Persistence of neuronal representations through time and damage in the hippocampus.

Title:

Persistence of neuronal representations through time and damag

e in the hippocampus.

Authors:

Walter G. Gonzalez, Hanwen Zhan

g, Anna Harutyunyan, and Carlos

Lois*

Affiliations:

Division of Biology and Biologi

cal Engineering, California Ins

titute of Technology, Pasadena, United States.

*Correspondence to: clois@caltech.edu.

Abstract:

Memories can persist for decades

but how they are stably encode

d in individual and groups of neurons is not

known. To investigate how a familiar environment is encoded in

CA1 neurons over time we implanted bilateral

microendoscopes in transgenic mice to image the activity of pyr

amidal neurons in the hippocampus over weeks. Most

of the neurons (90 %) are active every day, however, the respon

se of neurons to specific cues changes across days.

Approximately 40 % of place and time cells lose fields between

two days; however, on timescales longer than two

days the neuronal pattern changes at a rate of 1 % for each add

itional day. Despite continuous changes, field responses

are more resilient, with place/time cells recovering their fiel

ds after a 10-day period of no task or following CA1

damage. Recovery of these neuronal patterns is characterized by

transient changes in firing fields which ultimately

converge to the original representation. Unlike individual neurons, groups of neurons with inter and intrahemispheric

synchronous activity form stable place and time fields across d

ays. Neurons whose activity was synchronous with a

large group of neurons were more

likely to preserve their respo

nses to place or time across multiple days. These results

support the view that although task-relevant information stored

in individual neurons is relatively labile, it can persist

in networks of neurons with synchronized activity spanning both

hemispheres.

One Sentence Summary:

Neuronal representations in networks of neurons with synchroniz

ed activity are stable over

weeks, even after lack of

training or follo

wing damage.

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Main Text:

Memories are processed and stored by a complex network of neuro

ns across several circuits in the brain;

however, little is known about how

information is encoded and r

etained in these neurons for long periods of time.

Information could be stored at d

ifferent hierarchical levels, w

ithin individual neurons through modification of their

synapses, or distributed among many neurons in different brain areas including the hippocampus and cortex. The

hippocampus is known to play an essential role in the formation

of memories (

) and neurons in this brain area

show robust response to space (place cells), time (time cells),

or other task relevant cues (

–

). Many works have

studied how neuronal activity in the hippocampus changes during

learning (

), attention (

), and re-exposure

(

). However, what aspects of neuronal activity in the hippocampu

s persist during future visits to a familiar

environment, how is information encoded in group of neurons, an

d how lesions perturb the long-term maintenance of

these neuronal patterns remains poorly understood.

The question of how information is stably encoded in neurons in

the hippocampus remains a controversial

issue. Whereas extracellular r

ecordings show that place cells r

etain their fields from days to a weeks (

), calcium

imaging experiments show drastic changes of neuronal activity a

cross days (

–

). In freely moving mice place field

were also observed to be the sam

e across weeks bu

t head-fixed p

reparation show drastic changes between days (

–

). Considering that the hippocampus is necessary for the format

ion, but not for the long-term maintenance of

memories, it is possible that neuronal representations in the h

ippocampus may change over time as information is

transferred from other brain areas (

). Previous long-term imaging of neuronal activity have shown t

hat a large

number of pyramidal neurons in

CA1 are active during a 35 day p

eriod but only 31 % were active in each session and

only 2.8 ± 0.3% were active in all sessions (

). However, the inability to detect active neurons on consecuti

ve days

could be due to motion artifacts caused by the removal and reat

tachment of the microendoscope between days. In

addition, overexpression of GCaMP using AAVs can induce cell to

xicity or even death (

). To overcome these

potential limitations we built custom microendoscopes that were

chronically implanted and performed long-term

simultaneous bilateral imaging of hippocampal activity in freel

y moving Thy1-GCaMP6s mice

(Figure 1a, S1, and

see supplementary data)

(

–

). The combination of chronic implants, high sensitivity microe

ndoscope, improved

cel

l detection and registration w

ith the CNMFe software allowed

us to minimize motion artifacts and to increase the

reliability of long-term recordings

(Fig. 1b-c and S2)

(

). We imaged CA1 pyramidal n

euron activity for several

weeks through three situations, which we defined as follows: (i

) “learning”, during the initial 5 sessions of exposure

to a novel linear track with sugar water reward at the ends, (i

i) “re-exposure”, after a 10-day period during which the

animal was not exposed to the linear track, and (iii) “damage”,

following light-induced hippocampus lesion

(Fig. S1)

We observed robust single neuron activity across days for up to

8 months

(Fig. 1d-e and Video S1).

We did not notice

significant differences between hemispheres and unless stated otherwise, the values reported represent combined data

from both hemispheres. Across days, 88 ± 4 % of all neurons in

the imaging area were active in each session and 51

± 17 % were active every session (

Fig. 1f-g and S3

). Within a day, the vast majority of neurons (95 %) were activ

both while mice explored their home cage or while running in th

e linear track. However, in the same day between

environments (cage or linear tr

ack) or across days in the linea

r track, neurons displayed significant changes in their

firing rates (

Fig. S4

). Thus, minimizing motion artifacts using chronic implants and

improved signal extraction and

registration allowed us to observe that most CA1 neurons that a

re active one day are also active on subsequent days.

To acquire a comprehensive view of stability of neuronal repres

entations in CA1, we studied both cells that

were active in specific locations

of the maze when animals were

running (defined as “place cells”), and during periods

of immobility (defined as “time cells”) (

)

(Fig. S5-6)

. Previous work reported that place fields underwent drastic

changes across sessions, reaching

near random levels (85-75 % c

hange) on the following session (

). Under our

conditions, we observe that from one day to the next 54 ± 15 %

of neurons retained response to a field (defined here

as “recurrence”, n=56). Surprisingly, despite the initial abrup

t change from one day to the next (~46%), the drop in

field response for subsequent day

s was only an additional ~1 %

per day, reaching random levels after ~ 50 days (

Fig.

