A genetically tractable jellyfish model for systems and evolutionary neuroscience

A genetically tractable jellyfish model for systems and

evolutionary neuroscience

Brandon Weissbourd

a,b,c

Tsuyoshi Momose

Aditya Nair

a,b,c

Ann Kennedy

a,b,c,e

Bridgett

Hunt

a,b,c

David J. Anderson

a,b,c,f

Division of Biology and Biological Engineering

Howard Hughes Medical Institute

Tianqiao and Chrissy Chen Institute for Neuroscience, California Institute of Technology,

Pasadena, CA 91125, USA

Sorbonne Université, CNRS, Laboratoire de Biologie du Développement de Villefranche-sur-mer

(LBDV), 06230 Villefranche-sur-mer, France

Current address: Department of Neuroscience, Feinberg School of Medicine Northwestern

University, Chicago, IL 60611, USA

Lead contact.

Summary

Jellyfish are radially symmetric organisms without a brain that arose more than 500 million years

ago. They achieve organismal behaviors through coordinated interactions between autonomously

functioning body parts. Jellyfish neurons have been studied electrophysiologically, but not at the

systems level. We introduce

Clytia hemisphaerica

as a transparent, genetically tractable jellyfish

model for systems neuroscience. We generate stable F

transgenic lines for cell type-specific

conditional ablation and whole-organism GCaMP imaging. Using these tools and computational

analyses, we find that a diffuse network of RFamide-expressing neurons is functionally subdivided

into a series of anatomically cryptic, spatially localized subassemblies whose selective activation

controls directional food transfer from the tentacles to the mouth, revealing an unanticipated

degree of structured neural organization in this species.

Clytia

affords a platform for systems-level

studies of neural function, behavior, and evolution within a clade of marine organisms with

growing ecological and economic importance.

Correspondence: bweissb@gmail.com (B.W.), wuwei@caltech.edu (D.J.A).

Author contributions

B.W and D.J.A. conceived of the project and wrote the manuscript, with input from T.M., A.N., A.K., and B.H. B.W., D.J.A, and

B.H. designed and performed histology, behavior, and imaging experiments, and B.W., D.J.A, and T.M. designed and performed

experiments establishing transgenesis. B.W., A.N., and A.K. analyzed the data; A.N. performed NMF/ICA, subspace, and GLM

analyses, A.K. generated the neural network models.

Publisher's Disclaimer:

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our

customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review

of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered

which could affect the content, and all legal disclaimers that apply to the journal pertain.

Declaration of Interests

The authors declare no competing interests.

HHS Public Access

Author manuscript

Cell

. Author manuscript; available in PMC 2022 November 24.

Published in final edited form as:

Cell

. 2021 November 24; 184(24): 5854–5868.e20. doi:10.1016/j.cell.2021.10.021.

Author Manuscript

Graphical Abstract

In Brief

A jellyfish species is developed as a genetically tractable neuroscience model, in which the

application of GCaMP imaging and cell type-specific ablation has revealed spatially restricted

neuronal subnetworks controlling feeding behaviors.

Introduction

Jellyfish offer insights into the structure, function, and evolution of nervous systems: they

are apparent “living fossils,” whose last common ancestor with bilaterians emerged just

after the appearance of neurons (Figure 1A;

Arendt et al., 2016

;

Cartwright et al., 2007

Jellyfish use neurons homologous to our own (

Arendt et al., 2016

;

Bosch et al., 2017

), but

lack centralization, i.e., “brains”. How such organisms are able to feed themselves, navigate,

escape from predators, and even sleep (

Mackie, 2004

;

Nath et al., 2017

;

Lewis and Long,

2005

;

Meech, 2019

) in the absence of a central brain poses an important problem in the

field of evolutionary neurobiology, with implications for autonomous systems engineering

(

Nawroth et al., 2012

). Jellyfish are also attracting growing interest as critical components

of ocean ecosystems, in part due to jellyfish blooms and their negative economic impact

(

Condon et al., 2013

;

Graham et al., 2014

;

Hays et al., 2018

Weissbourd et al.

Page 2

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

Despite the importance of jellyfish to evolution, ecology, and economics, remarkably little

is known about the neural control of their behavior. Jellyfish neurons have been studied

extensively using single-unit electrophysiological recordings (

Meech, 2019

;

Satterlie, 2002

but systems-level analysis has been absent due to the lack of a genetically tractable model.

Attractive features of jellyfish for systems neuroscience include their small size, relative

planarity, and transparency, facilitating optical approaches (

Katsuki and Greenspan, 2013

;

Meech, 2019

;

Bosch et al., 2017

). Existing cnidarian genetic neuroscience models, such

Hydra

(

Dupre and Yuste, 2017

;

Tzouanas et al., 2021

;

Badhiwala et al., 2021

), lack a

jellyfish life cycle stage (

Künzel et al., 2010

;

Renfer et al., 2010

;

Wittlieb et al., 2006

Here we introduce the hydrozoan jellyfish

Clytia hemisphaerica

, originally established

to study early development and evolution (

Houliston et al., 2010

), as a genetic model

for systems neuroscience. The

Clytia

genome has been sequenced (

Leclère et al., 2019

)

and an atlas of its cell types generated using single-cell RNA sequencing (

Chari et al.,

2021

). CRISPR/Cas9-mediated gene knockout has been reported (

Momose et al., 2018

) but

transgenesis has not yet been established.

In this inaugural study, we describe the generation of stable

Clytia

transgenic lines

for population neural imaging and neuronal cell type-specific ablation. Using these tools,

we have investigated the neural control of feeding, in which captured food is vectorially

transferred from the margin of the umbrella to the central mouth. We find that directional

infolding of the umbrella is controlled by anatomically cryptic neural subassemblies that tile

the umbrella, uncovering a surprising degree of structural organization within a superficially

diffuse neural net. This work introduces

Clytia

as a genetically tractable model for systems

neuroscience, affording a platform for understanding the neural control of its decentralized

behavior and internal states, and for comparative studies across phylogeny.

Results

Jellyfish (medusae) are a free swimming life stage within the phylum

Cnidaria

and

subphylum

Medusozoa

(Figure 1A–D;

Leclère et al., 2019

Clytia

medusae are small

(~1mm–1.5 cm), optically transparent, and have approximately 10,000 neurons in a 1cm

adult (Figure 1C;

Chari et al., 2021

). Their anatomy exhibits the hallmarks of the hydrozoan

medusa body plan (Figure 1E): nerve rings (Figure 1F), circular and radial muscle

(Figure 1G), and a subumbrellar nerve net (Figure 1H). Importantly, the tri-phasic sexual

reproductive cycle (Figure 1D) can be recapitulated in the laboratory (Methods;

Houliston

et al., 2010

;

Lechable et al., 2020

). Following controlled fertilization, zygotes develop into

a planula (larval) stage. Planulae attach to a substrate (microscope slide) where they form

clonal polyp colonies that release free-swimming medusae. The entire life cycle takes ~6–8

weeks.

