Highly task-specific and distributed neural connectivity in working memory revealed by single-trial decoding in mice and humans

Working memory (WM), the capacity to briefly and intentionally maintain mental items, is key to successful goal-directed behaviour and impaired in a range of psychiatric disorders. To date, several brain regions, connections, and types of neural activity have been correlatively associated with WM performance. However, no unifying framework to integrate these findings exits, as the degree of their species- and task-specificity remains unclear. Here, we investigate WM correlates in three task paradigms each in mice and humans, with simultaneous multi-site electrophysiological recordings. We developed a machine learning-based approach to decode WM-mediated choices in individual trials across subjects from hundreds of electrophysiological measures of neural connectivity with up to 90% prediction accuracy. Relying on predictive power as indicator of correlates of psychological functions, we unveiled a large number of task phase-specific WM-related connectivity from analysis of predictor weights in an unbiased manner. Only a few common connectivity patterns emerged across tasks. In rodents, these were thalamus-prefrontal cortex delta- and beta-frequency connectivity during memory encoding and maintenance, respectively, and hippocampal-prefrontal delta- and theta-range coupling during retrieval, in rodents. In humans, task-independent WM correlates were exclusively in the gamma-band. Mostly, however, the predictive activity patterns were unexpectedly specific to each task and always widely distributed across brain regions. Our results suggest that individual tasks cannot be used to uncover generic physiological correlates of the psychological construct termed WM and call for a new conceptualization of this cognitive domain in translational psychiatry.


Introduction
and region, they were always significantly better than those of control classifiers trained with 222 shuffled labels (P < 10 -17 , t-tests) which, in turn, decoded indistinguishable from the 50% 223 chance level on average (Fig. 3b). To evaluate the generality of the obtained classifiers, we 224 assessed if they could also decode WM-based choices in data from other 5-CSWM 225 challenge protocols. Even though the resulting decoding accuracies were generally lower 226 compared to those achieved with data from the same protocol, they were still significantly 227 higher than those of classifiers trained with shuffled labels (Supplementary Fig. 5). These 228 analyses reveal that WM-based choice is encoded in LFP-based connectivity and activity 229 measures in individual trials and that such information is widely distributed across multiple 230 brain regions. or areas (grey), respectively to predict WM-based correct vs. incorrect choices in the DMTS 12 5-CSWM task (combined 1 s SP-SD, 5+2 s delay challenge, 2 sessions; red, b), the DNMTS 237 2-CSWM task (baseline, 2 sessions, green, b), the T-maze WM task with either 5s (solid 238 blue, 4 sessions, c) or 30s (dashed blue, 4 sessions, c). Thinner dotted lines show decoding 239 accuracies of corresponding classifiers trained with shuffled labels, remaining at chance 240 level (50%, orange). Classifiers trained with real labels perform better than those trained with 241 shuffled labels in all cases (P < 10 -17 , t-tests, not indicated). The accuracy of 80% is coloured 242 in purple to aid comparison. Numbers in coloured ovals indicate the rank of the prediction 243 accuracies achieved on average by using data from the respective connection or region. 244 Ranks have been generated from pairwise comparisons with Tukey post-doc tests 245 conducted after significant effects of connection/region in one-way ANOVAs (P < 0.0001 in 246 all cases); connections/regions that were not significantly different from each other were 247 assigned the same rank. Black stars in (b) indicate differences of accuracy values achieved 248 in the two operant tasks (Tukey post-hoc tests after significant effect of task-type in ANOVA 249 across all three tasks). Blue stars in (c) indicate pairwise differences between the two delays 250 (uncorrected t-tests). ** P < 0.01; *** P ≤ 0.001. Shaded regions around mean show s.e.m. 251 across 100 classifiers generated for each task and connection or region. 252 253 WM tasks 254 To investigate if this conclusion applies generally to rodent WM, we repeated the same 255 analysis for the operant DNMTS data (final baseline sessions, 2 s delay). In this case, 256 however, we obtained the maximum average prediction accuracy (86.1 %) from local dHC 257 activity, rather than PFC-MD (66.1%, lowest rank of all classifiers) or PFC-dHC (73.5%) 258