2a-c

). Lastly, these changes in neuronal representations decreased

as the animal became more familiar with the task

(after the “learning” phase (

Fig. S5

Repetitive exposure to the task could induce changes in neurona

l representation through continuous updates

in place and time fields due to m

inor changes in environment (i

.e. different personnel or odors in the room). To

investigate this possibility, we introduced a no-task period in

which trained animals were not exposed to the linear

track for 10 days. We then compared the changes in place and ti

me cells between animals not exposed to the track and

animals which were continuously exposed to the track. Following

re-exposure, place and time cells in sessions

separated by 11-14 days that included the 10 day period of no-t

ask changed their fields by a similar fraction as animals

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

continuously exposed to the task in sessions separated by 11 da

ys (

Fig. 2c-d and S7

). Thus, changes in place and

time cells happen independently of whether the animal is expose

d to the task.

Next, we analyzed whether the fields to which a neuron is respo

nsive changes between sessions. Response

fields were relatively similar acr

oss days (correlation of 0.7

± 0.3, n = 505, during the “trained” period between one

to 5 days apart) (

Fig. 2f-g

). Interestingly, in some sessions, we observed that 87 ± 8 % o

f cells changed the direction

of their fields by 180 º in the linear track, reversing the dir

ectional representation encoded in the previous session (

Fig.

). Rotations of fields happened simultaneously across hemispher

es and were accompanied by minimal changes in

animal behavior (

Fig. S8

). Similarities of fields across days were significantly lower

during periods of learning and

during the first two days following re-exposure to the task (

Fig. 2f-g).

During periods of learning or immediately

following re-exposure to an environment fields fluctuate, but o

nce the animal becomes familiar with the task, place

and time cells can ret

ain their fields even following a 10-day

period in which the animal is not exposed to the task.

To investigate whether CA1 representations are also resilient t

o brain damage we performed local lesions

induced by increasing the LED illumination power of the implant

ed microendoscopes (5 to 10-fold over the threshold

needed to visualize GCaMP) (

Fig. 3a

). High illumination intensity induces a local increase in the

tissue temperature

and affects neurons along a spectrum ranging from perturbing th

eir firing activity to triggering their death. One day

after light damage, we observed a

massive increase in synchroni

zed firing analogous to interictal discharges (

Fig. 3b

These abnormal bursts of activity recruited the majority of neu

rons in the field of view and were direction specific in

the linear track

(Fig. S9 and Video S2)

. During days with abnormal CA1 activity, the firing behavior o

f neurons

changed dramatically, and the

number of place and time cells in

creases significantly (

Fig. 3c and S9

). However,

during this period, place and tim

e field correlation across day

s decreased to near random levels

(Fig. 3d-f)

Interestingly, after 2 to 10 d

ays, abnormal activity ceased ove

rnight and in a similar fashion to what we observed in

the re-exposure experiments, a

fter recovery from damage, place

and time fields stabilized and a significant portion of

place and time cells were responsive to the same fields that t

hey had before the lesion (81 ± 11 % with correlation

above 0.7, compared to random p<10

-10

ranksum, n = 49 sessions)

(Fig. 3f)

It has been suggested that groups

of neurons with synchronized

activity form cell assemblies able to encode

learned representations for long periods of time. Cell assembl

ies encoding temporal and spatial cues have been

observed in the hippocampus (

–

). However, whether these assemblies develop during learning, h

ow many

neurons participate in them, for how many days they persist, an

d whether they can encode s

table information across

days is not known. We started analyzing the activity of pairs o

f neurons, the simplest level of synchronized groups of

cells. We observed that neurons in freely moving mice become mo

re synchronized with other neurons within and

across hemispheres as mice become familiar with the linear trac

k (

Fig 4a

). This increase in synchrony arises mainly

due to the activity of place and time cells. Synchronized pairs

of neurons also tended to make the same errors as the

animal performed the task, as illustrated in two scenarios. Fir

st, whenever one neuron in the pair failed to fire in its

field, 50 ± 30 % of the time the other neuron in the pair failed as well. Second, when the pair of neurons fired together,

their deviation from the field peak was highly correlated (0.71

± 0.14, p<10

-8

) (

Fi. 4b

). Across days, we observed that

the likelihood of a neuron maintaining its responsiveness to a

field on the following day was proportional to the degree

of synchrony it had with another neuron in a pair. Altogether,

these results support the notion that synchronized activity

does not occur simply by chance, and it may be responsible for

the stability of representations over time.

To explore this hypothesis, we analyzed correlations of neurona

l activity using graph theory to identify

whether neuron pairs belong to a larger cell assembly and wheth

er stable information could be encoded in these larger

neuronal networks (

). Network graphs show a clear behavior-dependent topology, evo

lving throughout periods of

learning and undergoing extensive reorganization upon transitio

n from exploring in the home cage to running in the

linear track

(Fig. 4c)

. Graphs revealed the presence of dense clusters comprised of n

eurons within and across

hemispheres whose neurons had preferences for specific behavior

(Fig. 4d and S10)

. Extraction of these clusters

using a Markov diffusion approach identified groups of neurons

(defined here as cell groups) encoding direction-

specific information about severa

l aspects of the task, includi

ng periods of running, immobility, drinking, and turning

(

Fig. 4d-f

) (

). Synchronized activity of the

cell groups were specific for t

he environment to which the animal was

exposed. Cell groups that had synchronous activity in the linea

r track became asynchronous in the home cage, and

vice versa (

Fig. S10

). Individual neurons developed their responsiveness to a field

within minutes of exposure to the

track. In contrast, cell groups gradually developed their field

s over two to four days during the learning phase.