Identification and genetic targeting of a neuronal subpopulation marked by RFamide

Neurons immunoreactive for an RFamide (RFa) peptide have previously been identified in

Clytia

(

Mackie et al., 1985

), and have been suggested to play a role in medusa defensive

and feeding behaviors (

Mackie, 2003

;

Satterlie, 2008

2002

;

Weber, 1989

). We confirmed

RFa immunoreactivity in the nerve net, mouth, nerve rings, and tentacles (Figure 2A–C and

Weissbourd et al.

Page 3

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

Figure S1A–G). RFa

neurons comprised a large subpopulation (~80%) of nerve net neurons

identified using tyrosinated tubulin, a generic neuronal marker (

Gröger and Schmid, 2000

;

Figure 2A; Figure S1A–B). They were radially oriented, with varicosities apposed to radial

muscle fibers (Figure 2B; Figure S1C–D). In contrast, RFa-negative neurons in the nerve

net were generally smaller and lacked a clear radial orientation (Figure 2A, arrow; Figure

S1A–B). RFa

neurons in other body parts were less abundant than in the nerve net (~10%

of total neurons in

Clytia

by scRNA-seq;

Chari et al., 2021

To gain genetic access to RFa

neurons, we established transgenesis in

Clytia

(Methods).

Plasmid and Tol2 transposase were co-injected into

Clytia

eggs (Figure S1H;

Koga et al.,

1996

;

Ni et al., 2016

) and polyp colonies screened for expression of an mCherry reporter.

Strongly expressing colonies were cultured until they produced F

medusae, which were

backcrossed to parental strains. ~100% of these backcrosses yielded germline transmission.

Stable F

progeny were maintained and expanded vegetatively as clonal polyp colonies

which released non-mosaic transgenic medusae that could be collected daily (Figure 1D;

Figure S1H).

We first ablated the RFa

neurons to determine the effect on the organism’s behavior. To

this end, we cloned 6.6kb of 5’ flanking DNA from the

Clytia RFamide

precursor gene and

inserted it upstream of a bi-cistronic construct encoding nitroreductase (

NTR

;

Curado et al.,

2008

;

Tabor et al., 2014

) and mCherry, separated by a 2A peptide (Figure 2D;

Daniels et

al., 2014

). Addition of the drug Metronidazole (MTZ) causes autonomous ablation of cells

expressing NTR, which converts MTZ to its toxic form (Figure 2E;

White and Mumm,

2013

The

RFamide

5’ genomic fragment successfully drove strong and specific mCherry

expression in the umbrellar RFa

network (~92% of RFa

neurons targeted, 100% of

targeted neurons were RFa

, Figure 2D). A 24-hr incubation in MTZ efficiently eliminated

RFa

neurons in the TG

RFa

∷

NTR-2A-mCherry

line (Figure 2D, F). Other neuronal populations

were intact, demonstrating that ablation of RFa

neurons was specific (Figure 2G). No

ablation was observed in controls with either MTZ or the NTR transgene omitted (Figure

2F).

RFamide neurons are required for feeding

The umbrella is involved in several behaviors, including swimming, feeding, and

defensive crumpling (

Hyman, 1940

;

Romanes, 1885

). While swimming and crumpling

utilize symmetrical umbrellar contractions, feeding employs asymmetric contractions that

vectorially transfer captured food (brine shrimp,

Artemia

) from contracted tentacles at the

umbrella margin to the elongated feeding organ (“mouth”), which extends from the center of

the umbrella (Figure 1C, E and Supplemental Video 1). Swimming stops (Figure S2F), and

the mouth bends (“points”) towards the infolding portion of the margin to receive the food

(Figure 3A–B; “pointing”). Food transfer is robust: 96% of first transfer attempts occurred

within one minute of prey capture, of which 88% were successful (Figure S2G–H). 96.3%

(52/54) of caught prey were eventually eaten.

Weissbourd et al.

Page 4

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

Remarkably, RFa

neuron ablation completely eliminated asymmetric inward folding and

hence the transfer of shrimp captured by the tentacles (Figure 2H). Ablation also prevented

folding induced by chemosensory stimulation using shrimp extract (Figure 2I). Thus, RFa

neurons are required for both food- and chemically-induced margin folding. In contrast,

swimming and crumpling were unperturbed (Figure 2J–K; Figure S2A–C), suggesting that

other neural cell types control these behaviors (

King and Spencer, 1981

Bath application of synthetic RFa peptide to transgenic medusae following RFa

neuron

ablation caused radial muscle contraction, confirming that the muscle was functionally

intact (Figure S2D). Local infusion of RFa but not control peptide into the subumbrella

caused local muscle contraction and margin infolding (Figure S2E). These data suggest that

local release of peptide from RFa

neurons activates the muscle, either directly or via an

intermediate cell population, to determine the site of margin folding.

Margin folding requires coordination between autonomously functioning body parts and is

influenced by internal state

These results prompted us to investigate margin folding behavior in more detail. First, we

tested the necessity and sufficiency of different body parts using surgical manipulations.

Mouth pointing towards the infolding margin was blocked if the umbrella was cut between

the mouth and the margin, while margin folding was not (Figure 3C). Following excision of

the mouth, the body swims, captures prey, and tries to pass prey to the hole where the mouth

formerly was (Supplemental Video 3). The mouth-less umbrella also performed margin

folding in response to shrimp extract (Figure 3J), which can trigger directional folding when

locally applied (Figure 3D; S2J; Supplemental video 2). Shrimp extract could also trigger

margin folding when tentacles and tentacle bulbs were removed (Figure 3K; S2K). Removal

of other body parts revealed a similar theme of modular functional organization, confirming

earlier studies (

Romanes, 1885

;

Passano, 1973

;

Quiroga Artigas et al., 2018

Clytia

also performed margin folding in the absence of any added stimuli (13.7±15% of time

observed, mean±SD), which was visually indistinguishable from evoked folding behavior

(Figure 3E, Figure S2I and Supplemental Video 1). To quantify margin folding and compare

spontaneous to prey-evoked behavior, we trained an automated classifier to discriminate

margin folding from swimming with high accuracy, using the major and minor axes of the

umbrella as features (Figure 3F–I). When trained on spontaneous folding alone, the same

classifier could identify episodes of induced folding evoked by live shrimp or shrimp extract,

with similar accuracy (Figure 3I).

To investigate whether margin folding behavior was a fixed-action stimulus-response reflex

or was modulated by internal state (

Anderson and Adolphs, 2014

), we examined the effect

of food deprivation. Animals starved for 24 hrs performed food passing significantly faster

than

ad libitum

fed controls, following prey capture (Figure 3L). In addition, spontaneous

folding behavior failed to occur when the animals were spawning (Figure 3M). Lastly,

when multiple shrimp were captured simultaneously they were transferred to the mouth

sequentially rather than coordinately, at a rate higher than chance (Figure S2L–M). Margin

folding is therefore modulated by at least two internal states (metabolic and reproductive)

and involves coordination across different sectors of the umbrella.

Weissbourd et al.

