The Roles of Left Versus Right Anterior Temporal Lobes in Conceptual Knowledge: An ALE Meta-analysis of 97 Functional Neuroimaging Studies

1Convergent evidence has implicated the bilateral anterior temporal lobes (ATLs), as critical neural substrates for the semantic representation of words, objects, people, and social concepts. Research attention has now shifted to exploring how conceptual knowledge is represented within this bilateral system, with a particular focus on the functions of the right and left ATLs. This large-scale meta-analysis focused on this issue and formally evaluated four prominent accounts from the literature. The “ATL hub-and-spoke” account proposes that the right and left ATLs represent conceptual knowledge in a unified manner as part of a bilateral, coupled system [thereby promoting robust representations: see )]. An extreme version of this account would predict no differences between the hemispheres; however, a more nuanced position holds that graded hemispheric specialization emerges as a consequence of differential connectivity. Conversely, a greater degree of specialization between the right and left ATLs has been proposed, reflecting (1) the modality of stimulus input, (2) the involvement of word retrieval or visual recognition in the task, or (3) the social content of the stimulus. The development of these four accounts has largely been in parallel. There is now a large body of published functional neuroimaging data that can be used to evaluate each of these theories. The key aims of this large-scale meta-analysis were, therefore, to provide a novel synthesis of the functional imaging literature in healthy individuals and a direct simultaneous evaluation of the differing accounts. The principal features of each hypothesis are briefly described below.

2We and others have proposed that the right and left ATLs work in tandem as a bilateral, partially redundant system (“ATL hub-and-spoke”:;;;; ). Interest in the ATLs as a conceptual region primarily stems from the study of semantic dementia (SD) patients who exhibit a selective yet progressive multimodal impairment of conceptual knowledge. In this patient group, atrophy is always bilateral (though often asymmetric;; ). Performance on semantic tasks in SD patients is correlated with the amount of atrophy and hypometabolism in both ATLs, centered on the ventrolateral surface, which directly mirrors findings from neurologically intact participants in PET or distortion-corrected fMRI studies. Unilateral damage to the same areas has much milder effects on semantic performance, which can be detected only if sensitive semantic tasks are utilized. This suggests that the semantic system can withstand the effects of unilateral damage more successfully than an equivalent amount of bilateral damage—the same conclusion as drawn from seminal primate studies of unilateral versus bilateral ATL resection. demonstrated this basic principle in a computational model and formal mathematical analysis which incorporated the assumption that the ATLs represent conceptual knowledge, bilaterally. In an extension of the hub-and-spoke computational model, the “semantic” hidden units in the model were divided into right and left “demi-hubs” (representing the right and left ATLs). When only one of the demi-hubs was damaged, the model's performance was only mildly compromised; when both demi-hubs were damaged, however, the model's semantic performance was severely impaired. Critically, this result held even when the total amount of damage was equated in the unilateral and bilateral lesions. Although primarily a bilateral model of semantic representation, importantly, this account does not preclude graded specialization in each demi-hub. Indeed, in two previous computational models, the left demi-hub was more strongly connected to speech output representations, with the consequence that damage to the left (in comparison with right) demi-hub produced more substantial deficits in picture naming despite equivalent levels of semantic impairment overall.

3Other researchers have proposed a more specialized organization of conceptual knowledge in left and right ATLs, based on performance differences between patients with left and right ATL damage. The “input modality” account emphasizes differences in ATL function based on the input modality of the task. Accordingly, verbal inputs (written or spoken words) are predicted to be associated with the left ATL and non-verbal inputs (pictures) with the right ATL. Evidence for this standpoint stems from SD patients, with bilateral yet asymmetric ATL damage. directly compared the performance of SD patients with R > L or L > R ATL damage on famous face versus written name recognition tasks. Performance on both tasks were impaired in R > L and L > R patient groups, compared with an older adult control group, and exhibited strong item association across face and name modalities (consistent with a bilateral model of semantic representation). However, the L > R group performed more poorly on the name recognition task relative to the R > L group, whereas the R > L group performed more poorly on the face recognition task relative to the L > R group. Similar conclusions have been drawn from studies that used voxel-based morphometry to relate behavioral performance on semantic tasks to the integrity of the ATL gray matter. For example, studied patients with progressive language deficits from mixed etiologies and correlated their performance on word and picture-based semantic tasks with the degree of damage in each voxel. Damage to both ATLs was negatively correlated with performance on both versions of the semantic task. In addition, damage to the left ATL was more strongly correlated with performance on the word-based version and right ATL damage with performance on the picture-based task.

4The word retrieval/visual recognition account predicts differences based on whether the task requires generation of a word based on semantic knowledge or access to semantic knowledge from the visual input (e.g., face recognition). On this view, word retrieval tasks (e.g., naming pictures) rely on the left ATL and other tasks (e.g., object recognition) are supported by the right ATL. This approach is again based on performance differences between patients with left and right ATL damage or resection. correlated patients' performance on naming and object recognition with voxel-based morphometry measures of temporal lobe integrity. Damage to the left ATL was correlated with scores on naming tasks, and damage to the right ATL was correlated with scores on visual recognition. Similarly, patients with unilateral lesions to the left ATL exhibit more severe naming deficits relative to patients with unilateral right ATL lesions. While it is possible to account for these data under the bilateral hub-and-spoke account (see above), the alternative account proposed by these researchers suggests that the left ATL is specialized for the process of lexical access from semantic knowledge, whereas the right ATL is specialized for visual recognition.

