Atypical neural topographies underpin dysfunctional pattern separation in temporal lobe epilepsy

Abstract

Episodic memory is the ability to remember events from our past accurately. The process of pattern separation is hypothesized to underpin this ability and is defined as the capacity to orthogonalize memory traces, to maximize the features that make them unique. Contemporary cognitive neuroscience suggests that pattern separation entails complex interactions between the hippocampus and neocortex, where specific hippocampal subregions shape neural reinstatement in the neocortex. To test this hypothesis, the current work studied both healthy controls and patients with temporal lobe epilepsy who presented with hippocampal structural anomalies. We measured neural activity in all participants using functional MRI while they retrieved memorized items or lure items, which shared features with the target. Behaviourally, patients with temporal lobe epilepsy were less able to exclude lures than controls and showed a reduction in pattern separation. To assess the hypothesized relationship between neural patterns in the hippocampus and neocortex, we identified the topographic gradients of intrinsic connectivity along neocortical and hippocampal subfield surfaces and determined the topographic profile of the neural activity accompanying pattern separation. In healthy controls, pattern separation followed a graded topography of neural activity, both along the hippocampal long axis (and peaked in anterior segments that are more heavily engaged in transmodal processing) and along the neocortical hierarchy running from unimodal to transmodal regions (peaking in transmodal default mode regions). In patients with temporal lobe epilepsy, however, this concordance between task-based functional activations and topographic gradients was markedly reduced. Furthermore, person-specific measures of concordance between task-related activity and connectivity gradients in patients and controls were related to inter-individual differences in behavioural measures of pattern separation and episodic memory, highlighting the functional relevance of the observed topographic motifs. Our work is consistent with an emerging understanding that successful discrimination between memories with similar features entails a shift in the locus of neural activity away from sensory systems, a pattern that is mirrored along the hippocampal long axis and with respect to neocortical hierarchies. More broadly, our study establishes topographic profiling using intrinsic connectivity gradients, capturing the functional underpinnings of episodic memory processes in a manner that is sensitive to their reorganization in pathology.

Li et al. study pattern separation in healthy individuals and in patients with temporal lobe epilepsy. In controls, neural activation follows topographic gradients and increases towards anterior hippocampal and transmodal cortices. This pattern is perturbed in patients, and correlates with their behavioural deficits.

Introduction

Episodic memories are records of experiences and events in our daily lives. Recent literature on the neural substrates of episodic memory posits a mechanism that involves recurrent connections between the hippocampus, which instantiates ‘coarse’ memory codes, and the neocortex, which provides specific features of the episode.¹ Memories can share features and, accordingly, it is necessary to be able to distinguish between memories with similar features. Pattern separation is believed to be a key contributor to this process, and it is understood to be the ability to orthogonalize similar external information into non-overlapping internal representations.² Studying pattern separation, as well as the principles governing its localization in the hippocampus and in large-scale neocortical networks, provides the opportunity to advance our understanding of the neural mechanisms contributing to human episodic memory and identify patterns of reorganization in disease. Here, we set out to identify such neural principles by studying both healthy individuals and patients with temporal lobe epilepsy (TLE), the most common drug-resistant adult epilepsy that is associated with pathological changes in the hippocampus.^3,4 Patients with TLE also present with deficits in episodic memory processing^5,6 and represent a candidate human disease model for studying the functional reorganization of the neural substrates, which accounts for episodic memory function in the presence of hippocampal pathology.^7,8

In healthy individuals, evidence for a hippocampal role in pattern separation comes from task-based functional MRI studies, which commonly probe the ability to discriminate previously visualized items and items that are similar, but not identical.^2,9,10 These findings are complemented by behavioural studies in patients with hippocampal lesions, including patients with TLE,¹¹ and functional MRI studies in non-epileptic patient populations.^12-15 In studying pattern separation, task-based functional MRI studies^16,17 have pointed to an implication of specific hippocampal subfields, notably the dentate gyrus and cornu ammonis 3 (CA3).^2,9 While not explicitly studied in the context of pattern separation, emerging literature indicates that hippocampal function may vary along its long axis,^18–21 with anterior portions often showing a contribution to self-generated and higher-order transmodal processes, while posterior segments participate in more detailed processes related to sensory information.^22–24

Contemporary views on the role of hippocampus suggest that pattern separation is achieved through the influence of hippocampal subregions on neural activity within the broader neocortex. Consistent with this view, prior investigations of the hippocampus emphasized its role as a hub that receives information from sensory and transmodal cortices and instantiates the formation of sparse codes for individual experiences that are then projected back.²⁵ Moreover, studies indeed implicate sensory and transmodal cortices in addition to the hippocampus in pattern separation.^17,26 These findings suggest that pattern separation involves subregional and network mechanisms, motivating research interrogating both levels of neural organization. Here, we capitalize on unsupervised machine learning techniques that can identify principal axes of subregional organization from resting-state functional MRI connectivity information and relate these axes to pattern separation activity in healthy individuals and patients with TLE. The applied framework visualizes smooth transitions (henceforth termed as gradients) in hippocampal and neocortical subregions in a fully data-driven framework. In the hippocampus, previous applications of these techniques identified a primary gradient that depicts the long axis, and a second gradient that follows medio-lateral infolding and subfield-to-subfield variations in connectivity.^22,24 In the neocortex, prior work in healthy adults has shown a principal functional gradient that differentiates sensory and transmodal systems, while a second gradient dissociates sensory-motor and visual cortices.²⁷ In other domains, the application of a gradient framework has brought substantial innovation to explaining the global patterns of functional MRI activity with respect to the balance between sensory and transmodal systems^28–30 and in studying the preferential susceptibility of transmodal areas in ageing^31,32 and neurodevelopmental conditions.^33,34 In episodic memory, this framework has not been applied yet. Here, the stratification of functional activations along hippocampal and neocortical gradients may provide a continuous and compact analytical space to interrogate subregional specialization and system-level integration in a unified manner, and to explore the specific contribution of the balance between sensory and transmodal processing to efficient pattern separation in both mesiotemporal and neocortical regions.

