Supramodal Shape Representation in the Human Brain

We can sense an object’s shape by vision or touch. Previous studies suggested that the inferolateral occipitotemporal cortex (ILOTC) implements supramodal shape representations as it responds more to seeing or touching objects than shapeless textures. However, such activation in the anterior portion of the ventral visual pathway could be due to the conceptual representation of an object or visual imagery triggered by touching an object. We addressed these possibilities by directly comparing shape and conceptual representations in early blind (who lack visual experience/imagery) and sighted participants. We found that bilateral ILOTC in both groups showed stronger activation during a shape verification task than during a conceptual verification task made on the names of the same manmade objects. Moreover, the distributed activity in the ILOTC encoded shape similarity but not conceptual association among objects. Besides the ILOTC, we also found shape representation in both groups’ bilateral ventral premotor cortices and intraparietal sulcus, a frontoparietal circuit relating to object grasping and haptic processing. In contrast, the conceptual verification task activated both groups’ left perisylvian brain network relating to language processing, and, interestingly, the cuneus in early blind participants only. The ILOTC had stronger functional connectivity to frontoparietal circuit than to the left perisylvian network, forming a modular structure specialized in shape representation. Our results conclusively support that the ILOTC selectively implements shape representation independently of visual experience, and this unique functionality likely comes from its privileged connection to the frontoparietal haptic circuit.


Introduction 36 37
Object properties can be accessed through multiple sensory channels. For instance, knowledge of an 38 object's shape can be acquired both by vision and touch. This brings up a critical question about the 39 cerebral architecture of object representation: Are shape representations derived from different senses 40 segregated from each other in the human brain or, alternatively, could the brain implement a shared 41 representation of object shape that is abstracted from the senses (e.g., Amedi  impaired shape discrimination but preserved texture discrimination performance (Milner et al., 1991;52 James et al., 2003), whereas lesions in the medial part of the visual cortex cause the opposite syndrome 53 (e.g., Cavina-Pratesi et al., 2010). 54 55 Studies on haptic shape representation highlighted the neural circuit in the ventral frontoparietal cortex. 56 Researchers found that the intraparietal sulcus (IPS; e.g., Amedi et al., 2001Amedi et al., , 2002Stilla & Sathian, 57 2008; Snow et al., 2015) and the ventral part of the premotor cortex (vPMC, e.g., Stilla & Sathian,58 2008; Snow et al., 2015) show greater activation when participants touch objects than textures. Lesions 59 in the superior parietal cortex and the adjacent IPS induce contralateral tactile agnosia characterized 60 Third, the ILOTC might engage in (visual) shape representation in the sighted but conceptual 112 representation in the early blind. Indeed, due to a lack of visual input since birth, the "visual" cortex in 113 the early blind, accompanied by its enhanced connectivity to high-order brain systems, could repurpose 114 its function for cognitive faculties distant from its native computation in vision, like language or 115 mathematics (see Bedny, 2017 for review). In line with this idea, it has been reported that the lateral 116 occipital cortex in the early blind, a region near the ILOTC, is more sensitive to semantic tasks 117 (Noppeney et al., 2003) and lexical semantics (e.g., Bedny et al., 2011) than in the sighted participants. 118 Moreover, the activity in the lateral occipital cortex in the early blind is more synchronized to the areas 119 in the perisylvian language network than in the sighted participants (Bedny et al., 2011). It is thus 120 possible that the ILOTC in the early blind implements conceptual instead of shape representation due 121 to functional reorganization (Bedny, 2017). 122 123 To address these unsolved questions comprehensively in a single study, we used functional Magnetic 124 Resonance Imaging (fMRI) to characterize the brain activity of sighted and early blind participants 125 when they were performing both shape and conceptual verification tasks on the same set of auditory 126 words referring to manmade objects. Importantly, we orthogonalized the pairwise shape similarity and 127 the pairwise conceptual association among the objects we selected (e.g., a "plate" is perceptual similar 128 to a "coin" in shape but is conceptually associated with a "fork" in function). Univariate and 129 representational similarity analyses (RSA; Kriegeskorte et al., 2008) were conducted to localize and 130 characterize the regions implementing shape and conceptual representation. Resting-state functional 131 connectivity (RSFC) was used to detect possible synchronization between the ILOTC and the 132 frontoparietal haptic network or the perisylvian language network. 133 If the ILOTC implemented supramodal shape representation, we would find the ILOTC showing 135 greater activation in the shape task than in the conceptual task in both sighted and early blind 136 participants, and the activity pattern in the ILOTC would encode objects' shape but not conceptual 137 properties. The ILOTC would have stronger connections to the frontoparietal haptic network than the 138 perisylvian language network. If the ILOTC represented objects' conceptual knowledge instead, we 139 would observe greater activation in the conceptual task than in the shape task in both sighted and early 140 blind participants, and the activity pattern in the ILOTC would encode objects' conceptual properties. 141 Alternatively, if the activation in the ILOTC depended on visual experience, we would discover the 142 ILOTC's involvement in shape processing/representation only in the sighted but not in the early blind 143 participants. If such "visual" ILOTC repurposed its function to conceptual processing in the early blind, 144 we could find the ILOTC's involvement in conceptual processing/representation only in the early blind 145 but not in the sighted participants. 146