Page 5

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

RFamide neurons are active during margin folding behavior in multiple contexts

To determine how RFa

neuronal activity correlates with margin folding behavior, we

generated a transgenic F

line bi-cistronically expressing the calcium indicator GCaMP6s

and mCherry (TG

RFa

∷

GCaMP6s-2A-mCherry

; Figure 4A;

Chen et al., 2013

). This transgenic

line also had a high efficiency and specificity of targeting RFa

neurons (~94% of RFa

neurons targeted, 100% of targeted neurons were RFa

). Because GCaMP imaging was

partially obscured by umbrellar expression of the endogenous

GFP1

gene (

Fourrage et

al., 2014

), we knocked out this gene using CRISPR/Cas9 (

Momose et al., 2018

) in the

RFa

∷

GCaMP6s-2A-mCherry

genetic background (Figure S3A).

We performed wide-field, two-color,

in vivo

imaging in

RFa

∷

GCaMP6s-2A-mCherry/+

;

GFP1

−/−

transgenic jellyfish, using either restrained or unrestrained preparations to balance

GCaMP signal extraction and naturalistic behavior (Figure 4B–C; Figure S3B–D). In

the unrestrained preparation, animals could behave freely in the small imaging chamber,

allowing identification and localization of active neurons within the subumbrellar nerve

net. However extraction of GCaMP traces from individual cells was not possible due to

movement (“Naturalistically behaving”, Figure 4H–J; Supplemental video 3). Preparations

in which the jellyfish were relatively motionless (“restrained”) using agarose embedding

allowed extraction of high-quality GCaMP traces from single neurons (Figure 5A–C; Figure

S4B; Supplemental video 5). Due to variation in the agarose-embedded preparations (see

Methods), in some animals there was both sufficient restraint to extract single-cell traces

and sufficient freedom of movement to identify apparent attempts at swimming or folding

(“Loosely restrained”, Figure 4D–G).

Movements interpreted as attempted “swimming” comprised high frequency, circumferential

contractions, while those interpreted as attempted “folding” comprised low frequency, radial

movements (Figure 4D; Figure S3E; Figure 3G–H). To quantify these behaviors in such

loosely restrained preparations, we extracted optic flow vectors across frames from the

imaging videos. Classifiers trained on these features could distinguish behaviors in held-out

test frames with high accuracy (swimming, accuracy = 0.85±0.04; folding, accuracy =

0.81±0.03, mean±SEM; Figure S3E). We used the mCherry channel to segment neurons and

exclude that GCaMP signals reflected motion artifacts (Figure 4E; Figure S3H–I; Figure

S4K).

In loosely restrained preparations in which both attempted margin folding and swimming

movements occurred, behavioral epochs and neural activity could be aligned to examine

their temporal relationship (Figure 4E–G). Patches of RFa

neurons in both the nerve rings

and net were strongly activated at the onset of folding, but not swimming (Figure 4E–F).

Neural activity in the patches was relatively synchronous (Figure 4E). Using population

neural activity, we trained a 3-way classifier to predict quiescence, swimming, or folding

and found that folding episodes could be predicted with high accuracy, while swimming

could not be distinguished from quiescence (Figure 4G). Principal component analysis

(PCA) confirmed that the largest source of variance in neural activity occurred during

folding (Figure S3F). These data indicate that RFa

neuronal activity is temporally and

specifically associated with margin folding behavior. The NTR ablation data (Figure 2H, I)

indicate, moreover, that this activity is likely a cause and not a consequence of this behavior.

Weissbourd et al.

Page 6

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

To examine the pattern of RFa

neuronal activation more broadly across the nerve net

under conditions where complete margin folding could be performed, we imaged GCaMP

in the unrestrained preparation (Figure 4H) in three contexts: (1) spontaneously, i.e., no

stimuli given (Figure S2I; Figure 3E), (2) with shrimp extract uniformly present, or (3) with

intact shrimp presented to the tentacles. We manually extracted body shape and identified

the location of active neurons during individual folding events. In both the spontaneous

and evoked contexts, RFa

neurons in the nerve net were activated in radially oriented,

wedge-shaped patches located at the epicenter of margin infolding (Figure 4I–J; Figure

S3H–K; Supplemental Video 4). Thus, RFa

neuronal activity is temporally correlated with

both evoked and spontaneous margin folding, and is spatially localized at the site of each

folding event (Figure 4K).

The observation of local ensembles of synchronously active RFa

neurons raised the

question of whether these neurons activate each other, as well as the radial muscle, via

release of RFamide peptide. To address this question, we perfused RFamide during GCaMP

imaging. RFamide addition did not detectably activate RFa

neurons (Figure S3G). This

suggests that the peptide may not be used for excitatory, inter-RFa

neurotransmission,

but does not exclude the possibility that these neurons communicate via an unidentified

co-transmitter or peptide.

Functional subdivisions within the umbrellar RFamide network

To further examine the pattern of ensemble activity in the RFa

system (Figure 4J–K),

we performed whole umbrella imaging in animals that were well restrained in agarose

(Figure 5A; Figure S4A–B). These preparations revealed spatially localized ensembles of

RFa

neurons that exhibited repeated events of spontaneous, synchronous activity (Figure

5B–C; Figure S4B–C). These ensembles appeared as radially oriented, wedge-shaped

populations that stretched between the margin and the mouth (Figure 5B; Supplemental

Video 5), similar to the pattern observed in naturalistically behaving animals (Figure 4).

The radial organization of ensembles was obvious along the diagonal of a neuronal activity

correlation matrix in which neurons were sorted by their relative angular position (Figure

5D). However, the correlation pattern was not a series of sharply defined blocks, but rather

patches of variable size with diffuse borders, indicating variability in ensemble membership

across individual events.

Closer inspection of the correlation matrix revealed structure on multiple spatial scales, with

very high correlations between neighboring neurons, as well as weaker, distance-dependent

correlations (Figure 5D; Figure S4D–E). To identify core groups of highly correlated

neurons active during repeated events of spontaneous activity, we used k-means clustering.

This revealed striking spatial groupings that roughly tiled the animal (Figure 5E). However,

ensemble membership between individual events was flexible, and many events were not

restricted to single k-means clusters (Figure 5F). This variability suggests that active

ensembles are not defined exclusively by deterministic connectivity rules (see below).

We next asked whether inter-episodic variability in ensemble membership reflected the

flexible recruitment of cells from neighboring clusters. We first plotted the activity of

individual neurons across multiple spontaneous events, grouping neurons by k-means cluster

Weissbourd et al.

Page 7

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

membership and sorting clusters by their relative radial positions (Figure 5G). We then

simplified the results by graphing which

clusters

were active during each event, not which

neurons

were active (compare Figure 5G to 5H). Next, we asked how much of the observed

activity could be explained by ensembles comprising single vs. multiple clusters.

Restricting ensembles to members of single k-means clusters explained much, but not

all, of the observed activity over time (F1 score: ~0.66, Figure 5I, violet bar). However,

allowing ensembles to flexibly incorporate multiple clusters better explained the observed

activity (F1 score: ~0.82, Figure 5H–I). Spontaneous events restricted to single clusters

were only slightly more frequent than those incorporating an additional cluster(s) (Figure

5J). Neighboring clusters were most often co-active (Figure 5K), although not all members

of such clusters participated in each event (Figure 5F–G). Thus, inter-episodic variation in

ensemble membership (Figure 5L) is best explained by a flexible incorporation of neurons

from adjacent clusters into ensembles initiated by a “core” k-means cluster.