5The third account suggests either that both ATLs are specialized for coding social concepts or that the right and left ATLs are differentially involved. Many researchers have noted that the ATLs are involved in social cognition in humans and primates. More recently, several research groups have proposed that part or all of the ATL codes social concepts, including person knowledge and emotional concepts. Deficits in social behavior are often observed in SD patients, including social awkwardness, person recognition deficits, and a loss of empathy. The current literature is inconsistent with regard to any laterality of social concepts across left versus right ATL regions. Clinically, it has been argued that the social impairments in SD patients are typically more severe, or more obvious, when atrophy is R > L. In a novel extension from these clinical findings to fMRI, demonstrated that activation associated with processing socially related words (e.g., “polite”) versus non-social words (e.g., “nutritious”) was localized to the right anterior superior temporal gyrus (STG). However, more recent fMRI studies of social processing and a direct replication of the Zahn et al. task found greater left than right ATL activations. Indeed, the potential role of left as well as right ATL in social concepts is emphasized by the study of, which, in a formal exploration, found social and behavioral deficits in L > R and R > L SD patients.

6Related to these results for social concepts, laterality effects have sometimes been reported for processing of faces and people's names. Specifically, it has been proposed that the left ATL is preferentially involved in processing names of people and the right ATL is preferentially involved in processing familiar faces. There is also an ongoing debate as to whether the ATLs preferentially process semantic knowledge for familiar people, which could be related to their high intrinsic social relevance.

7All four accounts described above draw heavily on evidence from patients with ATL damage. While patient studies have provided important insights into ATL function, there are limitations in their ability to distinguish between the roles of left and right ATL. Much of the evidence comes from SD patients, but these patients always have some degree of bilateral atrophy, precluding strong inferences about the function of each ATL. Functional neuroimaging studies in healthy participants provide an important additional source of constraint over the theories discussed above. While individual fMRI and PET studies provide support for each theory, when put side-by-side, a rather mixed picture emerges. For example, ATL activation is often left lateralized (L > R) following presentation of verbal information, but a corresponding R > L ATL activation for non-verbal inputs is not found, as ATL activation to these is typically bilateral. Similarly, PET studies suggest a left ATL bias for word retrieval tasks, but non-verbal semantic decisions tend to elicit bilateral activation. Likewise, as noted above, neuroimaging evidence is also inconsistent with respect to the involvement of left versus right ATLs in social concepts.

8Rather than picking individual studies, however, the neuroimaging literature on semantic processing is now sufficiently large that formal meta-analytic techniques can be applied to extract reliable trends and used to test the principal ideas from the four contrasting theories, directly. Specifically, the goal of the present study was to aggregate and analyze data from 97 neuroimaging studies of conceptual knowledge. We used activation likelihood estimation (ALE) analysis, a method that extracts coordinates from a set of neuroimaging studies and estimates the likelihood of activation across each voxel in the brain. The resultant “activation likelihood maps” can then be viewed on a standard brain. ALE also allows for formal subtractions between two maps to explore differences between conditions and to explore laterality effects. To compare the bilateral “ATL hub-and-spoke” and the three more specialized accounts of ATL function without any a priori assumptions, three analyses were designed to test the principal notions that differentiate the various hypotheses. With some inevitable simplifications, these map directly onto specific theories that have been articulated in the literature. In addition, they can be tested using a meta-analytic approach. In doing so, of course, the test analyses may not be entirely consistent with every aspect of each specific proposal in the literature, but they do allow a formal mapping and testing of the key hypotheses which, in turn, will inform more sophisticated explorations in future, targeted work.

9To explore right and left ATL function in conceptual knowledge, two separate literature searches were undertaken (the full list of studies are listed in Supplementary Table 1). First, 48 studies from a recent meta-analysis investigating the role of the ATLs in conceptual knowledge were re-analyzed. The inclusion criteria were: (1) PET or fMRI studies exploring semantic memory, (2) studies reported at least one peak in the ATLs (see below for definition); and (3) studies were published in peer-reviewed journals, in English, between January 1992 and March 2008. The exclusion criteria were: (1) studies which did not explicitly include the ATLs in acquisition or data analysis, (2) studies which focused on individual differences (e.g., sex or age differences), patients, task switching, priming or adaptation, language development, syntax, metaphoric or idiom comprehension, and bilingualism or working/episodic memory demands.

10A second literature search was conducted to investigate the role of the ATLs in “social semantics”, as this was not the focus of. Studies were defined as socially semantic if they conformed to one of three subcategories: (1) social concepts: Probing knowledge of socially relevant words or situations (e.g.,; ), (2) person knowledge: Person identification through faces, names, or voices (e.g.,;; ), and (3) emotion processing: Eliciting the concept of an emotion (e.g., ). The emotion subcategory was included to test for any differences between emotional versus social content in ATL function. The following search terms were entered into the Web of Knowledge (www.isiknowledge.com): [“fMRI” OR “PET”], [“anterior temporal lobes” OR “ATLs”] combined with [“social concepts,” “person recognition,” “face recognition,” “name recognition,” “voice recognition,” “emotion”]. The inclusion and exclusion criteria were matched to with the following amendments: (1) PET or fMRI studies which conformed to one of the three subcategories (social concepts, person knowledge, and emotion concepts), (2) published in peer-reviewed journals, in English, between January 1992 and September 2012. These criteria resulted in a cohort of 49 studies.

11ATL is a general term that has been used in slightly different ways by different researchers. Some have used the term in an inclusive fashion to describe areas subsumed by the typical site of atrophy in SD patients, including the temporal pole and the anterior portions of all temporal gyri. Others have used the term more selectively to specify temporal cortex anterior to the limen insula, anterior to the line y = 0 in the MNI space, or to include only tissue falling within the temporal pole proper. These latter definitions tend to exclude the tissue of the anterior inferior temporal gyrus (ITG) and anterior fusiform gyri, which recent fMRI and SD studies have shown to be critical in semantic cognition. In the context of this study, we wished to ensure that we included the anterior aspects of all the temporal gyri in equal amounts, so we constructed a plane perpendicular to the long axis of the temporal lobe (passing through the fusiform gyrus at y = −20, z = −30 and STG at y = 0, z = −5) and defined all temporal peaks anterior to this plane as falling within the ATL (Fig. 1). This region encompassed the temporal pole (BA 38) and anterior portions of the STG (BA 22), middle temporal gyrus (MTG; BA 21), ITG (BA 20), and fusiform gyrus (BA 20).

