A mechanism for hippocampal memory recall based on excitatory-inhibitory fluctuations in neocortex

The brain has a remarkable capacity to acquire and store memories that can later be selectively recalled. These processes are supported by the hippocampus which is thought to index memory recall by reinstating information stored across distributed neocortical circuits. However, the mechanism that supports this interaction remains unclear. Here, in humans, we show that recall of a visual cue from a paired associate is accompanied by a transient increase in the ratio between glutamate and GABA in visual cortex. Moreover, these excitatory-inhibitory fluctuations are predicted by activity in the hippocampus. These data suggest the hippocampus gates memory recall by indexing information stored across neocortical circuits using a disinhibitory mechanism.


INTRODUCTION 31
Memories are thought to be stored across sparse and distributed neuronal ensembles in the 32 brain 1,2 . To facilitate memory recall, activity across neuronal ensembles is selectively 33 reinstated to recover enduring representations of the past. This reinstatement is thought to be 34 mediated by the hippocampus, a brain region important for learning and memory 3 . 35 Anatomically, the hippocampus sits at the apex of a cortical sensory processing hierarchy 4 36 where inputs received by sensory cortices reach the hippocampus via the entorhinal cortex and 37 other relay regions, which in turn make widespread cortico-cortical connections that project 38 the hippocampal output back to neocortex 5,6 . This reciprocal anatomical connectivity equips 39 the hippocampus with the necessary architecture to coordinate activity with neocortex, thus 40 providing a 'memory index', or summary sketch, for information stored across distributed 41 cortical circuits 7-9 . Consistent with this view, during memory recall hippocampal reinstatement 42 predicts subsequent neocortical reinstatement 10 . 43 44 However, the mechanism that allows the hippocampus to coordinate reinstatement across 45 distributed neocortical circuits remains unclear. In animal models, neural circuit manipulations 46 suggest higher-order brain regions may modulate release of sensory information in neocortex 47 they could later recall the relevant auditory-visual association during a surprise post-scan 122 associative memory test performed after the inference task was completed ( Fig. 1e-g). Indeed, 123 performance on the associative memory test, that assessed memory for auditory-visual 124 associations learned on day 1, predicted performance on the inference test performed on day 3 125 (Fig. 1h). Consistent with previous neuroimaging data in humans and cellular recordings in 126 mice 33 , these behavioural findings suggest inferential choice during the inference test involves 127 associative recall of the intermediary visual cues. In this manner, the inference task provides a 128 suitable paradigm to investigate the neural mechanisms that underlie hippocampal-dependent 129 associative memory recall. 130 131

BOLD signal in the hippocampus and visual cortex is modulated during memory recall 132
To investigate the relationship between the hippocampus and neocortex during associative 133 memory recall we implemented a novel imaging sequence 36 , which enabled interleaved 134 acquisition of near-whole brain fMRI together with fMRS in V1 (Fig. 2a). This imaging 135 sequence thus provided a means to simultaneously measure both hemodynamic and 136 neurochemical changes during the inference task, in an event-related manner. 137 138 Using fMRI data from the interleaved sequence, we first identified brain regions modulated by 139 recall of a visual cue in response to the associated auditory cue presented during the inference 140 test (Fig. 1a). To obtain the most accurate estimate for associative memory recall, we 141 categorized trials post-hoc, using participants' behavioural performance from both the 142 inference test and subsequent post-scan associative memory test (Fig. 1c-e), which were highly 143 correlated across participants (Fig. 1h). Trials where participants made both the correct 144 inference and subsequently remembered the auditory-visual associations were classified as 145 'remembered'. Trials where participants made either the incorrect inference or subsequently 146 forgot the auditory-visual associations were classified as 'forgotten' (Fig. 2b, Supplementary  147 Table 2, Methods). Neural signatures acquired during the 'forgotten' trials thus provided a 148 control condition for those acquired during the 'remembered' trials. Consistent with previous 149 research investigating associative recall of visual cues 37,38 , we observed a significant increase 150 in BOLD signal in both the hippocampus and visual cortex on 'remembered' versus 'forgotten' 151 trials ( Fig. 2c; Supplementary Fig. 2). 152