To consolidate subregional and network perspectives of pattern separation, the current work examined how the spatial localization of task-related functional MRI activity is captured by hippocampal and neocortical topographic gradients. Our work was carried out in both healthy individuals and a well-defined cohort of patients with mesial TLE presenting with hippocampal lesions and impairment of episodic memory performance. Our primary objective was to demonstrate that topographic profiling can provide compact signatures of functional activation and functional reorganization in healthy and epileptic individuals, which can be successfully implemented to model inter-individual differences in pattern separation performance in both healthy and epileptic cohorts.

Materials and methods

Participants

We studied 26 healthy adults [15 males; 21–44 years; mean ± standard deviation (SD) age = 31.1 ± 7.31 years, 1/25 left/right-handed] recruited by advertisement, as well as 14 patients with drug-resistant TLE (six males; 19–56 years; mean ± SD years = 35.9 ± 14.66, 2/12 left/right-handed). Demographic and clinical data were obtained through interviews with patients and their relatives. Epilepsy diagnosis and lateralization of the seizure focus were determined through a comprehensive evaluation by the clinical team based on detailed history, neurological examination, review of medical records, prolonged video-EEG recordings, clinical MRI reading at 1.5 T, neuropsychology assessment in all patients and FDG-PET in a patient subgroup. There were no patients with mass brain lesions (malformations of cortical development, tumour or vascular malformations) or a history of traumatic brain injury or encephalitis. All patients had a left-sided seizure focus and a clinical and imaging characterization that was compatible with a mesial TLE subtype.

Although patients and controls did not significantly differ by age or sex, the former showed reduced hippocampal volumes in the left hemisphere compared with the controls, as well as increased inter-hemispheric volumetric asymmetry [Student’s t-test: t(38) = −3.91, two-tailed P < 0.001].

Our protocol received approval from the Research Ethics Board of the Montreal Neurological Institute and Hospital. All participants provided their informed consent in writing. See Supplementary Table1 for socio-demographic and clinical information on the patient and control cohorts included in this study.

Mnemonic similarity task

The mnemonic similarity task was adapted from a previous study³⁵ with stimuli taken from the original experiment (https://faculty.sites.uci.edu/starklab/mnemonic-similarity-task-mst/, accessed 24 July 2021) and administered inside the scanner. It consisted of two phases (Fig. 1A). In phase 1, participants were shown 64 colour photographs of everyday objects on a white background and gave an indoor-versus-outdoor judgement for each picture (lasting 2 s) via a button press. In phase 2, which occurred ∼7 min after the encoding task, participants engaged in a recognition memory test, during which they identified each item as ‘old’, ‘similar’ or ‘new’ via a button press. Trials in phase 2 consisted of 32 novel items that were not seen before (novel), 32 similar items that were similar to those seen in phase 1 (lure), and 32 old items that were exactly the same in phase 1 (repetition). There was a fixation period lasting 2–3 s between two consecutive items, resulting in a jittered interstimulus interval. The behavioural pattern separation (BPS) score was computed as the pattern separation rate P(similar|lure) corrected for similar bias rate P(similar|novel), thus resulting in BPS = P(similar|lure) − P(similar|novel).

Figure 1.

Mnemonic similarity task design and behavioural group comparison. [A(i)]. The mnemonic similarity task (MST) design and (ii) BPS responses in healthy controls. [B(i)]. Response proportions and (ii) BPS scores in both patients and controls and group comparison controlled for age and sex, showing reduced pattern separation in patients. HC = healthy controls; LTLE = left temporal lobe epilepsy.

Episodic task

All participants underwent a picture-based list paired associates learning task. The task involves two phases. In the encoding phase, participants had to memorize pairs of items shown simultaneously. Following a 10-min delay, the retrieval phase was administered. At each trial, participants had to identify the image that was originally paired with a given probe in a three-alternative forced choice paradigm. Items were pictures of everyday objects or animals, presented on a white background. We controlled the semantic relatedness of the items by computing the lexical similarity of words corresponding to the pictures³⁶ and similarity was chosen to be below 0.3 (range: 0–1). In the difficult retrieval condition, item pairs were only shown once during encoding (shallow encoding). In the easy retrieval condition, item pairs were displayed twice to ensure deeper encoding. There were 56 pseudo-randomized trials in total with 28 corresponding to pairs of images encoded only once (i.e. Epi-D) and 28 to pairs of images encoded twice (i.e. Epi-E).