Results 148
Behavior rating on shape similarity and conceptual association 150 151 In this study, we selected 21 Italian words, which referred to 21 manmade objects, as our stimuli. The 152 selection was mostly based on behavior ratings of object properties from an independent group of 153 sighted participants who did not take part in the fMRI experiment (N = 19; see Stimuli in Materials 154 and Methods about the stimulus selection procedure and criteria). To validate the rating results from 155 the stimulus selection stage and to verify whether the early blind population had similar shape and 156 conceptual knowledge as the sighted control, all participants who took part in the fMRI experiments 157 (N = 48) also rated the object properties of the stimuli selected. These participants formed into three 158 groups: 16 early blind (EB) participants, 16 gender-and age-matched sighted control (SC) participants, 159 and 16 independent sighted (IS) participants (see Participants in Materials and Methods for details). 160 161 Shape similarity and conceptual association were rated on a 7-point Likert scale in a pairwise manner 162 (see Procedures in Material and Methods about the rating procedure). We assessed the inter-rater 163 reliability within each group of participants using the intraclass correlation based on a mean-rating, 164 consistency, two-way random model (i.e., ICC(C,k)) (McGraw & Wong, 1996). Both shape rating 165 (ICC(C,k): 0.953-0.973) and conceptual rating (ICC(C,k): 0.984-0.985) showed "excellent" inter-rater 166 reliability (Koo & Li, 2016) (Supplementary Table 3). We averaged the rating scores within each group 167 and compared them across groups. Figure 1A illustrates that the rating scores on both object properties 168 were highly reliable across three groups (r(208) on shape similarity: 0.957-0.983; on conceptual 169 association: 0.982-0.984), and the pairwise shape similarity was orthogonal to the pairwise conceptual 170 association (r(208): 0.103-0.132). 171 We then averaged the pairwise rating scores of all the participants (N = 48) and calculated the 173 representational dissimilarity matrix (RDM) of shape similarity and conceptual association (i.e., 7 174 minus the mean rating score). The resulting two Model RDMs had comparable variance across pairs 175 of objects (shape similarity: variance = 2.163; conceptual association: variance = 2.498) and therefore 176 offered equated discovery possibilities when correlated with brain RDMs in the subsequent 177 representational similarity analysis. Figure 1B-C shows the organizational structure of the two RDMs, 178 where 21 items were grouped according to the clusters generated by the k-means clustering algorithm 179 (Lloyd, 1982;Vassilvitskii & Arthur, 2006), with the silhouette criterion used to decide the optimal 180 number of clusters (Rousseeuw, 1987). The shape similarity RDM fell into three clusters, 181 corresponding to square, round, and elongated objects ( Figure 1B). The conceptual association RDM 182 fell into seven smaller clusters, corresponding to different occasions in which objects were used 183 ( Figure 1C). For example, the two biggest clusters were related to eating and writing. The conceptual 184 rating results accorded closely with the teleological perspective, which suggests the essence of a 185 manmade object lies in its function, not its physical properties (e.g., Bloom, 1996). 186

187
Behavior rating on other object properties and confounding factors 188 189 Potential confounding factors were also considered. It has been reported that other properties of 190 manmade objects can also modulate brain activity, like object size (big vs. small; e.g., Konkle & Oliva,191 2012), toolness (tools vs. non-tool manmade objects; e.g., Chen et al., 2018), and contextual 192 association (strong vs. weak contextual association objects; e.g., Bar & Aminoff, 2003). These three 193 variables were rated on a 7-point Likert scale (see Procedures in Material and Methods about the rating 194 procedure). Supplementary Table 3 shows the inter-rater reliability within each group of participants. 195 The inter-rater reliability reached "excellent" on object size (ICC(C,k): 0.979-0.992) and varied from 196 "good" to "excellent" on toolness (ICC(C,k): 0.893-0.928). The inter-rater reliability on contextual 197 association differed between sighted and early blind groups. While sighted groups had a "good" to 198 "excellent" inter-rater reliability (SC: ICC(C,k) = 0.856; SI: ICC(C,k) = 0.919), the early blind group 199 only had a "moderate" one (EB: ICC(C,k) = 0.613). Such heterogeneity in the early blind might result 200 from a lack of instantaneous and global information about the environment from the visual input. 201 Table 1 shows the accuracy and reaction time (RT) across participants within each group in shape and 244 conceptual verification tasks. All groups of participants had near-ceiling accuracy on both tasks. The 245 shape verification task took about 130-200 ms longer than the conceptual verification task. We built a 246 linear mixed model to predict the RT in the correct trials with groups of participants (EB vs. SC) and 247 types of tasks (shape vs. conceptual tasks) as fixed effects variables and each participant as random 248 effects grouping factors. The analysis revealed a significant task effect (F(1,30) = 73.055; p < 0.001), 249 whereas neither the group effect (F(1,30) = 0.732, p = 0.399) nor the interaction effect between groups 250 and tasks (F(1,30) = 2.552, p = 0.123) were found significant. Such difference between shape and 251 conceptual tasks is in line with the evidence suggesting that retrieving specific semantic features (e.g., 252 shape knowledge) requires more time than general semantic knowledge (e.g., Hauk, 2016). 253 254 Shape compared to conceptual tasks engaged ILOTC in both EB and SC. 255

