The entorhinal-DG/CA3 pathway in the medial temporal lobe retains visual working memory of a simple surface feature
Curation statements for this article:-
Curated by eLife
eLife assessment
This useful study highlights the contribution of the medial temporal lobe (MTL), and the DG/CA3 hippocampal pathway in particular, to neural activity during the working memory delay period. The evidence supporting this is compelling, using diverse state-of-the-art approaches to neural data analysis and relating it to behavioural data. The work will be of significant interest to neuroscientists specialising in the research area of human working memory.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
Classic models consider working memory (WM) and long-term memory as distinct mental faculties that are supported by different neural mechanisms. Yet, there are significant parallels in the computation that both types of memory require. For instance, the representation of precise item-specific memory requires the separation of overlapping neural representations of similar information. This computation has been referred to as pattern separation, which can be mediated by the entorhinal-DG/CA3 pathway of the medial temporal lobe (MTL) in service of long-term episodic memory. However, although recent evidence has suggested that the MTL is involved in WM, the extent to which the entorhinal-DG/CA3 pathway supports precise item-specific WM has remained elusive. Here, we combine an established orientation WM task with high-resolution fMRI to test the hypothesis that the entorhinal-DG/CA3 pathway retains visual WM of a simple surface feature. Participants were retrospectively cued to retain one of the two studied orientation gratings during a brief delay period and then tried to reproduce the cued orientation as precisely as possible. By modeling the delay-period activity to reconstruct the retained WM content, we found that the anterior-lateral entorhinal cortex (aLEC) and the hippocampal DG/CA3 subfield both contain item-specific WM information that is associated with subsequent recall fidelity. Together, these results highlight the contribution of MTL circuitry to item-specific WM representation.
Article activity feed
-
-
Author Response
Reviewer #1 (Public Review):
The study by Xie et al., investigates whether the entorhinal-DG/CA3 pathway is involved in working memory maintenance. The main findings include a correlation between stimulus and neural similarities that was specific for cued stimulus and entorhinal-DG/CA3 locations. The authors observed similar results (cuing and region specificity) using inverted encoding modeling approach. Finally, they also showed that trials in which participants made a smaller error showed a better reconstruction fidelity on the cued side (compared to un-cued). This effect was absent for larger-error trials.
The study challenges a widely held traditional view that working memory and episodic memory have largely independent neural implementations with the MTL being critical for episodic memory but not for working …
Author Response
Reviewer #1 (Public Review):
The study by Xie et al., investigates whether the entorhinal-DG/CA3 pathway is involved in working memory maintenance. The main findings include a correlation between stimulus and neural similarities that was specific for cued stimulus and entorhinal-DG/CA3 locations. The authors observed similar results (cuing and region specificity) using inverted encoding modeling approach. Finally, they also showed that trials in which participants made a smaller error showed a better reconstruction fidelity on the cued side (compared to un-cued). This effect was absent for larger-error trials.
The study challenges a widely held traditional view that working memory and episodic memory have largely independent neural implementations with the MTL being critical for episodic memory but not for working memory. The study adds to a large body of evidence showing involvement of the hippocampus across a range of different working memory tasks and stimuli. Nevertheless, it still remains unclear what functions may hippocampus play in working memory.
We thank the reviewer’s positive appraisal of the current research, which adds to the growing research interest in the MTL’s contribution to WM.
Reviewer #2 (Public Review):
Xie et al. investigated the medial temporal lobe (MTL) circuitry contributions to pattern separation, a neurocomputational operation to distinguish neutral representations of similar information. This presumably engages both long-term memory (LTM) and working memory (WM), bridging the gap between the working memory (WM) and long-term memory (LTM) distinction. Specifically, the authors combined an established retro-cue orientation WM task with high-resolution fMRI to test the hypothesis that the entorhinal-DG/CA3 pathway retains visual WM for a simple surface feature. They found that the anterior-lateral entorhinal cortex (aLEC) and the hippocampal DG/CA3 subfield both retained item-specific WM information that is associated with fidelity of subsequent recall. These findings highlight the contribution of MTL circuitry to item-specific WM representation, against the classic memory models.