Because k-means forces neurons into single clusters, as an alternative way to

capture flexibility in inter-episodic ensemble membership, we used Non-negative Matrix

Factorization (NMF) followed by Independent Components Analysis (ICA) (

Lopes-dos

Santos et al., 2013

;

See et al., 2018

) to cluster active cells. In this method, individual

neurons can be members of more than one cluster. Among neurons assigned to NMF-ICA

clusters, the majority participated in only a single cluster (~89%), with fewer participating in

2 (~10%) or 3 clusters (<1%; n=489 neurons from 4 animals; Figure S4F–G). Thus, whether

cluster membership is defined rigidly (k-means) or more flexibly (NMF/ICA), clusters

emerge as principle units of spontaneous RFa

network activity, with episodic variability in

ensemble membership reflecting fluctuations in nearest-neighbor recruitment (Figure 5F).

A close examination of umbrellar anatomy did not any reveal identifiable structural

correlates on the scale of these subassemblies (Figure S4H–J), and no obvious structure

was evident in the mCherry channel (Figure S4K). Therefore, the spontaneously active

ensembles within the RFa

nerve net appear to be anatomically cryptic, at least at the level

of light microscopy.

Ring neurons act upstream of subjacent net neurons

Spontaneous activity in the nerve rings encircling the umbrella also showed spatial

clustering. Ring clusters were highly correlated with subjacent nerve net clusters (Figure

5M–N; Figure S4L–N). These observations suggested a possible flow of information

between the RFa

neurons in the nerve rings and net. We therefore examined the

directionality of this flow. Nerve net activation always had a nerve ring correlate, but

ring activity did not always coincide with a net event (Figure S5B). Correlated ring

activity was initiated prior to net activity during spontaneous events (Figure 5O; Figure

S5C). Furthermore, neurons in excised fragments of the ring responded to shrimp extract,

but neurons in net fragments only responded if the margin was attached (Figure S5D),

suggesting a dependence of net neurons on ring neurons for evoked activity. Consistent

with this, excising the margin (containing the rings, Figure 1E) eliminated evoked folding

behavior in the remaining umbrellar tissue (Figure S5A). Histology suggested that processes

Weissbourd et al.

Page 8

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

from ring and net neurons were intermingled at the margin (Figure S5E). Attempts to map

process origin by focal dye labeling were unsuccessful.

We further examined information flow by generating transverse incisions in the subumbrella

and monitoring whether spontaneous nerve net activity could be observed on the ring

(peripheral) side of the wound, the mouth (central) side, or both (Figure 5P). We observed

spontaneous activity peripheral but not central to the incisions (Figure 5P, Supplemental

Video 6). Thus, information flows from the rings into the nerve net, where it continues

centripetally towards the mouth (Figure 5Q).

The foregoing data suggested that radial umbrellar fragments containing tentacles, the mouth

and RFa

neurons in both the nerve rings and the net might perform food-passing behavior

autonomously. Indeed, such isolated wedge-shaped sectors were able to pass shrimp from

the margin to the mouth (Figure S5F; Supplemental Video 7).

Spontaneous and stimulus-evoked RFa

ensemble activity exhibits similar patterns

The foregoing analysis raised the question of whether stimulus-evoked ensembles can

emerge at arbitrary locations in the RFa

subnetwork, or whether they are spatially

constrained by the clusters revealed in our analysis of spontaneous events. To distinguish

these alternatives, we quantitatively compared spontaneous to evoked ensemble activity in

the same animal (Figure 6A–C; S5G–H), using the subspace alignment metric (

Elsayed et

al., 2016

;

Yoo and Hayden, 2020

). This metric tests the variance in evoked activity epochs

explained by the principal components of spontaneous epochs.

We observed high subspace alignment between spontaneous and evoked epochs (85 ±

1.21%, mean ± SEM, Figure 6D), which was similar to the alignment between two sets

of spontaneous epochs (83 ± 4%, mean+/−SEM) and was significantly different from

the alignment between randomly generated subspaces (Figure 6E; p=0.0079). These high

alignment indices suggest that evoked ensemble structure may be constrained by an intrinsic

pre-pattern exhibited during spontaneous episodes.

As an independent method for comparing spontaneous to evoked ensembles, we trained

generalized linear models (GLMs) to predict the activity of each imaged neuron using the

weighted activity of all other imaged neurons (

Efron et al., 2004

;

Mishchenko et al., 2011

;

Pillow et al., 2008

). GLMs trained and tested on spontaneous activity recapitulated the

activity of a given cell with high accuracy (83 ± 1.3% mean ± SEM; Figure 6F). More

importantly, GLMs trained on spontaneous activity and tested on stimulus-evoked neural

activity exhibited similarly high accuracy (80 ± 1.4% mean ± SEM, Figure 6F; Figure S5I).

These results further argue that spontaneous and evoked RFa

ensembles share common

structural constraints.

Modeling supports a partially structured nerve net organization

To examine possible circuit implementations of observed ensemble activity, we undertook

a modeling approach. We constructed a series of spiking neural network models, each of

which incorporated different assumptions about underlying connectivity (Figure 7A–D).

At one extreme, we assumed that connectivity strength between a given RFa

neuron and

Weissbourd et al.

Page 9

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

its neighbors declined continuously as a function of their angular distance (“continuous”;

Figure 7B). At the opposite extreme, we assumed all neurons were hard-wired into non

interacting subnetworks, with each neuron connected to all other members of its subnetwork

with equal strength (“structured”; Figure 7C). Finally, we generated hybrid models using

weighted combinations of “continuous” vs. “structured” connectivity (Figure 7D). For each

type of model, we tuned both the strength, and angular extent of, network connectivity to

best fit two observed statistics: the circumferential width of each spontaneous event, and the

percent of neurons recruited per event (see Methods; Figure S6A–E).

The best fit to the data was achieved using hybrid models combining both continuous

and structured connectivity (Figure 7D; Figure S6C–D). A model using only discontinuous

structures had the lowest performance (Figure 7C); some degree of continuous connectivity

was essential for the model to exhibit inter-episodic flexibility in event boundaries, as

observed experimentally. The relative contribution of structured vs. continuous connectivity

did not strongly effect performance, as long as structured connectivity was present (Figure

S6C–D). In all best-fit models, continuous connectivity was sparse and local, with each

neuron only forming synapses with a small number of its nearest neighbors (Figure S6C–E).

These findings support a model in which flexible ensembles are generated by combining

“core” networks with continuously graded, local connectivity.

The foregoing model assumes that local connectivity between net neurons, as well as input

from ring neurons, accounts for the synchronous activity of ensemble members. To evaluate

the relative contributions of ring

→

net vs net

→

net connectivity to this synchronicity, we first

trained GLMs using recordings that included both ring and net neurons (Figure 5). These

GLMs performed with high accuracy (76 ± 1.2% mean ± SEM; Figure S7A). Matrices

of fitted GLM weights between RFa

neurons, i.e. inferences of connection strengths

(

Mishchenko et al., 2011

), revealed that the GLMs used only sparse, local weights (Figure

7E–F), which was not due to GLM regularization (Figure S7B). This pattern of sparse

connectivity was similar to our best fit network models (Figure S6C

–D

To determine whether synchronous firing in the umbrellar net ensembles could be explained

purely by coordinated ring

→

net input, we digitally “ablated” intra-umbrellar network

connectivity from our GLMs, either after or before training but prior to testing (Figure

S7C–F). These “ablated” GLMs performed significantly less well than the “intact” GLMs

(Ablated, 54 ± 2%; Intact, 82 ± 0.6%; Figure S7G–J). These results strengthen the idea that

sparse local net

→

net connectivity contributes to the synchronicity of ensemble activity.