Specific connections and regions are engaged differently in distinct rodent
connections. Generally, in this task, local activities allowed relatively high decoding 259 accuracies (72-77% for PFC, MD, and vHC), while coupling metrics were significantly less 260 predictive (66-68%, P < 0.001, Tukey; except for dHC-PFC, Fig. 3b). 261 In reverse, trial-by-trial decoding of T-maze data achieved the highest average accuracies 262 (82-88%) when using connectivity data from either one of the four connections (with dHC-263 13 connections being most predictive), whereas local activities were significantly less predictive 264 (62-79%; P < 0.001, Tukey, Fig. 3c). However, decoding accuracies for information from all 265 4 connections decreased when analysing data from the 30 s delay challenge, in which these 266 animals also showed lower behavioural performance (Fig. 1h, 3c), suggesting that not only 267 task type but also task difficulty affect the information encoded in each connection. 268 Specific connections and regions are engaged differently in distinct phases of 269 the rodent DMTS 5-CSWM task 270 The prior analyses entail at least two conclusions: Firstly, WM-related information in a single 271 trial is not encoded in any single region or connection, although some of them bear higher 272 predictive power regarding WM-choice than others. Secondly, the predictive power of a 273 given region or connection is not uniform but strongly depends on the type of WM task, 274 indicating that different mechanisms and regions are engaged to solve distinct behavioural 275 demands. These conclusions re-emphasize the question as to what extent oscillatory 276 processes in distinct frequency bands, of a distinct biological type, or in a specific task phase 277 (encoding, delay, choice) can be regarded as correlates of WM (Supplementary Table 1). 278 To answer this question, we took advantage of the fact that a linear classifier reveals the 279 predictive power of each involved predictor variable according to its assigned weight. We 280 performed Bonferroni-adjusted t-tests comparing the weights for each connectivity variable 281 with the weights assigned by the classifiers trained on label-shuffled control data, and, 282 additionally, conducted t-tests comparing the amplitudes of each variable between correct 283 and incorrect trials that contributed to the classifiers. Variables for which both t-tests were 284 significant were considered as bearing WM-related information (indicated by colour in Fig.  285 4a). This analysis revealed a relatively small set of consistent WM-related feature-classes as 286 correlates of DMTS 5-CSWM, a majority of them in the -range ( representation of anticipated reward. Therefore, we replicated the decoding analysis for SP 320 choices for which animals also expect reward. For the SP, however, average decoding 321 accuracies werealthough still above the 50% chance level -considerably smaller, namely 322 64-68 % and 54-59 % for predictions based on connectivity and local activity, respectively 323 ( Supplementary Fig. 6a). While this result shows that the attentional element of the task is 324 more difficult to predict from the available parameters than WM choice, it also demonstrates 325 that the obtained CP prediction accuracy was not simply based on representations of motor-326 action (hole-poking), attention, or reward anticipation. We also repeated the decoding 327 analysis for CP choice with complete omission of all CP parameters. Even though average 328 decoding accuracies decreased significantly for some connections, including the most 329 predictive ones (PFC-dHC, 72%, PFC-MD, 73.9%) -but not for vHC-dHC (72%) -overall 330 accuracies remained far above those obtained from classifiers trained on shuffled labels, 331 and hence above chance level (P < 10 -30 and P < 0.002 for classifiers trained on connectivity 332 or local data, respectively; Supplementary Fig. 6b). Overall, these analyses demonstrate that 333 activities along distinct connections and in distinct frequency bands represent encoding (SP), 334 maintenance (delay), and recall (CP) of WM contents in the 5-CSWM task. 335

336
