Functional re-organization of hippocampal-cortical gradients during naturalistic memory processes

The functional organization of the hippocampus mirrors that of the cortex, changing smoothly along connectivity gradients and abruptly at inter-areal boundaries. Hippocampal-dependent cognitive processes require flexible integration of these hippocampal gradients into functionally related cortical networks. To understand the cognitive relevance of this functional embedding, we acquired fMRI data while participants viewed brief news clips, either containing or lacking recently familiarized cues. Participants were 188 healthy mid-life adults and 31 adults with mild cognitive impairment (MCI) or Alzheimer's disease (AD). We employed a recently developed technique – connectivity gradientography – to study gradually changing patterns of voxel to whole brain functional connectivity and their sudden transitions. We observed that functional connectivity gradients of the anterior hippocampus


Introduction
The hippocampus plays a central role in numerous cognitive functions including episodic and prospective memory ( Buzsáki and Moser, 2013 ;Scoville and Milner, 1957 ;Squire, 1992 ;Strange et al., 1999 ), spatial navigation ( Buzsáki and Moser, 2013 ;Maguire et al., 1998 ;Morris et al., 1982 ;O'keefe and Nadel, 1978 ), affective processing, and stress modulation ( Bannerman et al., 2004 ;Fanselow and Dong, 2010 ;Gray, 1982 ). The hippocampus is also vulnerable to degenerative processes, linking neurodegeneration to cognitive impairment in Alzheimer's disease ( Jack et al., 1999( Jack et al., , 2000, neuropsychiatric disorders ( Sapolsky, 2000 ;Sheline et al., 1996 ), multiple sclerosis ( Sicotte et al., 2008 ) and Parkinson's disease ( Camicioli et al., 2003 ). However, the hippocampus A prevailing view of the organization of the hippocampus posits functional specialization organized along its long axis ( Poppenk et al., 2013 ;Strange et al., 1999Strange et al., , 2014. For example, the anterior pole facilitates emotions and episodic memory whereas the posterior pole facilitates spatial navigation ( Plachti et al., 2019 ;Poppenk et al., 2013 ;Zeidman and Maguire, 2016 ). However, it is not clear whether functional differences along the antero-posterior axis reflect the existence of a discrete partition into segregated parcels or a continuous change along a smooth gradient ( Genon et al., 2021 ;Tian et al., 2020 ). Functional connectivity with cortical and subcortical structures appears to change gradually, rather than suddenly, along the long axis of the hippocampus ( Amaral and Witter, 1989 ;Strange et al., 2014 ;Vos de Wael et al., 2020 ). This gradient in functional connectivity predicts recollection ability ( Prze ź dzik et al., 2019 ) and gene expression ( Vogel et al., 2020 ). Behavioral profiling suggests a self-to world-centric gradient along the anterior-posterior axis, supporting the construction of abstract representations that vary from autobiographical memory (anterior) to spatial learning (posterior) ( Plachti et al., 2019 ). Hence functional specialization and external connectivity of the hippocampus appear to unfold along its long axis. However, the relationship of these gradients to anatomy and cytoarchitecture remain unclear. Notably, cytoarchitectural properties of the broader mesiotemporal lobe change smoothly along the isoto-allocortical direction, orthogonal to the antero-posterior alignment of the functional and connectivity gradients within the hippocampus ( Paquola et al., 2020 ;Vos de Wael et al., 2018 ).
The prevailing objective to partition the cartography of the cortex into distinct parcels ( Glasser et al., 2016 ) has likewise been complemented by a "gradient-based " view ( Margulies et al., 2016 ). This approach foregrounds a relatively small number of large-scale connectivity gradients upon which more fine-grained local architectures are superimposed. Integrative transmodal regions -angular gyrus, rostral anterior cingulate, posteromedial cortex, middle temporal gyrus, and superior frontal gyri -lie at the apex of the principal gradient, while the base is "tethered " to unimodal sensory cortices. A second gradient separates visual from auditory and sensorimotor cortex. Gradients appear to be a fundamental property of cortical organization ( Bernhardt et al., 2022 ), being present in functional and structural connectivity ( Margulies et al., 2016 ;Oligschläger et al., 2017 ;Vos de Wael et al., 2021 ), morphology and microstructure ( Demirta ş et al., 2019 ;Huntenburg et al., 2017 ;Paquola et al., 2019 ;Wagstyl et al., 2015 ).
