Supporting Online Material for - aav9199_Gonzalez

science

.sciencemag.org/content/

365

6455

821

/suppl/DC1

Supp

lementary

Material

for

Persistence of

euronal

epresentations

hrough

ime and

amage in the

ippocampus

Walter G.

Gonzalez

Hanwen

Zhang

Anna

Harutyunyan

Carlos

Lois

*Corresponding author

. Email

: clois@caltech.edu

Published

23 August

Science

365

821

(20

)

DOI:

10.1126/science.

aav9199

This PDF file

includes:

Materials and Methods

Supplementary

Text

Figs. S1 to S

Tables S1 to S

Captions for Movies S1 to S9

Reference

Other

Supp

lementary

Material

for this manuscript

include

the following

(available at

science

.sciencemag.org/content/

365

6455

821

/suppl/DC1

)

Movies S1 to S

(

mp4

)

Materials and Methods

Animals.

Male and female C57BL6J

-Tg

-Thy1

-GCaMP6s 6 to 20-week old (Jackson Labs

stock: 025776) were housed in a reverse 12 h light/dark photocycle and provided food ad

libitum. Mice were single housed post-surgery until the end of the experiment.

Experimental animals were s

elected randomly and included both sexes. All animal

procedures were approved and performed following institutional guidelines (Caltech

IACUC).

Unilateral/Bilateral endoscope implantation

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

before the surgery and placed into a stereotactic frame. The body temperature was

maintained with a passive heating pad at 37 ºC. Ketoprofen 5 mg/kg and buprenorphine

SR 1 mg/kg was subcutaneously injected prior to surgery. Bupivacaine 1 mg/kg solution

was added dropwise along the surgical incision prior to wound closure and animals were

maintained on ibuprofen 30 mg/mL (in the water) ad libitum for at least 3 days post-

surgery. Animals were in a recovery period for at least 4 weeks before attachm

ent of the

microendoscope.

Mice underwent unilateral or bilateral surgeries to implant GRIN lenses (1.8 mm

0.25 pitch 0.55 NA) directly dorsal to CA1. Before implantation, we performed a 1.8 mm

diameter craniotomy centered around the coordinates (relative to bregma: 1.8 mm and -

1.8 mm lateral; -2.0 mm posterior) using a FG1/4 carbide bur. Freshly prepared artificial

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 were quickly aspirated under continuous perfusion with aCSF.

Aspiration was stopped once a thin layer of horizontal fibers was left on the surface of the

hippocampus. The cortical cavity was perfused with 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

the cortex and insertion of the GRIN lens was performed in 10 minutes or less (in each

hemisphere) to prevent bulging of the hippocampus due to the decreased dorsal pressure.

Two skull screws were placed anterior to bregma (1.8/-

1.8 mm lateral; 1.0 mm anterior)

and both the screws and lens were secured with cyanoacrylate glue and dental cement.

The exposed end of the GRIN lens was protected with transparent Kwik

-seal glue and

animals were returned to a clean cage. Two weeks after the surgery, mice were

anesthetized with 1.0 to 2.0 % isoflurane, the glue covering the GRIN lens was removed

and a microendoscope was aligned with the GRIN lens. The miniature microscopes were

connected to a portable computer for live view of the fluorescence image which was used

to guide the final alignment and focal plane of the microscope and lens. The microscope

was permanently attached to the implant with 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 miniscope was measured to produce 500-700 μW at this setting. The power

measurements is done at the end of the GRIN lens facing CA1 using a power meter

(Thorlabs PM100D, S155C probe). The brain was illuminated for at least 30 minutes

during the foragin

g and linear track tasks. In one animal we performed local brain heating

on two sessions to increase the damage area (top row of Fig. 3A). The extent of the lesion

was assessed by the presence of abnormal burst of activity and by the presence of

continuously green cells on the following day. Sessions with abnormal activity were

selected based on the low place/time field correlation observed in Fig. S11 and by the

increase number of co

-active neurons, as shown in Fig. 3B. Three animals were exposed

to high LED power and all three developed abnormal activity in the form of large number

of co

-active neurons. However, the mouse with the most extensive damage required over

a month for the field of view to become clear enough to be registered. When this animal

was

exposed to the linear track it formed place/end cells but the similarity to the original

representation was lower than the other two mice.

Mouse behavior.

Mice were maintained on a reverse photo cycle and three days prior to the start of

behavior recording they were restricted to 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 cable (Mouser, 538-50MCX-

37). Animals were habituated to handling and being tethered to the cable for at least 3

days. During this time, mice were connected to the computer and allowed to explore their

home cage but not the linear track. After habituation, mice were placed in the middle of

the track and imaging was started within ~20 seconds of the mice being introduced in the

linear track. Mice were exposed to the maze every day during the learning period and

first 5 days following re-exposure. However, during the training period or 5 days afte

r re

exposure, recording sessions were at different day intervals ranging from 2-7 days. The

initial facing orientation of the mouse in the linear track was not controlled.