Montreal Cognitive Assessment

We administered the MoCA (Montreal Cognitive Assessment, version 7.1, original version) outside the scanner.³⁷

MRI acquisition

Magnetic resonance images were acquired using accelerated sequences on a Siemens Magnetom 3 T PrismaFit scanner with a 64-channel head coil. Two T₁-weighted images with identical parameters were acquired using a 3D MPRAGE sequence (repetition time = 2300 ms, echo time = 3.14 ms, inversion time = 900 ms, flip angle = 9°, field of view = 256 × 256; matrix size = 320 × 320 leading to a resolution of 0.8 × 0.8 × 0.8 mm³; ipat = 2; acquisition time = 6 min and 44 s). Resting-state functional MRI data were acquired using a 2D echo planar imaging (EPI) sequence (repetition time = 600 ms, echo time = 30 ms, flip angle = 50°, field of view = 240 × 240, matrix size = 80 × 80, leading to a resolution of 3 × 3 × 3 mm³; multiband acceleration factor = 6; acquisition time = 7 min and 6 s). Participants were instructed to keep their eyes open, fixate on a cross centrally presented on a white screen and not fall asleep. Task functional MRI used a 2D EPI sequence with parameters identical to the resting-state functional MRI (acquisition time = 5 min 37 s for phase 1, 8 min and ∼30 s for phase 2). Participants were scanned at both mnemonic similarity task phases 1 and 2.

MRI preprocessing

Structural MRI processing

Images underwent intensity non-uniformity correction and the two repeated scans were co-registered and averaged. Following brain extraction and segmentation of subcortical structures, preprocessed images were non-linearly registered to MNI152 space and cortical surfaces were extracted using recon-all in FreeSurfer 6.0.^38–40 To improve inter-individual correspondence, cortical surfaces were aligned to the hemisphere-symmetric Conte69 template.⁴¹ The hippocampus was automatically segmented into CA1–3, CA4-dentate gyrus and subiculum using our previously published and validated surface patch-based algorithm.^4,42 Hippocampal subfield surfaces were parameterized using a spherical harmonics framework (SPHARM-PDM). A Hamilton-Jacobi approach was used to create a medial sheet, representing the core of the subfield with a minimal partial volume effect. Spherical harmonics parameters on the outer hull were propagated to their corresponding medial sheet locations along a Laplacian field, thereby improving across subject correspondence.

Task-based functional MRI

We used SPM12⁴³ (version 6470, https://www.fil.ion.ucl.ac.uk/spm/software/spm12/, accessed 24 July 2021) for MATLAB for all task-based functional MRI processing. We first generated field maps from the AP-PA blip pairs and applied geometric distortion correction using the FieldMap toolbox.⁴⁴ Unwarped images were aligned to the first image using a rigid body transformation, followed by a linear co-registration of T₁-weighted structural images to the mean functional MRI image, using normalized mutual information as a cost function. The co-registered anatomical images were segmented into grey matter, white matter, CSF and non-brain tissue using New Segment and non-linearly registered to MNI152 space. Normalized functional images underwent Gaussian smoothing with a full-width at half-maximum of 8 mm.

Resting-state functional MRI

We showed our results based on normative gradients generated from the preprocessed Human Connectome Project dataset, as in prior work,^22,27 and for gradients based on our own resting-state MRI acquisitions. For the latter, we employed preprocessing methods described in prior work^45,46 and preprocessed the resting-state functional MRI using a combination of AFNI⁴⁷ and FSL.⁴⁸ In brief, the first five repetition times were discarded to allow for signal equilibrium, and posterior-anterior/anterior-posterior blipped scan pairs were used to correct for geometric distortions. Following motion correction, we warped the corrected images to the T₁-weighted space using a combination of rigid body and boundary-based registrations.⁴⁹ A high-pass filter was used to correct the time series for scanner drifts. Additional noise components were removed using ICA-FIX trained on in-house data from the same scanner.⁵⁰ The resting-state time series was sampled at each cortical vertex and along hippocampal subfields surfaces, as described previously.⁴ The cortical time series was mapped to the native surface and subsequently registered to the Conte69 template using tools from the connectome workbench.

Gradient identification

We followed previous approaches to identify neocortical²⁷ and hippocampal connectivity gradients²² (see Supplementary Fig. 1 for a schema). In the neocortex, the next steps were followed in order. First, we cross-correlated the resting-state functional MRI time series of all vertices for each participant to generate a cortex-wide functional connectome. Connectomes were Fisher r-to-z transformed and averaged across subjects. The group mean connectome was converted back to r-values using a hyperbolic tangent function that ensures scaling between −1 and 1. The top 10% connections for each row were retained and all others set to zero to provide a balanced representation of vertex-wise functional connectivity profiles; then, the row-wise cosine similarity matrix was calculated, representing the similarity of vertex-wise functional connectivity profiles. Next, we used diffusion map embedding to identify principal eigenvectors explaining spatial variations in connectivity,⁵¹ which, given their smooth appearance, are known as gradients in the literature. Hippocampal gradient computations were similar, except that connectivity patterns were calculated from hippocampal vertices to cortical parcels,⁵² and a normalized angle affinity matrix was computed to account for the asymmetric connectivity matrix.²² Neocortical and hippocampal gradients generated from our in-house data were furthermore aligned to the normative gradients obtained from the Human Connectome Project dataset using Procrustes rotation.⁵³ Although gradients can also in principle be generated for task functional MRI data, here we focused on building gradients from resting state functional MRI data, as it (i) anchored our gradients into the broader literature that derived these gradients mainly from resting state functional MRI and specifically prior work based on the Human Connectome Project dataset^22,27; (ii) was in line with conceptual and empirical accounts, suggesting that intrinsic functional gradients may serve as coordinate systems to represent brain organization and functional dynamics^54–57; and (iii) made these coordinates independent from the tasks being analysed. All gradient computations were performed using the BrainSpace toolbox, available at https://github.com/MICA-MNI/BrainSpace (accessed 24 July 2021).⁵⁸