The dependence of hippocampal function on its integration with the cortex suggests a mapping between hippocampal and cortical gradients, with the anterior head integrated with the self-centric apex of the cortical gradient and the posterior tail coupled to the world-centric sensorimotor base. While this mapping has been established in resting state connectivity ( Prze ź dzik et al., 2019 ;Vos de Wael et al., 2018 ), its presence and role during the execution of hippocampal-dependent tasks is not understood. Likewise, changes in hippocampal-cortical gradients in cognitive impairment -including amnesia -have not been studied. Here, we address these issues using 'gradientography', a method for characterizing how patterns of whole brain functional connectivity change across the subcortex ( Tian et al., 2020 ). When applied to resting state connectivity, this approach confirms the presence of a gradient along the anterior-posterior axis of the hippocampus mapping onto the principal cortical gradient. When taking into account the spatial geometry of the hippocampus -and in particular the narrowing of the hippocampus distal to the anterior pole -this approach suggests a discrete transition in the hippocampal gradient between its anterior and posterior ends ( Tian et al., 2020 ); effectively a "discrete step " superimposed on a background anteroposterior (AP) gradient.
Here, we study task effects on the AP hippocampal gradient in healthy midlife and older adults and those with cognitive impairment. We use psychophysiological interactions (PPI) to adapt gradientography to the study of the task-induced changes in hippocampal-cortical gradients during a memory retrieval task. Prior to imaging, participants viewed short news clips each comprising a newsreader's introduction to a news item. Following a short break, fMRI data were acquired while participants viewed the second half of these clips ( "continuing " task events) randomly intermixed with the second half of other uncued news clips ( "naïve " task events). The prior viewing of some (but not other) newsreader's introductions selectively embeds salient semantic and visual cues into naturalistic stimuli separated across a brief time interval (minutes), as a selective probe of the hippocampus and its functionally connected networks ( Ren et al., 2018 ). We first use PPI to disambiguate direct task effects on the cortex from those mediated by the hippocampus. We then apply gradientography to quantify task-related modulations in hippocampal-cortical gradients and their difference in those with early neurodegenerative changes.

Participants and protocol
The PISA cohort comprised a mid-life population enriched for high genetic risk of AD, derived from the Prospective Imaging Study of Ageing (PISA): Genes, Brain and Behaviour. Participants comprised 219 mid-life and older adults drawn from the Prospective Imaging Study of Ageing (PISA; Lupton et al., 2021 ), consisting of two cohorts: 188 in the healthy cohort (HC; 44 males, mean age: 61, standard deviation: 7, range 43-81) and 31 in the clinical cohort (CC; 15 males, mean age: 65, standard deviation: 7, range 51-77). The HC was enriched in participants at high genetic risk for AD, with participants recruited from the general community and screened for the presence of psychiatric or neurological disease (see Supp. Methods for recruitment, phenotyping and exclusion criteria). The CC participants were recruited from local memory clinics and comprised 13 meeting criteria for Mild Cognitive Impairment (MCI) and 18 with recent onset Alzheimer's disease (AD). These clinical participants had threshold to mild dementia severity (Clinical Dementia Rating, CDR of 0.5 or 1.0). All participants provided written informed consent and the PISA study protocol was approved by the Human Research Ethics Committee (HREC) of QIMR Berghofer Medical Research Institute and the University of Queensland.
Unless otherwise stated, the healthy cohort were subsampled (133 HC, 44 males, mean age: 63, standard deviation: 5, range 47-80) so that they were age and sex matched to the clinical cohort; (independent Ttest on mean age difference, p = 0.052, stat = − 2.0; Fisher exact test on sex difference, p = 0.15, odds-ratio = 0.53). Only the within-group contrast between the naïve and continuing tasks uses the full healthy cohort.
Functional MRI (fMRI) data were acquired while participants viewed brief news clips designed to assess episodic memory retrieval under naturalistic conditions ( Ren et al., 2018 ). In brief, prior to the fMRI session, participants viewed nine short videos (15 to 30 s duration) corresponding to the first half of news clips (including the newsreader's introduction). Approximately 20 min later, fMRI data were acquired during the presentation of the second half of these nine clips (continuing task), randomly intermixed with another nine clips which were viewed naivelythat is, without prior viewing of the first half of the news story (naïve task). All of these (second half) news clips showed journalists in the field, reporting on the stories introduced by the newsreader. Hence, the continuing task clips contained familiar (visual and semantic) cues primed by the prior viewing of the news clip introduction, whereas the naïve clips did not contain such cues (see Supp. Method for data acquisition and preprocessing).