Each behavior session consisted of a 10-minute recording of the mouse freely

moving in its home cage, without the cage lid or food dispenser, immediately followed by

a 20

-minute recording of the mouse running on a linear track. The mouse home cage is

rectangular 20 cm x 35 cm x 15 cm with transparent walls while the linear track is a 1.5

m x 12 cm x 15 cm made of white plastic. The linear track was cleaned with 70 %

ethanol before introducing another mouse. Three group of cues were place on both walls

of the linear track and consisted of black stripes (2 cm wide) at different angles (F

ig. S1).

The linear track was equipped with two automatic liquid dispensing ports at either end

delivering between 10-50 μL of sugar water (15% sucrose in DI water). The system

would require the animal to run to opposite ends of the track to receive water, a green

LED (right side) and a red LED (left side) indicated which port the mouse needed to run

towards. A beeping sound was played once the mouse activated the IR sensor. Delayed

reward experiments were performed by decoupling the delivery of sugar rewar

d from IR

sensor activation, thus the animal was required to wait for 5 seconds. The beeping sound

was not delayed. The position of the mouse was tracked by an ultra-

wide

-angle webcam

at 25 Hz and 640x360 pixels positioned about 1.8 meters above the maze. Mouse position

was extracted using OptiMouse (

Custom miniaturized fluorescence microscopes.

Miniaturized microendoscope were custom made following previous designs (

The main differences between the microendoscope used here and previous designs is the

reduction in size achieved by reverse engineering the CMOS sensor used in (

) and by

minimizing the wall thickness and footprint of the microendoscope body used in (

The CMOS sensor used here is adapted from a miniature CCTV marketed online as

model MC900 but can also be purchased under other names. As purchased, the video

camera circuit consists of two 1 mm thick PCB boards: a primary sensor PCB containing

an Omnivision 7960 CMOS sensor and some peripheral resistors and c

apacitors

connected through a 6 pin connector to a secondary PCB with an 24.545 MHz clock and

a voltage regulator circuit. These two circuits were carefully separated by desoldering the

6 pin connector. The clock and voltage regulator circuits can be connected to the sensor

PCB through cables as long as 20 cm but electrostatic noise can affect the clock output on

longer distances. Here we soldered a female 6 pin connector to the CMOS PCB and a

male connector to the clock and voltage regulator connectors. This approach allowed us

to decrease the weight of each microendoscope by as much as 200 mg during the period

the animal is not being recorded.

The clock and voltage regulator PCB were connected to an USB analog video to

digital converter. LED intensity was

controlled through the ADC output of a signal

generator. To decrease the footprint and weight we designed the body of the

microendoscope with wall thickness less than 1 mm. The body of the microendoscope

was machined from 1.25 x 2.5 x 2.5 cm

black acetal

(Delrin) blocks using a 5-

axis CNC

machine (PocketNC). All CAD designs and CNC trajectories were generated by Fusion

360 (AutoCAD) and are provided in the supplementary information. Light from the

excitation LED was filtered through a 470/40 nm bandpass e

mission filter, fluorescence

light was separated from the excitation light by a 495 nm long pass dichroic mirror and

filtered through a 520/50 nm bandpass filter (Chroma technologies). Light was collimated

onto the CMOS sensor through a 12.5 mm achromatic lens (Edmund Optics). The FOV

covered 600 μm x 479 μm at a resolution of 720 x 576 pixels, 0.83 μm/pixel (NBS1952

calibration target, Thorlabs).

Calcium imaging.

Video Acquisition:

signal

from the microendoscope CMOS sensor was acquired

using a UVC compliant USB analog video to digital converter (EasyCAP DC60). The

video feed was captured and saved by videoLAN media player (

www.videolan.org

) using

custom MATLAB scripts. Data acquisition was set at 25 Hz and display resolution at 720

x 576 pixels using a YUV4:2:2 codec and AVI file encapsulation. All three cameras were

synchronized to start 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 video 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

(www.ffmpeg.org

) controlled through custom MATLAB scripts. Down sampled videos

were motion corrected 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 concatenat

ed into a larger video and then motion corrected.

Signal extraction:

motion corrected videos were analyzed using CNMFe

(MATLAB variable names shown in parenthesis). We further down sampled our 4x

accelerated data with a 2

-fold spatial (ssub) and temporal down-

sampling (tsub)

(

The data was smoothed with a gaussian kernel of width between 6-8 pixels (gSig), and

neurons were constrained to a diameter of 30 pixels (gSiz). A ring model (bg_model) was

used for the background with radius of 30 (ring_radius). Neurons with spatial overlap

greater than 0.65 (merge_thr) and centroid distance less than 5 pixels (dmin) were

merged. Only ROIs with minimum peak-

to-noise ratio of 5 (min_pnr) and minimum

spike size of 3 (smin) were extracted. The foopsi deconvolution method by CNMFe was

employed in all datasets. In some cases we selected more stringent parameters to account

for different imaging quality across animals. To compare neuronal activity during home

cage exploration and running in the linear track in each session, 7450 frames of calcium

imaging during the 20

-minute linear track task and 3750 frames during home cage

exploration were analyzed simultaneously. Individual session analysis was used to

compare neurons active during home cage exploration and linear track as well as for

determining whether neurons were active every day using pixel intensity correlation. In

addition, field stability across days was determined from datasets where up to 200,000

frames of linear track or home cage data was analyzed simultaneously with CNMFe. The

resulting neuronal activity was also utilized to confirm registration procedures described

below. Data analysis was performed using the neural activity outp

ut (

neuron.S

) from

CNMFe.