I am a long-term memory researcher with expertise in representational similarity analysis, but not in inverted encoding modeling (IEM). Therefore, I cannot verify the correctness of these models and I will leave it to the other reviewers and editors. However, after an in-depth reading of the manuscript, I could evaluate the significance of the present findings and the strength of evidence supporting these findings. The conclusions of this paper are mostly well supported by data, but some aspects of image acquisition and data analysis need to be clarified.
We thank the reviewer for positive appraisal of the current study.
I would like to list several strengths and weaknesses of this manuscript:
Strengths:
• Methodologically, the authors addressed uncertainty in previous research resulting from several challenges. Namely, they used a high-resolution fMRI protocol to infer signals from the MTL substructures and an established retro-cue orientation WM task to minimize the task load.
• The authors selected a control ROI - amygdala - irrelevant for the experimental task, and at the same time adjacent to the other MTL ROIs, thus possibly having a similar signal-to-noise ratio. The reported effects were observed in the aLEC and DG/CA3, but not in the amygdala.
• Memory performance, quantified as recall errors, was at ceiling - an average recall error of 12 degrees was only marginally away from the correct grating towards the closest incorrect grating (predefined with min. 20 degrees increments). However, the authors controlled for the effects of recall fidelity on MTL representations by comparing the IEM reconstructions between precise recall trials and imprecise recall trails (resampled to an equal number of trials). The authors found that precise recall trails have yielded better IEM reconstruction quality.
• The author performed a control analysis of time-varying IEM to exclude a possibility that the mid-delay period activity in the aLEC-DG/CA3 contains item-specific information that could be attributed to perceptual processing. This analysis showed that the earlier TR in the delay period contains information for both cued and uncued items, whereas the mid-delay period activity contains the most information related to the cued, compared to uncued, item.
We thank the reviewer for highlighting the multiple strengths of the current study.
Weaknesses:
• The authors formulate their main hypothesis building on an assumption related to the experimental task. This task requires correctly selecting the cued grating orientation while resisting the interference from internal representations of the other orientation gratings. The authors hypothesize that if this post-encoding information selection function is supported by the MTL-s entorhinal-DG/CA3 pathway, the recorded delay-period activity should contain more information about the cued item that the uncued item (even if both are similarly remembered). Thus, the assumption here is that resolving the interference would be reflected by a more distinct representation in MTL for the cued item. Could it be the opposite, namely the MTL could better represent the unresolved interference, for example by the mechanism of hippocampal repulsion (Chanales et al., 2017). It could strengthen the findings if the authors comment on the contrary hypothesis as well.
We thank the reviewer for pointing out this interesting alternative hypothesis. Because of the different task design (e.g., over the course of learning vs. WM) and stimuli (e.g., spatial memory vs. orientation grating), it is hard to directly compare Chanales et al.’s findings with the current results. That said, we think the idea that the representation of similar information would lead to greater task demand on the MTL is consistent with our intuition regarding the role of the MTL in supporting the qualitative aspect of WM representation. We have now further discussed this issue in our revised manuscript to invite further consideration of the suggested alternative hypothesis,
“Our data suggest that this process would result in more similar and stable representations for the same remembered item across trials, as detected by multivariate correlational and decoding analyses in the current study. However, under certain task conditions (e.g., learning spatial routes in a naturalistic task over many repetitions), the MTL may maximally orthogonalize overlapping information to opposite representational patterns (hence “repulsion”) to minimize mnemonic interference (Chanales et al., 2017). It remains to be determined how these learning-related mechanisms in a more complex setting are related to MTL’s contributions to WM of simple stimulus features.”
• It is not clear for me why the authors chose the inverted encoding modelling approach and what is its advantage over the others multivoxel pattern analysis approaches, for example representational similarity analysis also used in this study. How are these two complementary? Since the IEM is still a relatively new approach, maybe a little comment in the manuscript could help emphasizing the strength of the paper? Especially that this paper is of interest to researchers in the fields of both working memory and long-term memory, the latter being possibly not familiar with the IEM.