256
We first contrasted the neural activity level between the shape and conceptual tasks. To remove the 257 domain-general RT effect, we modeled the trial-by-trial RT variability across two tasks in the first-258 level general linear model using both the variable epoch approach and the variable impulse approach 259 (Grinband et al., 2008). Figure 2 illustrates results while the domain-general RT effect was controlled 260 (vertex-wise p < 0.001, cluster-level family-wise error (FWE) corrected p < 0.05). 261 262 Figure 2A shows the contrast of shape task vs. conceptual task using all participants (N = 48). The 263 shape task and the conceptual task involved dissociable brain networks. The shape task activated 264 bilateral brain areas, including the ILOTC (i.e., the lateral part of the Broadman area (BA) 37), the 265 anterior IPS (aIPS), the posterior IPS (pIPS), the ventral part of the premotor cortex (vPMC), and the 266 inferior frontal sulcus. To verify whether the ILOTC activated in the shape task was the same region 267 as the LOtv reported in previous literature, we projected the peak coordinates of the LOtv from three 268 representative studies (i.e., Amedi et al., 2001Amedi et al., , 2002) to the brain surface and 269 found that these coordinates largely fell over the geometric gravity center of the ILOTC region. In 270 Figure 2A, we can identify two activity epicenters in the IPS-one was anterior, the other was posterior 271 and joined to the intraoccipital sulcus. 272

279
We then looked at the brain activation in EB and SC separately ( Figure 2B-C). Both EB and SC had 280 ILOTC activation in the shape task compared to the conceptual task. To confirm that the regions in the 281 ILOTC found in the two groups were the same, we calculated the overlap coefficient, i.e., the area of 282 the intersection region divided by the smaller area of the two regions. The overlap coefficient of the 283 left ILOTC was 100%, i.e., EB's ILOTC fell within SC's ILOTC. The overlap coefficient of the right 284 ILOTC was 81.5%. Consistent with the results pooling all participants (N = 48), SC also had significant 285 activation in bilateral aIPS, pIPS, and vPMC in the contrast between shape and conceptual tasks 286 ( Figure 2B). Although these regions did not survive the multiple comparison correction at the whole-287 brain level in EB ( Figure 2C), analyses using the significant areas in SC as ROIs showed bilateral aIPS, 288 bilateral pIPS, and the left vPMC in the EB also showed greater activation in the shape task than in the Both EB and SC activated the language network in the conceptual task. However, EB exhibited 294 reduced left lateralization than SC. To measure the extent of lateralization, we extracted the T scores 295 of the top 5% percentage of vertices showing the strongest activation in the contrast between the 296 conceptual task and the shape task within the language network, which was anatomically defined in 297 each participant's native space by combing bilateral STG, bilateral inferior parietal cortices (i.e., the 298 AG), and bilateral SMG in the DKT atlas ( Next, we directly contrasted the neural activity between EB and SC. As a sanity check, we first 309 compared the brain activity level in the shape and conceptual tasks to the resting state between EB and 310 SC. As both tasks included auditory input, the occipital cortex in EB should show enhanced activation 311 due to cross-modal neuroplasticity, and the results showed up as expected (Supplementary Figure 2; 312 vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). We then compared the activity level 313 between the shape and conceptual tasks between EB and SC. We found only one significant region in 314 the left cuneus near the parietooccipital sulcus ( Figure 2D; vertex-wise p < 0.001, cluster-level FWE 315 corrected p < 0.05). Region of interest (ROI) analysis showed that this region in EB had greater 316 activation in the conceptual task than in the shape task (t(15) = -3.447, p = 0.004), whereas in SC, it 317 showed an opposite pattern (t(15) = 3.213; p = 0.006). This finding demonstrates that the earlier "visual" 318 cortex in EB, not the ILOTC, might repurpose itself to higher-level cognitive functions like conceptual 319 processing. 320 Supplementary Figure 3A illustrates the RT effect across two tasks (N = 48; vertex-wise p < 0.001, 322 cluster-level FWE corrected p < 0.05). As expected, it involved both frontoparietal and cingulo-323 opercular networks underlying top-down control (Dosenbach et al., 2008). It also involved regions in 324 the default mode network, which could be because both the shape and the conceptual tasks require 325 mental simulation (Buckner et al., 2008). Intriguingly, contrasting the RT effect between EB and SC 326 revealed the lateral and ventral parts of the occipital cortex (Supplementary Figure 3B; vertex-wise p 327 < 0.001, cluster-level FWE corrected p < 0.05). These regions substantially overlapped with the LOC 328 involved in visual shape perception in the sighted population, suggesting a functional reorganization 329 of these regions in EB. Note that these regions did not overlap with the ILOTC. 330

331
Other object properties did not modulate ILOTC activity. 332

333
To investigate whether the other object properties modulated brain activity in the ILOTC, we 334 conducted a parametric modulation analysis. The set of the parametric modulators included the task 335 type (i.e., the shape task coded as 1 and the conceptual task coded as -1), the z-scores of the RT across 336 all the trials in each run, the rotated components corresponding to object size, toolness, touch 337 experience, word duration, and word frequency. Figure 3 presents the significant brain areas encoding 338 these parametric modulators (N = 48; vertex-wise p < 0.001, cluster-level FWE corrected p < 0.05). 339 340 When potential confounding factors were modeled, the difference between the two task types was still 341 preserved ( Figure 3A): The shape task activated bilateral brain areas, including the ILOTC, the aIPS, 342 the pIPS, and the vPMC. The conceptual tasks mainly activated the language network in the left 343 hemisphere, including the orbital part of the inferior frontal gyrus (IFG) (i.e., BA 47), the aLTC, the 344 posterior part of the STG (pSTG), the SMG, and the AG. Since these brain clusters were more discrete 345 than those reported in the univariate contrast reported in Figure 2 (with no control for alternative object 346 properties), we used the significant regions here to define the ROIs in the following analyses. No 347 regions showed significant differences between EB and SC. We also found the same region in the 348 cuneus when directly comparing EB and SC under a lower threshold (vertex-wise p < 0.001, 349 uncorrected). 350 351 Figure 3B-C shows the brain areas sensitive to the other object properties. The object size was mainly 352 localized to the three scene-selective regions-the transverse occipital sulcus, the parahippocampal 353 place area, and the retrosplenial cortex ( Figure 3B). It has already been reported that these areas also 354 prefer large nonmanipuable objects (e.g., Konkle  1F), we cannot distinguish between these two factors in this study. Moreover, we found a region in the 358 left ventral and medial temporal cortex (mainly in the BA 20), of which the level of activity negatively 359 correlated to touch experience ( Figure 3C), suggesting this region was sensitive to the novelty of 360 objects. We did not find any brain areas significantly modulated by toolness, which might result from 361 the lack of typical tools (e.g., hammer or scissor) in the stimuli. Directly comparing the effects of all 362 these parametric modulators between EB and SC also failed to reveal any significant brain regions.