Positron emission tomography (PET) data were used to quantify amyloid status (positive or negative). PET data were acquired on a Biograph mMR hybrid scanner (Siemens Healthineers, Erlangen, Germany) with 18 F-florbetaben, a diagnostic radiotracer which possesses a highly selective binding for -amyloid in neural tissue ( Fodero-Tavoletti et al., 2012 ;Rowe et al., 2008 ). The CapAIBL software ( Bourgeat et al., 2018 ) was used to quantify each image into centiloids (CL) allowing the classification of the participants as amyloid positive ( > 20 CL) or negative ( < 20 CL). Among the 159 healthy participants who had completed a PET scan, 15 were amyloid positive.
APOE genotype was determined from blood-extracted DNA ( Lupton et al., 2021 ). Among the 158 healthy participants whose APOE genotype was available, 85 had at least one ɛ 4 allele of the APOE gene.

Gradientography
To study functional connectivity gradient changes in the task fMRI data, we adapted a recently developed method to map the organization of the human subcortex from large-scale functional connectivity gradients ( Tian et al., 2020 ). In brief, principal components analysis (PCA) was first used on temporally concatenated fMRI signals to reduce the dimensionality of whole-brain activity. Whole-brain functional connectivity was then derived between cortex and all hippocampal and subcortical voxel by correlating their fMRI signals with the ensuing PCA components. Spatial gradients characterizing continuous modes of functional connectivity variation across the subcortex were then computed using Laplacian eigenmaps. Functional connectivity gradients were calculated across the ventral subcortex with subcortical boundaries defined according to a recently published functional subcortical atlas ( Tian et al., 2020 ). For the present purposes, we focus on the functional connectivity gradients in the hippocampus (results for the entire ventral subcortex are presented in Supplementary material).
The scalar values of these Laplacian eigenmaps (connectivity gradients) across the hippocampus capture whole brain connectivity of the subcortex. The rate of change in these connectivity gradients captures functional homogeneity: slow gradient changes represent locally shared patterns of whole brain connectivity whereas rapid changes indicate functional differentiation. The magnitude and direction of these changes was estimated by application of the Sobel gradient operator to each subcortical voxel. The magnitudes of these changes (or simply "gradient magnitudes ") capture gradual versus abrupt spatial changes (discontinuities) and their task modulation ( Tian et al., 2020 ).
Gradientography was previously employed to investigate variation in functional boundaries of the subcortex between rest and task conditions ( Tian et al., 2020 ). In this previous study, functional connectivity was disambiguated from the confound of task co-activation by regressing, task manipulations -represented as task block regressors convolved with hemodynamic response functions -from the BOLD signal prior to gradient analyses, consistent with previous work ( Cole et al., 2014 ). However, this approach does not disambiguate direct task effects in cortex from those that are mediated via interactions with hippocampus and is hence not considered here. To remedy this, we used psychophysiological interaction (PPI) ( Friston et al., 1997 ).
Gradientography with PPI: As in ( Tian et al., 2020 ), the BOLD signals for each individual were represented in a matrix of dimension T × M , where T denotes the number of time frames and M denotes the number of gray matter voxels (cortical, hippocampal and subcortical voxels). After using spatial PCA to reduce the dimensionality of this matrix to T x (T-1) , we performed a general linear model (GLM) predicting BOLD responses of the PCA components ( ( ) , ∈ { 1 , ... − 1 } ) as a function of the BOLD signal of subcortex voxels ( ( ) , ∈ { 1 , ..., } ), task manipulations ( , ∈[ na ïve, cont ]) convolved with the hemodynamic response function ( ) and the PPIs of these tasks with the deconvolved signal of the subcortex voxel ( ( * ) ) ( Gitelman et al., 2003 ), with 1 representing the collinearity between the BOLD responses of the PCA component and the subcortical voxel. Hence this approach allows disambiguation of the direct effects of the task on cortical responses 2 ( ) and 3 ( ∾ ) from those specifically mediated by task effects on subcortical voxels, 4 ( * ) and 5 ( ∾ * ) .
The T-statistic associated with the coefficient 1 represents the intrinsic (or task-free) functional connectivity between gray matter voxels and hippocampal or subcortex voxels whereas the T-statistics associated with each PPI coefficient ( 4 and 5 ) correspond to the modulation of functional connectivity for each task (naïve and continuing). Each subcortical voxel in each participant therefore possesses a unique connectivity map (or "fingerprint "): pairs of voxels with similar fingerprints are functionally connected to similar brain areas. As in ( Tian et al., 2020 ), the 2 coefficient was used to quantify the similarity in these connectivity fingerprints across the subcortex. This results in a symmetric matrix of dimension N * N for every individual and for each task ( Fig. 1 .1).