Moreover, the task information encoded in these groups did not

degrade over time (up to 35 days), even after a 10-

day period of no task (

Fig. 4g-i

). Thus, using graph topology we demonstrate that cell groups i

n CA1 can encode

persistent representations of the task across weeks even if the

activity of individual neu

rons varies over time.

The analysis of correlated neuronal activity of CA1 neurons has

been used to decode the behavior of the

animal or the response of neurons to a field (

), however, it is not known whether this activity can be used t

predict whether a cell will be responsive to a field into the f

uture (across days). Using graphs, we observed that the

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

likelihood that a neuron would maintain its responsiveness to a

field over multiple days was proportional to the extent

of connectivity in the graph (

Fig. 4j

). To test the hypothesis that graphs can be used to predict th

e responsiveness of

a neuron to a field we analyzed their topology and trained a de

coder to determine whether a neuron would be a place

or time cell in the future. Using this approach, we show that t

his synchrony-stability relationship can efficiently decode

which neurons in a session will be responsive to a field and pr

edict whether a neuron will maintain its field N sessions

apart, even after a period of 10 to 20 days

(Fig. 4k, see methods)

. Furthermore, a decoder trained to identify place

and time cells from graph topology in one trained animal in one

session can effectively id

entify place and time cells

in other animals simply by analyzing their graph structure (

Fig S11

). Thus, the features of a neuron in a graph are

sufficient to decode signatures that are specific to time cells

, specific to place cells, or even specific for cells that are

neither time or place cells, and t

hese signatures are common be

tween animals.

Discussion

In contrast to previous studies on the stability of CA1 represe

ntations, we observed that the vast majority of

neurons are active on most days but their firing rate changes a

cross sessions and tasks (

). We observe high stability

when analyzing the recurrence of place and time cells such that

even after 35 days 40 % of neurons were responsive

to a field (

). These differences could be due to better registration across

days thanks to the use of newer software

(CNMFe), a more sensitive CMOS sensor in the microscope, or the

use of transgenic animals i

nstead of viral vectors.

Our results indicate that hippocampal representations change dr

astically from one day to the next, but much more

slowly thereafter (an additional ~1% change per day), In addit

ion we found that the representations in CA1 were able

to recover after an extended period (10 days) without performin

g the task, or even after abnormal activity induced by

local lesions. These manipulati

ons revealed a common feature of

information persistence. In both cases, fields undergo

transient drifts and fluctuation ultimately converging to a neu

ronal representation similar to that present before

perturbations. This finding provides strong experimental eviden

ce for the presence of attractor-like ensemble

dynamics as a mechanism by which the representation of an envir

onment is represented in the hippocampus (

These results suggest a model with two complementary features. First, neuronal representations spontaneously change

over time, such that cells whose fields persist longer than 35

days are rare. Second, there are mechanisms that ensure

the persistence of representati

ons over short periods of time (

days) even if the animals are not training in the task, or

if the circuit is perturbed by lesions.

Unlike individual neurons that (on average) only retain informa

tion for 10 ± 5 consecutive days, cell groups

with synchronous activity encode stable spatial and temporal re

presentation for 35 days (the longest time that we

analyzed). Using synchrony to define functional connectivity an

d graph analysis, we demonstrate that groups of

synchronously active neurons provide information absent in indi

vidual neurons. The analysis of graph topology had

two key benefits. First, the features of a neuron in a graph a

re sufficient to decode signatures that are characteristic

of place cells, time cells or ne

ither, even without having any

additional information regarding the animal’s behavior.

Second, the likelihood that a neuron will maintain its responsi

veness to a field in the future is proportional to the extent

of connectivity of that neuron in the graph and can be predicte

d using graph topology. Overall, our findings suggest a

model where the patterns of activity of individual neurons grad

ually change over time while the activity of groups of

synchronously active neurons ensu

res the persistence of represe

ntations.

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Methods

Animals.

Male and female C57BL6J-Tg-Thy1-GCaMP6s 6 to 20-week old (Jack

son Labs stock: 025776) were housed

in a reverse 12 h light/dark photocycle and provided food and w

ater ad libitum. Mice were single housed post-surgery

until the end of the experiment. Experimental animals were sele

cted randomly and include both sexes. All animal

procedures were approved and performed following institutional guidelines (Caltech IACUC).

Bilateral endoscope implantation

Mice were anesthetized with a single dose of 100/10 mg/kg keta

mine/xylazine before the surgery and placed

into a stereotactic frame and body temperature was maintained w

ith a passive heating pad at 37 ºC. Ketoprofen 5

mg/kg and buprenorphine SR 1 mg/kg was subcutaneously injected

prior to surgery. Bupi

vacaine 1 mg/kg solution

added dropwise along the surgical incision prior to wound closu

re and animals were main

tained on ibuprofen 30

mg/mL (in the water) ad libitum for at least 3 days post-surger

y. Animals were in a recovery period for at least 4

weeks before attachment of the microendoscope.