Experimental design and statistical analysis

Pattern separation analysis in healthy individuals

Activation analysis was based on mnemonic similarity task phase 2 only. We employed an event-related functional MRI analysis and convolved trial-specific delta functions with the canonical hemodynamic response function (the sum of two gamma functions). For each participant, each of the two event types, (similar|lure) and (similar|novel), was modelled by a regressor of interest, and six motion parameters were included as nuisance regressors in a voxel-wised general linear model. Contrasts were created for each subject to address pattern separation, defined by (similar|lure) − (similar|novel). One sample t-tests were used to examine the effects at the second (i.e. group-wise) level. Contrasts for the two conditions at the first (i.e. subject-specific) level analysis and the t-value map for the one-sample t-test describing the group effects were mapped onto cortical and hippocampal surfaces using a boundary-based registration procedure.⁴⁹

We stratified findings based on previously mapped connectivity gradients in hippocampal²² and neocortical regions²⁷ to associate the spatial distribution of effects with respect to the main axes of functional organization. In brief, we computed Spearman correlations between pattern separation activation t-values and first/second gradients for the neocortex and hippocampal subfields, respectively. To account for spatial autocorrelations, we used non-parametric spin tests⁵⁹ implemented in BrainSpace.⁵⁸ Only positive t-values were considered.

In addition to group-wise analyses, we verified the results at the single subject level. To this end, we mapped subject-specific t-maps to the gradient space, thresholded these (t > 0) and assessed the correlation between individual activation patterns and neocortical as well as hippocampal gradients.

Comparison between healthy controls and patients with temporal lobe epilepsy

The patients with TLE underwent the same task- and resting-state functional MRI scans as the healthy controls. At the behavioural level, we compared the BPS scores between the patients and controls, adjusting for age and sex. We also compared each of the nine different item/response combinations (i.e. repetition items labelled as old/similar/new, lure items labelled as old/similar/new and novel items labelled as old/similar/new). Findings were corrected for multiple comparisons using the false discovery rate (FDR) procedure.⁶⁰

For each group, we had the contrast map representing the differences between the betas of conditions (similar|lure) and (similar|novel) mapped on neocortical and hippocampal surfaces, controlling for age and sex. For each group, we computed Spearman correlations between task-related t-values and functional gradients across neocortical and hippocampal vertices in both the left and right hemispheres, separately. As determined above, significances were determined using non-parametric spin tests that controlled for spatial autocorrelation.⁵⁹ Between-group comparisons in correlations were performed using non-parametric permutation tests, where the actual difference in correlation coefficients between groups was placed in a distribution of correlation coefficient differences after the group was randomly permuted 1000 times.

Associations to behaviour

Neocortical and hippocampal t-value maps in controls and patients were weighted by the corresponding gradient to generate a single scalar loading score. This provided a personalized loading score, with high loadings indicating that activity patterns followed the principal gradients and low loadings the opposite. We computed product moment correlations between both loading scores across all participants for both left and right first and second gradients and controlled for multiple comparisons using Bonferroni adjustment. We then calculated regressions between individual differences in residual BPS scores and neocortical as well as hippocampal loadings, respectively. In addition, we computed product moment correlations between individual residual BPS scores and mean hippocampal anterior/posterior t-values, by splitting the hippocampal mask along the y-axis into two equally long segments.

Data availability

Data to reproduce the main results will be made available on osf.io.

Results

Behavioural findings

In all participants, BPS scores were calculated as the pattern separation rate P(similar|lure) corrected for a similar bias rate P(similar|novel) based on the responses given in the second phase of the functional MRI task. In the healthy controls, the BPS scores followed an approximate normal distribution with mean ± SD = 0.39 ± 0.20 (Fig. 1A). In patients, we observed a lower score with mean ± SD = 0.18 ± 0.15. When comparing behavioural performance between patients and controls, we detected reduced pattern separation abilities in the patients. Indeed, while the controls successfully labelled 51.8 ± 16% of lure trials as ‘similar’, the patients did so for only 33.3 ± 16% trials [Student’s t-test; t(38) = 2.72, two-tailed FDR-corrected P = 0.02; Fig. 1B]. Critically, performances on other trial types did not differ across groups, and patients also showed intact pattern completion, i.e. labelling lure trials as ‘old’ (Supplementary Table 2). When the BPS score was calculated as BPS = P(similar|lure) − P(similar|novel), i.e. when scores were adjusted for the similar bias rate, patients still showed significant impairment compared to controls [Student’s t-test; t(38) = 2.73, two-tailed FDR-corrected P = 0.02]. Notably, we also calculated recognition discriminability, i.e. P(old|repetition) − P(old|novel), between patients and controls, but did not find a difference between TLE and controls groups [t(38) = 0.29, P = 0.39]. These findings suggest that impaired pattern separation in TLE is unlikely to be driven by impaired recognition memory. Findings were virtually identical after repeating between-group contrasts after controlling for age and sex.

