►
From YouTube: DevoWorm (2023, Meeting #11):
Description
Physical Modeling of Development updates, brief discussion of the APS meeting, instance segmentation discussion, GSoC and DevoLearn/ DevoGraph project updates (pull requests and open repositories), D-GNNs and graph schemas. Digits in a Dish, Curvatures (embryos) and Consendates (cellular compartments). Attendees: Richard Gordon, Sushmanth Reddy Mereddy, Susan Crawford-Young, Morgan Hough, and Bradly Alicea
A
B
A
B
E
D
D
B
A
B
D
A
I'm
just
trying
to
model
cells
with
it
and
trying
to
do
it
in
a
linear
manner.
No
Springs
I
ran
into
someone
who
said:
oh,
you
should
try
it
in
Morpho
and
I
said
well,
maybe
because
if
I
could
do
it
in
several
different
platforms
and
I
would
know
that
it
was
producing
the
same
thing.
At
least
he
said
well
give
it
to
me.
I'll
I'll,
see
if
it'll
fit
so
I
just
gave
in
my
stuff,
which
is
probably
the
wrong
thing
to
do,
giving
someone
all
your
research.
A
B
A
B
D
A
A
D
B
C
Yeah,
this
yeah
the
effort
of
micro
swimmers.
It's
you
know
they're
these
little.
What
are
they
like?
Cellular
Systems,
so
they're
living
systems,
but
they
can
study
swimming
at
very
small
scales,
so
they're
actually
getting
it
down
to
like
almost
like
the
Nano
scale
or
that
boundary
where
you
have
different
hydrodynamics
yeah.
B
B
Okay,
because
diet
sounds
way
more
than
water,
so
they'll
sink
in
the
Water
by
themselves,
all
right,
so
some
of
them
float
at
the
surface
because
they're
also
the
surface
parts
of
the
surface
are
hydrophobic,
and
so
they
will
actually
float
at
the
top
and
they
seem
to
move
around
even
if
they're
floating
at
the
top
yeah.
So
so
the
question
of
their
ability
to
swim
is
is.
D
A
No
I'm
sorry
I,
think
I
gave
I,
gave
him
the
ref
that
reference
because
of
the
waves
like
he
had
simulated
some
waves,
see
that's
what
I
mean
interesting.
It
simulated
waves
that
that
went
out
and
then
came
back
and
it
simulated
some
waves
that
went
out
and
dissipated
and
then
some
that
came
in
and
dissipated
like
we
had
a.
A
B
Yeah
in
these
different
traditional
ways,
we
had
two
categories
that
you
never
were
able
to.
Well,
we
don't
we
don't
know,
what's
the
difference
between
them,
some
waves
seem
to
stop
when
they
reach
a
certain
barrier
and
other
waves
seem
to
go
through
the
barrier.
Oh
okay,
okay,
Hannah
and
we
don't
know
the
difference
between
those.
A
A
D
A
And
also
yeah,
that's
my
progress.
My
last
I
was
putting
in
my
last
cylinder
and
I.
Wanted
it
to
do
this
and
attach
from
here
to
here,
and
what
did
it
do
I
did
this
or
it
did
this
and
it
won't.
It
won't
go
at
an
angle,
no
matter
what
I
do
I
can
tell
it
which
points
to
go
to
I
can
tell
it
what
angles
to
go
to,
and
it
won't
so
I,
don't
know
it's
a
goofy
program.
This
is
why
I
was
thinking.
Oh
morphle.
What
does
that
do?
C
D
D
A
D
F
D
D
F
F
Yeah,
this
is
the
you
know,
I
think
he
remembered
like
he
asked
an
update
like
centroids,
with
different
colors
and
membrane,
with
a
different
colors,
so
I
added
this
visualization
method.
This
is
how
yeah
how
this
is
important
upon
of
it.
I
added
this
also
like
comparison
of
two
different
frames.
This
is
original
microscope
image,
and
this
is
like.
F
F
B
F
Central,
okay
I
will
try
to
find
it
out
and
I
will
work
on
it.
Like
I
just
went
through
this
like
last
week
and
half
friend
segmentation
model
also
badly.