Animals underwent unilateral or bilateral surgeries to implant

1.8 mm GRIN lenses directly dorsal to CA1. Before

implantation, we performed a 1.8 mm diameter craniotomy centere

d around the coordinates (relative to bregma: 1.8

mm and -1.8 mm lateral ; -2.0 mm posterior) using a FG1/4 carbi

de bur. Freshly prepared ar

tificial cerebrospinal fluid

(aCSF) was applied to the exposed tissue throughout the surgery

to prevent dehydration. Using a blunt 26-gauge

needle, the dura, cortex, and portion of the corpus callosum we

re quickly aspirated under continuous perfusion with

aCSF. Aspiration was stopped once

a thin layer of horizontal fi

bers was left on the surface of the hippocampus. The

cortical cavity was perfused w

ith more aCSF and small pieces of

moist gelfoam were placed on the surface of the

craniotomy to prevent excessive bleeding while avoiding contact

with the surface of hippocampus. Once the surface

of the hippocampus was clear of blood, the GRIN lens was slowly

lowered in into the brain using a stereotaxic arm to

a depth of 1.30 mm below the surface of the skull. Removal of t

he cortex and insertion of the GRIN lens was performed

in 10 minutes or less (in each hemisphere) to prevent bulging o

f the hippocampus due to the lower dorsal pressure.

Two skull screws were placed ant

erior to bregma (1.8/-1.8 mm la

teral; 1.0 mm anterior) and both the screws and lens

were secured with cyanoacrylate and dental cement. The exposed

end of the GRIN lens was protected with transparent

Kwik-seal glue and animals were r

eturned to a clean cage. Two w

eeks after the surgery, mi

ce were anesthetized with

1.0 to 2.0 % isoflurane, the glue covering the GRIN lens was re

moved and a microendoscope was aligned with the

GRIN lens. The miniature microsc

opes were connected to a portab

le computer which provided live view of the

fluorescence image th

ough the endoscopic lens and guided the final alignment and focal plane of the microscope and

lens. The microscope was permanently attached to the implant wi

th dental acrylic and the focal sliding mechanism on

the microendoscope was sealed with superglue.

Local CA1 damage.

Damage was induced unilaterally by increasing the LED power of

the microendoscope to the maximum

allowable level. The power of the 470 nm blue LED on the minisc

ope was measured to produce 500-700 μW at this

setting, measured at the end of the GRIN lens facing the CA1 us

ing a power meter (Thorlabs PM100D, S155C probe).

The brain was illuminated for at least 30 minutes during the fo

raging and linear track task. Cell death was quantified

as the ratio of neurons active in the field of view prior and a

fter heat damage.

Mouse behavior.

Mice were maintained in a reverse photo cycle and three days pr

ior to initialization of behavior recording

they were set under a water restriction protocol in which they

receive 2.0 mL of water per

day. After three days of

water restriction, mice were brought to the recording room and

connected to a computer system through a commutator

with a 2.0-meter-long custom cab

le. Animals were habituated to

handling and being tethered to the cable for at least

days. During this time, mice

were connected to the computer a

nd allowed to explore their home cage but were not

exposed to the linear track. On t

he fourth day, behavioral reco

rding was initiated, mice grab

bed by the tail and placed

in the middle of the track and imaging was recorded within 20 s

econds of the mice being introduced in the linear track.

Mice were exposed to the maze every day during the learning and

first 5 days following re-exposure. However, during

the training period or 5 days after re-exposure, recording sess

ions were at different day intervals ranging from 2-5

days.

Each behavior session consisted

of a 10-minute recording of the mouse exploring its home cage without the

lid and food dispenser immediately followed (within 30 seconds)

by a 20-minute recording of the mouse running on

a linear track. The mouse home cage is rectangular 20 cm x 35 c

m with 15 cm transparent walls while the linear track

is a 1.5-meter-long track 12 cm wide with 15 cm high walls made

of white plastic. Three group of cues were place on

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

both walls of the linear track a

nd consisted of black stripes (

2 cm wide) at different angles. The linear track was

equipped with an automatic liquid dispensing port that would de

liver between 100-200 μL of sugar water (15% sucrose

in DI water) at the end of the track. The system would require

the animal to run to opposite ends of the track to receive

a water reward, a green LED (right side) and red LED (left side

) indicated which port was active. A beeping sound

was played once mice activated the IR sensor. Delayed reward ex

periments were performed by decoupling the delivery

of sugar reward from IR sensor activation, thus the animal was

required to wait for 5 seconds until sugar water was

delivered. The beeping sound was not delayed. The position of t

he mouse was tracked by an

ultra-wide-angle webcam

at 25 Hz and 640x360 pixels positioned about 1.8 meters above t

he maze. Mouse position was extracted using

OptiMouse (

Calcium imaging processing.

Video Acquisition:

signal from the microendoscope CMOS sensor was acquired though

t a UVC compliant

USB analog video to digital conv

erter. The video feed was captu

red and saved by videoLAN media player using

custom MATLAB scripts. Data acquisition was set at 25 Hz and di

splay resolution of 720 x 576 pixels using a

YUV4:2:2 codec and AVI file encapsulation. All three cameras we

re synchronized to star simultaneously and were

verified to have latencies smaller than 10 ms. Raw videos were

offline transcoded to lossless H.264 -MPEG-4 AVC

codec and MP4 encapsulation and the first 2 seconds of each vid

eo were deleted. Transcoded videos were filtered

using a high quality 3-dimensional low pass filter (

hqdn3d

) with spatial and temporal smoothing of 4x4 pixels and 2

frames, respectively. Denoised videos were then 4x down-sampled

by a moving window averaging of 4 frames. All

transcoding, smoothing, and down sampling was performed by the

open source program ffmpeg controlled through

custom MATLAB scripts. Down sampled videos were motion correcte

d using the recursive fast Fourier transform

approach provided in the MATLAB script

sbxalign.mat

from Scanbox. For batch analysis all 4x down-sampled videos

were concatenated into a larger

video, motion corrected, and an

alyzed with CNMFe.