Group and task comparison of gradient magnitudes: These steps yield task-related connectivity eigenmaps (gradients) and their spatial rates of change (gradient magnitudes) For group and task contrasts, we focus on the gradient magnitudes as they capture functional differentiation in whole brain connectivity by task effects or group differences: for example, differences in the gradient magnitude suggest changes in the extent of anterior versus posterior functional alignment between hippocampus and cortex. This was achieved by permutation testing across tasks or cohorts, respectively ( Fig. 1 .2). Note that by permuting gradient magnitudes between groups or tasks, this approach implicitly preserves the spatial correlations within the data. Task (for between-task permutation) or cohort (for between-cohort permutation) indices were permuted 1000 times for each comparison and the resulting gradient magnitude difference compared. When the magnitude difference obtained without permutation was in the lower 2.5% or the upper 2.5% of the permutation tests, the difference was considered significant. For ease of visualization, the difference measure was z -scored for each voxel in the subcortex based on the null distribution obtained with permutation. The z -scored measure is considered significant when it is outside + /two standard deviations.
Cortical projection: To visualize the functional integration between the hippocampus and the cortex, connectivity gradients for the left hippocampus (segmented according to the ( Tian et al., 2020 ) atlas) were projected onto the cortical surface on the basis of the cortical vertices with which they were most strongly associated ( Fig. 1 .3). To compute the connectivity matrix, the GLM coefficients 1 -representing the collinearity between the BOLD signal of each subcortical voxel and the PCA-transformed matrix -were projected onto the cortex by multiplying by the transpose of the PCA right singular matrix. Participants' projected coefficients were averaged for each cohort, CC and HC. Then, each voxel of the whole brain was associated with the voxel of the hippocampus with which it had the maximum mean projected coefficient.
The code for this study is available at https://github.com/ LeonieBorne/subcortex .

Results
Gradient eigenmaps and associated magnitudes: Gradientography yields a series of functional connectivity modes ranked in decreasing order of variance explained. Here we focus on the functional connectivity gradients in the hippocampus ( Fig. 2 ): Results for the entire ventral subcortex are presented in Supp. Fig. 1 .
Gradients I and II explain most of the variance in the Laplacian matrix, which then drops to a floor ( < 5% variance explained; Supp. Fig. 2). Both gradients exhibit a strong AP effect: For gradient I, this is superimposed on a strong left-right asymmetry, whereas gradient II is almost symmetric. The eigenmaps of these two gradients are visually similar between the healthy and clinical cohorts and across the task-free and two task conditions ( Fig. 2 ).
The magnitude of Gradient I possesses a peak between the anterior and posterior hippocampus ( Fig. 3 ), corresponding to a rapid change in hippocampus-to-whole brain task-free and task-related functional connectivity. This peak possesses a similar location and magnitude in both Gradient I and II (Supp. Fig. 4). (1) Whole-brain functional connectivity fingerprints are mapped for each subcortical voxel and individual. The general linear model coefficients ( 4 and 5 ) attributed to the psychophysiological interaction between the BOLD signal of each subcortical voxel and the tasks (continuing and naïve respectively) were used as connectivity matrices. The general linear model coefficient 1 for each subcortical voxel represents the task-free connectivity. The similarity in connectional fingerprints between pairs of subcortical voxels was measured using the 2 coefficient. The similarity matrices represent the cohort consensus for each condition (NAIVE: naïve task; CONT: continuing task; FREE: task-free).
(2) Spatial gradients characterizing continuous modes of functional connectivity variation across the subcortex were computed using Laplacian eigenmaps. The similarity matrices were transformed into sparse graphs with adjacency matrices. Eigenvalues and eigenvectors were computed for the graph Laplacian. The eigenvectors with the second and third smallest eigenvalues are called gradients I and II respectively. Each gradient provides a mapping of continuous variation in functional connectivity across the topography of the subcortex. (3) The connectivity gradient for the hippocampus is projected onto the cortical surface on the basis of the cortical vertices with which they are most strongly correlated (that is, each cortical voxel y is mapped to the hippocampal voxel with the maximum 1 ). Each cortical voxel is thus associated with the hippocampal voxel having the maximum task-free functional connectivity after projection. H, number of hippocampus voxels; N, number of subcortical voxels; M, number of whole-brain gray matter voxels; T, number of time frames.