12While the main focus of this meta-analysis was exploring ATL function, we also investigated the pattern of laterality in the rest of the brain. To do this, a separate analysis was conducted using all contrast coordinates without the ATL restriction. It is important to note that this analysis was based on the same 97 studies used for the ATL analysis. This means that it only contains studies that reported activations in both ATL and other regions. Investigations that reported non-ATL activations alone were not included in this analysis (which might shift the pattern of non-ATL activation likelihoods reported in this study).

13The cohort of studies was considered in three separate analyses, designed to test the predictions of the three specialized accounts. Each analysis divided the cohort into two different types of study proposed to be related to ATL lateralization. The number of studies in each analysis is listed in Table 1 below. Studies included in each analysis are listed in Supplementary Table 1.

14This analysis compared the probability of right and left ATL activation following “verbal” and “non-verbal” inputs. Studies were considered “verbal” if auditory or written words were presented. Studies were considered “non-verbal” if pictures or non-verbal auditory sounds were presented. Six studies were excluded from this analysis because they could not be classified into either “verbal” or “non-verbal” inputs. To investigate the reported left ATL bias for written information, the “verbal input” studies were further subdivided into “auditory words” and “written words.” The “non-verbal input” studies could not be split into “auditory” and “visual” studies because only three studies presented non-verbal auditory stimuli.

15This analysis compared the probability of right and left ATL activation during word retrieval versus other tasks. Studies were classified as requiring word retrieval if overt or covert speech generation, such as naming pictures or reading words or sentences, was required. “Other” tasks were those that required some other processing of the stimulus (e.g., lexical decision, semantic judgements, and item classification). One study was excluded from this analysis because no response was required. To investigate the reported left ATL bias for object naming specifically, the “word retrieval” studies were subdivided into “picture naming” and “word reading” studies.

16This analysis compared the probability of right and left ATL activation for “non-social semantic” and “social semantic” categories. “Non-social semantic” studies were taken from and included studies that investigated general semantic stimuli (e.g., tools, animals). “Social semantic” studies were taken from the second literature search and included studies which required judgements about socially relevant words, people, or emotions. To explore the hypothesized right ATL social bias, the “social semantic” studies were subdivided into “social concepts”, “person knowledge,” and “emotion concepts.”

17The following steps were carried out separately for the three analyses. First, to assess the incidence of unilateral and bilateral coordinates, each study was coded according to whether they reported (1) left ATL peaks only, (2) right ATL peaks only, or (3) bilateral ATL peaks. Chi-squared tests were then computed to test if the proportion of studies reporting a left ATL peak (i.e., unilateral left + bilateral) differed between the two types of study in each analysis (e.g., verbal vs. non-verbal input). This procedure was repeated for right ATL peaks.

18Analyses were performed with GingerALE 2.3 (available at http://brainmap.org/ale/ [date last accessed; 16 February 2015];; ). This version uses a random-effects analysis to minimize potential biases caused by within-experiment or within-subject effects. Spatial smoothing FWHM was determined based on the number of participants within each analysis and ranged from 8.66 to 12 mm. Talariach coordinates were converted to the MNI space using the Tal2MNI (SPM) transform in GingerALE. Areas associated with significant ALE values were plotted on the high-resolution MNI152 template brain using MRICron (http://www.nitrc.org/projects/mricron [date last accessed; 16 February 2015]).

19An overall activation likelihood map for all 97 studies was generated to show ATL coverage. This was thresholded using a false discovery rate (FDR) of P < 0.05 to correct for multiple comparisons. Then, to investigate shared versus specialized ATL function, three separate analyses were carried out (outlined above). In Analysis 1, studies were divided into two sets based on input modality, and the following analysis steps were performed: Separate activation likelihood maps were generated for each set of studies (i.e., verbal and non-verbal) and were thresholded at P < 0.05 (FDR-corrected) to correct for multiple comparisons.Subtraction analyses were conducted between these thresholded likelihood maps to investigate differences between verbal and non-verbal studies. Subtractions were thresholded at P < 0.05 (uncorrected) and were run using 5000 P-value permutations with a minimum cluster size of 200 mm. The conjunction image from this analyses represented regions which were equally likely to be activated by verbal and non-verbal studies.Laterality analyses were also performed to identify regions within the ATL in which activation was more likely in the left hemisphere than the right and vice versa. To achieve this, the “x” coordinates for the verbal studies were left–right reversed (i.e., were multiplied by −1). A new subtraction analysis was then carried out, contrasting these “x-reversed” co-ordinates with the original likelihood peaks, to highlight regions for which activation for verbal inputs in the left hemisphere was more likely than activation of the homologous region in the right hemisphere (and vice versa). The conjunction image from this analysis represented regions which were equally likely to be active in both hemispheres. This process was repeated for the non-verbal studies. Cluster volume sizes from the laterality analyses for each dimension were extracted and plotted separately to illustrate the volume of tissue associated with bilateral versus lateralized activation.For Analysis 2, the same steps were performed, except studies were divided based on whether they involved word retrieval or other processes. Finally, Analysis 3 repeated the same steps with studies divided according to semantic category (social vs. non-social).