Finally, we examined the relationship between GLM weights (connection strength), and

the angle of connections relative to the mouth. Interestingly, both ring

→

net and net

→

net

connectivity was primarily oriented radially, whereas ring

→

ring connectivity was primarily

horizontally/circumferentially (Figure 7G; net-net vs ring-ring, p = 3.5e–70; ring-net vs

ring-ring, p = 5.7e–15; net-net vs ring-net, p = 0.8). This suggests that the radial and

circumferential extent of ensembles may reflect intrinsic synaptic biases within the net and

ring RFa

subnetworks, respectively.

Weissbourd et al.

Page 10

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

Discussion

Here we introduce the jellyfish

Clytia hemisph

aerica as a model for systems neuroscience.

Clytia

combines optical accessibility and genetic tractability with a varied behavioral

repertoire. We report germline transgenesis in this organism and a systems-level

interrogation of neural activity.

Clytia

offers opportunities to study neural development,

function, evolution, and behavior in an organism that provides a window into the first

nervous systems on the planet.

A jellyfish model for neuroscience

Among cnidarian genetic models,

Clytia

is complementary in several respects. First and

foremost,

Clytia

has a medusa stage, while the others (e.g

Hydra)

are polyps (

Bosch et

al., 2017

). Understanding the particulars of jellyfish behavioral control is relevant to their

growing ecological and economic importance. Second, the medusa stage offers a relatively

rich behavioral repertoire in comparison to polyps (e.g.

Costello et al., 2021

). Finally,

Clytia

affords several advantages over

Hydra

as a genetically tractable model. It allows

routine generation of F

lines, whereas in

Hydra

only F

transgenics have been used for

neural imaging (

Dupre and Yuste, 2017

). Furthermore,

Clytia

husbandry permits inter-strain

genetic crosses, which are currently difficult to achieve in

Hydra

(

Klimovich et al., 2019

Finally, transgenesis is more easily scaled up in

Clytia

, where thousands of unfertilized eggs

can be generated for injection daily. In contrast, the

Hydra

lifecycle limits the availability of

embryos (Figure S1H;

Klimovich et al., 2019

Imaging results in

Clytia

provide an interesting contrast to those obtained in

Hydra

(

Dupre and Yuste, 2017

). The

Hydra

neural net is divided into several functionally

distinct subnetworks, whose activity is correlated with different types of polyp movement

(contraction, extension, bending) (

Dupre and Yuste 2017

). These subnetworks are not

spatially separated, but overlapping. As in

Hydra

, the

Clytia

umbrellar neural net comprises

multiple subnetworks, including the RFa

subnetwork. In contrast to

Hydra

, however, this

subnetwork is further parcellated into functional subassemblies (Figure 7H). These spatially

distinct subassemblies therefore indicate a higher order of neural structural organization in

Clytia

than in

Hydra

(Figure 7J). Further comparisons between

Clytia

Hydra

, and other

cnidarian model organisms should provide important insights into neural network evolution

(

Bosch et al., 2017

Decentralized, modular neural control of organismal behavior

A striking feature of

Clytia

behavior is its extreme functional modularity: for example, an

isolated mouth can ingest food, and margin folding can occur in a “mouth-less” umbrella.

Nevertheless, in intact jellyfish the mouth points towards the infolding margin during food

transfer, revealing coordination between these modules. Such coordination is also observed

during food-passing in wedge-shaped strips of

Clytia

umbrella containing the mouth (Figure

S5F). These data suggest that functionally autonomous behavioral modules (e.g., mouth,

umbrella) are combined to form coordinated “super-modules,” copies of which are arrayed

circumferentially around the umbrella. Such a mechanism accommodates the continuous

growth that jellyfish exhibit.

Weissbourd et al.

Page 11

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

If this hierarchical view is correct, coordinated behaviors in organisms lacking a central

brain may have emerged by duplication and modification of smaller autonomous modules to

form functionally interacting super-modules. How these interactions are achieved remains to

be determined. Over time, modification of super-modules might give rise to new structures

and organism-level behaviors. Such a nested modular organization could therefore be an

important substrate for the evolution of complex behaviors.

Clytia

have a remarkable ability to regenerate and recover behaviorally following injury

(

Kamran et al., 2017

;

Sinigaglia et al., 2020

). They also continuously integrate new stem

cell-derived neurons into their nervous system as they grow, without disrupting organismal

behavior (

Chari et al., 2021

). The local network interactions and modular organization

that we have described may facilitate such continuous growth and repair.

Clytia

affords a

genetically tractable model to investigate the dynamics of such regeneration in real time, a

process at the interface of neural development and systems-level function.

Limitations of the study

Our conclusions are largely based on analysis of spontaneous episodes of neural activity

in immobilized animals. These episodes likely correspond to spontaneous activity with

margin folding events observed in freely moving animals, since spontaneous activity without

folding was not seen. We cannot exclude that such “spontaneous” events are triggered by

microscopic exogenous stimuli, rather than endogenously generated. Furthermore, activity

in immobilized animals may exhibit subtle differences from that in freely moving jellyfish.

Our analyses relied on a finite number of spontaneous ensemble events (~100/animal);

significantly more events could change the observed correlation structure. Lastly, imaging

was performed in juvenile jellyfish: the functional organization described here may change

as the animal grows.

STAR Methods

Resource Availability

Lead Contact—

Requests for resources and reagents should be addressed to lead contact,

David J. Anderson (wuwei@caltech.edu).

Materials availability—

The plasmids and transgenic jellyfish generated in this study are

available upon request. Plasmids will also be deposited to Addgene.

Data and code availability

•

Source data reported in this paper will be shared by the lead contact upon

request.

•

Code used for analyses in this paper will be shared by the lead contact upon

request.

•

Any additional information required to reanalyze the data reported in this paper

is available from the lead contact upon request.

Weissbourd et al.

Page 12

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

Experimental model and subject details

Culture of the

Clytia

life cycle was carried out in accordance with published protocols

(

Lechable et al., 2020

), with the exception of the culture tank design. Circulating systems

used here had the same overall flow design as in

Lechable et al., 2020

, but used

modifications of zebrafish tanks (Pentair) for animal housing. Polyp slides were held in

glass slide racks (Fisher, cat#02-912-615) in small tanks with high water flow. Jellyfish

were cultured in large-sized tanks with a curved plastic insert placed at the back of the tank

with a nylon mesh outlet at the bottom. A slow drip from the circulating system into these

tanks allowed for water turnover without risking sweeping the jellyfish through the outlet.