Topographic profiles of pattern separation in healthy individuals

Our next analyses examined whether the process of memory retrieval in the presence of lures was linked to neural activity in the hippocampus. To achieve this aim, we performed a targeted analysis of neural activity in the hippocampus, controlling for a family-wise error (FWE) rate of 0.05 using small volume correction. In controls, conventional voxel-wised general linear models mapped pattern separation activation based on the contrast of (similar|lure) versus (similar|novel), similar to prior work,³⁵ and identified significant hippocampal activations with the most marked effects in anterior divisions (P < 0.005, uncorrected, cluster size >5; Fig. 2A, Supplementary Table 3 and Supplementary Fig. 2). Findings were virtually identical when additionally including time and dispersion derivatives in the voxel-wised general linear model (Supplementary Fig. 3). The following findings were all based on the simple model. Repeating this analysis in the neocortex highlighted activations in distributed cingulate, frontal, temporal and parietal areas, although findings were less robust (P < 0.005, uncorrected, cluster size >5; Fig. 2A).

Figure 2.

Mapping pattern separation activation to connectome gradient space. [A(i and ii)] Voxel-wise pattern separation mapping in controls (shown at P < 0.005 uncorrected, k > 5). Hippocampal findings survived FWE < 0.05 based on small volume correction. [B(i and ii)] Pattern separation activations are first mapped to hippocampal subfield surfaces and then to hippocampal connectivity gradient space. The first hippocampal gradient follows its long axis and the second its infolding.²² (iii) The correlation between gradient maps and activation maps can be calculated, and statistical significance can be determined using non-parametric spin tests that account for spatial autocorrelation.⁵⁹ (iv–vi) A similar analysis was performed for neocortical surfaces and gradients. The first neocortical gradient depicts a transition from unimodal to transmodal areas and the second gradient differentiates visual from somatomotor cortices.²⁷ [C(i–iv)] The analysis was repeated at a single subject level, assessing correlations between individualized activation patterns and functional gradients. Gradients in the figure are based on the Human Connectome Project dataset in accordance with prior work^22,27; for a replication based on sample specific connectivity gradients, see Supplementary Fig. 8.

Our analysis has so far demonstrated that retrieval in the face of lures activates regions of the anterior hippocampus and that patients with TLE are behaviourally impaired at this task. Next, we turned to the question of whether the topographic organization of the hippocampus and neocortex are functionally important for pattern separation during retrieval. To this end, we projected task-based functional MRI activation patterns to low dimensional manifold spaces spanned by the intrinsic connectivity gradients running along both hippocampal and neocortical subregions (Fig. 2B). Gradients in the main analyses were derived from resting-state functional MRI acquisitions based on the Human Connectome Project dataset, consistent with prior work.^22,27 In the Supplementary material, we also show consistency of our results when gradients were derived from our participants using the same methods.⁵⁸ In the hippocampus, the first gradient described the antero-posterior long axis that differentiates a more transmodal hippocampal head from the more visuo-spatial hippocampal tail, while the second gradient describes a medio-lateral differentiation.²² In the neocortex, the first gradient described a unimodal-transmodal hierarchy, while the second dissociates visual and somatosensory systems.²⁷

In the hippocampus, analysing the spatial correlation between task-based neural activations during pattern separation and the intrinsic functional connectivity gradients revealed a strong and significant correspondence. Specifically, task-based activation strength correlated strongly with the first principal gradient that differentiated anterior from posterior hippocampal divisions, with stronger activation in the transmodal anterior segments, with significances determined using non-parametric spin tests that control for spatial autocorrelations (r = 0.63; two-tailed spin-test P < 0.001). We also mapped individual subject activations (i.e. positive t-values obtained from the first level analysis) into gradient space and repeated the above analyses. Findings were consistent with the group level findings. Indeed, we observed that most participants showed a positive correlation between their activations and the first hippocampal gradient (73%), confirmed with a one-sample t-test to assess correlations different from 0 [one-sample t-test; t(25) = 2.75, two-tailed P = 0.01].

In the neocortex, we observed a marginal positive correlation between activation strength and the first unimodal-to-transmodal gradient (r = 0.14; two-tailed spin-test P = 0.1). These findings indicate that activations tended to be stronger at the transmodal apex compared to the unimodal end and that there is thus a similar organization in hippocampal and neocortical networks during pattern separation in healthy controls, with higher activations in transmodal systems (i.e. anterior hippocampus and default mode systems) compared to unimodal sensory systems.

Functional reorganization in patients with temporal lobe epilepsy

Having demonstrated that neural activations within the hippocampus during pattern separation adhere to the main axes of intrinsic connectivity organization, we next examined whether this becomes reorganized by pathology impacting the hippocampus. To achieve this goal, we repeated this analysis in the cohort of patients with TLE. Compared with controls, TLE patients presented with atypical activation-gradient relationships, specifically with respect to the first gradient in both hippocampal and neocortical areas (Fig. 3A and B, see Supplementary Fig. 1 and Supplementary Table 4 for voxel-wise findings). Formally, while activation of the left hippocampus (i.e. ipsilateral to the seizure focus) in patients was also positively correlated with the first gradient (r = 0.48, two-tailed spin-test P < 0.05), correlations were weaker than in controls (r = 0.64, two-tailed spin-test P < 0.001). A difference in correlations, i.e. a weaker correlation in patients ipsilateral to the focus, was confirmed using non-parametric permutation tests (two-tailed P = 0.05). Considering the right (i.e. contralateral) hippocampus, controls showed a similar positive alignment between functional activations and the first gradient (r = 0.56, two-tailed spin-test P < 0.05). As in the left (i.e. ipsilateral) hemisphere, correlations were lower in patients with TLE (r = 0.02, two-tailed spin-test P > 0.2). Indeed, non-parametric permutation tests confirmed that patients presented with weaker correlations than controls in the contralateral hippocampus (two-tailed P < 0.008).