We thank the reviewer for this suggestion. In principle, the IEM is a multivariate pattern classification analysis based on an encoding model. There is no fundamental difference between this approach and other machine-learning or classification approaches, except that the IEM is a more model-based approach and therefore can be more computationally efficient (see Xie et al., 2023 for a conceptual overview for multivariate analysis of high-dimensional neural data). The relationship between IEM and representational similarity is grounded in item-specific information that could lead to shared neural variance. How these two analyses are complimented each other is well characterized by a recent theoretical review (Kriegeskorte & Wei, 2021). The rationale is that trial-wise RSA reveals shared neural variance between items, implying the presence of item-specific information in the recorded neural data. And the IEM approach or other classification algorithms can more directly test this item-specific information under a prediction-based framework (e.g., train the data and test on a hold-out set). As a result, the findings of these two methods are correlated at the subject-level (Figure S4), which is important to note for the purpose of analytical reliability. Furthermore, using the IEM also allows us to compare our current findings with that from the previous research (Figure S3), addressing some replicability questions in the field (e.g., Ester et al., 2015).
We have clarified more on this issue in the paragraph when we first introduce IEM,
“To directly reveal the item-specific WM content, we next modeled the multivoxel patterns in subject-specific ROIs using an established inverted encoding modeling (IEM) method (Ester et al., 2015). This method assumes that the multivoxel pattern in each ROI can be considered as a weighted summation of a set of orientation information channels (Figure 3A). By using partial data to train the weights of the orientation information channels and applying these weights to an independent hold-out test set, we reconstruct the assumed orientation information channels to infer item-specific information for the remembered item – operationalized the resultant vector length of the reconstructed orientation information channel normalized at 0° reconstruction error (Figure S2). As this approach verifies the assumed information content based on observed neural data, its results can be efficiently computed and interpreted within the assumed model even when the underlying neuronal tuning properties are unknown (Ester et al., 2015; Sprague et al., 2018). This approach, therefore, complements the model-free similarity-based analysis by linking representational geometry embedded in the neural data with item-specific information under a prediction-based framework (Kriegeskorte and Wei, 2021; Xie et al., 2023). Based on this method, previous research has revealed item-specific WM information in distributed neocortical areas, including the parietal, frontal, and occipital-temporal areas (Bettencourt and Xu, 2015; Ester et al., 2015; Rademaker et al., 2019; Sprague et al., 2016), which are similar to those revealed by other multivariate classification methods (e.g., support vector machine, SVM, Ester et al., 2015). We have also replicated these IEM effects in the current dataset (Figure S3).”
Overall, this work can have a substantial impact of the field due to its theoretical and conceptual novelty. Namely, the authors leveraged an established retro-cue task to demonstrate that a neurocomputational operation of pattern separation engages both working-memory and long-term memory, both mediated by the MTL circuitry, beyond the distinction in classic memory models. Moreover, on the methodological side, using the multivariate pattern analyses (especially the IEM) to study neural computations engaged in WM and LTM seems to be a novel and promising direction for the field.
Thanks for the reviewer for this positive appraisal of the current study.
Reviewer #3 (Public Review):
This work addresses a long-standing gap in the literature, showing that the medial temporal lobe (MTL) is involved in representing simple feature information during a low-load working memory (WM) delay period. Previously, this area was suggested to be relevant for episodic long-term memory, and only implicated in working memory under conditions of high memory load or conjunction features. Using well-rounded analyses of task-dependent fMRI data in connection with a straightforward behavioural experiment, this paper suggests a more general role of the medial temporal lobe in working memory delay activity. It also provides a replication of previous findings on item-specific information during working memory delay in neocortical areas.
We thank the reviewer for highlighting the contribution of the current study to fill a gap in the literature.
Strengths:
The study has strengths in its methods and analyses. Firstly, choosing a well-established cueing paradigm allows for straightforward comparison with past and future studies using similar paradigms. The authors themselves show this by replicating previous findings on delay-period activity in parietal, frontal, and occipito-temporal areas, strengthening their own and previous findings. Secondly, they use a template with relatively fine-grained MTL-subregions and choose the amygdala as a control area within the MTL. This increases confidence in the finding that the hippocampus in particular is involved in WM delay-period activity. Thirdly, their combined use stimulus-based representational similarity analysis as well as Inverted Encoding Modeling and the convergence on the same result is encouraging. Finally, despite focusing on the delay period in their main findings, extensive supplementary materials give insight into the time-course of processing (encoding) which will be helpful for future studies.
We thank the reviewer for highlighting multiple strengths of this current study.