Dynamic increase in the ratio between glutamate and GABA in visual cortex during recall 154
We then asked whether associative memory recall of a visual cue is also accompanied by 155 changes in the ratio between glutamate and GABA ('glu/GABA ratio') in the visual cortex. 156 Using the interleaved fMRS data acquired in primary visual cortex (V1) (Fig. 2a,d), we 157 quantified the concentration of glutamate and GABA normalised to total Creatine (tCr) in an 158 event-related manner (Fig. 2b,e). We then used MRS-derived measures of glutamate and 159 GABA to estimate changes in glu/GABA ratio 39 (see Methods), where changes are evaluated 160 through assessment of the ratio of 'remembered' trials relative to 'forgotten' (as defined 161 above). In this manner, the 'forgotten' trials again provide a condition and stimulus-matched 162 control for data acquired during the 'remembered' trials. 163

164
To detect dynamic changes in glu/GABA ratio it was not appropriate to implement default 165 assumptions typically used to detect static estimates (see Methods). Namely, these default 166 assumptions assume the dynamic range of GABA is fixed by normalising GABA relative to 167 other more abundant metabolites. Here, to optimise our sensitivity to changes in glu/GABA 168 across conditions we removed these default constraints. Notably, while this approach leads to 169 higher GABA estimates, the uncertainty in the metabolite estimates were reduced 170 ( Supplementary Fig. 3). Moreover, our analysis controlled for any effect of metabolite scaling 171 by comparing the difference between two conditions ('remembered' versus 'forgotten'). 172 173 During recall, we observed an increase in glu/GABA ratio in V1 when comparing 174 'remembered' versus 'forgotten' cues ( Fig. 3a-b). Standard quality metrics indicated that our 175 data quality was comparable with those reported in previous studies 40-43 ( Supplementary Fig.  176 4, Supplementary Table 4). To control for any biases introduced by differences in the number 177 of 'remembered' versus 'forgotten' trials (Supplementary Table 5), we compared the group 178 mean metabolite change against a null distribution generated by permuting the identity labels 179 assigned to each trial. This analysis revealed a significant decrease in GABA and a significant 180 increase in glu/GABA ratio during memory recall (Fig. 3d-f). This change in glu/GABA ratio 181 was still observed when using performance on the inference task alone to categorise trials into 182 'remembered' and 'forgotten' (Supplementary Fig. 5). Furthermore, the increase in glu/GABA 183 ratio was not observed during periods immediately before or after recall ( Fig. 3a-b; 184 Supplementary Fig. 6). These findings cannot be explained by differences in data quality 185 measures between the 'remembered' and 'forgotten' conditions ( Supplementary Fig. 7). 7 found at higher concentration ( Supplementary Fig. 8). Thus, we propose this transient increase 188 in neocortical glu/GABA ratio reflects a mechanism for associative memory recall. 189

190
As an additional control, we assessed changes in glu/GABA ratio during a subset of 191 conditioning trials ( Supplementary Fig. 9a) that were interleaved with the inference test trials 192 during the MRI scan and shared the same temporal structure. Importantly, previous studies 193 suggest performance on conditioning trials is not hippocampal-dependent 33 . During the 194 conditioning trials, we observed no change in glu/GABA ratio during presentation of the visual 195 cue or outcome, relative to the ITI period ( Supplementary Fig. 9b-c). We next asked which brain regions coordinate this transient break in neocortical glu/GABA 199 ratio during memory recall. The hippocampus is a promising candidate, given this brain region 200 supports memory 3 and shows activity modulation during the inference test (Fig. 2c). To test 201 this possibility, we took advantage of our simultaneous fMRI-fMRS acquisition (Fig. 2a). We 202 hypothesized that the increase in hippocampal BOLD signal observed during recall (Fig. 2c) 203 should predict the increase in glu/GABA ratio observed in V1 (Fig. 3). In line with this 204 prediction, across participants the hippocampal BOLD signal negatively predicted the relative 205 concentration of GABA and positively predicted the increase in glu/GABA ratio in V1 206 ('remembered' versus 'forgotten' trials; Fig. 4a-b). Furthermore, across the imaged brain 207 volume (Fig. 2a), only the hippocampus significantly predicted the increase in V1 glu/GABA 208 ratio on 'remembered' versus 'forgotten' trials ( Fig. 4c). Finally, this relationship between the 209 hippocampus and glu/GABA ratio was specific to the recall period during the inference test 210 ( Fig. 4d, Supplementary Fig. 10). 211