To investigate if such phase-specific connectivity generalizes across tasks, we performed 337 the same analysis for the classifiers predicting performance in the operant DNMTS and the 338 T-maze assays. In both cases, considerably more parameters carried WM-related 339 information than in the 5-CSWM task ( Fig. 4a-b). Compared to DMTS WM, T-maze 340 rewarded alternation choice was predicted by a much larger number of predictors, with a 341 prominence of -and -range (as opposed to -range) variables, and a considerable 342 proportion of SP-parameters ( Fig. 4a-b). This pattern is reflective of the diversity of 343 connectivity parameters that have been associated with T-maze performance in prior studies 344 (Supplementary Table 1), and the relatively detrimental effects of optogenetic manipulations 345 in the SP and delay, as compared to the CP 33,34,46 . 346 Most astonishingly, the combined analysis of all three tasks revealed that none of the 347 specific connectivity parameters identified in one task bore significant predictive power in the 348 other two tasks, revealing a remarkable task-specificity of such parameters (Fig. 4a). It is, 349 however, possible that this finding is simply caused by a very conservative Bonferroni-350 adjustment of the P-value used as significance threshold (0.05/number of all connectivity 351 and activity variables combined; 0.05/1184 for the 5-CSWM, 0.05/888 for the T-maze and 2-352 CSWM). To test this possibility, we repeated the above analysis while relaxing this 353 adjustment incrementally over four orders of magnitude ( Fig. 4e-f). However, the number of 354 identified significant parameters, the relative contribution of different frequency bands, and 355 especially the extreme sparseness of overlap between task-specific predictors changed 356 relatively little ( Fig. 4e-f, Supplementary Fig. 7-8). This analysis also revealed that far more 357 connectivity parameters are identified according to their prediction weight than according to 358 their amplitude difference between correct and incorrect trials ( Fig. 4e-f). This suggests that 359 the classical approach of correlating behavioural performance with the amplitude of a given 360 metric (Supplementary Table 1) likely misses a sizable proportion of WM-related functional 361 connectivity. 362 To investigate potential differences or similarities between the time-course of individual 363 parameters during the task, we extracted those spectral connectivity parameters that were 364 predictive across all three assays albeit in different phases: directed dHCPFC -365 connectivity and intra-hippocampal -range coupling (Fig. 4a). Inspection of the time-course 366 of these parameters over the delay and CP revealed that they behaved rather differently in 367 the individual tasks: dHCPFC -connectivity showed a transient increase during the delay 368 of all three tasks, but only in the operant tasks a second increase occurred immediately after 369 correct choices (but not after incorrect choices; Fig. 5a-b). Intra-hippocampal -coupling 370 even showed a different time course in every task, including a correct choice-specific 371 decrease in the 2-CSWM delay which contrasted sharply with a steady rise during the T-372 maze delay ( Fig. 5a-b). Thus, even within the few predictor variables that are relevant across 373 all tasks, the actual physiological activity relating to the behaviour differed markedly. 374 Given these results, we directly tested the hypothesis that distinct activity patterns underlie 375 the different rodent WM tasks by rendering task-type a dependent variable: we trained 376 classifiers to decode which one of the three tasks a subject is currently conducting using 377 18 connectivity or local activity parameters from correct trials as input. Based on connectivity 378 data, task-type could be decoded with average accuracies of 97-99% when discriminating 379 between the two operant tasks (50% chance level) and with an accuracy >90-95% when 380 discriminating between all three tasks simultaneously (33.3% chance level; Supplementary 381 Common WM-related connectivity patterns shared across rodent tasks 405 To extract commonalities of connectivity between the three tasks, we aggregated predictive 406 non-directed (coherence, wPLI) and directed (GC, PDC) metrics (extracted from Fig. 4a) and 407 depicted their amplitude increases relative to the preceding ITI for each task phase (Fig. 5c-408 e). For the T-maze (the only task for which prior reference data exists), this revealed several 409 connectivity patterns associated before with rewarded alternation