Figure 3.

Perturbations in structure-function relations in neurological patients with hippocampal damage. (A) Voxel-wise pattern separation mapping in left TLE (shown at P < 0.005 uncorrected, k > 5). (B) Surface-based activations (i and iii) and correlations between pattern separation activations with both groups and connectome gradients (ii and iv) in the hippocampus. (C) Equivalent analysis in the neocortex. Findings show that the topography of functional activations in patients did not follow connectome gradients the same way as in controls, specifically with respect to the first gradient in hippocampal and neocortical regions. Gradients in the figure are based on the Human Connectome Project dataset in accordance with prior work^22,27; for a replication based on sample specific connectivity gradients, see Supplementary Fig. 9.

In the neocortex, patients with TLE showed an opposite correlation between activations and the principal gradient (left: r = −0.42, two-tailed spin-tests P < 0.001; right: r = −0.14, two-tailed spin-test P = 0.1) compared with healthy controls (left: r = 0.12, two-tailed spin-tests P = 0.1; right: r = 0.19, two-tailed spin-tests P = 0.1). Significant differences in correlations were confirmed via non-parametric permutation tests (two-tailed; left: P = 0.008, right: P = 0.2). Unlike controls, patients showed stronger activation at the unimodal relative to the transmodal end in both left and right hemispheres.

No significant between group differences were observed when comparing controls and patients with TLE for (old|repetition) contrasts both at the voxel-level and when comparing correlations between activity patterns and functional gradients (permutation tests: P > 0.1; Supplementary Fig. 5). In assessing the between group differences for the (new|novel) condition, we observed no differences in voxel space. However, an alteration in the correlation between hippocampal activation and connectome gradients was observed for the contralateral hippocampus in TLE relative to the right hippocampus in controls (Supplementary Fig. 6).

Analysis of hippocampal-cortical coupling and association with behaviour

Finally, we explored the coupling of hippocampal and neocortical findings and studied inter-individual differences in pattern separation ability. To this end, we weighted subject-specific t-values from the task-based functional MRI experiment relative to the neocortical and hippocampal principal gradients (Fig. 4). This procedure provided personalized loading scores for both the hippocampus and neocortex, with high loadings indicating that task-related functional MRI activity patterns followed the gradients and low loadings indicating the opposite. In the first series of analyses, we identified a correlation between hippocampal and neocortical loadings in both patients and controls for both gradients (first gradient: controls r > 0.7 patients r > 0.8; second gradient: controls r > 0.6 patients r > 0.9; all these correlations were significant at P < 0.05 after correction for multiple comparisons).

Figure 4.

Behavioural associations. (A) Correlation between hippocampal features and BPS score. (i) A subject-specific loading score was computed by weighting individual t-statistical maps from the pattern separation activation by normative gradient maps in the hippocampus. (ii) We then assessed the correlation between these loadings and BPS scores after controlling for age and sex for both hippocampal and neocortical regions. (B) Equivalent analysis for the neocortex. Gradients in the figure are based on the Human Connectome Project dataset in accordance with prior work^22,27; for a replication based on sample specific connectivity gradients, see Supplementary Fig. 10.

These findings indicate that that individuals with high hippocampal loadings (i.e. activity is clustered in more anterior regions) also showed high loadings in neocortical regions (i.e. activity is clustered in more transmodal default mode networks). Furthermore, these subject-specific loading scores were used as regressors for inter-individual differences in behavioural pattern separation score, controlling for age and sex across both patients and controls. Overall, left and right loadings were correlated to BPS performance for the hippocampus (r = 0.32, two-tailed P = 0.04), with moderate effects in the left hemisphere (r = 0.4) and small-to-moderate effects in the right hemisphere (r = 0.24). Considering neocortical loadings, we only observed marginally positive associations at uncorrected thresholds (both hemispheres: r = 0.26, two-tailed P = 0.10; left: r = 0.22, two-tailed P = 0.19, right: r = 0.31, two-tailed P = 0.05). Besides, left hippocampal activation within the anterior and posterior hippocampus was only marginally correlated with BPS scores, and correlation coefficients in both segments were similar. Findings were similarly weak for the right hippocampus (Supplementary Fig. 7).

Associations with episodic memory

We cross-validated our behavioural findings against the independently administered episodic memory task. First, BPS scores were indeed correlated with episodic memory scores (r = 0.488, two-tailed P = 0.002) and patients with TLE were significantly impaired compared to controls [Student’s t-test; t(38) = 2.75, two-tailed P = 0.009]. Furthermore, both hippocampal (r = 0.337, two-tailed P = 0.041) and neocortical loadings (r = 0.472, two-tailed P = 0.02) significantly correlated with episodic memory scores in both patients and controls. We furthermore computed product moment correlations between BPS scores and scores derived from the MoCA,³⁷ a widely used screening test administered outside the scanner. BPS scores for all participants in this study were positively correlated with their MoCA total score (r = 0.48, two-tailed P = 0.005) and also with short-term memory recall subscores (r = 0.32, two-tailed P < 0.05).