Data analysis

Neuronal activity extraction

: using CNMFe we extracted the background

subtracted raw calcium activity of each neuron (neuron.C_raw), the deconvoluted

calcium activity (neuron.C), and the neural activity (neuron.S). The footprint of a neuron

was recovered by selecting the contour from the neuron.Coor matrix generated by

CNMFe using a threshold of 0.6. This threshold selects the most intense pixels (60

percentile) as the contour of the neuron from the complete cont

our extracted by CNMFe.

The peak to noise ratio (PNR) was calculated by a built

-in function from CNMFe or by a

custom script, both approaches calculate the PNR by obtaining the maximum signal of

the deconvoluted calcium transients (neuron.C) and dividing it by the standard deviation

of the noise of each neuron. The noise level for each neuron was calculated by

subtracting the deconvoluted calcium activity from the raw calcium activity (noise =

neuron.C_raw – neuron.C). We noticed that not all regions of int

erest (ROIs) generated

by CNMFe can be considered neurons with certainty, some may be segmented dendrites

or background fluctuations (

Fig. S2b-c

). We removed these ROIs from the pool of

registered ROIs by selecting only those with areas between 30 to 250 pixel

and inverse

circularity less than 4.0. To determine these cut-

off parameters, we plotted the overall

distribution of PNR, Area, and circularity for each dataset and selected the values so that

only the most linear part of the distribution would be included in the data. The same

parameters were used for all animals. The stability of a neuron could also be affected by

the detection limit of our microendoscope, to avoid bias towards instability we only

analyzed neurons whose PNR was larger than 8. Thus, in concatenated datasets only

neurons satisfying these shape and signal constrain on at least one session were included

in our analysis. In individual datasets, when comparing home cage exploration and track

data, the neuron must pass such criteria at leas

t in one of the environments in one session.

Throughout the manuscript we refer to “firing rate” of a neuron. These values are not the

number of action potentials per second. Instead, “firing rate” here refers to the binarized

neural activity (neuron.S) so that any non-zero value was set to 1, summed over the

complete recording and divided by the recording time (20 minute track or 10 minute

home cage).

dentifying active cells

: a cell was defined to be active if the maximum amplitude

of the deconvoluted calcium activity (neuron.C) was larger than 3 standard deviation of

the noise of that neuron (neuron.C_raw minus neuron.C) in that session. The standard

deviation of the noise and maximum amplitude of the spike was calculated for each

session individually. Neuronal activity below 3 standard deviation (approximately 20 ±

11 % of all spikes) were not used to determine if a neuron was active or not but were

included in all other analyses. Alternatively, we have also generated motion corrected

images and videos of the correlated pixel fluctuations to support the results that most

ROIs have some level of activity on most days (

Fig. 1G, S4, and Movie 3, 4, 6

Correlation images provide an alternate visual inspection of activity per session and are

independent of the deconvolution ability of CNMFe or registration accuracy across days.

Identifying place cells

: response fields of place cells 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). The average firing rate of a neuron in each bin was calculated by the

sum of all

calcium activity in a bin divided by the amount of time the mouse spent in that bin. The

average firing rate of a neuron was then normalized by the total number of spikes in order

to generate normalized tuning profile of each neuron. Neurons wer

e classified as place

cells if: (1) the place field is at least 15 cm wide; (2) calcium transients were present > 30

% of individual lap traversals through the place field; and (3) the cell contains

significantly greater spatial information than chance. Spatial information is calculated

using (

) :

푆푆푆푆

휆휆

� 휆휆

푖푖

log

�

휆휆

푖푖

휆휆

�푃푃

푖푖

here λ is the overall average calcium transient rate of the cell, λ

is the average

calcium transient rate in spatial bin

and P

is the probability the mouse is in spatial bin

Chance level spatial information for each neuron is calculated by shuffling the time

stamps of the calcium transients and calculating the spatial information of the shuffled

neuronal activity, this is done for 1000 iterations. The spatial information of the cell is

considered significant if it is higher than 95% of the shuffled traces.

Identifying end cells

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

off once the animal activated the IR sensors at 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 32 frames or the time the animal

was immobile (velocity less 1 cm/sec, whichever was smaller) was used for analysis.

Neurons were classified as end cells if: (1) they fired at least 20 % of the times the animal

activated the water port; (2) the neuron fired 30% more within its field than outside

(within a field if defined as ± 10 % from the maximum of amplitude of the tuning curve);

and (3) the cells contain significantly greater information than that encoded in a dataset

where spikes where randomized. Information content was calculated using the same

equation above but using λ as the average activity during immobility, λ

is the average

acti

vity in frame

after activation of the water port. The variable P

in this case represent

the probability that the animal was not moving during frame

uantifying stability

: we investigated the stability CA1 representation by analyzing

neurons which were classified as place/end cells through several metrics including: (1)

cell overlap, (2) fraction consecutively active, (3) centroid shift, (4) directional stability,

(5) similarity, and (6) field correlation.