performance, including 410 vHC-PFC 33 and dHC-PFC 34 coupling during encoding and MDPFC -range activity during 411 maintenance across the delay 14,46 . Importantly, MD-PFC -range delay activity was also 412 20 seen in the other two WM tasks, although their directionality differed (MDPFC in the DMTS 413 task; PFCMD in the DNMTS task; Fig. 5c-e). Likewise, further task-independent 414 connectivity patterns emerged in this analysis: prominent vHC-PFC /coupling and vHC-415 dHC coupling in the CP, and MD-PFC  coupling in the SP (but with task-dependent 416 directionality; Fig. 5c-e). Some further patterns were shared only by the two DNMTS tasks, 417 e.g., directed PFC-dHC -and -connectivities in the SP and delay (Fig. 5d-e). At the same 418 time, this analysis also confirmed that the vast majority of WM-related connectivity was task-419 specific, especially when comparing the 5-CSWM DMTS to the other two tasks (Fig. 5c-e). 420 An important aspect of this analytical approach is that none of these individually highlighted 421 connectivity measures (Fig. 4a, 5c-e) is particularly predictive on its own: When performing 422 decoding analysis with reduced sets of predictor variablesstarting with the parameter with 423 the single highest weight and adding variables incrementallythe inclusion of several dozen 424 predictor variables was necessary to achieve maximum decoding accuracy (Fig. 5f). 425 Predictive power of local activity in a single area varies by task phase and type 426 Local oscillatory activity in the four analysed regions also allowed considerable prediction 427 accuracy in all three tasks -sometimes even exceeding that obtained from connectivity 428 metrics (Fig. 3b-c). Therefore, to reveal WM-related local activity metrics, we repeated the 429 prior weight-based analysis for the respective variables (power, local PAC). In the 5-CSWM 430 DMTS task only CP parameters, mostly in the /-range, were significantly associated with 431 WM (Fig. 6a). For the two other tasks, in contrast, significant predictors came from all three 432 phases and were somewhat less frequency-specific; the power of dHC-oscillations across all 433 frequency bands and phases constituted the most prominent cluster of choice-predictors in 434 both assays (Fig. 6a). In agreement with the high decoding accuracy obtained with local 435 activity (as opposed to connectivity) in the DNMTS 2-CSWM (Fig. 3b), many more significant 436 local predictor variables were found for this task compared to the other two, irrespective of 437 P-value threshold (Fig. 6b). Importantly, however, there was again hardly any overlap 438 between significant predictors from the three tasks. Also, while in DMTS WM all predictive 439 21 activity parameters had higher amplitudes in correct trials compared to incorrect trials, this 440 was not the case for the T-maze, where virtually all predictive hippocampal activity was 441 lower in correct trials compared to incorrect trialsonly PFC and MD power were higher in 442 correct trials (Fig. 6c-d). Hence, as observed in inter-regional connectivity, local activities 443 related to WM-choice were highly task-specific in multiple respects. during three types of WM assays whose trials were intermixed within a single test session: 465 identity-related WM (differentiating between identical and novel shapes), spatial WM, and 466 temporal WM (remembering the temporal order of two stimuli; Fig. 7a-b) 20 . For each of the 467 three tasks, we applied the same ML-approach as in mice, generating classifiers that use 468 activity data from four phases (SP, pre-cue-and post-cue delay phases, CP; see task 469 schedule in Fig. 7a) from only a single connection or region at a time. 470 Average decoding accuracies for trial-by-trial prediction of WM-choices were higher than 471 those achieved in mice, ranging consistently between 87-90% for predictions based on 472 connectivity and between 72-82% for predictions based on local activity, whereas 473 "predictions" based on shuffled control data remained significantly lower (P < 10 -40 , t-tests) 474 and were not different from chance level (50%; Fig. 7c). We also trained classifiers on the 475 combined data from all three inter-regional connections and three regionseither separately 476 for each task-type or combining all types of trials indiscriminately. For task-specific 477 classifiers, average decoding accuracies reached 87.6%, 90.8%, and 89.8% for identity-478 related, spatial, and temporal WM, respectively, i.e., no higher than