Signal extraction:

motion corrected videos were sav

ed as a matrix array and analyzed using CNMFe with a

2-fold spatial and temporal down-sampling (

). All 7450 frames of calcium imaging during the 20-minute line

track task or 3750 frames during

home cage exploration were ana

lyzed simultaneously. In some cases, were motion

artifacts were minimal, up to 200

,000 frames of data was analyz

ed simultaneously with CNMFe and these datasets

were utilized to confirm registration procedures described belo

w. All data analysis was performed using the

deconvoluted neural activity out

put (neuron.S) from CNMFe.

Cell registration:

accurate alignment of neurons across sessions spanning days to

months has noticeably

become a challenging aspect of calcium imaging. Taking from stu

dies using multiphoton imaging, registration

approaches using single photon e

pifluorescence involves alignme

nt of neurons between sessions based on their spatial

footprint. However, spatial footprint of extracted neurons usin

g microendoscopes can suffer from artifacts arising

from changes in firing rates, extraction algorithms, and motion

correction artifacts. These effects are compounded

when imaging areas with dense labeling as observed in transgeni

c animals. To overcome these limitations, we

employed an interleaved and batch data analysis approach in whi

ch a session of data was common between two days

to be aligned. This common session served the purpose of decrea

se noise in the extracted neuron footprints and

ovided firing data common to both datasets. First, the footpr

int extracted by CNMFe was thresholder so that only

the pixels above the 50

percentile formed th

e ROI footprint. Minor motion artifacts were corrected by

a fast Fourier

transform method using the PNR image output from CNMFe. The spa

tial correlation between all neurons with centroid

distances below 15 pixels between the two alignment datasets wa

s calculated. The correlation of the CNMFe

deconvoluted neural activity was also calculated for all neuron

pairs with centroid distance less than 15 pixels. Only

spatial and temporal correlation above 0.5 and 0.3, respectivel

y, were considered. An a

lignment coefficient was

calculated by multiplying the spatial correlation and temporal

correlation of every potential neuron pair. The alignment

coefficient was binned in 100 intervals and the probability dis

tribution calculated. A plot of the probability distribution

showed a clear bimodal distribution

(Fig. S2f)

, a threshold of approximately 0

.5 was selected manually. All n

euron

pairs below this threshold were deleted, and the remaining were aligned based on a iterative selection of pairs with

the highest alignment coefficient. This approach was validated

by aligning a video to itself, where we observed that

including temporal information increased aligning accuracy by a

bout 10% (see supplementary information). Here, we

take advantage of small motion artifacts across days in order t

o motion correct and analyze several days together in

one animal. This ensures, that even if the neuron decreases is

firing rate significantly, CNMFe will still draw a ROI

around the neuron and extr

act its firing activity.

Data analysis

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Identifying place cells

: place fields were extracted by identifying periods when mice

ran continuously faster

than 3 cm/second for more than 0

.4 seconds. Together, these thresholds eliminate periods

during which mice were

grooming, rearing, or turning. The length of the linear track was divided into bins spanning 3 centimeters (50 bins),

neuronal activity at the ends of

the maze (7 cm) w

ere not inclu

ded in the analysis. The average firing rate of a neuron

in each bin was calculated by the sum of all calcium activity i

n a bin divided by the amount of time the mouse spent

in that bin. The average firing rate of a neuron was then norma

lized by the total number of spikes in order to generate

normalized tuning profile of each neuron. Neurons were classifi

ed as place cells if: (1) the place field is 15 cm wide;

(2) calcium transients were present > 30 % of the time the mous

e spent in the place field; and (3) the cell contains

significantly greater spatial in

formation than chance. Spatial

information is calculated using (

푆퐼 ൌ

휆

෍휆

௜

log

ଶ

൬

휆

௜

휆

൰푃

௜

where λ is the overall average cal

cium transient rate of the ce

ll, λ

is the average calcium tra

nsient rate in spatial bin

and P

is the probability the mouse is in spatial bin

. Chance level spatial information for each neuron is calculate

d by

shuffling the time stamps of th

e calcium transients and calcula

ting the spatial information

of the shuffled trace. This

is done for 1000 iterations and the spatial information of the

cell is considered significant if it is higher than 95% of

the shuffled traces.

Identifying time cells

: the linear track was equipped with a LED light that would tur

n off once the animal

activated the IR sensors near the water reward port. The ON/OFF

transition of the LED in the behavior video was

extracted and used as a timestamp. The LED timestamp was set as

time = 0 and a window of 31 frames or the time

the animal was immobile (velocity less 1 cm/sec), whichever is

smaller, was used for analysis. Neurons were classified

as time cells if: (1) they fire at least 20% of the times the a

nimal activated the water port; (2) the neuron fired 30%

more within its field (20 % of time immobile); and (3) the cell

s contain significantly great

er information than chance.

Temporal information was calculated using the same equation abo

ve but using λ as the average activity during

immobility, λ

is the average activity in frame

after activation of the water port. The variable Pi in this ca

se represent

the probability that the animal was not moving during frame

Network graphs

: an adjacency matrix was genera

ted by calculating the pairwise

Pearson correlation between

all neurons during a session. Only statistically significant (p<0.05) correlation values above 0.10 were used to build

the adjacency matrix. The weight

of the edge between two neuron

s was set to be equal to the correlation coefficient.

The networks are plotted with the network graphical tool (Gephi

0.9.2,

www.gephi.com

) using the built-in ForceAtlas2

layout. Modularity was calculated with Markov Diffusion approac

h with a diffusion time of one using MATLAB

scrips as previously described (

). Cell groups are defined as cells making up each module extra

cted by the

modularity calculation. Only the largest 11 modules were analyz

ed. The degree, average clustering coefficient,

eigenvector centrality, and PageRank were calculated using buil

t in and custom MATLAB functions.