Cell overlap

is the fraction of place/end

cells on one day that remain classified as

place/end cells N days apart. In this approach, for each animal we consider all possible

session pairs that are separated by N days. For instance, session 1 and 2 (N=1), 3 and 7

(N=4), 10 and 16 (N=6), etc. We the

n calculate the fraction of place/end cells that remain

classified as such for each session pair at each N day interval and pool the data across

animals. It is expected that this approach would yield more samples on sessions 1 day

apart than on 50 days apart, which leads to larger error towards long timescales in Fig.

2A. Periods of learning, trained, and re-

exposure were analyzed simultaneously. Damage

and recovery periods were not included.

Fraction consecutively active refers to the number of neurons that were classified as

place/end cells on consecutively recorded sessions (Fig. 2A, right panel). For each neuron

we calculated all consecutive sessions it was classified as a place/end cell and counted

how many neurons were place/end cells on one session only, two sessions, three session,

etc. For instance, a given place/end cell could have a response field on one session, lose it

the next session, comeback for 3 sessions consecutively, lose it for one day, and then

become responsive again for 4 more sessions consecutively. Such a neuron would be

classified as being three and four sessions consecutively active. The probability

distribution for 2 to 40 consecutive sessions was then calculated. Neither of these two

metrics are sensitive to changes in response

field position, thus a place cell changing

directional preference or peak of activity in the linear track on two sessions would be

considered to be overlapping and consecutively active.

Centroid shift was defined as the difference between the centroid pos

ition of

place/end cells in two sessions at N days intervals. The centroid position was obtained by

averaging the normalized response field of each place/end cell above a threshold of 0.5.

When neurons had multiple peaks in their response fields, we only considered the

centroid closer to the reward port the animal was running towards. We then plotted the

distribution of centroid shifts at 1, 10, and 20-

day intervals as well as the fraction of

place/end cells shifting their centroids by less than 10% as a function of day intervals. For

place cells, 10 % is equal to 5 bins (15 cm) and for end cells is equal to 3 frames (480

ms). The cell overlap, fraction consecutively active, and centroid shift were compared to

a distribution in which the same number of place

/end cells randomly gained/lost a

response field at random positions in the maze or time after arrival at the reward port.

Directional stability

is the fraction of neurons (shown in %) retaining a response

field with the same directional preference between

two sessions. Directional stability was

calculated by finding all place/end cells that had retained the same directional preference

between two sessions and dividing by the session with fewer number of place/end cells.

Similarity

: for any two sessions se

parated by N days we found the place/end cells

that retained their classification as place/end cells with the same directional preference.

For every place/end cell we calculate the correlation of their response field between the

two sessions N days apart a

nd counted what fraction of these neurons had a correlation

above 0.4. The threshold of 0.4 was selected based on the average correlation observed in

naïve animals exposed to the linear track.

The same calculation was performed for

place/end cells randomly

selected between two sessions at the same interval. The field

similarity was obtained by subtracting the fraction of neurons with correlation above 0.4

in the random sample from the actual data.

Field correlation

was calculated similarly to the field similarity but instead of

calculating the fraction above 0.4 correlation we calculated the median correlation across

all neurons retaining their classification as place/end cells with the same directional

preference b

etween two session. In this case, the random correlation level was not

subtracted but it is shown in

Fig. 2f and 3f

Identifying rotated sessions

: we noticed that in some sessions the direction

selectivity of a large number of place/end cells would change abruptly. That is, place/end

cells with response fields when the animal runs in the right direction and arrives at the

right side of the maze in one session would change this rightward preference to a leftward

preference in the next session, and vice vers

a. To quantify the magnitude of this effect we

defined a session to have undergone a rotation using two methods. Method 1: a

representation in a session was classified to be rotated if (1) place/end cells in one session

with preference in one direction would be classified as place/end cells with preference in

the opposite direction on the next session; (2) the number of place/end cells changing

their directional preference was larger than the number maintaining their preference; and

(3) reassigning the dire

ctional preference of place/end cells in a rotated session improved

the directional stability by 2

-fold or more. This definition was used in

Fig. S9a

Alternatively, we also calculated changes in directional preference by considering only

place/end cells t

hat retain a field response across two consecutive days. Method 2: (1) we

determined if a place/end cell had a left, right, or bidirectional firing preference; (2)

changes in classification from place to end cell or vice versa were not considered as a

chan

ge in directional preference; and (3) only cells with fields on two consecutive

sessions or more were analyzed. Cells that underwent changes in classification from

place to end cells were not excluded from the analysis. Next, in each two consecutive

sessio

ns we determined the number of place/end cells changing the direction of their field

and divided it by the total number of cells passing our criteria. We refer to this parameter

as the flipping ratio, which fluctuates from 0 to 1 with a clear bimodal distr

ibution (

Fig.