what could be achieved 479 by connectivity data from the single best connection in each task (Fig. 7d). However, 480 23 decoding accuracy dropped to 79.4% if task-paradigms were intermixed (Fig. 7e) suggesting 481 that functional connectivity is, at least partially, task-specific. Task-specific prediction 482 accuracies of up to 91% could also be obtained without including CP connectivity measures 483 ( Supplementary Fig. 10). Furthermore, in two cases, an average prediction accuracy of up to 484 81% could even be achieved if using connectivity data from only a single task phaseeither 485 the SP in temporal WM or the post-cue delay in spatial WM ( Fig. 7f; Supplementary Fig. 10). 10 for analysis but using only predictors from single task-phases. 516 WM-related functional connectivity is highly task-specific and broadly 517 distributed in humans 518 In order to identify possible correlates of WM in humans, we analysed the prediction weights 519 of the individual connectivity metrics similarly as for the mouse dataset, again extracting 520 WM-related metrics based on the two criteria of prediction weight and a different amplitude 521 of the metric in correct trials compared to incorrect trials (Bonferroni-adjusted t-tests). As in 522 mice, WM-related measures (185 out of 1344 connectivity predictor variables) were widely 523 distributed across connections, frequency bands, and metric types. When inspecting the 524 matrix of significant predictors more closely, some regularities emerged (Fig. 8a): First, WM-525 related activity was highly task-specific with 88% of significantly WM-related connectivity 526 metrics being relevant in only a single paradigm. Only a single metric was predictive in all 527 three paradigms -OFCPFC post-cue delay -PDC. The principle task-specificity was 528 maintained also with relaxed P-value thresholds ( Supplementary Fig. 11-12). Second, by far 529 the mostand the most common -predictors emerged in the -band, irrespective of 530 significance threshold (Fig. 8a-b, Supplementary Fig. 11, 13-14). The -band -in contrast to 531 mice -contributed almost no WM-related variables (only one each in spatial and temporal 532 WM, confined to the OFC-MTL connection). Also, the -band bore relatively few WM-related 533 connectivity parameters, and these were mostly relevant for spatial WM and to a lesser 534 extent for identity WM, but hardly for temporal WM. Furthermore, --PAC appeared rather 535 relevant (as found in the same data before 20 ) in all three types of tasks, especially identity-536 related WM. Third, changes of a metric relative to the ITI before each trial were rarely 537 predictive. Finally, despite the relatively high decoding accuracy achieved for temporal WM 538 ( Fig. 7c-d), the number of connectivity metrics related to this WM-type was considerably 539 smaller (18 out of 1344 measures) than for the other two (72 and 95) and there was hardly 540 any overlap between these metrics and those relevant for the other two WM-paradigms (only 541 3 each, mostly in the -band; Fig. 8a-b). In summary, the analysis in humans confirms the 542 high task-specificity, and broad anatomical and frequency-range distribution of WM-related 543 neural activity already seen in mice. showing all connectivity predictor variables that contributed to the classifiers shown in (8d) 547 and were significantly associated with WM-performance according to their weight and 548 27 differences between correct and incorrect CP (see Results) in the paradigms coded by 549 colour (see legend on the right). For , mean amplitude (m), peak amplitude (p), and 550 frequency of peak (f) are shown, while for all other variables only the mean amplitude is 551 used. The -band contributed three predictors each as this frequency was split into a high-552 and low--range in addition to using the whole range (30-100 Hz). See Supplementary Fig.  553 11 for the same analysis with relaxed P-value correction and Supplementary Fig. 13-14  Common -band connectivity across human WM tasks 568 To scrutinize this conclusion, we searched for commonalities between tasks by aggregating 569 predictive non-directed (coherence, wPLI) and directed (GC, PDC) metrics and the multiple 570 measures within the -and -bands (extracted from Fig. 8a), and depicted their amplitude 571 change relative to the preceding ITI for each task phase (Fig. 8e-g), as previously done for 572 the mouse dataset (Fig. 5c-e). In this analysis, OFC-PFC -coupling during encoding and 573 directed OFCPFC/MTL -connectivity during the post-cue delay emerged as common 574 patterns present in every task. There were also more commonalities between spatial and 575 28 identity WM, namely -coupling between all three regions that was elevated throughout 576 encoding and delay phases and then decreased below ITI-levels in the CP (Fig. 8e-f). 