = 5.64, p = 0.08, ηp2 = 0.20). Critically, there was a significant interaction between site and task (F(1,22) = 8.22, p = 0.027, ηp2 = 0.27), suggesting that the between-task difference in reading versus naming was greater for pIFC than that for FG (i.e., +42 vs +12).
Second, the inhibitory effects of SMG stimulation were +12 for reading and +28 for naming at 100 ms after word onset. We also compared these effects of SMG stimulation with those of pIFC stimulation using a 2 × 2 ANOVA with site (pIFC and SMG) and task (reading and naming). The main effects of task and site were both nonsignificant (F(1,22) = 4.52, p = 0.135, ηp2 = 0.17; F(1,22) = 4.62, p = 0.129, ηp2 = 0.17). However, the critical task × site interaction was again significant (F(1,22) = 21.38, p = 3.95 × 10−4, ηp2 = 0.49), suggesting that the effect of task difference was significantly greater for pIFC than for SMG (i.e., +42 vs −16). Third, we further compared the between-task differences in behavioral effects between FG and SMG using a 2 × 2 ANOVA with site (FG and SMG) and task (reading and naming) and confirmed that the main effects and their interactions were all nonsignificant (F(1,22) = 0.06, p = 1.000, ηp2 < 0.01 for task; F(1,22) = 0.15, p = 1.000, ηp2 = 0.01 for site; F(1,22) = 6.12, p = 0.065, ηp2 = 0.22 for interaction).
Collectively, these results therefore replicate the effects of pIFC stimulation on the early stage of reading observed in Experiment 1 and further show that TMS of the same region has a specific impact on oral reading at 100 ms even when compared with the object naming task requiring overt spoken production. In turn, since reading and naming share neurocognitive components of spoken production other than fast phonological conversion of visual inputs (Moore and Price, 1999; Price and Mechelli, 2005), the present results allowed us to refute the remaining possibility that the task-specific effects observed in Experiment 1 reflected between-task differences in behavioral response modality.
Time range of the early ERP response during reading
In Experiments 1–2, we found that the earliest effects of TMS emerged in FG at 100 ms after word onset, much faster than predicted from ERP/MEG data. This unexpected finding may indicate that the left FG is already active for visual word processing ∼100–120 ms after word onset because neural effects induced by single-pulse TMS dissipate in 10–20 ms (Ilmoniemi et al., 1997; Walsh and Cowey, 2000). To examine this possibility, we compared the timing of the earliest TMS effects (100–120 ms after word onset in Experiments 1 and 2) with the time range of the N170 ERP component associated with visual word processing reported in the ERP literature (Table 3).
We first examined the existing literature of N170 to estimate the onsets of N170 recorded from the left posterior temporal region. On the initial visual inspection of ERP waveforms, we found that category-specific neural response diverged between words and symbols (or consonant strings) at ∼150 ms in most of those studies that directly contrasted words with other stimuli (Maurer et al., 2005, 2008; Cheviet et al., 2022). In other studies, however, the onsets of N170 could be hardly identified on the basis of such visual inspection alone, because these onsets were invariably overlapped by the offsets of preceding P1 response, such that it is almost impossible to determine when ERP response returns to the baseline.
We then examined the onsets and offsets of the time window used for measuring the N170 response in the same studies. This is because we reasoned that these values should represent the time range where N170 is generally thought to occur relative to word onset. By calculating the mean (SD) of the onset and offset times across the 23 ERP studies (Table 3), we found that the N170 component has been isolated within a ∼60 ms time window from 147 (15) ms to 216 (30) ms after word onset.
Since the observed earliest effect of FG stimulation should last during 100–120 ms after word onset (see Results and Discussion), we further examined whether the upper edge of the time range (i.e., 120 ms) fell below the reported onsets of N170 [i.e., 147 (15) ms]. We confirmed that this difference in timing was significant even when the between-study variability in onset time is considered (p = 0.036; Z test; one-tailed), suggesting that the observed effect of TMS occurred before the known onset of N170.