A constant speed 5rpm DC motor (Uxcell) attached to the lid of a multi-well tissue culture

plate was then used to create a constant circular current. A dimmer switch was used to tune

the rotation speed of the motor, with reduced speeds as the jellyfish grew. Jellyfish used

for transgenesis, and all polyps, were maintained in these circulating systems, while smaller

jellyfish were maintained in beakers. In beakers, current was generated using stirring with a

DC motor, as above. All artificial sea water for culture and experiments was made using Red

Sea Salts (Bulk Reef Supply, cat# 207077) diluted into building deionized water to 36ppt.

Unless otherwise indicated, behavior experiments were performed using sexually mature

animals of the Z4B strain of

Clytia

, which are female. Transgenesis was performed by

crossing Z4B females to Z13 males. For generating experimental F

lines, NTR and GCaMP

lines were backcrossed to Z4B. A single F

polyp colony was then chosen to maintain

for each strain to control for genetic background. Clonal experimental animals were then

collected from these polyp colonies. Experiments and culture were performed at room

temperature.

Method Details

Histology—

For antibody staining,

Clytia

were fixed for 2h at room temperature in 4% PFA

in 0.2um filtered artificial sea water (Red Sea Salts, Bulk Reef Supply, cat# 207077, diluted

into building deionized water to 36ppt). They were then washed 3x in PBS followed by

blocking for 1h in PBS with 0.1% Triton (PBST) and 10% normal donkey serum (NDS).

Animals were then incubated for 1–3 nights in primary antibody with 5% NDS in PBST at

4°. Primary antibodies used in this study were: anti-FMRF (Immunostar, cat# 20091), anti

Tyrosine Tubulin (Sigma cat# T9028–100UL), 647-conjugated anti-

Tubulin, clone DM1a

(Millipore Sigma cat# 05–829-AF647), and anti-

Tubulin (YL1/2; Thermofisher cat# MA1–

80017, Figure S5E). Following primary antibody incubation, one short (~5min) and then

3–4 long (~30min+) washes were performed in PBST, and then animals were transferred

into secondary antibodies and/or Phalloidin-488 (Thermo Fisher, cat# A12379) for 2h at

room temperature or overnight at 4°. Secondary antibodies: donkey anti-rabbit conjugated

to Alexa Fluor 647 or 488 (Jackson ImmunoResearch cat# 711-605-152 and Thermofisher

cat# A-21206), or donkey anti-mouse conjugated to Alexa Fluor 647 (Thermofisher cat#

A-31571). Animals were stained with DAPI (BD cat#564907) and mounted onto glass

slides for imaging. For staining shown in Figure 1F, animals were dehydrated stepwise

into methanol and then rehydrated prior to the blocking step. Quantification of overlap

related to images in Figure 2 was from at least 3 separate locations/each from at least 3

animals.

In situ

hybridization was carried out as described in (

Chari et al., 2021

), including

Weissbourd et al.

Page 13

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

the RFamide (pp5) probe, which was the same as the one used in that study. Imaging of

histological specimens was carried out using an Olympus FV3000 confocal. Multicolor

micrographs shown throughout are pseudocolored composites, with brightness and contrast

adjusted individually for clarity and aesthetics using ImageJ (NIH;

Schneider et al., 2012

Cloning—

To generate the Actin::mCherry plasmid used to establish transgenesis, codon

optimized mCherry cDNA was designed using COOL (

Yu et al., 2017

). The ACT2 promoter

was cloned from upstream of a non-muscle actin gene (XLOC_011689) using primers:

TTTGCTGCGTACAACAACAACGACC and TCGACTTATGTCCTGATAGTTCGGA. The

3’UTR used in all constructs was from a different actin gene (XLOC_021750)

and was amplified using primers: CCAACAGATGTGGATCTCCAAACA and

ACTGGAAGCCTGAGTTCCATCAAA. This was assembled into the pT2AL200R150G

backbone (

Urasaki et al., 2006

; licensed under MTA - N° K2010–008).

Other

Clytia

transgenesis constructs were based on the miniTol2 backbone,

a gift from Dr. Stephen Ekker (Addgene plasmid # 31829). To generate

RFamide::NTR-2A-mCherry, the RFamide (XLOC_019434) fragment was amplified

from

Clytia

genomic DNA using the following primers, ATCCCCATCCGCCATCTTTG,

GTTGTGTTCTTTCTTGATTTGATGG, and inserted into the miniTol2 backbone using

In-Fusion Cloning (Takara). This miniTol2-RFamide backbone was then used to insert

different effectors, always using In-Fusion Cloning, following digestion with Spe1. To

generate RFamide

∷

NTR-2A-mCherry and RFamide

∷

GCaMP6s-2A-mCherry: epNTR was

amplified from the pCS2-epNTR plasmid, a gift from Dr. Harold Burgess (Addgene plasmid

# 62213); both GCaMP6s and the 2A peptide used in this paper was derived from AAV

hSyn1-GCaMP6s-P2A-nls-dTomato, a gift from Jonathan Ting (Addgene plasmid # 51084).

Transgenesis—

In order to establish and optimize transgenesis, we first used the actin

promoter (see Cloning), which we found to drive strong, widespread expression in planula

and in polyp tentacles. This enabled accurate estimates of efficiency during the critical early

life stages following injection (Figure S1H). Using Actin-mCherry to test strategies, we

established a protocol that now enables routine, efficient transgenesis, using microinjection

of Tol2 transposase protein together with circular plasmid DNA into unfertilized eggs.

Collection of eggs and sperm, and microinjection, was carried out as previously described

(

Momose et al., 2018

). Briefly,

Clytia

medusa spawn ~2 hours after the onset of light.

In order to collect eggs and sperm, animals were transferred into either dishes (for the

females) or 6-well plates (for the males) within the first hour of light onset. After spawning,

eggs were immediately collected and injected with a mixture of 6.25ng/ul Tol2 transposase

protein and 10ng/ul plasmid DNA using a Femtojet (Eppendorf). Pulled glass capillaries

were used for microinjection (WPI, TW100F-4). Rather than use the ‘inject’ function on

the Femtojet, injections were carried out by puncturing eggs and allowing the backpressure

in the capillary to fill to the desired volume (~1/4–1/3 egg diameter). Tol2 protein was

produced by Creative Biomart using a plasmid generously provided by Dr. Stephen Ekker

(

Ni et al., 2016

). Protein was then stored at −80 as 5ul aliquots and thawed directly prior to

injection.

Weissbourd et al.

Page 14

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

Immediately following injection – and within an hour of spawning (i.e. 3h after light

onset) – eggs were fertilized and allowed to develop overnight into planula larvae. They

were then transferred into a 12-well plate of artificial sea water containing penicillin/

streptomycin, which prevents early metamorphosis of the planulae into polyps. On the

second day following injection, we checked the expression of plasmids in the planulae under

a fluorescent microscope to ensure that they were capable of driving sufficient expression.

Importantly, at this stage expression is not dependent on integration. Planulae were then

induced to metamorphose into primary polyps onto glass slides (Ted Pella, cat #260439)

using a synthetic GLWamide neuropeptide (produced by Genscript; see

Lechable et al., 2020

for details on culturing across the life cycle).

Primary polyps were hand-fed mashed shrimp until they began to grow into a colony.