Behavior decoding

: mouse position was decoded by a generalized linear model usin

g the

glmfit

function with

a normal distribution in MATLAB. Mouse position was discretized

into 50 intervals and periods of mobility and

immobility were decoded. The linear model was first trained wit

h 50% of randomly selected place/time cell activity

and mouse position and then tested on the remaining 50%. For ti

me-lapse decoding we used the decoder trained in

one session (M) to decode the position M+N sessions apart using

the activity of the same place/time cells on day N.

Decoding mouse position using cell groups was performed in a si

milar fashion, but instead of using place/time cell

activity we used the integrated deconvoluted neuronal activity

of neurons making up each cell group. Thus, we had

11 inputs (for each cell group)

and one target (mouse position)

as a training set, 70 % of

the dataset was used for

training. Time-lapse decoding was performed by training a decod

er on one sessions and then testing it on session N

using cell group activity as an input. Cell group activity was

determined at sessions N by first calculating the maximum

projection of the adjacency matri

x across 5 trained sessions (

5-10), calculating the cell groups using Markov diffusion.

The IDs of neurons making each

cell group were use to calculate

their integrated group activity in all other session.

In addition, to account for the n

onlinear and asymmetric nature

of fields encoded by cell groups, we also performed

position decoding using a nonlinear fitting neural network (

fitnet

function in MATLAB). This network was set to have

50 hidden layers and trained with 70% of randomized cell group

activity and mouse position, validated with 15 % and

tested with 15%. The network trained in one session was used to

decode N sessions apart. In all, cases we also trained

the decoders with a randomized dataset for comparison. Error re

ported represent the median ± sem across N sessions

apart including periods of learning and no task.

Time/place cell decoding

: whether a neuron was place/time cell was decoded from neurona

l activity by (1)

calculating the intrahemispheric correlation in one session, (2

) generating a graph and calculating its topological

properties, (3) training a fitting neural network with half the

data, (4) use the decoder trained on one session to decode

another session. Graph topology metrics for each node included:

degree, modularity, eigenvector centrality, PageRank,

calcium activity, and the assemblies to which a node was connected. The calcium activity was defined as the integrated

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

intensity of a neuron in a session minus the median integrated

intensity of all neurons. All metric, excluding

modularity, calcium activity, and node connectivity were normalized so that the sum in a session is equal to one. These

metrics were the input to the neural network decoder and whether a neuron had (1) or not (0) a field was used as the

training target. In addition, we

trained a decoder simultaneous

ly with 10 sessions of data from one mouse and used to

decode which neurons were place

and time cells in other animals

only using their graph topology.

Decoding stability of a cell

: we predicted whether a place/time cell observe today will los

e its field in N days

by first randomly selecting 50 % of the sessions in one mouse a

nd calculating all pairwise session intervals between

them. We trained one fitting neur

al network for each set of ses

sions separated at a N interv

al, generating N decoders.

The inputs consisted of graph topology metrics calculated from

brain activity (the same parameters used above)

recorded in the sessions at N int

ervals. The target dataset was

a matrix listing whether a neuron became responsive or

unresponsive to a field (logical 1) or the neuron retained its field or lack thereof (logical 0) between two sessions at a

N interval. We then tested each decoder on the remaining 50 % o

f the sessions with intervals ranging from 1 to N. We

filtered the decoder’s output so

that the highest outputs above

the median were considered positive.

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

References

1. W. B. Scoville, B. Milner, Los

s of Recent Memory After Bila

teral Hippocampal Lesions.

Journal of

Neurology, Neurosurgery & Psychiatry

, 11–21 (1957).

2. N. J. Cohen, L. R. Squire, Preserved learning and retention

of pattern-analyzing skill in amnesia: dissociation

of knowing how and knowing that.

Science

210

, 207–210 (1980).

3. J. O’Keefe, J. Dostrovsky, The hippocampus as a spatial map. Preliminary evidence from unit activity in the

freely-moving rat.

Brain Research

, 171–175 (1971).

4. A. Czurkó, H. Hirase, J. Csicsvari, G. Buzsáki, Sustained a

ctivation of hippocampal pyramidal cells by ‘space

clamping’ in a running wheel.

European Journal of Neuroscience

, 344–352 (1999).

5. D. Aronov, R. Nevers, D. W. Tank, Mapping of a non-spatial dimension by the hippocampal–entorhinal

circuit.

Nature

543

, 719–722 (2017).

6. M. R. Mehta, C. A. Barnes, B. L. McNaughton, Experience-dep

endent, asymmetric expansion of hippocampal

place fields.

Proceedings of the National Academy of Sciences

, 8918–8921 (1997).

7. M. A. Wilson, B. L. McNaughto

n, Dynamics of the hippocampal

ensemble code for space.

Science

261

1055–1058 (1993).

8. N. T. Agnihotri, R. D. Hawki

ns, E. R. Kandel, C. Kentros, T

he long-term stability o

f new hippocampal place

fields requires new protein synthesis.

Proceedings of the National Academy of Sciences

101

, 3656–3661

(2004).

9. C. G. Kentros, N. T. Agnihotri, S. Streater, R. D. Hawkins,

E. R. Kandel, Increased Attention to Spatial

Context Increases Both Place Fi

eld Stability and Spatial Memory

Neuron

, 283–295 (2004).

10. J. Debiec, J. E. LeDoux, K. Nader, Cellular and Systems Re

consolidation in the Hippocampus.

Neuron

527–538 (2002).

11. L. T. Thompson, P. J. Best, Long-term stability of the pla

ce-field activity of single units recorded from the

dorsal hippocampus of freely behaving rats.

Brain Research

509

, 299–308 (1990).

12. Y. Ziv

et al.

, Long-term dynamics of CA1

hippocampal place codes.

Nature Neuroscience

, 264

(2013).