To explore the effect of spatial smoothing, we repeated the gradientography pipeline after reducing the size of the data smoothing kernel from 6 mm to 4 mm FHWM. We observe that the total variance explained is considerably smaller (from approximately 25-30% at 6 mm down to 12-16% at 4 mm) and the relative variance explained by the two principle gradients switches order. This may reflect the lower signalto-noise ratio in subcortical regions due to the surface head coil, with resulting noisier data when smoothed at 4 mm FHWM, consistent with the comparison of different smoothing kernels in resting state fMRI ( Tian et al., 2020 ). Despite these differences, the global pattern and associated magnitudes of the gradients are preserved, including the existence of a faster gradient transition between the anterior and posterior hippocampus (Supp. Fig. 5). We hereafter focus on gradients following smoothing at the default of 6 mm.
The location of the peak in the gradient magnitude between the anterior and posterior hippocampus is consistent across both smoothing kernels and with the boundary previously identified using resting state gradientography ( Tian et al., 2020 ). We next used permutation testing to study task effects and group differences in the location and extent of this magnitude peak across tasks and groups.
Task comparison: We used PPI to contrast hippocampal gradients between the naïve and continuing task conditions. The gradient peak between the anterior and posterior hippocampus is visually similar between the two tasks ( Fig. 3 ). However, between-task permutation tests show that in the healthy cohort, the continuing task has a significantly stronger gradient I across the anterior hippocampus than the naïve task ( Fig. 4 ). This implies that the pattern of whole-brain functional connectivity changes more quickly from the anterior toward the posterior hippocampus in the continuing compared to the naïve task -that is, in the presence of familiar cues in the naturalistic stimuli. This occurs bilaterally and extends across much of the anterior extent of the hippocampus. This task-related difference in gradient magnitude is not significant for gradient I in the clinical cohort ( Fig. 4 ) nor for gradient II in either cohort (Supp. Fig. 4). Fig. 2. Gradients eigenmap. The eigenmaps of the two gradients explaining the most variance are projected onto the hippocampus surface for each cohort and for each condition (continuing task, naive task and task-free). The percentage of variance explained is indicated below each gradient. The colormap limits are different for each projection: they range from 0 to the maximum eigenmap value (i.e. from left to right, CONTINUING: 0.054, 0.042, 0.052, 0.040; NAIVE: 0.053, 0.042, 0.052, 0.038; TASK-FREE: 0.056, 0.046, 0.052, 0.042). Alternative views of the first eigenmap are shown in Supp. Fig. 3.   Fig. 3. Gradient I magnitude. The magnitude of Gradient I is projected onto the hippocampus surface for each cohort and for each condition (continuing task, naive task and task-free). The 1st and 3rd columns correspond to the anterior view of the hippocampus and the 2nd and 4th columns to the posterior view. The gradient magnitudes are similar between conditions. On each projection, there is a peak in magnitude on the boundary between the anterior and posterior hippocampus. This peak appears to have a more posterior location for the clinical cohort. Corresponding results for Gradient II are provided in Supp. Fig. 4.  Fig. 4. Between-task difference in Gradient I magnitude. For each voxel, the difference in magnitude of Gradient I between the naïve (NAIVE) and the continuing (CONT) tasks was z -scored relative to the null distribution from the between-task permutation ( n = 1000). The z -scored difference is here projected onto the anterior (columns 1 and 3) and posterior (columns 2 and 4) view of the hippocampus for each task. The threshold of the color palette corresponds to + /-two standard deviations of the z -scored difference. For HC, the magnitude of Gradient I is significantly higher in the anterior hippocampus for the continuing task than for the naïve task. Corresponding results for Gradient II are provided in Supp. Fig. 4. We next contrasted the task-free and task-related functional connectivity gradients. In the clinical cohort, the anterior/posterior gradient magnitude is substantially stronger at the boundary of the anterior and posterior hippocampus in both task-related conditions than for the taskfree regressor (Supp. Fig. 6). The opposite effect occurs in the healthy cohort, although in a spatially heterogeneous pattern.
Group comparison: On visual inspection, the peak in functional connectivity gradient between the anterior and posterior hippocampus appears to be located in a more posterior position in the clinical than the control cohort ( Fig. 3 ). Between-cohort permutation testing shows that this cohort-wise location shift is statistically significant for the gradient I on the left hippocampus across both tasks ( Fig. 5 ). This shift in the location of the magnitude in the left hippocampus is even more pronounced for gradient II (Supp. Fig. 4). There is no discernible task-related effect and no change in the right hippocampus. The magnitude of the task-free connectivity does not differ across groups.