Weaknesses:
While the evidence generally supports the conclusions, there are some weaknesses in behavioural data analysis. The authors demonstrated fine stimulus discrimination in the neural data using Inverted Encoding Modeling (IEM), however the same standard is not applied in the behavioural data analysis. In this analysis, trials below 20 degrees and trials above 20 degrees of memory error are collapsed to compare IEM decoding error between them. As a result, the "small recall error" group encompasses a total range of 40 degrees and includes neighbouring stimuli. While this is enough to demonstrate that there was information about the remembered stimulus, it does not clarify whether aLEC/CA3 activity is associated with target selection only or also with reproduction fidelity. It leaves open whether fine-grained neural information in MTL is related to memory fidelity.
We thank the reviewer for this cautious note. As the current task is optimized to reveal the neural representation during visual WM and as our participants are cognitively normal college students, participants’ behavioral performance in the current experiment tends to be very good (Figure 1). This leaves us relatively small variation to further probe the behavioral outcomes of the task. We have recently generalized our findings using intracranial EEG and confirmed that trial-by-trial mnemonic discrimination during a short delay is indeed associated with the fidelity of item-specific WM representation (Xie, Chapeton, et al., in press).
We have further discussed this issue in the revised Discussion,
“… These two approaches are therefore complementary to each other. Nevertheless, these analyses are correlational in nature. Hence, although fine-grained neural representations revealed by these analyses are associated with participants’ behavioral outcomes (Figure 4), it remains to be determined whether the entorhinal-DG/CA3 pathway contributes to the fidelity of the selected WM representation or also to the selection of task-relevant information. Strategies for resolving this issue can involve generalizing the current findings to other WM tasks without an explicit requirement of information selection (e.g., intracranial stimulation of the MTL in a regular WM task without a retro-cue manipulation, Xie et al., in press) and/or further exploring how the frontal-parietal mechanisms related to visual selection and attention interact with the MTL system (Panichello and Buschman, 2021).”
Moreover, the authors could be more precise about the limitations of the study and their conclusions. In particular, the paper at times suggests that the results contribute to elucidating common roles of the MTL in long-term memory and WM, potentially implementing a process called pattern separation. However, while the paper convincingly shows MTL-involvement in WM, there is no comparison to an episodic memory condition. It therefore remains an open question whether it fulfils the same role in both scenarios. Moreover, the paradigm might not place adequate pattern separation demands on the system since information about the un-cued item may be discarded after the cue.
We thank the reviewer for this cautious note. We have now included a more detailed discussion on this issue.
In the Discussion,
“To more precisely reveal the MTL mechanisms that are shared across WM and long-term memory, future research should examine the extent to which MTL voxels evoked by a long-term memory task (e.g., mnemonic similarity task, Bakker et al., 2008) can be directly used to directly decode mnemonic content in visual WM tasks using different simple stimulus features.”
-
eLife assessment
This useful study highlights the contribution of the medial temporal lobe (MTL), and the DG/CA3 hippocampal pathway in particular, to neural activity during the working memory delay period. The evidence supporting this is compelling, using diverse state-of-the-art approaches to neural data analysis and relating it to behavioural data. The work will be of significant interest to neuroscientists specialising in the research area of human working memory.
-
Reviewer #1 (Public Review):
The study by Xie et al., investigates whether the entorhinal-DG/CA3 pathway is involved in working memory maintenance. The main findings include a correlation between stimulus and neural similarities that was specific for cued stimulus and entorhinal-DG/CA3 locations. The authors observed similar results (cuing and region specificity) using inverted encoding modeling approach. Finally, they also showed that trials in which participants made a smaller error showed a better reconstruction fidelity on the cued side (compared to un-cued). This effect was absent for larger-error trials.
The study challenges a widely held traditional view that working memory and episodic memory have largely independent neural implementations with the MTL being critical for episodic memory but not for working memory. The study adds …
Reviewer #1 (Public Review):
The study by Xie et al., investigates whether the entorhinal-DG/CA3 pathway is involved in working memory maintenance. The main findings include a correlation between stimulus and neural similarities that was specific for cued stimulus and entorhinal-DG/CA3 locations. The authors observed similar results (cuing and region specificity) using inverted encoding modeling approach. Finally, they also showed that trials in which participants made a smaller error showed a better reconstruction fidelity on the cued side (compared to un-cued). This effect was absent for larger-error trials.