20Separate activation likelihood maps were generated for each set of studies (i.e., verbal and non-verbal) and were thresholded at P < 0.05 (FDR-corrected) to correct for multiple comparisons.

21Subtraction analyses were conducted between these thresholded likelihood maps to investigate differences between verbal and non-verbal studies. Subtractions were thresholded at P < 0.05 (uncorrected) and were run using 5000 P-value permutations with a minimum cluster size of 200 mm. The conjunction image from this analyses represented regions which were equally likely to be activated by verbal and non-verbal studies.

22Laterality analyses were also performed to identify regions within the ATL in which activation was more likely in the left hemisphere than the right and vice versa. To achieve this, the “x” coordinates for the verbal studies were left–right reversed (i.e., were multiplied by −1). A new subtraction analysis was then carried out, contrasting these “x-reversed” co-ordinates with the original likelihood peaks, to highlight regions for which activation for verbal inputs in the left hemisphere was more likely than activation of the homologous region in the right hemisphere (and vice versa). The conjunction image from this analysis represented regions which were equally likely to be active in both hemispheres. This process was repeated for the non-verbal studies. Cluster volume sizes from the laterality analyses for each dimension were extracted and plotted separately to illustrate the volume of tissue associated with bilateral versus lateralized activation.

23Figure 1 shows the activation likelihood map for all 97 studies. Activation likelihood was greatest in the superior, lateral portions of the right and left ATLs, with low activation likelihood in the ventral ATLs. The ventral ATLs suffer from severe signal distortion and dropout in standard gradient-echo EPI fMRI and this, as well as other methodological factors, probably account for their under-representation here [see )]. One consequence of this inherent limitation in the literature is that studies using non-verbal pictorial stimuli, which are likely to rely heavily on the inferior aspects of the ATLs, may be under-represented. Recent studies that have used distortion-corrected fMRI to improve signal have demonstrated robust activation in ventral ATL regions for a range of semantic tasks. This aligns with evidence from PET studies and with a recent study of hypometabolism in SD that linked the patient's semantic deficits with dysfunction centered on the anterior fusiform. It is likely that the paucity of activation observed in this area in the current meta-analysis reflects the reliance on (non-distortion-corrected) fMRI in the majority of studies, highlighting the need for future studies to take steps to improve imaging sensitivity in this region. For present purposes, however, this means that our analyses are necessarily focused on the lateral and superior aspects of the ATLs.

24Next, the 97 studies were divided according to the predictions of the three specialized accounts of ATL function (“input modality,” “word retrieval,” and “semantic category”). To summarize what follows: (1) in each analysis, there was a high likelihood of bilateral ATL activation; and (2) variation within each analysis, reflecting the key feature of each specialized account, generated considerable common (bilateral) activation and relatively little specific activation. Beyond the overall bilateral distribution of activations, there was evidence of secondary specializations between and within the ATLs. Activation likelihood peaks for each ALE analysis are reported in Supplementary Table 2.

25The bar chart in Figure 2A divides studies according to whether they reported unilateral left, unilateral right, or bilateral ATL peaks. The majority of studies across the input modalities reported bilateral peaks. Chi-squared analyses showed no significant differences between the proportion of “verbal” and “non-verbal input” studies reporting left ATL co-ordinates (χ2 = 1.84, P = 0.18) or in the proportion reporting right ATL coordinates (χ2 = 1.52, P = 0.22).

26In the ALE analysis, both the “verbal input” and “non-verbal input” studies produced predominantly bilateral and highly overlapping activation, encompassing STG (BA 38) and MTG (BA 21; Supplementary Table 2).

27Figure 2B shows the subtraction analysis, which reveals regions for which “verbal input” studies were more likely to produce activation than “non-verbal” (in light blue) and vice versa (yellow). The conjunction analysis (green) indicates areas with equal activation likelihood for both sets of studies. The majority of activated cortex, in both ATLs, was equally activated by both input modalities: This common area encompassed bilateral portions of STG (BA 38), extending along MTG (BA 21). Specialized regions for “verbal inputs” showed greater activation likelihood dorsally, along left STG (BA 38) and right MTG (BA 21), whereas specialized regions for “non-verbal inputs” showed greater activation likelihood more ventrally, focused in the left temporal pole (BA 38) and a smaller cluster in right parahippocampal gyrus (BA 34; Supplementary Table 3).

28Figure 2C shows the laterality analyses for “verbal” and “non-verbal” inputs. The laterality analyses tested for regions in which there was greater activation likelihood in one hemisphere compared with the homologous region in the opposite hemisphere and, conversely, for those in which activation was equally likely in both hemispheres. The bar charts plot the volumes of clusters arising from these analyses. For both “verbal inputs” and “non-verbal inputs,” activation in the left and right was equally likely for large areas of superior, lateral ATLs (purple). Smaller, unilateral effects were observed across both sets of studies (Fig. 2C—red and dark blue and see also Table 2); however, the bar charts illustrate that these unilateral clusters were much smaller than the more prominent bilateral clusters. For “verbal inputs,” there was limited evidence for a small degree of left ATL specialization, as predicted by the input modality hypothesis. For “non-verbal inputs,” the corresponding prediction of right ATL specialization was not supported.

29Finally, the sensory input subanalysis, in which verbal studies were subdivided into written versus auditory words, revealed some differences between these two stimulus classes. While both written and spoken words showed bilateral activation across STG (BA 38) and MTG (BA 21; Supplementary Fig. 1A and Table 4), there was a tendency for this activation to be more focused on the left ATL in the case of written words.

30The bar chart (Fig. 3A) indicates within “word retrieval” studies, there were similar numbers of bilateral and left hemisphere-only studies. In contrast, the majority of “other” studies reported bilateral peaks. Chi-squared analyses showed no significant differences between the proportion of “word retrieval” and “other” studies reporting left ATL co-ordinates (χ2 = 2.47, P = 0.12), or right ATL coordinates (χ2 = 3.25, P = 0.07).