We also contrasted healthy individuals according to their amyloid and APOE status. Despite a reasonably well-powered contrast for APOE status ( N = 85 APOE 4 carriers; N = 73 non-carriers), there is at best an equivocal effect near the transition of the right anterior hippocampus (Supp. Fig. 7). There is no substantial effect of amyloid status across either task within the hippocampus (Supp. Fig. 8), although we note weaker power for this contrast ( N = 15 HC A ( + ); N = 144 HC A (-)).
Cortical projection: To study the cortical projection of the hippocampal eigenmode, we first identified the maximum functional mapping between all cortical and all hippocampal voxels. For each cortical voxel y i , we identified the hippocampal voxel x j with the maximum taskfree GLM coefficient 1 . Group-wise, second-level inference within the healthy cohort showed that all cortical voxels possess a statistically significant mapping to one or more hippocampus voxels (Supp. Fig. 9,10) following correction for all possible pairwise coefficients (see Supp. Methods -Corticohippocampal functional networks). The corresponding functional cortico-hippocampal maps for each of the task-associated PPI coefficients are weaker, but still encompass extensive cortical networks. Approximately 66.8% of cortical voxels possess a statistically significant psychophysical interaction for the naive task (coefficient 5 ) , increasing to 75.3% of cortical voxels for the continuing task (PPI coefficient 4 ; FDR corrected, q = 0.05).
We next used this mapping to project the hippocampal gradient eigenmap onto the cortex. The corresponding task invariant projection of Gradient I shows that progressively more anterior areas of the hippocampus project to progressively more ventromedial parts of the prefrontal and cingulate cortices. The same pattern is observed when approaching the temporal pole and the precuneus ( Fig. 6 a,b). Notably, this cortical projection shows a high similarity to connectivity gradients previously derived from cortical surface connectivity ( Margulies et al., 2016 ). We used nonparametric (wavelet-based) resampling to formally Fig. 5. Between-cohort differences in Gradient I magnitude. For each voxel, the difference in magnitude of Gradient I between the clinical (CC) and healthy cohorts (HC) was z -scored relative to the null distribution of the between-cohort permutation ( n = 1000). The z -scored difference is here projected onto the anterior (columns 1 and 3) and posterior (columns 2 and 4) view of the hippocampus for each task. The threshold of the color palette corresponds to + /-two standard deviations of the z -scored difference. For both continuing and naive tasks, the functional boundary between the anterior and posterior hippocampus is significantly more posterior for the CC than for the HC. There are no significant group differences in the corresponding task-free-derived gradient magnitude.
identify cortical regions that specifically map to the anterior and posterior extremes of the hippocampal gradient (see Supp. Method -Wavelet resampling). We observe that projections of the anterior extreme of the hippocampus gradient reside in the anterior pole of the temporal cortex, the angular gyrus, rostral anterior cingulate cortex, posteromedial cortex, and mid insula gyrus ( Fig. 6 c). The first four of these are key regions of the default mode network and mirror the peaks in gradients derived purely from cortico-cortical functional connectivity ( Margulies et al., 2016 ). Here, we also find that the insula maps to the anterior pole of the hippocampal gradient. In contrast, the cortical projection of The eigenmap for Gradient I in anatomical space shows the anterior-posterior organization. (b) right: The gradient eigenmap projection onto the cortical surface shows that the progressively more anterior hippocampus (yellow) projects to progressively more ventromedial parts of the prefrontal and cingulate cortices. The same progressive projection pattern is observed when approaching the temporal pole and the precuneus (posteromedial cortex) (c) Nonparametric wavelet resampling reveals specific projections from the anterior extreme of the hippocampal eigenmap to the anterior pole of the temporal cortex, the angular gyrus, rostral anterior cingulate cortex, posteromedial cortex, and mid insula gyrus. Color depicts the strength of statistical association (log( p )). Thresholded maps are provided in Supp. Fig. 13. the posterior tail of the hippocampus is weaker and more diffuse than the anterior pole (Supp. Fig. 11). The strongest projections include regions within the ventral attention network ( Vossel et al., 2014 ) such as the middle and inferior frontal gyrus, and the temporoparietal junction. Weaker projections of the posterior tail of the hippocampus include unimodal cortical regions such as the visual, auditory and somatomotor cortex, again reflecting the pattern of cortical-derived gradients.
The two cohorts and the two tasks show similar projection patterns (Supp. Fig. 12). Of note, the projections for the CC show greater spatial variability, but this is largely due to the smaller sample size of the CC, as revealed by downsampling the size of the HC to match that of the CC (Supp. Fig. 12).