The study challenges a widely held traditional view that working memory and episodic memory have largely independent neural implementations with the MTL being critical for episodic memory but not for working memory. The study adds to a large body of evidence showing involvement of the hippocampus across a range of different working memory tasks and stimuli. Nevertheless, it still remains unclear what functions may hippocampus play in working memory.
-
Reviewer #2 (Public Review):
Xie et al. investigated the medial temporal lobe (MTL) circuitry contributions to pattern separation, a neurocomputational operation to distinguish neutral representations of similar information. This presumably engages both long-term memory (LTM) and working memory (WM), bridging the gap between the working memory (WM) and long-term memory (LTM) distinction. Specifically, the authors combined an established retro-cue orientation WM task with high-resolution fMRI to test the hypothesis that the entorhinal-DG/CA3 pathway retains visual WM for a simple surface feature. They found that the anterior-lateral entorhinal cortex (aLEC) and the hippocampal DG/CA3 subfield both retained item-specific WM information that is associated with fidelity of subsequent recall. These findings highlight the contribution of MTL …
Reviewer #2 (Public Review):
Xie et al. investigated the medial temporal lobe (MTL) circuitry contributions to pattern separation, a neurocomputational operation to distinguish neutral representations of similar information. This presumably engages both long-term memory (LTM) and working memory (WM), bridging the gap between the working memory (WM) and long-term memory (LTM) distinction. Specifically, the authors combined an established retro-cue orientation WM task with high-resolution fMRI to test the hypothesis that the entorhinal-DG/CA3 pathway retains visual WM for a simple surface feature. They found that the anterior-lateral entorhinal cortex (aLEC) and the hippocampal DG/CA3 subfield both retained item-specific WM information that is associated with fidelity of subsequent recall. These findings highlight the contribution of MTL circuitry to item-specific WM representation, against the classic memory models.
I am a long-term memory researcher with expertise in representational similarity analysis, but not in inverted encoding modeling (IEM). Therefore, I cannot verify the correctness of these models and I will leave it to the other reviewers and editors. However, after an in-depth reading of the manuscript, I could evaluate the significance of the present findings and the strength of evidence supporting these findings. The conclusions of this paper are mostly well supported by data, but some aspects of image acquisition and data analysis need to be clarified. I would like to list several strengths and weaknesses of this manuscript:
Strengths:
• Methodologically, the authors addressed uncertainty in previous research resulting from several challenges. Namely, they used a high-resolution fMRI protocol to infer signals from the MTL substructures and an established retro-cue orientation WM task to minimize the task load.
• The authors selected a control ROI - amygdala - irrelevant for the experimental task, and at the same time adjacent to the other MTL ROIs, thus possibly having a similar signal-to-noise ratio. The reported effects were observed in the aLEC and DG/CA3, but not in the amygdala.
• Memory performance, quantified as recall errors, was at ceiling - an average recall error of 12 degrees was only marginally away from the correct grating towards the closest incorrect grating (predefined with min. 20 degrees increments). However, the authors controlled for the effects of recall fidelity on MTL representations by comparing the IEM reconstructions between precise recall trials and imprecise recall trails (resampled to an equal number of trials). The authors found that precise recall trails have yielded better IEM reconstruction quality.
• The author performed a control analysis of time-varying IEM to exclude a possibility that the mid-delay period activity in the aLEC-DG/CA3 contains item-specific information that could be attributed to perceptual processing. This analysis showed that the earlier TR in the delay period contains information for both cued and uncued items, whereas the mid-delay period activity contains the most information related to the cued, compared to uncued, item.Weaknesses:
• The authors formulate their main hypothesis building on an assumption related to the experimental task. This task requires correctly selecting the cued grating orientation while resisting the interference from internal representations of the other orientation gratings. The authors hypothesize that if this post-encoding information selection function is supported by the MTL-s entorhinal-DG/CA3 pathway, the recorded delay-period activity should contain more information about the cued item that the uncued item (even if both are similarly remembered). Thus, the assumption here is that resolving the interference would be reflected by a more distinct representation in MTL for the cued item. Could it be the opposite, namely the MTL could better represent the unresolved interference, for example by the mechanism of hippocampal repulsion (Chanales et al., 2017). It could strengthen the findings if the authors comment on the contrary hypothesis as well.