31In the ALE analysis, both the “word retrieval” and “other” studies produced predominantly bilateral and highly overlapping activation, encompassing STG (BA 38) and MTG (BA 21; Supplementary Table 2).

32Figure 3B shows the subtraction between “word retrieval” and “other” tasks. Equal activation likelihood for both study types was shown around left STG (BA 38) and, to a lesser extent, bilateral MTG (BA 21). Specialized regions for “word retrieval” showed greater activation likelihood along left STG (BA 38), extending into the left temporal pole and right MTG (BA 21), whereas those for “other” tasks showed greater activation likelihood along right MTG (BA 21), extending toward the temporal pole bilaterally (Supplementary Table 3).

33Figure 3C shows the laterality analyses for “word retrieval” and “other” tasks. For “word retrieval,” left and right activation was equally likely in some areas of STG (BA 38) and MTG (BA 21); however, these bilateral regions were outweighed by a large area of STG that showed increased activation likelihood in the left ATL compared with the right ATL, and a smaller, more anterior STG region that showed the opposite effect (Table 2). In all, the volume of the left ATL cluster far exceeded that of the right ATL or the bilateral clusters, indicating that tasks involving word retrieval activate large areas of the left ATL, to a greater degree than the right. This provides support for the notion that word retrieval tasks rely heavily on the left ATL. For “other” tasks, the majority of cortex was equally likely to be active in the left and right ATLs (purple), with smaller unilateral effects observed in both hemispheres (Table 2). The bar chart illustrates that these unilateral clusters were outweighed by the prominent bilateral cluster.

34The secondary analysis subdivided the “word retrieval” studies into reading versus naming studies and showed that reading studies principally activated the left ATL, whereas naming studies produced more bilateral activation (Supplementary Fig. 1B and Table 4). Therefore, as we also observed in the previous analysis, there was a tendency for processing of written words to produce more left-lateralized activation.

35The bar chart (Fig. 4A) indicates that the majority of studies across semantic categories reported bilateral peaks. There was, however, some asymmetry among studies that reported unilateral peaks, with non-social studies tending to report left rather than right peaks and social studies tending to report right rather than left peaks. Chi-squared analyses showed that “non-social semantic” studies were indeed more likely to elicit left ATL peaks than “social semantic” studies (χ2 = 5.55, P = 0.02), whereas “social semantic” studies were more likely to produce right ATL coordinates than “non-social semantics” (χ2 = 6.83, P = 0.01).

36In the ALE analysis, both the “non-social semantic” and “social semantic” studies produced predominantly bilateral and highly overlapping activation, encompassing STG (BA 38) and MTG (BA 21; Supplementary Table 2).

37Figure 4B shows the subtraction between “non-social” and “social semantics.” Equal activation likelihood for both study types encompassed large parts of bilateral MTG (BA 21) and STG, extending into the temporopolar cortex. Specialized regions for “non-social semantics” showed greater activation likelihood in the left STG (BA 38) and a smaller part of ITG (BA 20). In contrast, specialized regions for “social semantics” showed greater activation likelihood in more polar and more medial areas, encompassing left STG (BA 38), bilateral hippocampus, and right ITG (BA 20; Supplementary Table 3).

38Figure 4C shows the laterality analyses for “non-social semantics” and “social semantics.” For both the “non-social and “social semantic” studies, bilateral ATL activations (purple) were most prominent. Again, smaller unilateral effects were observed in the right and left ATLs for both study types (Table 2). For the non-social studies, there was a combination of bilateral clusters and a left (superior) ATL cluster, suggesting a degree of leftward bias in an otherwise bilateral picture. For social studies, the bulk of the ATL tissue showed equal activation likelihood in left and right. There was, therefore, limited support for the idea that social concepts preferentially activate the right ATL.

39Across all three social subcategories (“social concepts,” “person knowledge,” and “emotion concepts”), activation likelihood was bilateral and overlapping, encompassing portions of the STG (BA 38) and MTG (BA 21; Supplementary Fig. 1C and Table 4).

40This analysis investigated effects of each study type across the whole brain to contrast against the previous ATL-focused analyses (Fig. 5). Considering input modality first, there was specialization for “verbal inputs” in areas of the left temporal and prefrontal cortex, predominantly in the left hemisphere. In contrast, “non-verbal inputs” (dominated by picture-based studies) were more likely to activate the posterior ventral temporal lobe bilaterally, although this was stronger in the right hemisphere. Specialized regions for “word retrieval” were found almost exclusively in the left hemisphere along the length of the STG extending into the inferior frontal gyrus, whereas “other” tasks showed greater activation likelihood in right ATL regions and in the orbitofrontal cortex. Greater activation likelihood for “non-social semantics,” relative to social semantics, was observed in large areas of temporal and prefrontal cortex, predominantly in the left hemisphere. Specialization for “social semantics” was evident in the superior temporal poles and temporoparietal junction bilaterally. In summary, in contrast to other parts of the brain, the ATLs tended not to show hemispheric specialization along any of the three specialized accounts of ATL function, reinforcing the idea that the two ATLs operate as an integrated, bilateral system.

41In addition, the whole-brain analysis revealed that studies of every type reliably activated the inferior prefrontal cortex (in particular, pars orbitalis) and the posterior MTG, suggesting that these areas are also involved in semantic cognition. A detailed consideration of the function of these areas is beyond the scope of the present study; however, both have been linked to top-down executive influences on the retrieval and manipulation of semantic knowledge.