S9b

). Sessions with flipping ratio higher than 0.5 were classified as being rotated

Fig

S9c.

The two methods identified 98 % of the same sessions (90 and 92 sessions out of

330 sessions).

Network graphs

: an adjacency matrix was generated by calculating the pairwise

Pearson correlation between neuronal activity (neuron.S) of all neurons during a 20

minute linear track or 10

-minute home cage recording. Only statistically significant

(p<0.05) correlation va

lues above 0.10 were used to build the adjacency matrix. The

correlation threshold was also tested at 0.15 or 0.05 and similar results were observed

(Fig. S21)

. The weight of the edge between two neurons was set to be equal to the

correlation coefficient. The networks are plotted with the network graphical tool (Gephi

0.9.2, www.gephi.org

) using the built

-in ForceAtlas2 layout. Modularity was calculated

with a Markov diffusion built

-in function from Gephi using a diffusi

on time of 0.7. Cell

groups are defined as cells making up each module extracted by the modularity

calculation. All the modules or the 11 largest modules (whichever was smaller) were

analyzed and a neuron was classified as belonging to one of the cell groups or none.

Unless stated otherwise, cell group participation was not changed across days and

neurons were assigned to belong to the same cell group across days. The connectivity of a

node in a graph was calculated by counting how many edges a node made wi

th all 11

modules. The cell group connectivity was defined as the fraction of edges from node A to

other nodes residing in the same cell group divided by all edges from node A. Metrics

including degree, average clustering coefficient, eigenvector centralit

y were calculated

using built in and custom MATLAB functions.

Synthetic dataset:

it is possible that the observed correlations of neuronal activity

could be explained by random coincidence of activity arising from field overlap between

place/end cells. T

o test if random firing of neurons within their field could lead to high

levels of correlation, we generated a synthetic dataset sharing similar properties as the

original. In this synthetic dataset the spikes of all neurons from the original dataset were

randomly shuffled in time. Spikes in the synthetic dataset were generated with a uniform

random distribution so that: 1) in each 20 minute session, the real and synthetic neuron

maintained the same number and amplitude of firing events as the original; and 2) the

position in the maze were a spike occurred was preserved. This was achieved by

discretizing the mouse position in the maze into 450 bins, assigning a value between 1

and 449 when running right and a value of

-1 to -

449 when running left. At the ends of

the maze, when the animal is not moving, the position was assigned to be -

450 at the left

side of the maze, +450 at the right. Then, for each spike of each neuron we did the

following: 1) find in which frame the neuron fired, 2) get the position in the maze when

that firing occurred (using values defined above), 2) find all the times the animal was at

that position in the maze during that session, 4) randomly reassign the spike to one of the

times the animal was in that position but the neuron was not active. This approach would

not significantly affect neurons that fired exclusively at one position every time the

animal was there. However, these cells are rare, since even robust place cells would show

a variability in the location along the maze where they fire.

Linear decoding of behavior with place/end cells:

The position of a mouse on frame

on day

was decoded using the neuronal activity (neuron.S) of place/end cells in a

moving window between frames

-5 and

. This resulted in an n x 5 training

matrix (n

being the number of place/end cells and their signal in 5 frames). The training matrix was

used as an input into a generalized linear model (

glmfit,

normal distribution) with the

mouse position at frame

as the target output. The mouse position was discretized in 50

bin intervals, each bin is 3 cm. The linear model was trained with randomly selected 10

minutes of place/end cell activity and mouse position. Decoding mouse position on the

same day was performed with the remaining 10 minutes of dat

a. For time

-lapse decoding

the linear decoder was trained with the full 20 minutes of place/end cell activity on day

A. The linear model was then used to decode mouse position on day B using the neuronal

activity on day B of the same place/end cells used t

o train the decoder on day A. The

decoder was also tested on a random dataset in which the neural activity of each neuron

was randomly shuffled in time.

Linear decoding of behavior with cell groups:

Decoding mouse position using cell

groups was performed in a similar fashion, but instead of using place/end cell activity we

used the summed neuronal activity (

neuron.S

) of neurons making up each cell group. Cell

group activity was determined at session N by: (1) calculating the maximum projection of

the adjac

ency matrix across 5 trained sessions (sessions 5

-10), (2) calculating the cell

groups using Markov diffusion on the maximum projection of the adjacency matrix, and

(3) assigning the same cell group to all sessions. The IDs of neurons making each cell

grou

p were used to calculate their integrated group activity in all other sessions. Thus, we

used the

neuron.S

output from CNMFe, identified cells in a group, and integrated the

neuronal activity for each group. Resulting in a 11 x 5 training matrix composed of signal

from 11 cell groups in a window of 5 frames and one target (discretized mouse position)

as a training set. Distribution of training data in time

-lapse decoding using cell groups

was performed similarly to place cell decoding. In both decoding approaches, the output

of the linear decoder was smoothed by moving window average of 5 frames. The mean

absolute error between the decoded and real mouse position was used to quantify the

accuracy of the decoder. The accuracy of time-

lapse decoding was calcu

lated by taking

the median across all mice in both hemispheres.