Discussion
Fluent reading is known to rely on the distributed left-hemisphere network that links posterior brain regions involved in visual and phonological processing with the inferior frontal region for speech production. Given the fact that reading speed in proficient readers is very fast, i.e., 330 ms per word (Brysbaert, 2019), this cortical reading network should be fully active within the comparable timescale after the visual exposure to print. However, there seems to remain a substantial gap between the behavioral measurements of fluent reading and the known temporal dynamics in the neural components of reading (i.e., FG, pIFC, and SMG) in the left hemisphere. As argued below, this may be attributed to the possible difference in timing between actual neuronal firing and peak response latencies estimated from ERP/MEG waveforms.
In Experiment 1, we applied single-pulse TMS to the left pIFC, FG, and SMG while participants were engaged in three different behavioral tasks, oral reading, semantic judgment, and perceptual judgment. While the left postero-inferior frontal lobe is known to comprise structurally and functionally heterogeneous subregions (Amunts et al., 2010; Hagoort, 2014), our pIFC target is located at its ventral premotor sector in the inferior frontal/precentral junction and associated with speech motor codes (Hickok and Poeppel, 2007; Hickok, 2012). We found that TMS of the left pIFC interfered with oral reading at 100 and 150 ms after word onset, whereas no such effects of interference appeared during semantic and perceptual judgment at the same time points. Post hoc between-task comparisons further confirmed that pIFC stimulation caused a specific impact on oral reading at 100 ms after word onset. In Experiment 2, we also found that TMS of pIFC has a specific impact on oral reading at 100 ms as compared with the object-naming task involving overt spoken response. This latter finding not only replicates the observed effects of pIFC stimulation on the early stage of reading in Experiment 1 but also corroborates the notion that early left inferior frontal activation mediates the rapid computation of speech motor codes because reading and naming share neurocognitive components for spoken production except the rapid computation of phonological codes (Moore and Price, 1999; Price and Mechelli, 2005). Since the functional disruption induced by single-pulse TMS is estimated to last 10–20 ms (Ilmoniemi et al., 1997; Walsh and Cowey, 2000), this finding suggests that the early activation in the left pIFC specifically contributes to spoken production at 100–120 ms after word onset, thus concurring with previous MEG studies showing that early pIFC response in the same time window mediates speech production codes (Pammer et al., 2004; Cornelissen et al., 2009; Wheat et al., 2010; Klein et al., 2015). Moreover, the double dissociation between pIFC and other sites observed in Experiments 1 and 2 suggests that the fast pIFC activity plays a causal role in generating the articulatory codes during speeded reading, rather than exerting top-down modulatory influences over the occipitotemporal regions in visual perceptual processing (Woodhead et al., 2014).
These findings are seemingly at odds with the previous TMS study by Wheat et al. (2013) that reported no behavioral effects of TMS in the same early period. It is possible, however, that early pIFC response per se could be stronger and more detectable in the regular kana script than in other writing systems because of its very high correspondence between character and sound. This is because it is well known that visual words are pronounced faster when the spelling corresponds to the sound (e.g., “hint”) than when otherwise (e.g., “pint”; Feldman and Turvey, 1980; Frost et al., 1987). Such speed advantage in reading is thought to arise because phonological conversion from print is more straightforward and efficient for regular words. In the same vein, (1) phonologically regular Italian words are read aloud faster as compared with more irregular English words (Paulesu et al., 2000), and (2) Japanese words are pronounced faster when written in regular kana script than when in more irregular kanji script (Nakamura et al., 2007; Yoshihara et al., 2017).
The observed task-specific impact of pIFC stimulation on oral reading can be further explained by the proposal that the fast activation of the left inferior frontal region arises from a sublexical link between orthographic and articulatory motor codes during reading (Klein et al., 2015). This is because (1) such sublexical activation can facilitate oral reading but should be irrelevant to perceptual judgment and (2) it should be also insufficient for semantic judgment which requires higher-order, lexico-semantic activation at the whole-word level (see below for further discussion on semantic judgment). Taken together, the present results show that early inferior frontal activation specifically and causally contributes to the sublexical activation of speech motor codes during reading.