13. J. D. Zaremba

et al.

, Impaired hippocampal place cell

dynamics in a mouse model of

the 22q11.2 deletion.

Nature Neuroscience

, 1612–1623 (2017).

14. T. Hainmueller, M. Bartos, P

arallel emergence of stable an

d dynamic memory engrams in the hippocampus.

Nature

, 1 (2018).

15. N. B. Danielson

et al.

, Sublayer-Specific Coding Dynamics during Spatial Navigation a

nd Learning in

Hippocampal Area CA1.

Neuron

, 652–665 (2016).

16. S. M. Zola-Morgan, L. R. Squi

re, The primate hippocampal formation: evidence for a time-limited role in

memory storage.

Science

250

, 288–290 (1990).

17. H. A. Zariwala

et al.

, A Cre-Dependent GCaMP3 Reporter Mouse for Neuronal Imaging In

Vivo.

J. Neurosci.

, 3131–3141 (2012).

18. D. J. Cai

et al.

, A shared neural ensemble links

distinct contextual memories encoded close in time.

Nature

534

, 115–118 (2016).

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

19. H. Dana

et al.

, Thy1-GCaMP6 transgenic mice for neuronal population imaging i

n vivo.

PLoS ONE

e108697 (2014).

20. K. K. Ghosh

et al.

, Miniaturized integration of

a fluorescence microscope.

Nature Methods

, 871 (2011).

21. W. A. Liberti, L. N. Perkins, D. P. Leman, T. J. Gardner, An open source, wireless capable miniature

microscope system.

J. Neural Eng.

, 045001 (2017).

22. E. A. Pnevmatikakis

et al.

, Simultaneous Denoising, Deconvolution, and Demixing of Calciu

m Imaging Data.

Neuron

, 285–299 (2016).

23. P. Zhou

et al.

, Efficient and accurate extract

ion of in vivo calcium signals

from microendoscopic video data.

eLife

, doi:10.7554/eLife.28728.

24. C. J. MacDonald, K. Q. Lepage, U. T. Eden, H. Eichenbaum,

Hippocampal “Time Cells” Bridge the Gap in

Memory for Discontiguous Events.

Neuron

, 737–749 (2011).

25. K. Kay

et al.

, A hippocampal network for spatial coding during immobility an

d sleep.

Nature

531

, 185

(2016).

26. K. D. Harris, J. Csicsvari, H. Hirase, G. Dragoi, G. Buzsá

ki, Organization of cell assemblies in the

hippocampus.

Nature

424

, 552 (2003).

27. V. Itskov, C. Curto, E. Pasta

lkova, G. Buzsáki, Cell Assem

bly Sequences Arising from Spike Threshold

Adaptation Keep Track of

Time in the Hippocampus.

J. Neurosci.

, 2828–2834 (2011).

28. E. Pastalkova, V. Itskov, A. Amarasingham, G. Buzsáki, Int

ernally Generated Cell

Assembly Sequences in

the Rat Hippocampus.

Science

321

, 1322–1327 (2008).

29. Y. N. Billeh, M. T. Schaub,

C. A. Anastassiou, M. Barahona, C. Koch, Revealing cell assemblies at multiple

levels of granularity.

Journal of Neuroscience Methods

236

, 92–106 (2014).

30. R. J. Low, S. Lewallen, D. Aronov, R. Nevers, D. W. Tank,

Probing variability in a cognitive map using

manifold inference fr

om neural dynamics.

bioRxiv

, 418939 (2018).

31. L. Meshulam, J. L. Gauthier, C. D.

Brody, D. W. Tank, W. B

ialek, Collective Beha

vior of Place and Non-

place Neurons in the H

ippocampal Network.

Neuron

, 1178-1191.e4 (2017).

32. A. Samsonovich, B. L. McNaughton, Path Integration and Cog

nitive Mapping in a Continuous Attractor

Neural Network Model.

J. Neurosci.

, 5900–5920 (1997).

33. Y. Ben-Shaul, OptiMouse: a comprehensive open source progr

am for reliable detection and analysis of mouse

body and nose positions.

BMC Biology

, 41 (2017).

34. W. E. Skaggs, B. L. McNaughton, K. M. Gothard, in

Advances in Neural Information Processing Systems 5

S. J. Hanson, J. D. Cowan, C. L. Giles, Eds. (Morgan-Kaufmann,

1993; http://papers.nips.cc/paper/671-an-

information-theoretic-approach-to-deciphering-the-hippocampal-c

ode.pdf), pp. 1030–1037.

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Figure 1

Long-term simultaneous bilateral

imaging of CA1 activity in fr

eely moving transgenic mice

(a)

Rendering of custom endoscope used in the bilateral configurati

on.

(b)

Chronic implants show small field of view

(FOV) displacement across weeks.

Dashed line represents period

of no task (10 days).

(c)

Multiple sessions (25)

combined and analyzed simultaneously using CNMFe (see methods).

(d)

Raster plot of deconvoluted neuronal

activity in the right (top) and left (bottom) hemisphere of a t

rained freely moving mouse. Neurons are ranked from

high to low peak-to-noise ratio (PNR). Between 754-1480 neurons

were recorded per hemis

phere, median 1236, in 4

bilateral mice and 5 with unilateral microendoscopes). (

) Maximum intensity projection in the field of view in one

animal recorded for 8 months. (

f-g

) Intensity correlation map of a single session (

left

) and a small region of interest

(green rectangle) showing the pe

rsistence of activity throughou

t 25 sessions (

right

). (

) Activity distribution of the

right hemisphere showing that th

e majority of neurons are activ

e on most sessions (n=8 mice).