Nonlinear decoding of behavior with cell groups:

To account for the nonlinear and

asymmetric nature of fields encoded by cell groups we also performed position decoding

using a nonlinear input

-output time series neural network (

timedelaynet

). This network

was set to have 10 hidden layers and a delay of 800 ms (5 frames). The network was

trained with 70% of randomized cell group activity (11 inputs) and mouse position,

validated with 15 % and te

sted with 15% of the remaining data. The network trained in

one session was used to decode N sessions apart.

Decoding behavior with graph topology:

we also decoded mouse position by

approximating the mouse position based not on which neurons are active but rather on

which functionally connected neurons in a graph are active. For day A we extracted

where in the linear track each neuron was active and calculated the preferred firing

position of each neuron by taking the median position of all its spikes. Thus, each neuron

was assigned a positional preference in the maze ranging from 0 to 50 (3 cm bins). Then,

the neuronal activity (neuron.S) was used to calculate the pair

-wise neuronal correlation

and build a network graph where functional connections between two neurons was

proportional to the Pearson’s correlation of their activity. The mouse position was

decoded at frame i by integrating the activity of each neuron between frames

and

-5.

Neurons with nonzero neuronal activity in the 5 frame window were c

onsidered active.

We then found the nearest neighbor of each active neuron in the graph using Dijkstra’s

Shortest Path First algorithm in which the weight of an edge between two nodes is set as

the distance. Only near neighbors that are also active in fram

-5 and within a radius

of 0.5 were considered. We modified each neuron’s preferred firing position in the time

window by averaging it with the preferred firing position of the nearest neighbors. This

approach ensures that the decoded position from a neuron that is highly interconnected

would be weighed more than the position decoded from a neuron with sparse

connectivity in the graph. Finally, the mouse’s decoded position at frame

was

determined to be the median of the preferred firing position of all neurons active and their

neighbors. We compared the performance of all decoders with a randomized dataset in

which spikes were randomly shuffled in time. The error reported represent the mean

absolute error ± sem between the mouse position and the decoded position. All periods

including learning, trained, and re

-exposure were decoded simultaneously. Sessions with

rotated representations significantly decreased the decoders efficiency b

ut were not

removed from the analysis. Diagrams showing different decoding approaches are shown

in Fig. S14.

Classifying cells using graphs

: here we seek to identify whether classification of

neurons as place/end cells or neither can be achieved on a day based solely on synchrony

and graph topology without any behavioral input. To achieve this, we first obtained the

cell group classification of each neuron from the network graph using the Markovian

diffusion approach (see above). We also calculated the cel

l group connectivity of each

neuron, a metric which determines how many edges a node makes with its own cell group

divided by all the edges it makes with all other neurons. Cell group connectivity is

determined from the graph connectivity in each session. The cell group classification of a

neuron was updated if in one session its cell group connectivity towards another cell

group increased above 0.5, otherwise it remained in the same cell group. We observe

approximately 80 % of place cells do not drift into other cell groups. We then assigned

end cells an index of 1, place cells an index of 2, and cells with no statistically significant

response field an index of 0. Cells classified as place and end cells were not analyzed. On

one day, we ranked each cell gr

oup based on how many place/end cells each cell group

contained. We assigned the two cell groups with the most place cells as a place cell

group, and the two with the most end cells as the end cell group. On another session we

identified the cells belongin

g to the original place/end cell groups as place/end cells.

Cells that had a cell group connectivity of 0 or were not part of the place/end cell group

were set as having no response field. Note that behavioral information only from one

trained day is neces

sary in order to assign a label to each cell group. We compared the

accuracy of classification by comparing how many correct classifications arise by using

firing rate. In this approach, we discretized firing rate into 11 bins and performed the

same analys

is on these binned firing rates as we did with cell groups.

Decoding centroid shifts using graphs

: we observe that response fields of place/end

cells drift from day to day. Here we investigated whether these drifts can be decoded

from changes in graph to

pology without the need of behavioral information. First, we

calculated the centroid of all place cells on each session, end cells and cells without a

response field were not included. We then calculated the network graphs from neuronal

activity of all neu

rons in two sessions separated by N days, session A and B. From

neuronal activity on session A we identified the nearest neighbors in the graph within a

radius of 0.5 of all cells having a nonzero cell group connectivity. For each neuron,

regardless of its

classification as place/end cell, we then calculated the mean centroid

position of all its place cell neighbors. On session B, we performed the same analysis

using neuronal activity from that session but the centroid of the place cell neighbors was

the same as that determine in session A. Thus, every cell in session B was assigned a

neighbor centroid as long as it had neighbors who were place cells on the session used as

training. The error of this decoding approach was determined by taking the mean absolute

deviation of the decoded centroid and the experimentally determined response field

centroid on session B. Without knowledge of behavior on session B, the best estimate is

to assume that response fields do not drift. Thus, we compared our decoder with the error

observed if fields were maintained constant from session A to B.