DISCUSSION 213
The hippocampus is thought to provide an index for memories stored across distributed 214 neocortical circuits 7-9 . However, the mechanism by which hippocampal activity is coordinated 215 with neocortex to facilitate memory recall has remained unclear. Here, using time-resolved 216 fMRI-fMRS in humans, we show that recall of a visual cue is accompanied by a dynamic 217 increase in the ratio between glutamate and GABA in visual cortex. This transient increase in 218 glu/GABA ratio in visual cortex is selectively predicted by activity in the hippocampus. 219 Accordingly, we propose the hippocampus gates recall of memories stored across distributed 220 neocortical circuits using a disinhibitory mechanism (Fig. 4e). This mechanism may explain 221 how a memory index represented by the hippocampus selectively releases otherwise dormant 222 representations stored across distributed neocortical circuits. 223 224 Memory recall via a disinhibitory mechanism may be supported by neural circuits identified in 225 rodents, where glutamatergic projections from higher-order or interconnected brain regions 226 have the capacity to instantiate highly specific disinhibition in cortical circuits 13 for prolonged periods, the exact E/I ratio is highly dynamic. Evidence in humans, animal 246 models and theoretical models together suggest overall proportionality between excitation and 247 inhibition is maintained to hold memories in a silent and dormant state 30-32,51 , thus protecting 248 memories from interference caused by new learning 46,52 . Within this framework, memories 249 must be released from inhibitory control to permit recall. While the precise mechanism may 250 vary across brain systems and circuits, our data suggest disinhibition in V1 can release 251 excitatory ensembles from balanced inhibition. Moreover, at the microcircuit level, 252 disinhibition during memory recall has previously been identified following fear promoting initial encoding of memory 11,45 , disinhibition may play a significant role in 255 facilitating release and recall of previously learned but latent cortical associations. sensitive methods can be employed to relate fMRS measures to physiological parameters. Here, 298 by implementing an inference task in VR, we operationalize memory recall using the exact 299 same paradigm previously employed in rodents 33 . Therefore, in addition to engaging attention 300 and memory-dependent inference, "opening the box" to find a reward in the VR environment 301 approximated the process of rodents finding a reward from a dispenser in a 3D environment. 302 By using VR, the findings presented here may be compared to data acquired in animal models 303 in ongoing future research. In this manner, VR paradigms in humans may provide a basis from 304 which to gain insight into the cellular and circuit mechanisms that underlie macroscopic studies, together with our data, these suggest fMRS is highly sensitive to detecting task-318 relevant dynamic changes in glutamate and GABA 71 . For example, in the lateral occipital 319 complex fMRS demonstrates differences in glutamate in response to presentation of objects 320 versus abstract stimuli 57 , and in the left anterior insula fMRS reveals a transient increase in 321 glutamate with exposure to painful stimuli 40 . fMRS-derived glutamate is even sufficiently 322 sensitive to detect repetition suppression effects in the lateral occipital complex 43 , mirroring 323 analogous effects reported in fMRI 72 . Here, we further illustrate that within a 3 second window 324 delineated by the question period in the inference task, the temporal resolution of fMRS is 325 sufficient to relate transient changes in glutamate and GABA to memory performance. fMRS 326 therefore provides a promising tool to capture real-time, task-relevant changes in 327 neurometabolites, on a time scale equivalent to task-based fMRI. Assessing whether the 328 temporal resolution of fMRS can be further improved will likely prove an important step in 329 refining fMRS in the future.

412
Notably, there was no significant effect of sex on behavioural performance (Supplementary Table 1

419
Trials were categorised as 'remembered' if participants correctly inferred the appropriate outcome during the inference test and 420 subsequently recalled the auditory-visual association in the post-scan memory test. Trials were categorised as 'forgotten' if 421 participants incorrectly inferred the appropriate outcome during the inference test or subsequently forgot the auditory-visual 422 association in the post-scan memory test. c During the question period in the inference test ( Fig. 1c-

477
The environment included a square-walled room with no roof (Fig. 1b). To help evoke the experience of 3D space

485
the VR environment participants were exposed to a range of different sensory stimuli, in accordance with the 486 three-stage inference task described below.

489
In the VR environment (Fig. 1b) humans performed an inference task (Fig. 1a). The rationale for using an 490 inference task to assess mechanisms responsible for associative memory was three-fold. First, evidence in both 491 humans and mice shows that performance on this inference task requires associative memory recall 33 . Second, in 492 mice, inference, but not first-order associative recall, is hippocampal dependent 33-35 , thus providing an opportunity 493 to investigate hippocampal dependent associative memory recall. Third, the task can be deployed across humans 494 and rodents, which may allow future investigation of the cellular mechanisms that underlie non-invasive measures 495 reported here.

