A neurofeedback paradigm to promote SWRs.

(A) Schematic of the neurofeedback (NF) protocol: while subject’s nose remains in the neurofeedback port, SWRs (yellow) are detected in real time. Top trace: raw CA1 LFP; bottom trace: ripple filtered (150-250 Hz) CA1 LFP. During the neurofeedback interval, the first event exceeding a set standard deviation (sd) threshold (red line) triggers the delivery of a sound cue and food reward to the rat.

(B) Top-down view of the maze environment. Reward ports are indicated by rectangles.

(C) The rules of the spatial memory task for each of the behavioral cohorts.

(D) Experimental timeline for each behavioral cohort.

Neurofeedback training enhances SWR rate during targeted interval.

(A) Example CA1 raw LFP traces and ripple filtered LFP traces (150-250Hz) with SWRs highlighted in yellow, from the time spent at a center port on a neurofeedback (NF) trial (top) and a delay trial (middle) from a manipulation subject and a trial from a control subject (bottom).

(B) SWR rate calculated in 0.5 s bins during the pre- and post-reward periods, aligned to the time of reward delivery (dashed lines) for subjects in the manipulation (top row) and control (bottom row) cohorts. Trigger SWRs on NF trials are excluded from the rate calculation, and time bins with fewer than 100 trials contributing data are not shown. Vertical bars indicate S.E.M.

(C) SWR rate calculated during the pre-reward period (left), post-reward period (middle) and for the total time at the center port (right). Manipulation cohort n = 1892, 684, 1157, and 1602 NF trials and 2022, 640, 1201, and 1552 delay trials; control cohort n = 2490, 2629, 2027, and 3021 trials. For the pre-reward period, manipulation cohort ranksum comparisons between NF and delay trials: p = 4.382x10−258, 7.111x10−83, 5.689x10−214, and 3.285x10−191. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 1.126x10−16; manipulation cohort delay trials vs control cohort trials: p = 0.009. For the post-reward period, manipulation cohort ranksum comparisons between NF and delay trials: p = 3.646x10−127, 0.038, 6.538x10−11, and 2.768x10−23, respectively. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.142; manipulation cohort delay trials vs control cohort trials p = 0.691. For pre+post combined, manipulation cohort ranksum comparisons between NF and delay trials p = 3.324x10−64, 3.136x10−40, 2.996x10−69, and 9.066x10−50, respectively. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.237; manipulation cohort delay trials vs control cohort trials: p = 0.504.

(D) Count of SWR events detected during the pre-reward period (left), post-reward period (middle) and for the total time at the center ports (right). Trial n are the same as in (C). For the pre-reward period, manipulation cohort ranksum comparisons between NF and delay trials: p = 3.133x10−96, 8.936x10−37, 1.266x10−106, and 6.024x10−65. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 6.790x10−7; manipulation cohort delay trials vs control cohort trials: p = 0.0018. For the post-reward period, manipulation cohort ranksum comparisons between NF and delay trials: p = 1.797x10−137, 6.100x10−10, 6.688x10−13, and 2.152x10−130. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 2.361x10−12; manipulation cohort delay trials vs control cohort trials p = 5.820x10−9. For pre+post combined, manipulation cohort ranksum comparisons between NF and delay trials: p = 1.534x10−19, 5.828x10−14, 6.735x10−32, and 0.083. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials p = 0.366; manipulation cohort delay trials vs control cohort trials: p = 0.299. (E), Dwell time post-reward. Trial n are the same as in (C). Manipulation cohort ranksum comparisons between NF and delay trials: p = 1.397x10−23, 8.588x10−49, 6.180x10−4, and 7.127x10−257. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 4.075x10−18; manipulation cohort delay trials vs control cohort trials: p = 1.835x10−18.

For C-E, all within-subject ranksum p-values are corrected using the Benjamini-Hochberg method and all groupwise comparisons are performed using linear mixed effects models (see Methods).

Speed does not account for differences in SWR rate.

(A) Mean head velocity (smoothed), during pre-reward time at the center ports. Trial n are the same as in (E); manipulation cohort ranksum comparisons between NF and delay trials: p = 2.447x10−223, 2.760x10−85, 3.247x10−80, and 9.077x10−36. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 1.260x10−7; manipulation cohort delay trials vs control cohort trials: p = 1.121x10−4.

(B) Mean head velocity for subsets of trials: the quartile of trials with the lowest mean head velocities are shown for the control cohort and for delay trials for the manipulation cohort. The quartile of trials with the highest mean head velocity are shown for neurofeedback trials. Manipulation cohort n = 473, 171, 289, and 400 NF trials and 505, 160, 300, and 388 delay trials; control cohort n = 624, 794, 423, and 913 trials. Manipulation cohort ranksum comparisons between NF and delay trials: p = 1.342x10−160, 1.053x10−55, 9.488x10−98, and 4.751x10−130. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.989; manipulation cohort delay trials vs control cohort trials: p = 4.196x10−4.