Predicting future response field drifts

: here we investigate whether future changes in

response fields are to some extent encoded during performance of the task. In this case,

we us

ed neuronal activity and behavioral data from one session to make a prediction

about another session N days in the future. We hypothesized that if a place cell has near

neighbors in a graph whose response field centroid is different than its own then it is

more likely to drift its field in the direction of the average centroid of its neighbors. Thus,

using data from session A we subtracted the near neighbor centroids of a place cell from

its own response field centroid. We then plotted all the actual field drift between two

session as a function of the neighbor

-node differences and fitted the data to a linear

equation. Linear fitting was performed on less than 10% of randomly selected data from

all the mice. The linear relationship allowed us to make predictions about future field

drifts using data from one session. We compared the linear predictions to that obtained by

the linear relationship between field drift and firing rate, random guess of the field drift,

or assuming no field drift. In all cases the si

mple neighbor

-node linear relationship

outperformed the other prediction models.

Histology

Mice were perfused transcardially with 4 % paraformaldehyde in phosphate

buffer saline solution at pH 7.4 (PBS). Brains were extracted and fixed overnight at 4 ºC

in the same PFA solution. Brains were embedded in 3 % agarose and sectioned in a

vibratome into 80 μm thick sections. Slices were washed in PBS for 30 minutes at room

temperature, permeabilized for 15 minutes in PBS+0.3 % Triton X100, and washed again

wit

h PBS. Slices were blocked for at least 1 hour at room temperature in PBS+10 % fetal

bovine serum solution (FBS) before incubation with primary antibodies in PBS+1% FBS

overnight at 4 ºC. Chicken polyclonal anti

-GFP (1:1000 dilution, Aves Lab, GFP

-10-

10)

and mouse monoclonal anti

-RGS14 (1:250 dilution, UC Davis, 75-

170) antibodies were

used to label GCaMP6s positive or CA2 neurons, respectively. After incubation

overnight with primary antibodies, slices were washed three times with PBS and

incubated for 90 minutes with secondary antibodies (Alexa Fluor 488 goat anti

-chicken,

and Alexa Fluor 555 anti

-mouse, Invitrogen) diluted to 1:1000 in PBS+1% FBS. Lastly,

sections were mounted on a glass slide with Fluoromount (F4680, Fluoromount Aqueous

Mounting Medium).

Quantification and statistical significance

All statistics were done with nonparametric and parametric approaches either

using a Wilcoxon rank-

sum test, two

-way ANOVA, or a 1000 iterations of a bootstrap

shuffling procedure. Correlation p-

values were calculated using a two

-sided t

-test. The p

values for the linear regression fits (R

-squared values) were obtained by testing whether

the model fits better than a constant (F

-statistic vs constant model). Measurement of

effect size were performed with the par

ametric Cohen’s d metric or the nonparametric

Cohen’s U3

(see supplementary Table S1 and S2)

. Values are shown as median ±

standard deviation (SD) unless stated otherwise. We did not record from 3 CA1 regions

after the no

-task period and data from these mi

ce were only included when analyzing

learning and trained periods.

Data availability

Sample raw data and processed data is available at

(

https://doi.org/10.22002/d1.1229

). Custom MATLAB scripts, fusion360 C

AD files for

the design of the custom microendoscope, and CNMFe parameters used are available at

(

https://doi.org/10.22002/d1.1229

Supplementary

Text

Cell Registration across days

Accurate alignment of neurons across sessions spanning days to months has

noticeably become a challenging aspect of calcium imaging. Taking from studies using

multiphoton imaging, registration approaches using single photon epifluorescence

involve alignment of neur

ons between sessions based on their spatial footprint. However,

the spatial footprint of extracted neurons using microendoscopes can suffer from artifacts

arising from changes in firing rates, extraction algorithms, GRIN lens aberration, and

motion correct

ion artifacts. These artifacts can be compounded when dense cellular

labeling, as observed in transgenic animals, is analyzed. To overcome these limitations,

we employed two novel approaches: 1) interleaved analysis and 2) batch analysis.

CellReg registration assigns scores to each registration, thus providing a useful

metric to asses registration confidence (

)

. However, validating registration across days

using such approaches is not possible since no ground truth is available. Alternatively,

the statistical nature of the CNMFe and CellReg algorithms makes it possible to analyze

how minor changes in neuronal footprint can affect signal extraction and alignment. We

explored the effect of analysis noise by using a 20-

minute recording of CA1 activity in a

mouse running in the linear track. We generated two videos, one is the original 20-

minute

video, and another is a temporally concatenated version of t

he original video (40 minute

long). We motion corrected and analyzed both videos separately using CNMFe as

described in the methods section. Temporal concatenation of the video increases the

number of frames and changes the statistics used by CNMFe to extr

act neuron footprints,

thus leading to minor changes in the footprints. CNMFe effectively extracted similar

fraction of ROIs from both videos (649 in original vs 646 in concatenated). The footprint

extracted by CNMFe was thresholded so that only the pixels

above the 60

percentile

formed the ROI footprint. Next, we tested whether CellReg was able to register ROIs

from these two videos using the spatial correlation of their footprint. Using a 0.5

confidence level in CellReg we found that CellReg registered 85 % of ROIs (549 out of

646) and at a confidence of 0.95 the fraction dropped to 76 % (492 out of 646). The

misaligned ROIs also led to 15 % (0.5 confidence) and 24 % (0.95 confidence) increase

in the number of ROIs. Thus, we observe that minor fluctuations introduced by CNMFe

can lead to inaccurate registration of neurons across days when only using footprints. To

investigate if cell registration across days could be improved we tested two additional

method of analysis.