Discussion
A core computational principle of the hippocampus is to bind cortical representations ( Whittington et al., 2022 ), a function that requires a topological mapping between its internal organization and its interactions with cortex ( O'keefe and Nadel, 1978 ). Here, we show that the task-related functional connectivity of the hippocampus is organized along an anterior-posterior gradient which maps onto a corresponding cortical gradient with apices in core regions of the default mode network and tails in primary sensory cortices. The hippocampal gradient exhibits a transition in its connectivity fingerprint between the anterior head and posterior tail, which increases in magnitude in the presence of familiar cues and shifts toward the posterior tail in those with cognitive impairment. These findings add to our emerging knowledge of hippocampal and cortical interactions by highlighting the task-relevance of connectivity gradients and their role in neurodegenerative disorders.
Converging evidence suggests that the hippocampus links the disparate representations of a salient stimulus across the cortex ( Olsen et al., 2012 ), enabling their joint reactivation in response to later mnemonic cues ( Kumaran et al., 2016 ;McClelland et al., 1995 ;Teyler and DiScenna, 1986 ). Prior to scanning, our participants viewed newsreaders' introductions to brief news items, hence imbuing latent visual and semantic cues in the second half of these clips when they were later viewed. These cues are latent in the sense that there is no direct audiovisual replay of the naturalistic perceptual stimuli, but rather indirect cues given implicit semantic meaning by the recently viewed narrative preface. The presence of these cues increases the salience and allocentric reference of the news clips viewed in the continuing task, which are absent in the naive task condition ( Buzsáki and Moser, 2013 ). These cues in turn increase the functional integration of the anterior hippocampus with relevant cortical networks including the default mode, the insula and medial parietal memory networks ( Ren et al., 2018 ), consequentially sharpening the functional gradient at the transition to the posterior hippocampus.
We used the framework of psychophysiological interactions to decompose cortical activity into task-invariant and task-modulated functional connectivity with the hippocampus. These analyses show how task-mediated effects superimpose on a widespread task-invariant functional embedding of cortex and hippocampus. Application of gradientography reveals how these task effects modulate a functional gradient along the AP axis of the hippocampus that maps onto an extensive functional gradient across the cortex. Of note, the task-evoked hippocampal gradients presently observed are similar to the resting state gradient recently reported ( Tian et al., 2020 ), with a comparable peak in gradient magnitude between anterior and posterior hippocampus. Likewise, the cortical projection of this gradient mirrors that of purely cortical gradients, although the mid-insula is an additional inclusion of the projection of the anterior extent of the hippocampus gradient. This perspectiveof task-dependent modulation of hippocampal and cortical connectivity gradients -departs from the classic view of discrete and disparate taskevoked activations, adding to our knowledge of the complex organization of distributed, spontaneous brain activity and its task modulation ( Salehi et al., 2020 ).
We recruited clinical participants from local memory clinics comprising those with MCI as well as those meeting threshold (mild) criteria for AD (CDR of 0.5 or 1.0). Our objective was to characterize cortical function in those with relatively mild cognitive impairment -regardless of specific diagnostic thresholds -in most activities of daily living. We observed a posterior shift in the location of the gradient transition in the left hippocampus, present in both task conditions in these clinical participants, but not in the task-free connectivity. Hence, in the presence of a task-related stressor, the region of the anterior hippocampus with a shared connectivity fingerprint extends further in the posterior direction in this cohort. Given the relative maintenance of cognitive and behavioral function in our clinical cohort -compared to well established AD -this increase in extent of task-related functional connectivity is consistent with longstanding notions of compensatory cortical responses to maintain function despite incipient pathology ( Amieva et al., 2014 ;Grady et al., 2003 ;Springer et al., 2005 ). In this regard, it is notable that functional connectivity gradients encompass extensive cortical networks, potentially facilitating substantial compensatory resources through modest compensatory changes in the hippocampus. The left-sided nature of our effect is consistent with previous reports of emergence of hippocampal asymmetry during the progression of AD ( Wachinger et al., 2016 ). Both APOE and amyloid status were also available on 158 of the healthy participants, indicating a higher risk of future neurodegeneration. We observed a patchy and somewhat equivocal effect for APOE status predominantly in the right hippocampus (Supp. Fig. 7) but no clear effect of amyloid status across either task. Our future longitudinal follow-up of this cohort may reveal stronger effects that are currently latent.