497
The task was adapted from associative inference and sensory preconditioning tasks described elsewhere 33,83,84 and 498 involved 3 stages performed across 3 consecutive days, respectively (Fig. 1a,c). The first and second stages were 499 performed outside the scanner while the third stage was performed inside the scanner (Fig. 1c). At the start of the 500 experiment the pairings between auditory, visual and outcome cues were randomly assigned for each participant.

502
On day 1, participants performed the 'observational learning' stage ( Fig. 1a), during which participants were 503 required to learn at least 40 (out of 80 total) auditory-visual associations via mere exposure. In total, there were 4 504 visual cues, each associated with 20 different auditory cues. Auditory cues constituted 80 different complex 505 sounds (e.g. natural sounds or those produced by musical instruments) that were played over headphones. Visual 506 cues constituted 4 different unique patterned panels which could appear on the walls of the environment (Fig   507   1a,b,e). To control for potential spatial confounds, two of the visual cues were always presented on the same wall, 508 the assignment of which was randomized for each participant. The two remaining visual cues were 'nomadic', 509 meaning that with each presentation they were randomly assigned to one of the four walls.

511
Training during the observational learning stage occurred within the VR environment and was divided into 8 sub-512 sessions. In each sub-session, participants controlled their movement within the VR environment and were 513 presented with 20 trials in which 10 different auditory-visual associations, different in each sub-session, were 514 each presented twice, in a random order. On each trial an auditory and visual cue were presented serially and 515 contiguously: 8 s auditory cue followed by 8 s of the associated visual cue, followed by an ITI of 5 s 516 ( Supplementary Fig. 1a). Participants were given the choice to repeat the sub-session if they so wished. After the observational learning test, 1 auditory cue from the sub-session was presented, followed by presentation of 4 520 different visual cues (Supplementary Fig. 1b). Participants were instructed to select the visual cue associated with 521 the auditory cue using a button press response within 3 s, and only at the end of the test were participants given 522 feedback on their average performance. Each auditory cue in the sub-session was presented 2 times. Participants 523 were required to repeat training in the VR environment (including the observational learning test) until they 524 obtained at least 50% accuracy for auditory-visual associations in the sub-session (chance level: 25%).

526
After obtaining at least 50% accuracy on the observational learning test for each sub-session, participants were 527 given an 'overview' memory test (Supplementary Fig. 1b). The memory test had the same format as the 528 observational learning test used for each sub-session, except that it included all 80 auditory cues, each of which 529 was presented 3 times. Training on the observational learning stage was terminated when participants reached 530 >50% accuracy on this 'overview' memory test (Supplementary Fig. 1e). If participants failed to reach >50% 531 accuracy, training in the VR environment was repeated for those sub-sessions with poor performance. Those 532 participants that failed to reach >50% accuracy on the 'overview' memory test (n=3) did not proceed to day 2 and 533 were thus not included in the experiment.

535
On day 2, participants performed the 'conditioning' stage (Fig. 1a), during which they learned that two of the four 536 visual cues (set 1) predicted delivery of a rewarding outcome (virtual silver coin, as above) on 80% of trials, while 537 the other two visual cues (set 2) predicted delivery of a neutral outcome (virtual wood-chip, as above) on 100%

545
Training during the conditioning stage occurred within the VR environment and on each trial, participants were 546 presented with a visual cue and outcome which were presented serially and contiguously: visual cue (8 s) followed 547 by outcome delivery to a wooden box (available for 6 s) (Supplementary Fig. 1c). Participants were instructed to 548 only look in the wooden box after the visual cue was presented and instructed to leave the wooden box before the 549 next trial. The inter-trial interval (ITI) was 2 s.

551
Learning during the VR conditioning training was monitored using a conditioning test coded in Matlab 2016b 552 using Psychtoolbox (version 3.0.13). On each trial of the conditioning test, participants were presented with a still 553 image of a visual cue before being asked to indicate the probability of reward using a number line (Supplementary 554 Fig. 1d). Participants were given 3 s to respond and were only given feedback on their average performance at the 555 end of the test. Participants were required to repeat the VR conditioning training and conditioning test until they 556 performed the test with 100% accuracy (Supplementary Fig. 1f).