I
want
to
show
you,
because
if
he
looks
at
them
from
the
side,
they
would
look
like
three-dimensional
object
that
consists
of
a
stack
of
closed
lines.
Yeah,
therefore,
that
object,
if
it
can
be
called
one
object,
is
Central.
F
Okay,
okay,
I
have
worked
on.
My
instant
segmentation
model
like
I,
was
trying
to
build
the
instant
segmentation
model,
which
can
segment
it.
These
are
like
results
of
it.
I
have
plotted
them
on
images,
3D
viewer
to
see
which
you
have
mentioned
last
week.
This
is
how
are
the
results
except
this
cell?
All
other
cells
are
segmented
properly.
You
'd
see
them.
This
is
segmented
in
like
real
language.
B
F
F
Calculating
and
trying
to
calculate
the
volume
also
for
this
is
all
by
using
like
first
calculating
the
area
of
it,
so
you
already
have
yeah
you
already
have
the
data
then
stack
for
itself.
So
if
you
get
the
centroid
of
the
volume
that
would
that
will
do
it,
yeah
that
could
okay
I'll
give
it.
We
can
get
centroids
of
each
cell.
That
would
be
the
nucleus
and
the
volume
also
we
can
calculate
how
means
like
we
will
take
the
potency
wheel.
F
F
B
F
F
B
F
B
B
F
B
The
Village
should
be
changing
in
time,
for
example,
is
that
a
four-dimensional
structure
which
is
hard
to
visualize,
and
we
also
use
it?
Okay,
so
in
a
sense
where
I'm
trying
to
ask
questions
like?
Is
there
a
relationship
between
the
lineage
Tree
in
three
dimensions
over
time
and
the
construction
of
of
the
embryo.
B
F
F
B
B
D
F
Looking
on
it
first
ever
creating
a
pipeline
of
this
instant
segmentation
modem
after
that,
whatever
data
is
extracted
and
giving
it
to
the
data
analysis
I'm
trying
to
build
it
out.
How
can
the
repo
privately
of
some
reasons,
software,
like
proposal,
submission
I,
will
keep
it
open
and
we
can
get
it
so
I
can
show
you
like
I
just
thought:
yeah
I
haven't
pushed
before.
F
F
F
F
F
F
B
B
A
F
B
F
B
F
Okay,
well
it
really
twists
yeah
it
twists.
Your
brain
I
will
I
think
by
next
meeting.
I
will
label
it
I
will
label
these
images
like
on
a
3D
viewer
like
this
is
appearance
when
this
ZIP
file
is
running
right
right
here,
I
will
just
create
a
3D
kind
of
thing
between
when
we
rotate
it.
We
can
see
like
which
is
parent
cell,
which
is
what
the
ministry
of
it
and
try
to
make
it.
I
just
have
my
midterm
exams
and
showings
really
going
to.
F
Work
on
some
project
actually
visiting
Indian
villagers
and
finding
what
their
problems
trying
to
make
a
research
paper
product
like
by
solving
them.
This
is
a
C20
I.
You
know
anyone
know
about
G20,
like
government
of
India
some
initiative
is
there
going
on
as
a
part
of
it.
Our
college
is
like
going
to
Villages
Indian
villages.
We
are
visiting
them.
We
are
finding
the
local
problems
around
here.
F
D
F
Yeah
he
liked
the
idea
he
told
like
segmentation.
Results
are
good,
but
he
wrote
just
told
me
to
calculate
volume
properly
of
each
cell,
actually,
demographic,
so
volume
of
itself,
so
I
am
working
on
it.
Actually,
he
gave
me
this
idea
like
calculating
area
of
it's
an
offer
that
keep
them
in
a
frames.
Video
friends,
you
will
get
the
volume
of
itself.
F
F
C
Yeah,
it's
been
out
of
date
for
a
while.
We
just
been
kept
working
on
it
and
you
know
we
we
built
that
a
while
back,
but
you
know
it
always
needs
to
be
updated.
So
yeah
please
do
this
is
on
GitHub,
I
o.
So
this
is
a
different
repo.