ILOTC represented shape similarity, not conceptual association in both EB and SC. 371 372
We then used representational similarity analysis (RSA) to investigate whether the ILOTC identified 373 above implemented shape representation (Figure 3, left and right panels corresponding to left and right 374 ILOTC). A three-way mixed ANOVA was first performed between Groups (EB vs. SC), Tasks (shape 375 vs. conceptual tasks), and Representations (shape similarity vs. conceptual association). The Groups 376 factor was between-subject, whereas Tasks and Representations were within-subject factors. In 377 bilateral ILOTC, we only found a significant effect in the representational content and a significant 378 interaction between Tasks and Representations (Table 2). 379 380 Figure 4A illustrates the RSA results in bilateral ILOTC across all participants (N = 48). Bilateral 381 ILOTC represented shape similarity in both the shape task (left ILOTC: t(47) = 10.367, p < 0.001; 382 right ILOTC: t(47) = 7.705, p < 0.001) and the conceptual task (left ILOTC: t(47 = 4.066), p < 0.001; 383 right ILOTC: t(47) = 3.209, p = 0.002). The shape representation was stronger in the shape task than 384 in the conceptual task (left ILOTC: paired t(47) = 5.183, p < 0.001; right ILOTC: paired t(47) = 3.776, 385 p < 0.001). We found no clear evidence that bilateral ILOTC represented the conceptual association in 386 either the shape or the conceptual tasks-only the conceptual effect in the left ILOTC in the conceptual 387 task was marginally significant (t(47) = 2.123, p = 0.039). No significant difference was found in 388 conceptual representation between shape and conceptual tasks (left ILOTC: paired t(47) = 0.558, p = 389 0.580; right ILOTC: paired t(47) = 0.395, p = 0.695). 390 391 Figure 4B highlighted that the population without visual experience (i.e., the EB) showed a largely 392 similar pattern. Bilateral ILOTC represented shape similarity in the shape task (left ILOTC: paired 393 t(15) = 4.568, p < 0.001; right ILOTC: paired t(15) = 3.610, p = 0.003), whereas their shape 394 representation in the conceptual task was less evident (left ILOTC: paired t(15) = 1.220, p = 0.241; 395 right ILOTC: paired t(15) = 1.852, p = 0.084). The paired t-test revealed a significant difference 396 between the two tasks in the left ILOTC (paired t(15) = 3.361, p = 0.004) but not in the right ILOTC 397 (paired t(15) = 1.466, p = 0.163). No evidence supported bilateral ILOTC represented conceptual 398 association in either shape or conceptual tasks (t(15) < 1.282, ps > 0.219). 399 400 We also investigated whether bilateral ILOTC in EB and SC share a matched shape representation 401 ( Figure 4C). By doing so, we measured the within-group coherence-the correlation between each We also used the RSA to investigate whether the brain areas sensitive to the conceptual task in the 447 univariate analyses represented multivariate conceptual association ( Figure 5). Interestingly, although 448 all these regions showed significantly stronger univariate activation in the conceptual task than in the 449 shape task, only the left AG represented the conceptual association in the conceptual task across all showing stronger activation in the shape tasks or in the conceptual task. It shows that the brain areas 473 sensitive to the shape task and those sensitive to the conceptual task belonged to separate network 474 modules in both EB and SC. Figure 6D further compares the mean RSFC across all the pairs among 475 the shape-sensitive regions, among the conceptual-sensitive regions, and between the shape-and the 476 conceptual-sensitive regions. In both EB and SC, the mean RSFC within the shape module (EB: paired 477 t(15) = 10.650, p < 0.001; SC: paired t(15) = 9.563, p < 0.001) and within the conceptual module (EB: 478 paired t(15) = 10.024, p < 0.001; SC: paired t(15) = 8.014, p < 0.001) was significantly stronger than 479 the mean RSFC between the two network modules. whether the "visual" cortex would repurpose its function for conceptual representation due to early 488 visual deprivation (Bedny, 2017). We found that bilateral ILOTC, a region that overlaps with the LOtv 489 (Amedi, 2001;), together with bilateral aIPS, pIPS, and vPMC, showed 490 greater activation when people process shape rather than conceptual attributes of the same name of 491 objects, and their activity pattern encoded shape similarly but not conceptual association among 492 objects. In contrast, regions in the left perisylvian area, including the orbital IFG, the aLTC, the pSTG, 493 the AG, and the SMG, showed greater activation in the conceptual task than in the shape task. RSFC 494 analysis further demonstrated that shape-and conceptual-relevant regions formed distinct brain 495 networks. Interestingly, in all the above results, visual experience had little influence-EB and SC had 496 similar activity profiles and connectivity patterns. 497