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Figure 2

Spatial and temporal representati

ons persist through periods o

f no exposure to the task

(a)

The

probability of a neuron with place

and time field in one sessio

n having any place/time information N days apart. Solid

lines represent the median and shadow the 95 % bootstrap confid

ence interval. Random level represent overlap if

neurons randomly became responsive to a field on the following

day.

(b)

Distribution of neurons retaining their field

response for N consecutive sessions.

(c)

Quantification of cell overlap 1 day apart (65 ± 14 %, n = 193

sessions) and

10 days (54 ± 12 %, n = 97, ranksum-test) when continuously exp

osed to the track.

(c)

Quantification of cell overlap

1-4 sessions apart (65 ± 14 %, n = 193 sessions) and 1-4 sessions days (54 ± 12 %, n = 97, ranksum-test) when

including a 10-day period of no task.

(e)

Response fields of six place cells across 45 days. Black recta

ngles show

sessions in which 87 ± 8 % of fields changed direction. Green r

ectangle are the two sessions after re-exposure.

(f)

Pair-wise field correlation of n

eurons in the right hemisphere

with a field response between sessions. Colors represent

the median value. Numbers indicate the session. Sessions were g

rouped as indicated by the dashed rectangle and

letters.

(g)

The similarity between fields in each group of sessions was ob

tained by subtracting the correlation obtained

by random aligning neurons (n = 7, right hemisphere, each dot i

s the median between two sessions, p< 10

-4

marked

by *** ).

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Figure 3. Tissue heating induces

direction specific bursting ac

tivity and unstable fields followed by partial

recovery of the original representation.

(a)

Intensity correlation of a 20-minute session before, during, a

nd after

tissue damage induced by high illumination intensity (DPD, days

post damage).

(b)

Number of neurons active

simultaneously (within 160 millis

econds) during each session (n

umber shows median of the largest 90% bursts). Each

day is marked by a vertical red line.

(c)

Single neuron activity across days while the mouse runs in the linear track

(black traces) showing changes

in activity induced by tissue da

mage (red rectangle). The ho

rizontal green line shows

the 3 standard deviation of the background fluctuations. Top tw

o neurons recover from the damage, middle neuron is

highly active during abnormal activity, and the last two neuron

s become inactive following damage.

(d)

Activity

distribution of six place cells s

howing the high instability of

fields during abnormal activity, colors as in figure 2.

Black rectangle shows session

s with abnormal activity.

(e)

Pair-wise correlation between sessions of neurons which

retained their fields (right hemisphere). (f) The similarity between fields in groups of sessions before, during, and after

the lesion (n = 2 mice, each dot

is median of session pair, num

bers are median, p< 10-4 marked by ***).

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

Figure 4. Synchronous activity in group of cells encode stable

representations of the task.

(a)

number of inter-

and intra-hemispheric synchronous pairs as a function of traini

ng (4 mice, each dot represe

nts a session in one mouse,

all significant p<10

-6

(b)

(Top) Neuron pairs are co-active 90 ± 3 % (orange distribution

) of the time and fail to fire

synchronously 50 ± 20 % of the time (blue distribution). Synchr

onized pairs also show correlated fluctuations of their

fields (green distribution, 0.77 ± 0.17, p<10

-8

). (Bottom) The number of sessions they remain responsive a fie

ld is

proportional to the level of synchrony in the pair (R-square 0.

35, p< 10

-4

(c)

Graph topology in the linear track and

home cage of one mouse during learning, trained, and re-exposed periods (data from one hemisphere used, lines

represent correlation > 0.15, neurons shown as nodes).

(d)

Graph colored by the median

location where each neuron

(node) was active during a 20 minute session in the linear trac

k. Inset shows the same graph but colored by cell groups

(see methods, right hemisphere). Graph of neuronal activity in

the linear track colored by

(e)

experimentally

determined field responses and

(f)

by the largest six cell groups identified by Markov diffusion.

The size of each node

corresponds to the degree (numbe

r of connecting lines) of each

node.

(g)

(Top panel) Integrated activity of neurons in

a cell group during a linear track session and (bottom panel) a

ctivity each time the mouse ran through the cell group’s

receptive field.

(h)

Persistence of a cell group identified by neuronal activity from days 5-6 (shadows represent

deviation during a session).

(i)

Mouse position in the linear track decoded using all place/tim

e cells (132 to 639

neurons, errors 10-19 cm using generalized model), or the summe

d activity of all neurons in 11 cell groups (error 13-

17 using a generalized model or a 12-15 nonlinear fitting neura

l network). The performance of decoders using cell

group activity outperformed place/time cell decoders on long-ti

mescales (20+ sessions apart, random vs place/time

decoder p=0.210 and random vs cell group decoder, p<10

-5

, Friedman-test).

(i)

Activity of one cell group across 35

days (number indicate day, red traces indicate sessions with ro

tated representations).

(j)

Graph colored by the fraction

of sessions a neuron was classified as time or place cell, larg

er nodes have larger number of synchronous pairs. Top

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;

inset shows the presence of unstable nodes with low connectivit

y, bottom inset shows quantification of degree and

stability across 25 sessions (red

dots mark median, R-square XX

(k)

A decoder trained with graph topology metrics

on one day can identify 71 ± 11 % of all place/time cells the n

ext day and 59 ± 10 % after 30 days in the same mouse

(left) with similar res

ults obtained when a decoder trained wit

h one mouse is used to identify place/time cells in other

mice (n = 7, right). It also identifies 86 ± 2 % of neurons wit

h no field responses. (Right) Neurons which will become

responsive or unresponsive to a field between two re-exposures

to the track N sessions apart can be decoded from

graph metrics 46 ± 5 % of the time.

(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint

http://dx.doi.org/10.1101/559104

doi:

bioRxiv preprint first posted online Feb. 24, 2019;