558
Finally, on day 3, participants first repeated the conditioning test. Participants then entered the 7T MRI scanner 559 and performed the 'inference test' (Fig. 1a, c- 560 9a) (see fMRI-fMRS scan task below). Immediately after exiting the scanner, participants were given a surprise 561 associative memory test to assess which auditory-visual associations they remembered and which they had 562 forgotten (Fig. 1e). The memory test was equivalent to the test performed on day 1 during the observational 563 learning ( Supplementary Fig. 1b), with 3 trials for each auditory stimulus. Performance on auditory-visual 564 associations was categorised as correct if participants scored 3/3 for that auditory cue on the subsequent surprise 565 memory test. Performance on auditory-visual associations was categorised as incorrect if participants scored 0/3 566 or 1/3 for that auditory cue on the subsequent surprise memory test (i.e. no different from chance). Trials where 567 participants scored 2/3 were not categorised as either correct or incorrect due to their ambiguity. The behavioural 568 performance measured on the post-scan associative memory test (Fig. 1f)

577
The inference test was incorporated into the fMRI-fMRS scan task. This provided an opportunity to measure 578 neural responses to associative memory recall required for inferential judgements. The scan task included two different trial types: inference test trials (Fig. 1d) and conditioning trials (Supplementary Fig. 9a). For both types 580 of trial participants viewed a short video taken from the VR training environment. The videos were presented via 581 a computer monitor and projected onto a screen inside the scanner bore. On each trial the duration of the video 582 was determined using a truncated gamma distribution with mean of 7 s, minimum of 4 s and maximum of 14 s.

583
During the inference test trials, the video of the VR environment was accompanied by an auditory cue, played 584 over MR compatible headphones (S14 inset earphones, Sensimetrics). Visual cues were not displayed during these 585 trials: the auditory cues were presented in isolation. At the end of the video, participants were presented with a 586 question asking: 'Would you like to look in the box?', with the options 'yes' or 'no' (Fig. 1d). Participants were 587 required to make a response within 3 s using an MR compatible button box and their right index or middle fingers.

588
No feedback was given. To infer the appropriate outcome participants were instructed to use the learned structure 589 of the task. The inference test thus provided an opportunity to investigate memory recall: to infer the correct 590 choice participants needed to recall the appropriate visual cue associated with the auditory cue (Fig. 1g).

591
Conditioning trials were interleaved with inference test trials to minimise extinction effects. During conditioning 592 trials, the video of the VR environment orientated towards a visual stimulus displayed on one of the four walls 593 ( Supplementary Fig. 9a). At the end of the video, participants were presented with a still image of the associated 594 outcome for that visual cue (Supplementary Fig. 9a). After each trial (inference or conditioning) a cross was 595 presented in the centre of the screen during an inter-trial interval of varying length, determined using a truncated 596 gamma distribution (mean of 2.7 s, minimum of 1.4 s, maximum of 10 s).

598
To control for potential confounding effects of space, each video during the inference test involved a trajectory 599 constrained to a 1/16 quadrant of the VR environment, evenly distributed across the different auditory cues. Across 600 conditioning trials, each visual cue was presented 16 times, once in each possible spatial quadrant. The fMRI-601 fMRS scan task was evenly divided across 2 scan blocks, each of which lasted 15 minutes. The fMRI-fMRS scan 602 task was then repeated (2 more scan blocks) using a higher quality multiband fMRI sequence (not reported here). images was also acquired (TE1 = 4.08 ms, TE2 = 5.1 ms, whole-brain coverage, voxel size 2 × 2 × 2 mm 3 ).

628
In addition to the fMRI-fMRS sequence acquisition, an additional set of fMRI data (reported elsewhere 33 and not 629 shown here) was acquired using a multiband EPI sequence (50 1.5 mm thick transverse slices with 1.5 mm gap, 630 in-plane resolution of 1.5 × 1.5 mm 2 , TR=1.512 s, TE= 20 ms, flip angle = 85°, field of view 192 mm, and multi-631 band acceleration factor of 2). To increase SNR in brain regions for which we had prior hypotheses, both the 632 fMRI sequences were restricted to partial brain coverage (Fig. 2a,