42This study tested four accounts of the roles of right and left ATLs in supporting conceptual knowledge. While some researchers hold that the ATLs form an integrated bilateral system for representing knowledge, others have proposed a degree of hemispheric specialization, organized by input modality, word retrieval, or semantic category. We tested these predictions in an ALE meta-analysis of 97 neuroimaging studies. The most striking finding was that, predominately, bilateral overlapping activations were observed across all three proposed dimensions of specialization. Secondary to this bilateral pattern, there were two more subtle graded hemispheric specializations: (1) A left hemispheric bias for tasks requiring word generation and (2) some evidence that written word input is more likely to activate the left ATL. In contrast, there was no evidence of hemispheric specialization for social versus non-social concepts. Although extending further into anteromedial ATL regions, activation likelihood for social concepts was firmly bilateral.

43None of the existing accounts of ATL function can fully account for these results. At face value (though see the “Caveats” subsection below), the very clear bilateral activation likelihood maps found in this study seem to be inconsistent with notions of strong specialized distinctions between the left and right ATLs, as would be expected on an extreme version of some theories. Likewise, the presence of second-order graded specialization refutes an extreme, undifferentiated bilateral account in which there are no differences between the left and right ATLs, whatsoever. Instead, we argue that the results fit best with a neurocomputational framework for ATL function (summarized in Fig. 6), which incorporates a basic principle of bilateral representation, but allows for some graded functional specialization emerging as a result of asymmetric connectivity between ATL subregions and primary input/output areas.

44The main result from the current study is that conceptual knowledge of various types activates an ATL system that is both bilateral and transmodal. The role and potential importance of a transmodal representational hub has been discussed and computationally implemented in previous studies. In short, a transmodal hub supports the complex, nonlinear computations required to integrate multiple sources of verbal and non-verbal information into generalizable coherent concepts. Why is a bilateral hub beneficial for conceptual representation? One potential advantage of such a system is that it is more resistant to damage and, indeed, both human and primate data indicate that unilateral ATL damage/resection has much less of an effect on semantic performance than bilateral damage. While the consequences of unilateral damage are certainly not trivial, bilateral damage is typically necessary to produce the profound deficits seen in, for example, SD. Using a neurocomputational dual hub-and-spoke model of semantic representation, provided a computational demonstration and explanation for this advantage by showing that bilateral lesions produced greater semantic impairments than unilateral lesions, even when the total amount of damage was controlled for. During learning, this model was able to use two transmodal “demi-hubs”—analogous to left and right ATLs—to mediate between knowledge of objects coded in different sensory modalities. Following unilateral damage to one “demi-hub”, semantic performance was supported by high fidelity and strong propagation of activation between the undamaged “demi-hub” and the various input/output units. Consequently, this strong and accurate activation propagation compensated, at least in part, for the weakened and distorted activations within the damaged demi-hub. In contrast, bilateral damage to both demi-hubs created representational distortion and activation weakness across the entire semantic system, resulting in much worse performance. Although explored in the context of semantic representation, the advantages for having a bilateral, partially redundant system have also been highlighted in other cognitive domains such as episodic memory and visual recognition—suggesting that this may be a more general neurocomputational principle.

45One remaining question related to a bilateral conceptual knowledge system is why some studies report ATL effects limited to one hemisphere, or report highly specific deficits following unilateral damage. In the context of neuropsychology, a small number of case studies have reported deficits following unilateral lesions which are specific to one modality. While these studies appear to provide support for a more lateralized view of ATL function, one explanation for the apparent sparing of other aspects of knowledge may be that the methods of testing were not sensitive enough. The importance of maximizing testing sensitivity, particularly in patients with unilateral damage, has been highlighted in two recent studies. found that patients with unilateral ATL resection show subtle transmodal semantic deficits, which were generally observed as selective slowing of reaction times and only resulted in increased error rates in particularly demanding semantic tasks (e.g., understanding low frequency and abstract words). Similarly, a number of neuroimaging studies included in this meta-analysis only reported peaks in one ATL (e.g.,;;; ). If the semantic system is truly bilateral, why do some semantic studies observe unilateral activation? Setting aside the influence of using written words as stimuli or requiring word retrieval (see below), a potential answer is that any individual study is susceptible to false positives and false negatives. While false positives can be rigorously controlled through good statistical practice, false negatives are more difficult to avoid and become more of a danger when stringent whole-brain corrections are performed. In the minority of studies that reported unilateral ATL activation, it is entirely possible that subthreshold activation was present in the opposite ATL. Indeed, our meta-analysis suggests that it is unwise to conclude that a particular task activates one ATL and not the other, on the basis of a single study, unless activation levels in each ATL are compared with one another statistically. The experience from reviewing all of the studies included in this meta-analysis suggests that such tests are rarely performed [but see ) for an exception].

46Beyond the overall bilateral transmodal pattern of ATL activation, we observed clear evidence for two graded functional specializations between the ATLs: Studies presenting written words or requiring word retrieval were more likely to activate the left ATL than the right. These results, particularly those in relation to the left ATL and word retrieval, align closely with repeated observations of greater word-finding difficulties in patients with left greater than right ATL damage (e.g., left > right in neurodegenerative bilateral diseases or in comparisons of left versus right ATL resection cases:;;;; ). Likewise, the left bias for written word processing found in this meta-analysis reproduces the results of spoken (bilateral) versus written (left) word processing found in a previous MEG study. This latter result is important, methodologically, given that neuroimaging studies of semantic processing commonly use written word stimuli. This is perhaps unsurprising given that written words allow a full range of concepts to be probed and are trivially easy to source when constructing experiments but, in doing so, the resultant neuroimaging data may be artificially left-biased.