498
Our results thus favor the hypothesis suggesting the ILOTC implements supramodal shape 499 representation and reject the alternative hypotheses that such activation depends on visual imagery or 500 conceptual processing. These results echoed various perspectives suggesting object representation in 501 the brain is organized according to properties, not modalities (e.g. In contrast to the view that lLOTC implements supramodal shape representation, one could argue that 505 this region might represent visual shapes in the sighted and haptic shapes in the early blind. Testing 506 this possibility using fMRI is challenging as it is difficult to distinguish supramodal representation and 507 visual representation derived from visual imagery triggered by touch in the sighted participants. One 508 option would be to examine whether the sighted patients with bilateral lesions in the ILOTC have both 509 visual and tactile shape agnosia or only visual shape agnosia. Unfortunately, the two existing cases of 510 bilateral ILOTC lesions cannot convincingly answer this question. One case is patient D. . Besides resorting to rare patient cases, another seemingly plausible option would be selective 520 transcranial magnetic stimulation (TMS) over bilateral ILOTC to evaluate whether it interferes with 521 both haptic and visual shape tasks or only visual shape tasks. However, the pitfall is that even if TMS 522 over bilateral ILOTC does disrupt haptic shape tasks (e.g., longer reaction time), such disruption might 523 be mediated by the disruption in visual imagery, a strategy sighted participants would adopt to facilitate 524 haptic shape tasks (e.g., Lederman et al., 1990;Zhang et al., 2004). 525 526 While conclusive proof is still warranted, converging evidence supports the role of ILOTC in 527 supramodal shape representation in the sighted population. On the one hand, the ILOTC (mainly in 528 BA 37) is anterior to the lateral occipital cortex (LO, mainly in BA 18), which is engaged in visual 529 shape perception. According to embodied semantic theories (Barsalou et al., 2003) and the "anterior 530 shift" phenomenon noted first by Thompson-Schill (2003), the associate cortex anterior to each 531 sensorimotor area can gradually capture the regularities of the activity patterns in its nearby 532 sensorimotor cortices induced by different exemplars of the same concept (e.g., different exemplars of 533 an apple) and generate a schema-like representation as the sensorimotor knowledge of that concept 534 (e.g., the typical color, shape, and action related to an apple). In line with this hypothesis, the region 535 representing objects' color knowledge is localized to the fusiform gyrus anterior to the color perception Converging the two groups of evidence described above-the position in the ventral visual pathway 555 and the connection to the frontoparietal haptic circuit, it appears parsimonious to postulate that the 556 ILOTC act as an operator bridging visual and haptic shape representations. As the shape representation 557 in the ILOTC is sensorimotor-derived, the format of such representation would still be analogical 558 instead of amodal symbolic, which is usually supported by the language system (e.g., the symbol of a 559 "ring" associated with the symbol of "round") (e.g., Paivio, 1986;Bi, 2021). Nevertheless, the 560 supramodal shape representation in the ILOTC should be more abstract than the topographic 561 representation in the earlier sensorimotor cortex (e.g., retinotopy) and can be shared across modalities. 562 Recent studies have shown that the activity pattern in the LO-the visual shape perception region 563 posterior to the ILOTC-were significantly correlated to shape curvatures (Vernon et al., 2016). It is 564 thus possible that the ILOTC might also represent curvature information but independently of specific 565 senses. 566 567 As for the neural representation of functional knowledge, contrasting the conceptual task with the 568 shape task revealed the left perisylvian regions related to linguistic processing (Figure 2 and Figure 3), 569 implying that function knowledge is supported by the language system. This result is supported by a 570 recent massive study with 136 acute left-hemisphere stroke patients . They found 571 that the deficit in tool selection (e.g., choosing the nail for the hammer) was specifically related to 572 lesions in the left perisylvian regions, mainly including the whole length of the lateral temporal lobe 573 and the anterior IFG. The language system might provide a symbolic format of representations, which 574 can better capture the abstract "associations" among holistic concepts. It contrasts with the analogical 575 format of representation grounded in the sensorimotor system (as discussed for the shape 576 representation in lLOTC), which can better reflect the "similarity" in one particular semantic feature. 577 Such findings suggest that function is not an explicit object property, which can be directly derived 578 from sensorimotor experience-we cannot reduce an objects' function to what it looks like and how it 579 is manipulated; it must therefore rely on some sort of abstract/linguistic coding. 580 The differences between these two neural coding mechanisms may explain the discrepancy in the RSA 582 results between shape similarity and conceptual association. RSA assumes that the representational 583 content can be inferred from the distributed activity pattern across cortical surfaces. The most 584 definitive evidence supporting this assumption comes from the primary sensorimotor system following 585 a topographic organization (e.g., retinotopy), where the input and output information is transparently 586 projected to the cortical surface. Since the shape representation (e.g., curvatures) is transited and 587 abstracted from the topographic representation (Vernon et al., 2016), the activity pattern across the 588 cortical surface in the shape-relevant regions would still be informative. However, in the linguistic 589 system, the representation is presumed to be coded in the format of "arbitrary" symbols, where the 590 linguistic sign (e.g., word forms) bears no obvious resemblance to the content signified. The content 591 represented in the language system thus is not directly transited or abstracted from the word form 592 representations in the sensorimotor cortex and might not be transparently reflected on the activity 593 pattern across the cortical surface. Our results confirmed this hypothesis. Whereas all the shape- Our study also reveals crucial neuroplastic principles about how the "visual" cortex reorganizes its 607 function after vision loss. In the high-order visual cortex, where brain areas receive not only visual 608 input but also information from other sensorimotor systems, brain functions are likely to be resilient 609 to vision loss through compensation. The most well-documented example is the region hMT+/V5, a 610 highly specialized area for visual motion processing. This region also has a direct white-matter 611 connection to the planum temporale specialized in auditory motion processing (Gurtubay-Antolin et 612 al., 2021) and preferentially responds to moving auditory and tactile stimuli in the early blind (e.g., in EB showed greater activation to the conceptual task than the shape task, whereas the same 624 conceptual preference can only be observed in the left perisylvian linguistic areas in SC (Figure 2). 625 Similarly, the lateral occipital cortex and the posterior fusiform gyrus in EB-two "earlier" regions 626 along the visual processing stream than the ILOTC-showed a domain-general RT effect, which is 627 typically observed in the frontoparietal and cingulo-opercular areas in SC (Supplementary Figure 2). 628