Discussion

The ability to retrieve a specific memory from our past depends on a successful distinction of the relevant event from episodes with similar features, a process which is assumed to be instantiated by the hippocampus in conjunction with distributed neocortical networks. Here, we identified compact signatures of hippocampal and neocortical functional activation during pattern separation in healthy individuals. We also evaluated perturbations in these signatures in a cohort of epileptic patients with hippocampal pathology. Our patients with TLE indeed presented with both episodic memory deficits and structural anomalies to the hippocampus, making this cohort a candidate human model to study the reorganization of functional memory networks. Our work capitalized on connectome dimensionality reduction techniques to identify topographic connectivity gradients in the hippocampus and neocortical regions and used these as intrinsic coordinate systems to map functional activation patterns. This approach established that task-based activations during a pattern separation paradigm in healthy individuals were successfully captured by intrinsic functional connectivity gradients in the hippocampus, which recapitulates its long axis,^22–24 and marginally in the neocortex, where functional activity tended to increase towards the transmodal apex.²⁷ Studying a cohort of patients with mesial TLE, we then showed that these topographies can also help to conceptualize profiles of disease-related functional network alterations in a compact reference frame. Specifically, we identified a reduced concordance between connectome gradients and functional activations in hippocampal as well as neocortical networks, relative to controls. Notably, our approach delivered a personalized functional MRI signature score, indexing the alignment of functional activations to connectome gradients, which was found to be associated with behavioural performance in both patients and controls. Collectively, these findings identify topographic underpinnings for episodic memory processing in humans and chart potential mechanisms governing functional reorganization in neurological patients who present with damage to pivotal nodes in the human memory network.

Recent years have seen an increase in the application of manifold learning techniques to capture smooth interregional transitions—also referred to as gradients—in brain function, microstructure and connectivity.^{54,58,61–65} Leveraging resting-state functional MRI, these techniques have visualized connectivity gradients in both hippocampal subregions,^22–24 as well as neocortical areas.^27,62 In contrast to conventional whole-brain voxel-wise activation analysis¹⁷ or region of interest approaches that focus on the whole hippocampus or its subfields,⁹ gradient mapping techniques capture salient organizational axes of intrinsic brain function in a low-dimensional coordinate system governed by connectivity. This estimate is potentially less affected by subtle inter-individual variations and challenges in reliably defining specific subregions such as hippocampal subfields.^66,67 Specifically, the need to define discrete boundaries, which can be difficult and rater-dependent when the boundaries are drawn in demarcating borders between individual subfields, was sidestepped by capitalizing on gradients. Gradients have increasingly been used to stratify stratification to assess changes in functional activation patterns during specific tasks,^29,68 to visualize spatial trends of neurodegenerative deposits and cortical thinning in ageing,^31,32 and to assess molecular underpinnings of neuroimaging findings.^63,69 More generally, the application of gradients provided a novel perspective on pattern separation processes, because the connectome-embedding techniques we applied placed hippocampal subregions in the broader cortical landscape. This approach, thus, complements approaches focusing on specific hippocampal subregions, by contextualizing them in the cortex as a whole. Collectively, our work provides a novel method for dissociating anterior-posterior and medial-lateral differences and provides novel evidence for an association between pattern separation activation and hippocampal long axis location that allows this important aspect of cognition to be contextualized at both macro and microscales of investigation.

Here, we projected functional MRI activations from the pattern separation paradigm into this intrinsic connectivity space in both neocortical and hippocampal regions. We used an adapted version of the mnemonic similarity task, as it allows for sensitive functional mapping and the calculation of BPS scores at the group and single subject levels.^35,70 Prior task-based functional MRI analyses identified both hippocampal and neocortical activations during pattern separation.^2,9,17,26,71 Compared to our work, these studies generally carried out more conventional region-of-interests analyses, either at the level of the entire hippocampus ^2,17 or hippocampal subfields,^2,16,71 or they ran unconstrained voxel-wise analysis.^26,71 Our work established that gradient stratification of activation during pattern separation shows strong alignment with the principal hippocampal gradient and a marginal alignment with the neocortical gradient, anchoring our findings to established models of neocortical hierarchies ^72–74 and hippocampal long axis specialization.^{18,20–24,75,76} Thus, our findings provide a novel compact way to conceptualize functional participation of mesiotemporal and neocortical subregions in the processes that guide memory retrieval.