Next, we also found in Experiment 1 that TMS of FG disrupted both oral reading and semantic judgment at 100 ms after word onset, suggesting that this early activation of FG mediates an initial stage of visual word-form analysis commonly required for spoken production and semantic processing. Given the short longevity of neural effects induced by single-pulse TMS (10–20 ms; Ilmoniemi et al., 1997; Walsh and Cowey, 2000), the present findings suggest that visual word-form processing is already initiated in FG ∼100–120 ms after word onset. This interpretation may seem at odds with the fact that ERP response in FG (“N170”) peaks at ∼170 ms after word onset (Brem et al., 2006, 2009). In supplemental analysis of the ERP literature, we confirmed that N170 has been isolated within a ∼60 ms time window from 147 to 216 ms during reading and that the observed earliest effect of FG stimulation occurred before the known onset of N170. A possible account for the apparent discrepancy is that TMS can effectively suppress the early part of neural activation driven by visual stimuli which may precede the highest peak of bell-shaped ERP responses. Indeed, some intermediate outputs of early orthographic analysis are likely to be already present in FG around this timing, because early neural responses from the same area can distinguish orthographically typical letter strings from atypical ones at 100 ms after stimulus onset (Hauk et al., 2006).
It is also important to note that the earliest effects of TMS emerged simultaneously at 100 ms for both pIFC and FG in Experiments 1 and 2, i.e., well before intermediate outputs of early orthographic processing are generated in the left inferior temporal cortex at ∼160 ms after stimulus (Hauk et al., 2006). The present findings therefore suggest that pIFC starts to respond to print independently of the visual word-form system in FG and adjacent posterior temporal cortex that is estimated to yield final outputs only at ∼200 ms (Hauk et al., 2006; Dufau et al., 2015). Moreover, coupled with the finding that the earliest effect of TMS appeared in SMG only later at 150 ms (Table 1), the present data collectively suggest that the pIFC is not necessarily functioning downstream of the FG and SMG as assumed in neurocognitive models of reading (Fig. 1).
It might be argued, however, that early pIFC response occurs after higher-order lexico-semantic activation in the posterior temporal cortex because TMS of FG disrupted semantic judgment as early as 100 ms after word onset. Yet this interpretation is problematic particularly because the earliest effect of pIFC stimulation was found for oral reading but not for semantic judgment at the same time point. Conversely, if pIFC is indeed activated exclusively downstream of FG during visual word processing, then it should be expected that the earliest effect of pIFC stimulation appeared only after that of FG. Obviously, however, this is not supported by the finding that the effects of TMS emerged simultaneously in pIFC and FG at 100 ms after word onset (Table 1). Moreover, the middle part of the left FG targeted in the present study is thought to store prelexical orthographic representations of written words (Cohen et al., 2002; Dehaene et al., 2002) and thus probably insensitive to their lexico-semantic status (Vinckier et al., 2007; Lochy et al., 2018). It is thus highly unlikely that early pIFC activation arises via lexico-semantic activation in the posterior brain regions.
While the articulation codes associated with the inferior frontal region are estimated to be activated ∼300 ms after stimulus onset during visual object naming (Schuhmann et al., 2009; Shinshi et al., 2015), the observed fast activation of pIFC during reading is likely to arise from the direct neural connectivity between occipitotemporal and frontal cortices which is thought to develop with reading expertise (Hannagan et al., 2015). In fact, a direct connection between orthographic codes and speech motor codes has been incorporated in some cognitive models (Coltheart et al., 2001) and associated with the left inferior front-occipital fasciculus by diffusion tractography studies (Vandermosten et al., 2012; Vanderauwera et al., 2018). While it has remained largely unknown how such direct neurocognitive pathway contributes to skilled reading, these neuroimaging data and the present results collectively suggest that it plays a specific and causal role in speeded spoken production as early as 100 ms after word onset. This sublexical neural pathway for pronunciation, although very rapidly activated by print, may be seemingly “dormant” or play only a limited role in proficient adult readers who primarily rely on the more effective whole-word recognition system in the left FG during fluent reading (Cohen et al., 2002; Dehaene et al., 2005). However, this sublexical phonological conversion system may play a greater role in less proficient readers who rely on the serial conversion from letters to sound, because such slow and effortful word processing is known to appear when the expert VWFA system is not yet fully functioning during early literacy development (Sowden and Stevenson, 1994; Aghababian and Nazir, 2000; van den Boer and de Jong, 2015) or when it is compromised by brain damage (“letter-by-letter reading”; Cohen et al., 2003; Fiset et al., 2006; Ablinger et al., 2014). Based on those neuroimaging data and the present TMS results, we therefore propose that the seemingly dormant, fast sublexical pathway for pronunciation is nevertheless fully functioning in literate adults.