577 Strikingly in fact, all significant predictors from the -range in these tasks showed 578 elevated amplitudes during encoding and delay, but decreased amplitudes during the CP, 579 compared to their amplitude in the preceding ITI ( Fig. 8e-g). Only  and -PAC predictor 580 variables increased in amplitude during the CP in spatial WM (Fig. 8f). Finally, even with this 581 aggregated analysis, not a single connectivity pattern that was shared between any two 582 tasks emerged outside the -band in any task phase (Fig. 8e-g). This analysis suggests that 583 human WM generally relies on anatomically broad, task phase-specific modulation of -584 connectivity between several brain regions irrespective of task, while the engagement of 585 oscillatory coupling in other frequency bands is task-specific. 586 587

588
Here we demonstrate that WM-related choices can be predicted trial-by-trial in mice and 589 humans using linear decoding of high-dimensional arrays of LFP-based measures of inter-590 regional connectivity or local activity (the electome). The high decoding accuracies of around 591 90% (compared to a chance level of 50%) achieved in both species are remarkable 592 considering the spatially coarse nature of the extracted neural signal, the shortoften sub-593 seconddata traces used to calculate predictors, the intrinsic variability caused by merging 594 data from all analysed individuals with varying electrode placements, and the lack of precise 595 neuronal information as encoded in spike trains of individual neurons 10,21,33,41,47 . Using the 596 trial-by-trial predictive power of physiological activity as the indicator for its association with 597 WM 42 , this approach enabled a largely unbiased top-down analysis to reveal an 598 unexpectedly rich pattern of frequency-specific connectivity changes during individual 599 phases of distinct WM assays in mice and humans. 600

29
The comparative analysis of multiple WM assays -including those that allow control over 601 basic motivational and attentional parameters -provides a unique advantage in that 602 neurophysiological activity patterns which might be truly relevant to WM may be isolated. In 603 this way, we could reveal MD-PFC  and -range coupling during memory encoding and 604 maintenance, respectively, as well as vHC-PFC and vHC-dHC -coupling during retrieval 605 as common connectivity patterns across all three rodent tasks -although the vast majority of 606 connectivity proved to be task-specific (Fig. 5c-e). In humans, -band connectivity across all 607 analysed connections was commonly linked to WM-choice across tasks, while WM-related 608 connectivity in other frequency bands was mostly task-specific (Fig. 8). 609 Against a backdrop of widely varying assumptions about which kind of neural connectivity 610 underlies WM (see Introduction), our analysis was initially motivated by the possibility to 611 extract "true" anatomical and frequency-related WM-correlates using the predictor weights 612 generated by the linear classifiers that decode WM-based choices with high accuracy. Our 613 results, however, refute some implicit key assumptions of this endeavourand, by 614 extension, of many prior investigations of WM-correlates: First, there is no singleor small 615 set ofanatomical regions or connections, types of directional information transmission, or 616 frequency bands that can be regarded as a unique WM correlate. Indeed, previously 617 suggested "correlates", especially in the rodent literature (Supplementary Table 1), could  618 appear as such only because the sum total of connections and measures investigated in 619 each study was small (streetlight effect), as opposed to the 1184 and 1584 metrics analysed 620 here in mice and humans, respectively. In our study, virtually every analysed frequency 621 band, metric, connection and region bore some predictive power regarding WM-mediated 622 choice. 