Mashed shrimp were generated by drawing brine shrimp into a 10ml syringe with a blunt

tipped needle attached and then expelling them while pressing the end of the needle against

the bottom of a small dish or beaker. Once colonies had 3 or more polyps, they were

screened for transgenic expression, with all but the most highly and broadly expressing

polyp colonies removed. These colonies were then allowed to grow until they began

releasing jellyfish. Jellyfish were then collected, raised to maturity, and backcrossed to

parental strains in order to generate stable F

colonies (see “Experimental Model...” section,

above). F

s were seeded sparsely on slides, and were screened for several criteria: strong

expression of the transgene, strong polyp colony growth and health, and the ability to release

healthy jellyfish that were able to reach maturity. F

colonies were maintained as single

clonal colonies per slide, and once the best colony was identified, the rest were thrown away

and that colony was expanded. This allows for the same genetic background to be used

across experiments, as these colonies then release clonal experimental jellyfish as needed.

Behavioral analysis and NTR ablations

Acquisition.:

Most behavioral analysis was carried out on videos acquired through an

Olympus stereoscope (SZX2). Videos were manually annotated using the BENTO analysis

suite (

https://github.com/annkennedy/bento

;

Segalin et al., 2020

). The exceptions are videos

for tracking the major and minor axis (Figure 3F–I), and for NTR swimming controls

(Figure S2A–C), both of which used a white LED tracing pad as a backlight rather than

the stereoscope (Amazon ASIN# B01M26S3VY) and used automated rather than manual

tracking (see below). All videos were acquired using Flea3 or Grasshopper USB3 cameras

from FLIR, and the manufacturer’s acquisition software (FlyCapture).

Mouth pointing:

Mouth pointing shown in Figure 3B–C was assayed by pinning animals

to sylgard-coated plates (Dow Corning) using stainless steel minutien pins (Fine Science

Tools). This prevented the margin from getting close to the mouth during a folding event,

which would have the potential confound of causing directional mouth pointing by direct

sensory stimulation of the mouth. The subumbrella was then wounded only on the right side

of the animal to compare pointing in the intact (Figure 3B) versus the wounded (Figure 3C)

direction. Having the internal control of the wounded direction ruled out the possibility that

the mouth and margin were responding to a shared, directional sensory stimulus, or that

the folding of the margin itself was a directional sensory stimulus for the mouth (e.g. by

Weissbourd et al.

Page 15

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

creating fluid flow). Videos were acquired at 15fps, and mouth pointing events 30s before

and after a margin folding event were manually annotated using these videos. If the mouth

was leaning in one direction at the start of the epoch, that lean was treated as a baseline for

further pointing. All margin folding events were spontaneous (i.e. no stimuli were delivered)

to avoid possible shared, directional sensory stimulation of the margin and mouth.

Sensory stimuli.:

Shrimp extract used in all stimulation experiments was generated by

homogenizing brine shrimp using a syringe with a blunt tipped needle followed by filtration

through a 40μm cell strainer. Experiments used either this 40μm filtered extract, or extract

that was passed through an additional 0.2μm filter. Mechanical stimuli in Figure S2J was

delivered by gently touching the tentacles and margin using a glass Pasteur pipette. For

experiments in which shrimp extract was used to trigger behavior in freely moving animals,

animals were transferred into 6-well plates and 20ul of 40μm filtered shrimp extract was

added to the well.

Directional folding and ablation experiments:

Directional folding and ablation

experiments shown in Figure 3: animals were pinned to sylgard plates, as described above

for mouth pointing experiments. ~5ul of 0.2μm filtered shrimp extract was then pipetted

directly onto either the top, bottom, left, or right portion of the margin. The timing of

margin folding events from all quadrants, and the locations of sensory stimulation, were

then manually annotated and compared. For physical ablation experiments, body parts were

cut off using spring scissors (Fine Science Tools) or, to remove the mouth, by creating

hole-punches using a blunt-tipped needle.

Automated behavior tracking:

Automated behavior tracking of the major and minor axes,

shown in Figure 3, was performed using custom Matlab (Mathworks) software. Briefly,

following background subtraction, a convex hull was found around the dark pixels using

the “regionprops” function, and the major and minor axes of the hull were extracted. To

achieve high quality tracking, tentacles were trimmed and jellyfish were pinned to sylgard

coated plates (Dow Corning) using a single stainless steel minutien pin (Fine Science Tools)

through the mouth. This maintained their orientation relative to the camera. To behaviorally

distinguish swimming from spontaneous margin-folding, we created binary classifiers

using the length of the major and minor axes of jellyfish and the ratio between these

measurements. Support vector machines (SVMs) with linear kernels were trained using

these three features on equal samples of video frames where animals were either swimming

or performing spontaneous margin-folding behavior, using cross-validation from data across

animals. Classifier performance is reported as the average model accuracy across validation

folds. Chance level performance was obtained by shuffling the identities of swimming and

margin-folding behavioral frames before forming the same analysis. To assess the similarity

between different forms of margin-folding, we tested the performance of SVMs trained to

distinguish swimming vs. spontaneous margin-folding on margin-folding triggered by either

shrimp extract or live shrimp (vs. swimming), using the same features of tracking data.

Comparison of fed to starved and spawning animals:

Comparison of fed to starved and

spawning animals shown in Figure 3. Jellyfish were split into two cohorts in the morning

Weissbourd et al.

Page 16

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

and one cohort was fed

ad libitum

during the day while the other was not (~24h starvation).

Animals were then split into the wells of 6 well plates and passing times individually tested

by manually placing shrimp in their tentacles. For comparison of animals during versus after

spawning, jellyfish spawn ~2h after the onset of light. Mature medusa were recorded from

~1h50min-2h after light onset, and the number of spontaneous folding events was manually

quantified. Egg release could be observed in each of the experimental animals.

Passing of multiple shrimp,:

Passing of multiple shrimp, shown in Figure S2L: animals

were fed many shrimp simultaneously such that most tentacles captured a shrimp nearly

simultaneously. The time and direction of folding onset was then manually recorded for

each, distinct shrimp passing event until most shrimp had been passed. A behavior-triggered

average was then performed for each folding direction versus folding from any other region.

These comparisons were then combined across folding directions.

NTR ablations:

animals were incubated overnight in 10mM Metronidazole (Fisher, cat#

ICN15571005;

Curado et al., 2008

). The following morning, animals were washed several

times in clean artificial sea water and behavioral testing was performed on the same day. For

crumpling behavior, the subumbrella was gently poked with a glass pipette and crumpling

duration was manually quantified. For control experiments shown in Figure S2, automated

behavior tracking was performed as above, with the centroid of the convex hull used to

calculate location, velocity, and turning angle.

The RFamide peptide:

The RFamide peptide (QWLNGRF-amide) and scrambled-sequence

control (FRGNLWQ-amide) used in Figure S2 and S3 were synthesized by Genscript and

bath applied to a final concentration of 100um, or injected at 100um.