Furthermore, we found that oral reading was disrupted by SMG stimulation delivered at 150 ms after word onset in Experiments 1 and 2. On the one hand, this finding well concurs with a body of neuroimaging studies showing that SMG mediates phonological processing during reading (Petersen et al., 1988; Pugh et al., 2001; Raizada and Poldrack, 2007). On the other hand, however, our results also suggest that neural response in SMG starts much faster than thought previously, because the same region associated with print-to-sound conversion has been shown to respond ∼300 ms after stimulus onset (Khateb et al., 1999; Pammer et al., 2004; Dien, 2009). Coupled with the other early effects observed in pIFC and FG, our results overall indicate that neural systems for reading act much faster than assumed by most neurocognitive models of reading derived from ERP/MEG data. The precise temporal dynamics during the reading would be of critical importance for understanding the neurophysiology of dyslexia as well as for developing the neuromodulation methodology for effective intervention (Costanzo et al., 2019; Marchesotti et al., 2020). Future research using high temporal resolution methods, such as electrocorticography and electrical cortical stimulation, is needed to further elucidate the precise spatiotemporal dynamics in the entire neural network for visual word processing.
Lastly, we also observed that SMG stimulation produced significant effects of interferences on semantic judgment when delivered at 150 and 200 ms in Experiment 1. This finding might be taken as suggesting that SMG is involved not only in reading aloud but also in semantic processing. However, this interpretation is probably premature given the well-established role of the SMG in print-to-sound conversion in visual word processing (Petersen et al., 1988; Pugh et al., 2001; Raizada and Poldrack, 2007). Indeed, SMG subregions are now thought to contribute to multiple levels of phonological processing, from grapheme–phoneme conversion to whole-word reading up to phonological working memory (see Turker and Hartwigsen, 2021 for review). A more likely account for this finding is that semantic analysis of written words may rely on the SMG for phonological computation to a larger extent when the stimuli are presented in a highly regular script (e.g., Japanese kana) than when otherwise. This interpretation is supported by (1) neuropsychological data suggesting that reading comprehension of kana words mainly relies on the dorsal neural pathway including the SMG (Iwata, 1984) and (2) neuroimaging data showing that visual words activate the left SMG more strongly when written in kana than when in logographic kanji even during semantic judgment (Nakamura et al., 2005). Thus, the observed effects of left SMG stimulation should be interpreted as arising at the stage of print-to-sound conversion which occurred during the semantic analysis of kana words.
In conclusion, the present TMS data provide the first causal evidence showing that the early activation of the left inferior frontal cortex contributes specifically to speeded oral reading. As proposed by previous ERP/MEG studies, this fast response to print is likely to reflect rapid activation of articulatory codes, which probably relies on the enhanced long-distance connectivity between occipitotemporal and frontal cortices that developed with the extensive experience in reading. Our results further suggest that neurocognitive components of reading can respond to print ∼30 ms faster than thought previously, but not necessarily in an ordered cascade as assumed by cognitive models of visual word processing.
Data Availability
Behavioral data and individual E-field maps analyzed in the present study are available at the OSF (https://osf.io/pa3jw/). Other behavioral and MRI datasets are archived in the Section of Systems Neuroscience, National Rehabilitation Center for Persons with Disabilities and also available from the corresponding author on reasonable request. All codes used for data analysis are available from the corresponding author on reasonable request.
Footnotes
We thank Tomoaki Komatsu for his assistance in MRI scan acquisition. This work was supported by the Japan Society for the Promotion of Science (Grants-in-Aid for Scientific Research, 20K22263 and 23KJ2187 to T.U., 20K11176 to K.T., 19H03992 and 23H03262 to K.N.) and the Japan Agency for Medical Research and Development (21dk0310115j0001 to K.N.).
The authors declare no competing financial interests.
- Correspondence should be addressed to Kimihiro Nakamura at nakamura-kimihiro{at}rehab.go.jp