Our findings in healthy controls suggest increased functional activity towards anterior hippocampal segments and densely connected transmodal association cortices during pattern separation. In the hippocampus, prior work has already investigated subregional specialization with respect to hippocampal subfields, particularly CA3/dentate gyrus. However, to our knowledge, these processes have not been studied extensively relative to the hippocampal long axis, which is considered as a crucial dimension of hippocampal subregional organization, particularly in relation to function.^18,21,22 Our results seem to echo findings from other memory tasks that suggest a distribution of neural processing across different hippocampal segments and different levels of the cortical hierarchy. Indeed, a prior study using a scene repetition task observed distinct familiarity and repetition-related recognition signals in the CA3/dentate gyrus,¹⁶ which were dissociated along the hippocampal longitudinal axis. Further work associated gist memory with increased anterior hippocampal activity.⁷⁷ An established model of hippocampal long axis specialization¹⁸ posits that hippocampal representations gradually vary along its long axis, with broader anterior representations and sharper posterior representations; this differentiation may be relevant in the context of the pattern separation, where similar items need to be adequately discriminated from old items. Hippocampal subregions are likely to carry out these computations in concert with other brain networks.^17,26 Consistent with these prior studies, we also observed tendencies for neocortical participation in pattern separation. These were localized in both higher-order (e.g. default mode) apex regions in prefrontal and midline cortices and also in inferior temporal regions and occipital-temporal areas that participate in ventral visual streams. The interplay between hippocampal activation shifts along the long axis, and recruitment of sensory as well as integrative cortical areas may reflect the balanced contribution of different networks to memory. This involves both sensory^78,79 and higher order transmodal cortices involved in self-related cognition,⁸⁰ mental time travel^81,82 and cognitive control.⁸³ These macroscale findings could potentially reflect computational accounts justified by evidence in animals,⁸⁴ whereby hippocampal nodes invoke a reverse hierarchical series of cortical pattern association networks implemented through hippocampal-cortical connections, to perform a form of pattern generalization and retrieve complete patterns stored in higher order areas. Although prior studies have not examined the profile of connectivity gradients during pattern separation, recent studies have suggested that dissociations between brain activity linked to more general memory processes have a similar spatial pattern. Two prior studies in healthy individuals demonstrated that when individuals use information that is represented in memory rather than in immediate sensory input to guide decision making, a shift in the locus of brain activity is observed away from the visual cortex towards transmodal systems, principally the default mode network.^29,85 Similar shifts in neural activity from sensory to transmodal regions occur when associative decisions are primed with affective or spatial cues consistent with this decision.⁸⁶

Several experimental studies in non-human animals⁸⁷ and behavioural assessments in neurological patients^11–15 have demonstrated deficits in pattern separation associated to damage of the hippocampus. Here, we studied a relatively small, yet well-characterized cohort of patients with left TLE who presented with variable degrees of mesiotemporal pathology, and we used an identical neuroimaging and topographic profiling framework as in our healthy controls. The focus of the current study was on left TLE patients, given previously established differences in functional activation patterns of the left and right mesiotemporal lobe during memory activation paradigms⁸⁸ and prior work suggesting more marked network alterations in left relative to right TLE patients.⁸⁹ We observed an association between pattern separation and activity in bilateral hippocampal subregions, compatible with prior studies.^2,9,90 Notably, our findings in patients suggest that epileptic pathology in the left hippocampus is sufficient to induce bilateral functional reorganization and evoke behavioural deficits in pattern separation performance, and both macro- and microscale connectivity gradients provide an important way to explore this issue. In this context, it is interesting to note that pathological and neuroimaging findings in TLE generally confirm anterior-to-posterior gradients in the degrees of structural pathology, with most marked findings in the hippocampal head and a less affected hippocampal tail.^4,91,92 At the behavioural level, our young and middle-aged patients with TLE had a lower pattern separation scores than healthy controls, giving more ‘old’ responses and fewer ‘similar’ responses to lure items. However, performance was unimpaired for novel and repeated items. This behavioural deficit is generally consistent with mnemonic findings in the broader neurocognitive literature on TLE. For example, patients with TLE typically do not perform as well as healthy individuals on visual and verbal memory tasks,^93–96 and recent behavioural findings have also demonstrated impaired spatial pattern separation performance in this condition.¹¹ At the level of brain activation, we observed tendencies for decreased ipsilateral hippocampal functional activations, together with activity increases in the contralateral hemisphere in our patients compared with controls. Overall bilateral alterations in hippocampal function, together with compensatory involvement of neocortical networks in TLE patients with left-side hippocampal anomalies is consistent with prior findings,^97,98 and may indicate a broadening of the memory networks in patients with hippocampal damage. Patients with left TLE showed postoperative reorganization to the ipsilateral posterior hippocampus that correlated with a decline in postoperative verbal memory.^99,100 Patterns of functional alterations in patients may reflect a less efficient architecture and relate to episodic memory deficits. The novel application of topographic profiling in this study contributed to this growing understanding by showing specifically that neural activity patterns in patients atypically anchors to topographic gradients in both hippocampal as well as neocortical networks. Here, findings were indicative of a weaker concordance in both the left/ipsilateral and right/contralateral hippocampus in patients relative to controls, suggesting an overall attenuation of gradient-activation relationships in the ipsilateral and contralateral hippocampi in TLE. These changes occurred together with likely compensatory gradient-activation reorganization in neocortical networks. In addition to these group-level findings, we computed a subject-specific gradient loading score, an index of alignment of subject-specific functional activations with principal gradients. Notably, this approach revealed that individual behavioural separation ability was positively correlated with both left hippocampal loadings, and marginally with bilateral neocortical loading scores, in a sample composed of both patients with TLE and healthy controls. These findings further support that stratification of brain functional activity from connectivity-based dimensions may provide a principled and compact way to obtain brain measures that are behaviourally and clinically relevant.

Glossary

BPS

behavioural pattern separation

TLE

temporal lobe epilepsy