47We propose that these hemispheric specializations could arise in a bilateral semantic system as a consequence of asymmetric connectivity [for implemented computational models of this proposal, see ) and )]. These differences in connectivity could follow from two potentially linked sources. The first is asymmetry in the key white matter pathways. Both the ventral language pathway (tracts coursing through the extreme capsule complex) and the dorsal language pathway (arcuate fasciculus) are asymmetrically biased toward the left hemisphere. The second source would arise from left-lateralized neural systems for speech production, leading to greater functional connectivity between them and the left ATL. Even if the white matter structural connectivity was equal across left and right hemispheres, functional connectivity would still become left ATL-biased if the speech output system is predominantly rooted in left hemisphere structures. It is likely that these same neurocomputational principles can be extended to explain the left ATL bias for written word semantic tasks, as well. It is already established that there are graded left–right differences in posterior ventral occipito-temporal (vOT) for different categories of visual stimuli, including a left bias for written words. Given that the vOT provides the visual input to a transmodal ATL hub, it is likely that any asymmetries in the input will have a corollary effect on ATL activation, making it left-biased. The source of the left vOT bias for written word processing is debated. It seems to emerge during the process of learning to read (such that face recognition, for example, then becomes more right lateralized as a consequence: ). There is an intriguing possibility, therefore, that the left bias for written words also relates directly to speech production; in learning to read, children are initially taught to read aloud (i.e., “speak the words”). Thus, as per the biases in connectivity for (semantically driven) word retrieval, written word recognition and reading aloud might also become left biased.

48In contrast to word retrieval and written word processing, we found no support for a corresponding right ATL bias for either non-verbal information or socially related concepts; instead both were distinctly bilateral. Again, this might reflect the symmetry of the relevant connections from visual association cortex into the left and right ATL. Thus, for example, the inferior longitudinal fasiculus (the principal white matter bundle connectivity occipital and ATL regions) shows no hemispheric asymmetry. Accordingly, visual perceptual information may be better thought of as being graded along the posterior–anterior axis, with greater specialization in the posterior temporal cortex because of increased proximity to the primary visual cortex. For the formation of social and emotional concepts, inputs from frontal and limbic regions—primarily delivered through the uncinate fasciculus—will be important. The laterality of the uncinate fasciculus is currently unclear as, although one postmortem study found this tract to be larger in the right than left hemisphere, recent diffusion-weighted imaging studies have failed to replicate this asymmetry, convincingly.

49Finally with regard to left–right ATL differences, we should note that there does appear to be one disparity between the functional neuroimaging results (as analyzed here) and neuropsychology literature. As described above, the two literatures align very clearly with regard to the leftward bias for speech production/confrontational naming. Given that both L > R and R > L SD patients are reported to have altered social behavior and neuropsychiatric features when formally tested, the bilateral activations for social concepts, found in this meta-analysis, also seem to be consistent with the neuropsychological literature. In contrast, the lack of a rightward activation bias for face-based stimuli is potentially surprising given reports of face recognition deficits in patients with chronic or progressive damage to the right ATL. A recent study directly compared left and right ATL-resected temporal lobe epilepsy patients on various aspects of face processing. In keeping with the current meta-analysis and other findings (see above), the left ATL-resected patients were much more anomic than their right ATL counterparts. The reverse was also true, however, for face familiarity and identification; although the left ATL cases were mildly impaired when compared with controls, the right ATL group was significantly worse. Future targeted research is needed to explore this puzzle in more detail. As noted below, bilateral activation might be expected if different components of a task (e.g., face recognition vs. name retrieval) were supported separately by each hemisphere. An alternative hypothesis arises from considering some details in the results, which may prove critically important. Briefly, as described above, following the convergent inputs and outputs of the ventrolateral ATL region, the resulting representations will tend to be transmodal in nature. The face recognition deficits of the right ATL cases in the study, however, were not transmodal in form. Instead, the patients demonstrated a classical prosopagnosic impairment in which identification of familiar people was impaired when presented (visually) as a face but was much better when the same information was probed from a different modality (e.g., the spoken name). This suggests some kind of disruption between visual input and the semantic system rather than a deficit within the semantic system itself. Following the impact that differential neural/functional connectivity can have on performance (see above), it may be possible that the right ATL patients demonstrated poor visual face recognition because the visual input (from vOT regions) is right biased even if person knowledge itself is bilaterally distributed across the ATLs (the latter would be consistent with the fact that multiple modalities contribute to our knowledge of familiar people). If this hypothesis is correct, then the right ATL patients' prosopagnosia would follow from the same neurocomputational principle of differential connectivity which we and others have used to explain the left ATL bias for speech production. A caveat to this discussion is that did not test person recognition from any other modalities (e.g., voices). It is therefore not clear that the right ATL patients had a purely prosopagnosic impairment. Indeed, a review recent has shown that, in the cases where faces and voices are tested together, right ATL patients often exhibit a transmodal recognition deficit, rather than a purely visual deficit affecting face recognition.

50We also found graded specializations within the ATLs and again, these gradations were subtle and secondary to the overall picture of bilateral and transmodal activations. Specifically, activation extended dorsally toward STG when inputs were verbal and extended ventrally toward ITG when inputs were non-verbal (primarily pictures). This finding is consistent with the results of recent individual fMRI studies and a previous meta-analysis. In addition, activation extended further into the temporopolar cortex when stimuli had social connotations. Like inter-ATL differences, we propose that these intra-ATL gradations arise from differential connectivity with primary sensory and limbic regions (Fig. 6). The ATLs are richly connected to widely distributed structures, including primary sensory regions in posterior temporal and occipital lobes, medial structures (limbic system, olfactory cortex, and episodic systems), and frontal systems, implicated in social behavior, valence, executive function, and various aspects of language processing. Variations in the degree to which particular ATL subregions are connected to each of these areas could account for these patterns of graded specialization [see ); ); ); )]. Although not applied to the ATL region per se, computational model of semantic processing demonstrated the principle that, while an entire representational region may take part in semantic computations, subregions will become differentially important for divergent subsets of semantic activities depending on their distance/strength of connection to the critical input/output cortical areas.