629
To conclude, our study identified dissociable brain networks representing objects' shape and 630 conceptual knowledge. The bilateral ILOTC-IPS-vPMC circuit represented shape knowledge, and the 631 left perisylvian circuit related to language processing represented conceptual knowledge. Relying on 632 data collected in EB, we highlighted that the ILOTC represents shape knowledge independently of 633 visual experience. We argue that the ILOTC implement a supramodal shape representation by virtue 634 of its position in the ventral visual pathway and its strong connections to the IPS-vPMC circuit 635 involved in haptic processing, and such sensorimotor-derived representation differs from the 636 disembodied representation supported by the language system in their representational formats. 637

Materials and Methods 639
Participants 641 Forty-eight native Italian speakers with no history of neurobiological or psychiatric disorders 643 participated in the fMRI experiment. Thirty-two participants were sighted, and sixteen participants 644 were early blind. Further recruitment of the blind participants was stalled due to Covid-19 restrictions.  Table 1 shows the demographic information of the early blind and their matched 658 sighted control. In each matched pair, the gender was the same, and the age difference was no more 659 than three years. All blind participants were blind since birth except for three participants, who also 660 had visual trouble since birth but fully lost their vision at eight months, two years, and four years. 661 These participants' data did not differ from those of the other blind participants. 662 The ethical committee at the University of Trento approved the experimental protocol in this study. 664 All participants provided written informed consent and were paid for their time. As a starting point, we preselected a set of Italian words referring to 60 everyday manmade objects 691 based on our subjective impressions so that, among these objects, the shape similarity did not always 692 correlate to the conceptual association. For example, a plate ("piatto") is perceptually similar to a coin 693 ("moneta") but conceptually relates to a fork ("forchetta"). 694 695 Next, we recruited 19 sighted native Italian speakers (age: M = 25.4, SD = 3.6) who did not participate 696 in the fMRI experiments to rate the shape similarity and the conceptual association among the 60 697 objects. As pairwise rating among numerous items is time-consuming (60 objects requires 1770 pairs 698 of comparison), we adopted the multi-arrangement method (Kriegeskorte & Mur, 2012). By doing so, 699 participants arranged Italian words on a computer screen by mouse drag-and-drop operations in two 700 45 min task sessions. The closeness among the words was required to reflect shape similarity in the 701 shape task session and conceptual association in the conceptual task session. Participants were 702 instructed to disregard other object properties like color and size. The pairwise dissimilarity matrix of 703 shape and conceptual information was estimated as the weighted mean of the scale-adjusted on-screen 704 distances from individual arrangements. We averaged the ratings across participants and obtained a 705 mean pairwise dissimilarity matrix for shape and conceptual information, respectively. 706 707 Then, these participants rated the potential confounding factors, i.e., object size (big vs. small), 708 toolness (tools vs. non-tool manmade objects), and contextual association (strong vs. weak contextual 709 association objects). Participants were instructed to rate these three unidimensional variables by sliding 710 a horizontal slider from left to right on a computer screen. To assess the variance in familiarity across 711 objects, participants also rated each object on a 7-point Likert scale about the degree to which they 712 knew its typical shape and primary function (1: do not know it at all; 7: know it very well). We also 713 conducted a telephone interview with 16 early blind participants (eight females; age: M = 33.0, SD = 714 6.6; six of the participants took part in the fMRI experiment). We let them perform the same shape and 715 conceptual familiarity rating tasks and asked them whether they had ever touched the objects. We 716 averaged the rating score across participants to obtain a mean rating score for each object and each 717 rating task. 718 719 After that, we selected 21 from the 60 Italian words based on the above ratings. This set of words met 720 the following criteria: (1) Both sighted and early blind participants knew each object's typical shape 721 and primary function. The shape and the conceptual familiarity rating scores were higher than 5.8 to which they knew its typical shape and primary function (1: do not know it at all; 7: know it very 746 well). They also rated how frequently they touched each object (1: have never touched it before; 7: 747 touch it every day). We then explained the items of which either shape or conceptual familiarity rating 748 score was below 6 points to ensure that all participants knew each object's typical shape and primary 749 function. Supplementary Text 1 shows the survey questions of these ratings. 750