Interleaved analysis

: in this approach a session of data was common between two

days to be aligned. This common session served the purpose of decreasing noise in the

extracted neuron footprints and provided firing data common to both datasets. Minor

motion artifacts between two sessions wer

e corrected by a fast Fourier transform method

using the PNR image output from CNMFe. The spatial correlation among all neurons

with centroid distances below 15 pixels between the two alignment datasets was

calculated. The correlation of the CNMFe deconvol

uted 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, respectively, were considered. An alignment

coefficient was calculated by multiplying the spa

tial correlation and temporal correlation

of every potential neuron pair. The alignment coefficient was binned in 100 intervals and

the probability distribution calculated. A plot of the probability distribution showed a

clear bimodal distribution (Fig. S2f), a threshold of approximately 0.5 was selected

manually. All neuron pairs below this threshold were deleted, and the remaining were

aligned based on an iterative selection of pairs with the highest alignment coefficient.

This approach was validated by a

ligning a video to itself, where we observed that

including temporal information increased aligning accuracy by about 10%.

Batch analysis:

Here, we take advantage of small motion artifacts across days in

order to motion correct and analyze several days together in one animal. This ensures,

that even if the neuron decreases its firing rate significantly, CNMFe will still draw an

ROI around the neuron and extract its firing activity or show that it is not active. Because

in some cases the FOV drifts in a non

-rigid manner, we restricted our analysis to less than

30 sessions. In some figures we show the correlation image of neurons per session

(Fig.

1g and S4a)

. These images were obtained by analyzing videos of CA1 calcium activity

during home cage exploration

and linear track (30

-minute videos). The resulting

correlation image from CNMFe was then motion corrected using a Fourier transform

method as mentioned above.

Using interleaved analysis, we investigated whether using footprint and the activity

profile of a neuron can enhance the registration across days. CellReg uses two metrics

(spatial correlation and centroid distance) in order to build a distribution of potentially

same or different cell pairs. However, these metrics are not fully independent, potentia

lly

leading to significant overlap and high uncertainty in the registration. Registration can be

improved by temporally concatenating two days and setting the activity of a neuron as a

constraint for alignment, a metric which is independent and orthogonal to both centroid

and shape. However, this requires the ability to analyze multiple days simultaneously,

which is something we can easily achieve using chronic implants thanks to the minimal

drift across consecutive days. To test the effectivity of this app

roach, we performed the

same analysis as above but including in the analysis that neurons should not only look

similar (0.5 spatial correlation) and be located in the same region (<15 pixels centroid

distance) but should also have similar activity (0.3 tem

poral profile). Using this method,

we can align 93 % of all the ROIs (599 out of 646 ROIs), leading to 7% increase in the

number of total ROIs. There were 80 ROIs not satisfying our shape and PNR constrain

(see above) in the original video and 66 in the concatenated version. After removal of

these ROIs, the fraction aligned using our method (90 %) or CellReg (82 %) did not

change dramatically. Using our method, we calculated that analysis of two versions of the

same data by CNMFe resulted in footprints with

areas differing by as much as 19 ± 18

pixel

We tested whether the improvement persisted when analyzing multiple sessions. In

this case, we analyzed three 20

-minute recordings of CA1 activity in the linear track

separated by 1 day

(Fig. S2a)

. Two sets of recordings were generated. Set A contains

temporally concatenated recordings of day 1 and day 2 (in that order) and set C contains

recordings from day 2 and 3

(Fig. S2d).

The recording of day 2 is common to both sets

and the calcium activity of such neur

ons can be assigned as a constrain on the alignment.

In batch analysis we concatenated all three videos into a 60

-minute video (Set D)

and analyzed it with CNMFe. Set D can be used as validation, in this set we observe that

most (97 %) ROIs are active in all 3 days and there is a total of 731 ROIs

(Fig S2d)

CNMFe extracted 702 ROIs in Set A and 700 ROIs in Set C. CellReg is able to align 74

% of these ROIs at 0.5 confidence (521 out of 700) and 63 % at 0.95 confidence (442 out

of 700). Due to the misalignment, a total of 881 ROIs are detected in both sets, an error

of 21 % when compared to the 731 ROIs detected by CNMFe. Using the interleaved

method, we can align 87 % of the ROIs (607 out of 700) and identify a total of 794 ROIs

(13 % error compared to 731 ROIs by CNMFe). CNMFe found that 97 % of ROIs are

active all three days, the interleaved approach identifies 88 %, and CellReg found 72 %.

All three approaches lead to the same answer; most neurons are active all three days and