• It is not clear for me why the authors chose the inverted encoding modelling approach and what is its advantage over the others multivoxel pattern analysis approaches, for example representational similarity analysis also used in this study. How are these two complementary? Since the IEM is still a relatively new approach, maybe a little comment in the manuscript could help emphasizing the strength of the paper? Especially that this paper is of interest to researchers in the fields of both working memory and long-term memory, the latter being possibly not familiar with the IEM.Overall, this work can have a substantial impact of the field due to its theoretical and conceptual novelty. Namely, the authors leveraged an established retro-cue task to demonstrate that a neurocomputational operation of pattern separation engages both working-memory and long-term memory, both mediated by the MTL circuitry, beyond the distinction in classic memory models. Moreover, on the methodological side, using the multivariate pattern analyses (especially the IEM) to study neural computations engaged in WM and LTM seems to be a novel and promising direction for the field.
-
Reviewer #3 (Public Review):
This work addresses a long-standing gap in the literature, showing that the medial temporal lobe (MTL) is involved in representing simple feature information during a low-load working memory (WM) delay period. Previously, this area was suggested to be relevant for episodic long-term memory, and only implicated in working memory under conditions of high memory load or conjunction features. Using well-rounded analyses of task-dependent fMRI data in connection with a straightforward behavioural experiment, this paper suggests a more general role of the medial temporal lobe in working memory delay activity. It also provides a replication of previous findings on item-specific information during working memory delay in neocortical areas.
Strengths:
The study has strengths in its methods and analyses. Firstly, …Reviewer #3 (Public Review):
This work addresses a long-standing gap in the literature, showing that the medial temporal lobe (MTL) is involved in representing simple feature information during a low-load working memory (WM) delay period. Previously, this area was suggested to be relevant for episodic long-term memory, and only implicated in working memory under conditions of high memory load or conjunction features. Using well-rounded analyses of task-dependent fMRI data in connection with a straightforward behavioural experiment, this paper suggests a more general role of the medial temporal lobe in working memory delay activity. It also provides a replication of previous findings on item-specific information during working memory delay in neocortical areas.
Strengths:
The study has strengths in its methods and analyses. Firstly, choosing a well-established cueing paradigm allows for straightforward comparison with past and future studies using similar paradigms. The authors themselves show this by replicating previous findings on delay-period activity in parietal, frontal, and occipito-temporal areas, strengthening their own and previous findings. Secondly, they use a template with relatively fine-grained MTL-subregions and choose the amygdala as a control area within the MTL. This increases confidence in the finding that the hippocampus in particular is involved in WM delay-period activity. Thirdly, their combined use stimulus-based representational similarity analysis as well as Inverted Encoding Modeling and the convergence on the same result is encouraging. Finally, despite focusing on the delay period in their main findings, extensive supplementary materials give insight into the time-course of processing (encoding) which will be helpful for future studies.Weaknesses:
While the evidence generally supports the conclusions, there are some weaknesses in behavioural data analysis. The authors demonstrated fine stimulus discrimination in the neural data using Inverted Encoding Modeling (IEM), however the same standard is not applied in the behavioural data analysis. In this analysis, trials below 20 degrees and trials above 20 degrees of memory error are collapsed to compare IEM decoding error between them. As a result, the "small recall error" group encompasses a total range of 40 degrees and includes neighbouring stimuli. While this is enough to demonstrate that there was information about the remembered stimulus, it does not clarify whether aLEC/CA3 activity is associated with target selection only or also with reproduction fidelity. It leaves open whether fine-grained neural information in MTL is related to memory fidelity.Moreover, the authors could be more precise about the limitations of the study and their conclusions. In particular, the paper at times suggests that the results contribute to elucidating common roles of the MTL in long-term memory and WM, potentially implementing a process called pattern separation. However, while the paper convincingly shows MTL-involvement in WM, there is no comparison to an episodic memory condition. It therefore remains an open question whether it fulfils the same role in both scenarios. Moreover, the paradigm might not place adequate pattern separation demands on the system since information about the un-cued item may be discarded after the cue.
-
-