Orthogonal decompositions of structural and functional neuroimaging data do typically yield symmetric plus asymmetric leading modes ( Tokariev et al., 2019 ), as we indeed observe here. Through their relative contributions to the original data (through summation or subtraction) such complimentary modes can account for shared bilateral effects versus unilateral hemispheric dominance. Human hippocampi do show a degree of left-right structural asymmetry in health ( Woolard and Heckers, 2012 ) and during neurodegenerative disorders ( Wachinger et al., 2016 ). While we focus on the gradient magnitudes, future work could explore the relative importance of these two modes across different tasks. However, in contrast to the robustness of the gradient magnitude, spatial smoothing strongly influenced the relative variance explained of these two modes, suggesting further validation is required before this potential approach could be exploited.
There are several caveats to our study. First, the gradients observed here align with the AP axis of the hippocampus and as such do not engage with other facets of hippocampal organization, including the influence of subfield divisions. Hippocampal functions reflect a composite of connectivity gradients and subfields ( Genon et al., 2021 ;Vos de Wael et al., 2018 ) whose disambiguation is unlikely to be evident at the relatively coarse scales of acquisition resolution and smoothing width employed in the present study. We do, however, note that the hippocampus is considerably longer ( ∼4-5 cm) than our primary smoothing kernel (6 mm FWHM) and moreover, connectivity gradients remain clearly present when the data are smoothed at 4 mm FWHM. While these observations are reassuring, imaging at high field strength combined with techniques that eschew the need for spatial smoothing may assist here. Second, we apply gradientography in three-dimensional (voxel) space, treating the hippocampus as a discrete subcortical volume -consistent with much of the prevailing analysis of hippocampal structure and function. However, the human hippocampus is more accurately a folded archicortical sheet, continuous with the adjacent medial temporal lobe ( DeKraker et al., 2021 ). In addition to incorporating this important feature of hippocampal structure, a surface-based treatment of hippocampal gradients would also render the analysis of its associated magnitude to a more parsimonious rate of change on a two-dimensional sheet. Third, we focus on group and task differences, using permutation to perform inference. Future work is required to determine if individual differences in connectivity gradients covary with differences in task performance. Likewise, using a larger clinical cohort, it would be possible to study hippocampal functional connectivity gradients for the influence of genetic and biological risk factors for dementia, as well as protective factors supporting cognitive reserve ( Grady, 2013 ;Stern, 2003 ). Further development is also required for statistical inference -for example to correct gradient magnitude effects when there is a contiguous clusterlike excursion from the null as we see in our data (e.g. Figs. 4 , 5 ). Finally, we note that while the study of functional gradients across brain systems has already revealed promising new principles of functional organization, there remains considerable validation to be achieved, including the relationship of connectivity gradients to cytoarchitecture and their robustness across different preprocessing and analytic choices ( Bernhardt et al., 2022 ;Genon et al., 2021 ).
The hippocampus underpins relational binding, forming "rapid, continuous, and obligatory associations among disparate elements across space and time " ( Olsen et al., 2012 ). In this vein, hippocampal neurons detect cognitive boundaries as they occur in naturalistic stimuli and reinstate neural states in response to subsequent cues (J. Zheng et al., 2022 ). Recent theories propose that these hippocam-pal functions are supported via a 'conceptual model' of cortex, performing a computational integration of related cognitive representations ( Whittington et al., 2022 ;Zeidman and Maguire, 2016 ). Here, we find that visual and semantic cues that link present with recent narratives lead to a modulation of the functional embedding of hippocampal activity in cortical systems, organized along hippocampal-cortical connectivity gradients. Intriguingly, neurophysiological recordings reveal that theta oscillations precede continuously along the long axis of the hippocampus as traveling waves ( Lubenov and Siapas, 2009 ) suggesting that hippocampal gradients shape local information propagation ( Kleen et al., 2021 ). Waves of activity also characterize cortical states across a wide variety of cognitive states and tasks , including working memory ( Sreekumar et al., 2021 ). Coupling between cortical and hippocampal waves propagating along their principal connectivity gradients is hence a candidate mechanism for the spatiotemporal integration of cortical and hippocampal activity.

Data availability
De-identified data from the Prospective Imaging Study of Ageing (PISA) will be made available to other research groups upon request. Due to privacy, confidentiality and constraints imposed by the local Human Research Ethics Committee, a "Data Sharing Agreement " will be required before data will be released. Due to ethics constraints, data will be shared on a project-specific basis. Depending on the nature of the data requested, evidence of local ethics approval may be required.
The code for this study is available at https://github.com/ LeonieBorne/subcortex .