51Figure 6 attempts to integrate some of the known long-range connectivity patterns to the ATL and to sketch the resultant graded representational hub centered on the ventrolateral ATL bilaterally [see ) for a more detailed consideration of this idea and the intra- and inter-ATL connectivity]. On this view, the dorsal ATL's partial specialization for auditory-verbal stimuli emanates from stronger connections to the auditory association cortex in posterior STG and speech production areas in the inferior frontal gyrus (via the middle longitudinal fasciculus; see Fig. 6, green). Conversely, greater specialization for visual stimuli in more ventral areas would arise from inferior longitudinal fasiculus connections between the occipital and occipito-temporal regions (Fig. 6, blue). Likewise, the extension of the ALE maps for social concepts into anteromedial temporopolar regions would fit with stronger connections (via the uncinate fasiculus; Fig. 6, yellow) to the more medial (limbic; Fig. 6, yellow/purple) temporal lobe regions as well as orbitofrontal and insular areas [for studies of ATL functional and structural connectivity in human and non-human primates, see ); ); )]. This latter proposal resonates closely with previous suggestions from fMRI that temporopolar regions are crucial in the representation of social and emotionally laden abstract concepts due to strong connectivity with limbic and orbitofrontal regions.

52In this study, we have focused on data arising from functional imaging studies in healthy individuals, rather than the effects of neurological damage. The principal advantages of functional neuroimaging are (1) they provide data about the organization of the unimpaired cognitive system, whereas brain damage can induce plasticity and compensatory processes that result in functional reorganization in patients and (2) they provide coverage of the whole brain (though see next paragraph for some limits to this), whereas lesion studies only provide data on the areas that happen to be damaged in a particular patient cohort. Conversely, the chief advantage of lesion studies is that they can demonstrate the necessity of a brain area for a particular function, whereas activations are correlational in nature. The tasks used in both types of study are also typically composed of multiple cognitive processes, which might draw on different neural resources. This point is particularly important when we consider the word retrieval tasks investigated as part of this study. Naming a picture is a complex process that involves considerable semantic processing in identifying the object as well as in retrieving its name. Therefore, even if there was a strongly specialized system (in which the left ATL was only involved in name retrieval and the right only in identifying the object), the multiple cognitive components that underpin this task could give rise to bilateral activation. Individual studies can be designed to tease apart the contributions of brain regions to these different processes (by, for example, devising a control task thought to require the same cognitive processes with the exception of the name retrieval) and/or might require MEG or electrophysiological techniques with better temporal resolution than fMRI. This level of experimental control is impossible, however, in a large-scale meta-analysis that includes data from a diverse set of studies that have employed very different paradigms and stimuli. Conversely, the great advantage of meta-analyses is that they allow hypotheses to be tested on a much larger dataset than could be collected in a single study, thus reducing the probability of false positives or false negatives.

53It is important therefore to be mindful of the advantages and limitations of each technique and to seek convergence in the conclusions offered by each. The present results offer a good deal of convergence. As discussed earlier, the pattern of overall bilateral recruitment of the ATLs across different stimulus types is consistent with the notion of a bilateral distributed system, which was developed based on observations of the consequences of unilateral versus bilateral ATL damage. Similarly, the left ATL activation bias for word retrieval tasks is consistent with findings in various patient groups, including the consequences of left-dominant ATL atrophy in SD and unilateral resection in patients with temporal lobe epilepsy. There do remain some areas of apparent disagreement between the results of this neuroimaging meta-analysis and the consequences of ATL damage. Most notably, many clinical studies have reported that abnormalities in social behavior are more pronounced in patients with greater right ATL damage. In contrast, our meta-analysis provided no evidence that healthy individual preferentially activates the right ATL when processing social concepts. This suggests that there may be other factors influencing the association of abnormal social behaviors with right ATL damage. For example, patients with right ATL damage might have a more severe level of impairment generally or they may tend to have greater damage to other brain areas, such as orbitofrontal cortex.

54The chief advantage of the meta-analysis approach is that it identifies results that are consistent across a large number of studies and is therefore resistant to weakness that are associated a particular study. It remains susceptible, however, to weaknesses that are systematic across many studies. The omission of the ventral ATL is one such systematic weakness. As described earlier, the ventral ATL has been poorly sampled in many previous neuroimaging studies due to various technical limitations, and this may disproportionately affect studies investigating visual processing [see ) and )]. For this reason, it has not appeared in this and other meta-analyses of semantic processing. In contrast, PET and fMRI studies that have avoided these technical problems have reliably observed considerable semantically related, multimodal activations in ventrolateral ATL regions (with peaks centered on the anterior fusiform/ITG;;;;; ). In addition, the same region has recently been shown to code semantic structure using multivoxel pattern analysis and is the peak region when correlating the degree of semantic impairment in SD patients against the distribution of FDG hypometabolism. For this reason, we have included the ventral and lateral ATL in the representational hub in Figure 6 (white).

55Supplementary material can be found at http://www.cercor.oxfordjournals.org/ online.

56G.E.R. was supported by a PhD studentship from EPSRC and a President's Doctoral Scholarship from the University of Manchester. The research was supported by an MRC Programme Grant (MR/J004146/1) to M.A.L.R., a Manchester Mental Health Social Care Trust Fellowship to P.H., and a Wellcome Trust Institutional Strategic Support Fund (ISSF) award (097820) to the University of Manchester. Funding to pay the Open Access publication charges for this article was provided by an RCUK block grant to the University of Manchester.