(C) SWR rate for the pre-reward period for the subset of trials included in G, showing that even in neurofeedback trials with equal or higher velocities than delay or control trials, the SWR rate remains significantly elevated. Trial n are the same as in (G); manipulation cohort ranksum comparisons between NF and delay trials: p = 3.923x10−11, 1.155x10−8, 1.902x10−18, and 5.081x10−6. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 1.019x10−6; manipulation cohort delay trials vs control cohort trials: p = 7.958x10−3.

Neurofeedback preserves replay content.

(A) The 2D maze is linearized to 1D for decoding efficiency; movement trajectory for an example neurofeedback (NF) trial is shown in green. Times of SWR events are highlighted in yellow, and the small letters indicate the SWRs that are shown in (C).

(B) Decoding during movement times shows a close correspondence between decoded position and the subject’s real position.

(C) Several examples of decoding during SWRs, drawn from the example trial above. These include local events (two left), remote events (four middle), and one uninterpretable event (right).

(D) The fraction of all SWRs per behavioral epoch that contain interpretable spatial content during pre-reward period at center ports. Manipulation cohort n = 37, 14, 18, and 26 behavioral epochs per subject for each trial type and control cohort n = 24, 32, 23, and 33 behavioral epochs. Manipulation cohort ranksum comparisons between NF and delay trials: p = 0.1813, 0.5978, 0.0273, and 0.6738. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.5523; manipulation cohort delay trials vs control cohort trials: p = 0.2098.

(E) Rate of remote replay events during pre-reward period. Manipulation cohort n = 1843, 558, 1011, and 1513 NF trials and 1982, 535, 1038, and 1447 delay trials; control cohort n = 2058, 2509, 1879, and 2795 trials per subject, respectively. Manipulation cohort ranksum comparisons between NF and delay trials: p = 6.507x10−126, 9.703x10−30, 2.991x10−79, and 6.293x10−55, respectively. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 9.073x10−4; manipulation cohort delay trials vs control cohort trials: p = 0.4304.

(F) Rate of remote replay events during the pre- and post-reward periods combined. Trial n are the same as in (E). Manipulation cohort ranksum comparisons between NF and delay trials: p = 7.666x10−47, 1.573x10−6, 1.044x10−16, and 5.768x10−6. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.6644; manipulation cohort delay trials vs control cohort trials: p = 0.1477.

(G) Generalized linear model coefficients quantify the extent to while replay of an arm is modulated by its behavioral relevance. Manipulation cohort n = 1661, 392, 866, and 1281 NF trials and 1705, 367, 894, and 1213 delay trials per subject; control cohort n = 1458, 1636, 1464, and 2181 trials, respectively.

For panels D-F, all ranksum p-values are corrected using the Benjamini-Hochberg method and all groupwise comparisons are performed using linear mixed effects models (see Methods).

Neurofeedback does not alter behavioral performance.

(A) During trials when the subject is searching for a new goal location, we quantify the fraction of trials in which the subject chooses an arm that has not yet been sampled. Manipulation cohort ranksum comparisons between neurofeedback (NF) and delay trials: p = 0.9562, 0.7509, 0.7509, and 0.9562, respectively. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.5425; manipulation cohort delay trials vs control cohort trials: p = 0.0069.

(B) The fraction of redundant search trials per epoch which are NF (vs delay). Two-sided sign test vs 0.5: p = 0.1102, 0.2668, 1, and 1, respectively.

(C) During trials when the subject has discovered the goal arm, the fraction of subsequent trials in which the goal arm is visited (“correct” choice). Manipulation cohort ranksum comparisons between NF and delay trials: p = 0.7202, 0.7202, 0.8123, and 0.7202. Inset: Groupwise comparisons. Manipulation cohort NF trials vs control cohort trials: p = 0.6946; manipulation cohort delay trials vs control cohort trials: p = 0.2857.

(D) The fraction of error repeat trials which are NF trials. Two-sided sign test: p = 0.3770, 0.7744, 1, and 0.0227, respectively.

For all panels, manipulation cohort n = 37, 14, 18, and 26 behavioral epochs per subject, respectively, and control cohort n = 24, 32, 23, and 33 behavioral epochs. For (A) and (C), all ranksum p-values are corrected using the Benjamini-Hochberg method and all groupwise comparisons are performed using linear mixed effects models.