GCaMP imaging acquisition, processing, and data analysis—

For highly

restrained imaging experiments, ~4mm

Clytia

medusa were embedded in agarose, as

follows. 3–4%, Type VII-A, low-melting point agarose (Sigma cat# A0701–25G) was

first made in artificial sea water, with particular care to avoid evaporation. Agarose was

then aliquoted and kept in a heat block set at 50-degrees until ready to use in screw-top

tubes. Single tubes were then removed from the heat block and vortexed occasionally until

reaching nearly room temperature. Medusa were then added into the tube, gently mixed,

and rapidly transferred to a glass-bottomed dish (Ted Pella, 14036–20). They were then

quickly spread out to make them as planar as possible before being briefly touched to a cold

object to rapidly cool and harden the agarose. Agarose was then covered with a thin layer of

mineral oil (Sigma M5310–1L) to avoid evaporation during imaging experiments. Following

agarose embedding we observed a spectrum of animal restraint. This could be controlled,

to a degree, by the extent of mixing of the jellyfish in the agarose prior to transferring

to a glass bottom dish. The most well restrained animals were chosen for the analyses

shown in Figure 5. Experiments in which both behavior and GCaMP traces could be

analyzed, shown in Figure 4, were from animals in which agarose embedding was performed

but significant freedom of movement was still visible. Having acquired videos, we then

retrospectively chose animals that had the highest possible movement while still allowing

extraction of traces by circling ROIs (see below). There was consistently a relatively small

Weissbourd et al.

Page 17

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

portion of the animal in the field of view in these experiments. For imaging experiments in

naturalistically behaving jellyfish, animals were placed into a glass depression slide, and a

coverslip was gently placed on top, using small amounts of Vaseline applied to each corner

of the coverslip as a spacer. The coverslip was then gently pressed down to create a small

chamber in which the jellyfish could still perform behaviors. Since the curvature of the

jellyfish in the preparation prevented the entire animal from being in focus, the nerve net

nearest to the mouth was focused on (rings and ring-adjacent net were out of focus and not

measured). Chambers of this type were also used for peptide application during GCaMP

imaging (Figure S3) and for application of shrimp extract to pieces of the ring or net (Figure

S5D). In those experiments, flow-through was achieved using a Kimwipe (Kimberly-Clark

Professional) applied to the edge of the coverslip.

Video acquisition and calcium trace extraction.:

GCaMP imaging experiments always

had synchronous acquisition of the red and green channels. This was achieved using an

Olympus BX51WI microscope with two Photometrics Prime95B cameras connected using

a W-View Gemini-2 Optical Splitter from Hamamatsu. Acquisition was controlled with

Olympus Cellsens software. Images were acquired with downsampling during acquisition

(to 600×600 pixels) and then further processed using ImageJ software, as follows (NIH;

Schneider et al., 2012

). First, images were re-sized to 400×400 pixels. An average Z

projection was then performed, and each frame in the video stack was divided by this

projection to generate a normalized intensity over time for each pixel. If needed, a spatial

filter was applied using ImageJ’s Bandpass Filter function to remove spatial light artifacts,

e.g., from movement of the mouth, with custom parameters tuned to each video. Regions

of Interest (ROIs) were then circled using the ImageJ ROI Manager, and the average pixel

intensity within, and location of, these ROIs was then exported for further processing

using Matlab (Mathworks). Signal from both the red and green channels were acquired

simultaneously using two cameras in all experiments. We then could use the same ROIs as

were used to extract GCaMP traces to extract fluorescence over time from the red channel

to ensure that our downstream analyses are not the result of imaging or motion artifacts.

Running the same analysis pipeline on traces extracted from the red channel did not result in

correlated activity structure, shown in Supplemental Figure 4K. Red channel traces were not

directly used for normalization of the GCaMP traces.

Spike inference and population activity metrics.:

The raw traces, behavior-triggered

averages, and behavior classifiers shown in Figure 4, and the relationship between ring and

net neurons shown in Figure 5O, were computed using smoothed and z-scored GCaMP

traces. For event-based analyses in Figures 5–6, and for computing correlation between

neurons (where long-timescale GCaMP fluorescence changes affect results), we instead

inferred peak times of neuronal events, and performed analyses on these inferred peaks to

remove the effects of underlying noise or drift in the raw trace. Peaks were detected in one

of two ways: either using Matlab’s “findpeaks” function or using spike inference from the

CNMF_E software package (

Zhou et al., 2018

) with parameters manually adjusted for each

trace, and inferred events manually validated. Following spike detection, the first spike in

a bout of inferred spikes was used as the timing of the activity. For computing correlation

between neurons, events were smoothed by sequential convolution with box filters of width

Weissbourd et al.

Page 18

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript

3-seconds and 1.5-seconds. A value of “1” was then added back in at the time of the inferred

peak.

Detecting swimming events using optical flow.:

Optical flow analysis shown in Figure S3

was performed using the Horn-Schunck method (Matlab function opticalFlowHS, default

parameters) applied to cropped video frames of the nerve ring in the red channel. For Figure

S3, we trained a classifier to distinguish video frames of manually annotated behaviors

from all other frames in a trial, where input to the classifier was the mean orientation and

magnitude of the optic flow vectors for each frame. We used a binary boosted decision

tree classifier, trained using the LogitBoost algorithm with cross-validation. Classifier

performance is reported as the average model accuracy across validation folds for each

animal.

Predicting behavioral events from population activity.:

Classifier analysis comparing

GCaMP to behavior was performed in Matlab. Given a set of GCaMP traces, we used the

first 1500 frames for training and held the remaining frames out for testing; this ranged

from 15%–36% of the data in the training set, depending on the length of the recording. We

trained a 3-way, Random Forest classifier with 60 bags, and used out-of-bag prediction to

distinguish between epochs of margin folding, swimming, and quiescence.

Analysis of margin folding events.:

For the loosely-restrained experiments in Figure 4,

and imaging-with-stimulation experiments shown in Figure S5, neurons that responded to

the stimulus and body shape were manually annotated using the ImageJ ROI Manager, and

Adobe Illustrator, by comparing the fluorescence intensity of neurons on a frame before

initiation of a folding event to a frame during the folding event (Figure 4I; Figure S3H). To

generate a polar histogram of margin folding directions (Figure 4), we defined the folding

axis as the line between the mouth and the center of the inward fold, and calculated the

angular position of active neurons relative to this axis.

Reconstructing population events using k-means clusters.:

In Figure 5, we asked what

fraction of variance in spontaneous neural activity could be explained by activation of a

small, fixed number of neural ensembles. First, we defined a population event as more

than 2 neurons being active (using the single frame of peak detection, described above)

within a 2-second window, and generated a matrix of neuronal participation in all observed

ensembles for each animal. To identify a set of fixed neural ensembles, we performed

k-means clustering on this matrix; we tested a range of values of k, and used the silhouette

metric to pick the value of k that best captured the observed data. For each animal, we

then generated a set of all possible combinations of up to 4 ensembles being coactive.

For each observed event from that animal, we found the generated pattern of k-means

cluster activation that minimized the Hamming distance between the observed activation and

the cluster-based activation. This yielded a new matrix of events recreated using k-means

clusters (shown in Figure 5H), which was compared to the observed events matrix to

generate the F1 scores in Figure 5I.

NMF/ICA detection of cell ensembles.:

While k-means clustering requires each cell to

only be counted as a member of a single ensemble, we also wanted to visualize predicted

Weissbourd et al.

Page 19

Cell

. Author manuscript; available in PMC 2022 November 24.

Author Manuscript