751
During the fMRI scanning, we presented audio stimuli using Psychotoolbox-3 752 (http://psychtoolbox.org/). The sound was delivered through in-ear headphones. Before the formal 753 scanning, we adjusted the volume for each participant so that they could hear the pronunciation clearly 754 under the scanning noise but did not feel too loud. To ensure both sighted and blind participants 755 received the same input during the scanning, we blindfolded all participants and turned off the lights 756 in the scanning room. 757

758
The scanning session included one resting-state run at the beginning (8 min), ten task-state runs (5 min 759 30 s each), and one run collecting T1 weighted images after the first five task-state runs. During the 760 resting-state run, participants were instructed to keep their heads still, not fall asleep, and not think 761 about particular things. During the task-state runs, participants performed verification tasks on the 762 words they heard. 763 764 Each task-state run was divided into two even blocks. One corresponded to the shape verification task, 765 and the other corresponded to the conceptual verification task. The order of the two task blocks was 766 interleaved across runs within each subject, and the order in the first run was counterbalanced across 767 subjects within the early blind and the sighted control group. Each block started with a 10 s rest, 768 followed by a 20s task probe. In the shape verification block, we instructed participants to think 769 carefully about objects' shape ("Pensa attentamente alla forma") and judge whether they were 770 elongated ("allungato"), angular ("angolare"), hollow ("cavo"), circular ("circolare"), and discal 771 ("discoidale"). In the conceptual verification block, we instructed participants to think carefully about 772 objects' function ("Pensa attentamente alla funzione") and judge whether they were used for eating 773 ("per mangiare"), writing ("per scrivere"), sleeping ("per dormire"), lighting ("per illuminazione"), and 774 purchasing ("per fare acquisti"). These five shape and conceptual verification tasks were randomly 775 assigned to each participant's first five task-state runs, and the second five task-state runs repeated 776 these tasks in the same order. In this way, gaps between the same tasks were evenly distributed, and 777 the same tasks could not be repeated in close time proximity. Participants made a yes/no judgment by 778 pressing buttons using their right index/middle fingers. To counterbalance the motor effects of 779 different fingers, we instructed each participant to switch the correspondence between yes/no 780 judgments and index/middle fingers in the second five runs. 781 782 Each block included 21 trials after the task probe, with 21 words presented once. Each trial started 783 with a 100 ms beep to capture participants' attention, followed by a 300 ms silence and an auditory 784 word (word duration: M = 662 ms, SD = 165 ms). The stimulus onset asynchrony was jittered as either 785 5 s or 8 s-eleven trials lasted 5 s, and ten trials lasted 8 s. The order of the words and the jitter intervals 786 were randomized for each block. Participants were instructed to press buttons within 5 s after the 787 stimulus onset. The reaction time was measured as the interval between the stimulus onset and the 788 button press. 789 790 After the fMRI scanning, participants also rated object properties. For pairwise shape similarity and 791 pairwise conceptual association, we adapted the paradigm for both sighted and blind populations by 792 presenting the stimuli in the auditory modality. In each trial, participants heard two words in sequence 793 and rated on a 7-point Likert scale (for shape rating, 1: not similar at all, 7: identical in shape; for 794 conceptual rating, 1: not associated at all, 7: strongly associated). Both rating tasks consisted of 210 795 trials covering all the possible object pairs. For the other three object properties as potential 796 confounding factors, participants rated item-wise on a 7-point Likert scale. They were object size (1: 797 as small as a needle, 7: as big as a television), tool-ness (1: nontools like a lamp, 7: tools like a hammer), 798 and conceptual association (1: weak contextually associated like a cellphone, 7: strong contextually 799 associated like a bowling ball). Supplementary Text 1 shows the survey questions of these ratings. 800

801
Behavior analysis 802 803 For pairwise shape similarity and conceptual association ratings, we averaged the rating scores across 804 all participants who took part in the fMRI experiment and calculated the Model RDMs for the 805 following representational similarity analysis (i.e., 7 minus the mean rating score). To investigate the 806 organizational structure of the two Model RDMs, we performed the clustering analysis using the k-807 means clustering algorithm (Lloyd, 1982;Vassilvitskii & Arthur, 2006). The maximum number of 808 iterations was 10000, the number of times to repeat clustering using new initial cluster centroid 809 positions was 100, and the silhouette criterion was adopted to decide the optimal number of clusters in 810 the range from 2 to 10 (Rousseeuw, 1987). We conducted this analysis using the kmeans and 811 evalclusters function in Matlab 2021. 812

813
For the ratings on other object properties (i.e., object size, contextual association, and toolness) and 814 touch experience, we averaged the rating scores across all participants to obtain a mean vector for each 815 rating item. These mean rating vectors, together with word duration and word frequency, constituted 816 the potential confounding factors. To investigate the effect of these factors in the subsequent 817 parametric modulation analysis, we orthogonalized these unidimensional variables using principal 818 component analysis. Varimax rotation was applied to increase the interpretability of components, and 819 five rotated components of which the eigenvalues were greater than 1 were selected. The principal 820 component analysis was performed using the principal function in the R package psych 2.1.9. The GLM for the univariate contrast analysis involved three events-the shape task, the conceptual 866 task, and the task probe. The duration of shape and conceptual tasks was set as each trial's RT, and the 867 duration of task probes was set as the auditory period before each block introducing the task ahead. 868 The resulting boxcar function was convolved with a canonical hemodynamic response function (HRF). 869 In this way (i.e., the variable epoch approach), the trial-by-trial RT variability was modeled (Grinband 870 et al., 2008). To further control the domain-general effect of RT across the two tasks, we also used 871 stick functions to model the trial-by-trial RT variability. We pooled the trials in the two tasks together, 872 modulated the amplitude of sticks by the mean-centered RT, and convoluted the RT-modulated stick 873 function with the canonical HRF (i.e., the variable impulse approach). The resulting RT variable was 874 involved in the GLM as one regressor. We contrasted the shape task, the conceptual task, and the RT 875 regressor to the resting state and contrasted between shape and conceptual tasks. The obtained 876 combined beta images were used in the second-level analysis. 877