623 Second, there is no single behavioural WM-task that could be regarded as representative of 624 the generic psychological construct termed "working memory" in order to allow the 625 identification of the neurophysiological correlate of that construct. The latter is illustrated by 626 the enormous variability in the patterns of predictive connections and metrics across task-627 30 paradigms in both species. In other words, a physiological variable that correlates with 628 choice accuracy in the T-maze represents a neurophysiological correlate of T-maze 629 performance, but not necessarily of WM. The same principle applies to our cross-species 630 comparison, as the uncovered candidates for a generic (task-independent) WM correlate 631 originated from different frequency bands in humans () than in rodents (). A 632 translational implication of these findings is that it is likely impossible to define 633 neurophysiological underpinnings of "working memory" as a uniform psychological construct. 634 However, the existence of certain tasks that represent such a psychological function (i.e., 635 that engages a physiological mechanism that is central to all WM tasks across species) is an 636 implicit key assumption of the Research Domain Criteria (RDoC) approach, which envisions 637 to use those representative paradigms in search of WM-enhancing cellular and molecular 638 targets 48 . Our data suggest that the key target variable in the preclinical discovery of WM-639 enhancing compounds might be the appropriate regulation of -range connectivity (given its 640 importance for human WM) rather than behavioural performance in any particular rodent 641 task. 642 The analysis of the human dataset, in particular, paints a rather different picture of what a 643 correlate of WM could beat least when searched for in LFP-data. In all three tasks, 644 prediction accuracy calculated from connectivity (as opposed to local activity) metrics was 645 not only very high, but it was also roughly equal between the three analysed connections, 646 even though these are anatomically quite distinct. This was the case even with the limited 647 analysis incorporating only SP or post-cue delay connectivity as predictors of spatial WM. 648 These findings may be taken as an indication that WM-related information is extremely 649 broadly distributed, and -rather than specific activity located in a certain connection or 650 region -it is the ability to manipulate information flow across brain regions as such, that 651 determines task performance 49,50 . This model is in line with the ever growing list of brain 652 areas that are implicated in WM, including the superior frontal 51 , anterior cingulate 52,53 and 653 sensory cortex 54,55 , ventral tegmental area 28 , and the nuclei of the midline and anterior 654 31 thalamus 4,9,34 , and the concept that different areas may be involved depending on the 655 strategy used to solve the task 54 . An unexpectedly broad anatomical representation of 656 sensory and behavioural information across the brain has recently been uncovered by 657 decoding activity of individual neurons in multiple cortical areas 56,57 , and our decoding 658 analysis of LFP-based connectivity in cognition underscores this phenomenon. 659 In conclusion, our multi-area decoding approach and cross-task cross-species comparative 660 analysis revealed not only a rich functional connectivity supporting WM, exceeding the WM-661 associated connectivity described before. It also demonstrated an unexpected task-and 662 species-specificity of WM-related neural coupling that raises substantial caution regarding 663 the predictive translational value of each assay and demands to re-think our search for 664 physiological WM-correlates. 3.4 mm for single and 3.8-3.9 mm for dual electrodes below pia). In most mice, dual 683 electrodes were used for PFC and vHC, whereby the second electrode was placed about 684 0.5 mm higher than the stated distance from pia. Both hemispheres were implanted at 685 roughly equal proportion. Stainless steel screws (1.2 mm diameter, Precision Technologies, 686 UK) were implanted in the contralateral hemisphere ca. 1 mm from the midline above the 687 cerebellum (AP -5.5) for ground and above the anterior frontal cortex (AP +4.0) for additional 688 reference (used for the analysis in the open-field test, but not for the present analysis), and 689 where connected with a 120 µm PTFE-insulated stainless steel wire (Advent Research 690 Materials Ltd., UK; Fig. 1b). All electrode wires were connected to pins in a dual-row 6-pin or 691 8-pin connector (Mill-Max, UK). Electrode placements were determined post-mortem from 692 electrolytic lesions made under terminal ketamine/medetomidine anaesthesia followed by 693