653
Spectra were then analysed in an event-related manner. For each participant, the preprocessed spectra were first 654 assigned to the tone/question/ITI periods by aligning the time stamps for the spectra to the time stamps for each 655 event recorded during the inference task. Then, spectra acquired within the tone/question/ITI periods were 656 selected for analysis. Next, these selected spectra were separated into two categories according to task 657 performance, 'remembered' or 'forgotten' (Fig. 2b, see Trial categorisation during the inference test), before 658 being analysed using LCModel. Participants (n=1) with less than 8 spectra for either the 'remembered' or 659 'forgotten' conditions were excluded from the fMRS analysis, as previous studies report minimal change in test-

689
Further, to control for differences in the number of 'remembered' and 'forgotten' spectra, we compared the group number generator. The relative metabolite concentrations for each condition were then estimated in LCModel and 694 the difference between conditions computed. The group mean for each permutation was then added to the null 695 distribution. The difference between 'remembered' and 'forgotten' conditions derived from the unshuffled data 696 was then compared against the null distribution generated from the shuffled data ( Fig. 3d-  analyses as the quality of fMRI data in the fMRI-fMRS sequence was too poor to ensure reliable pre-processing.

716
For the first-level analyses, three different GLMs were used.

730
In the second and third GLMs, the same EVs were included, however the first 8 EVs accounted for the auditory 731 cue period in the inference test (second GLM), or the inter-trial interval in the inference test (third GLM). In both 732 cases, the EVs were divided according to performance of the subject ('remembered' or 'forgotten'), as in the first 733 GLM.

735
Univariate fMRI analysis and statistics

736
Using the output of the GLMs we assessed the difference in the univariate BOLD response between 'remembered' 737 and 'forgotten' trials during the inference test (as defined in Fig. 2b

739
('forgotten'), using the first GLM (see above). The resulting contrast images ('remembered'-'forgotten') for all 740 participants were entered into a second-level random effects 'group' analysis. We set the cluster-defining 741 threshold to p<0.01 uncorrected before using whole-brain family wise error (FWE) to correct for multiple 742 comparisons, with the significance level defined as p<0.05 (Fig. 2c, Supplementary Table 3).

744
Assessing the relationship between fMRI and fMRS

745
To assess the relationship between event-related hippocampal BOLD signal and event-related fMRS measures 746 from V1, we used an anatomical ROI for the hippocampus (Fig. 4a). Capitalising on variance across participants, 747 the relationship between the BOLD signal for 'remembered'-'forgotten' within this ROI was compared with 748 equivalent changes in glutamate, GABA and glu/GABA ratio using a Spearman rank correlation. To assess the 749 selectivity of these effects to the recall period (question) during the inference test, control analyses were performed Next, to assess the relationship between fMRS and the BOLD signal across the entire imaged brain volume, we 754 repeated the second-level random effects 'group' analysis using the output of the first GLM, but now included 755 group-level covariates for the change in glutamate and GABA for 'remembered'-'forgotten' (i.e. Fig. 3a), along 756 with 2 'nuisance' regressors that accounted for unwanted variance attributed to differences in age and sex. To 757 identify brain regions where the BOLD signal for 'remembered':'forgotten' predicted changes in glu/GABA ratio,

758
we contrasted the explanatory variables on the covariates for glutamate and GABA (glutamate -GABA) to 759 generate a single contrast to test statistical significance. We set the cluster-defining threshold to p<0.01 760 uncorrected before using whole-brain family wise error (FWE) to correct for multiple comparisons, with the 761 significance level defined as p<0.05 (Fig. 4c, Supplementary Table 6).

763
To visualize the time course of fMRI and fMRS across the inference test trials, we estimated a moving average

769
'remembered' and 'forgotten' spectrum were calculated for each time bin, and the ratio estimated to give a 770 measure of 'remembered':'forgotten' for both glutamate and GABA (Fig. 3c, Fig. 4d, Supplementary Fig. 10c-

773
For the fMRI, for each participant, and for each time bin during the inference test trial, the time course of the 774 preprocessed BOLD signal was extracted from the hippocampal ROI (Fig. 4a) and from two control ROIs defined 775 using a 12 mm sphere within our partial epi volume (Fig. 2a). The first control region was positioned at the 776 junction between parietal and occipital cortex ('parietal-occipital cortex') while the second control region was 777 positioned within the brainstem (Supplementary Fig. 10c                         in the spectrum 91 . This phenomenon is most discernible on the strongest singlets such as total creatine (tCr). To assess the 1162 reliability of our fMRS measures we therefore quantified the difference in line width between our conditions of interest. These