878
The GLM for the parametric modulation analysis only involved two conditions-the trials (i.e., shape 879 and conceptual tasks pooled together as one condition) and the task probes. The duration of trials was 880 set as its RT, and the duration of task probes was set as the auditory period before each block 881 introducing the task ahead. We modulated the condition of the trials with a set of parametric variates, 882 including the task type (i.e., the shape task coded as 1 and the conceptual task coded as -1), the z-883 scores of the RT across all the trials in each run, the rotated components corresponding to word 884 duration, word frequency, object size, toolness, and touch experience. The option for orthogonalizing 885 modulations in the SPM was turned off (Mumford et al., 2015). We contrasted each parametric 886 modulator to zero. The obtained combined beta images were used in the second-level analysis. 887

888
The GLM for the RSA involved each word in each task as a separate condition and the task probes as 889 one condition. Following Kriegeskorte et al. (2008), we concatenated ten runs to improve the reliability 890 of the model estimation. The duration of trials was set as its RT, and the duration of task probes was 891 set as the auditory period before each block introducing the task ahead. The trial-by-trial RT variability 892 across two tasks was also modeled using the variable impulse approach. We contrasted each word in 893 each task to the resting state. The obtained T images instead of the beta images were used in the 894 following RSA (Misaki et al., 2010). 895

896
Representational similarity analysis 897 898 The RSA was conducted among the 21 object conditions within shape and conceptual tasks separately. 899 It included two steps of correlations (Kriegeskorte et al., 2008). In the first-order correlation, we 900 calculated the Spearman distance of the activity patterns across vertices between each pair of 901 conditions and obtained a 21 × 21 Neural RDM for a particular region. In the second-order correlation, 902 we correlated the Neural RDM and each Model RDM (i.e., shape similarity and conceptual association) 903 across the 210 pairs using Spearman correlation. The resulting correlation coefficients were Fisher z-904 transformed using the inverse hyperbolic function. 905

906
The ROI-based RSA focused on two sets of ROIs derived from significant brain areas in the second-907 level of the parametric modulation analysis (see below). The shape ROIs were bilateral and had 908 significantly greater activation in the shape task than in the conceptual task-the ILOTC, the aIPS, the 909 pIPS, and the vPMC. The conceptual ROIs were left-lateralized and were significant in the opposite 910 contrast-the orbital IFG, the aLTC, the pSTG, the AG, and the SMG. In cases when clusters were 911 stuck together under the conventional threshold (vertex-wise p < 0.001, cluster-level FWE corrected p 912 < 0.05), we raised the vertex-wise threshold until they were isolated. To investigate whether the ILOTC in EB and SC represented the same content, we compared the inter-922 subject neural RDM correlation within the same group (i.e., EB-EB and SC-SC) and between different 923 groups (i.e., EB-SC). The within-group inter-subject correlation was calculated in a leave-one-subject-924 out manner. The neural RDM of the ILOTC of one participant was correlated to the mean neural RDMs 925 of all the other participants within the same group across the 210 object pairs. This procedure ended 926 up with 16 correlation coefficients for each group. The between-group inter-subject correlation was 927 calculated in two steps. First, the neural RDM of the ILOTC of one participant in one group was 928 correlated to the mean neural RDMs of all the participants in the other group across the 210 object 929 pairs, which generated 16 correlation coefficients for each group. Second, we averaged the correlation 930 coefficients from the EB and SC participants in the same pair to obtain 16 between-group correlation 931 coefficients. These correlation coefficients were calculated using Spearman's correlation and were 932 Fisher z-transformed. 933 To provide a planar visualization about the representational pattern in bilateral ILOTC in the shape 935 task, we performed the multidimensional scaling analysis using the mdscale function in Matlab 2021. 936 The input dissimilarity matrix was the mean Euclidean distance between each pair of conditions 937 averaged across all participants (N = 48). We used the squared stress, normalized with the sum of 4 th 938 powers of the dissimilarities, as the goodness-of-fit criterion to minimize. 939 The correlation coefficients were Fisher z-transformed before the second-level analysis. The ROIs 952 used in the seed-based RSFC and the interregional RSFC also came from the parametric modulation 953   3. Neural correlates of task types and other object properties (vertex-wise p < 0.001, clusterlevel FWE corrected p < 0.05). Dots in colors denote the location of LOtv in three representative studies. (A) Neural correlates of task types (the shape task coded as 1 and the conceptual task coded as -1). (B) Neural correlates of object size. Activations in the significant brain areas positively correlated with object size, i.e., larger objects induced higher activation. (C) Neural correlates of touch experience. Activations in the significant brain area negatively correlated with touch experience, i.e., less-touched objects induced higher activation.  . RSA results of conceptual association in the brain areas with greater activation in the conceptual task than in the shape task. (A) Brain areas with significantly greater activation in the conceptual task than in the shape task defined in Figure 2A. (B) RSA results of these conceptualrelevant areas in shape and conceptual tasks. *: p < 0.05, **: p < 0.01.