This
isn't
the
the
Devo
learn.
This
is
the
Devo
worm
GitHub
and
it's
like
the
GitHub.
I
o
repository
there.
So
that's
where
that's
hosted.
F
D
C
C
Yeah
do
it
when
you
can
actually
it's
fine,
it's
been
outdated
for
this
long,
it's
you
know,
but
yeah,
that's
great!
Well!
Thank
you
for
that
update.
That's
a
lot
of
work,
it's
great
work,
and
so
I'm
gonna
share
my
screen.
So
if
you're
applying
to
gsoc,
we
are
at
the
first
day
of
being
able
to
submit
proposals.
So
if
you
want
to
submit
your
proposal,
you
know
this
will
be
the
first
day
you
can
that
it's
open
to
submit.
C
So
it's
20
hours
a
week
for
12
weeks
and
you
know
so
keep
that
in
mind
when
you're
writing
your
schedule
out.
So
this
is
the
evil
learn
repository
here.
This
is
the
organization
that
we're
working
with
for
this
gsoc.
We
have
the
Divo
learn
repository,
which
has
the
evil
learn
software
in
it.
This
is
the
software.
This
is
the
first
software
package
that
we
developed
in
the
Stevo
learn
repository,
so
this
was
developed
in
2019,
2020
2021..
C
So
in
that
period
we
built
this
pre-trained
model.
It's
built
on.
You
know
it's
a
deep
learning
model,
it's
built
on
python.
You
have
some
example
notebooks,
and
this
is
where
we're
able
to
get
some
capabilities
for
segmenting
cells.
Although
you
know,
we've
been
working
on
refining
how
that's
handled
pointed
out.
This
is
a
semantic
segmentation
which
is
actually
useful.
For
you
know,
finding
things
have
categories
like
tissues
or
a
specific.
C
You
know
cell
types
or
cell
groups,
but
you
also
want
to
have
instant
segmentation
where
you
have
individual
cells
that
may
be
unlabeled
or
are
just
kind
of
like
in
a
early
early
stage
embryo,
where
there's
not
a
lot
of
where
there's
no
tissue
differentiation.
So
we
have,
you
know,
training
data
for
this.
We
have
you
know
you
can.
C
The
idea
is
to
take
a
set
of
images
that
you
can
get
on
the
web,
or
you
know
from
an
experiment
that
you're
running
you
can
plug
it
into
this
tool,
and
then
you
can
get
segmented,
ideally
three-dimensionally.
But
you
know
two
Dimension.
Usually
images
are
required
in
two-dimensional,
slices
from
the
dorsal
to
the
ventral
side
of
the
embryo,
and
you
take
those
slices
and
you
can
analyze
them
getting
segmented
cells,
getting
volumes
getting
centroids
and
then
having
those
data
outputted
as
a
table
or
in
a
tabular
format.
You
can
then
use
to
build.
C
C
B
C
C
That
point
spread
function
right.
So
that's
yeah,
that's
so
you
know
we
have
a
lot
of
stuff
in
place
and
of
course
it
can
always
be
improved
upon.
That's
what
we're
doing
here
with
some
of
the
pull
requests
that
have
been
issued
here.
We
have,
let's
see
a
couple
that
are
closed.
I
think
sushmoth
has
been
working
with
Maya
Mayo
dab
was
the
person
who
first
put
this
organization
repository
together.
C
Then
his
brother,
my
knock
work
done
in
21
and
then
in
you
know,
in
22
we
worked
on
this
a
bit
with
respect
to
the
devil
graph
project
and
now
we're
working
on
this
this
year,
so
such
month
has
been
working
with
mayok
on
this
mayok
is
the
maintainer
of
this
repo.
So
you
might
want
to
be
if
you're
interested
in
contributing
you'll
want
to
be
in
communication
with
him,
but
we
have
yeah.
We
have
a
lot
of
I,
don't
think
we
have
any.
C
C
Or
a
0.3.0,
you
know
two
releases,
so
we
have
yeah
we're
at
0.3.0
right
now,
and
this
was
a
long
a
while
back.
We
were
going
to
do
an
update
last
year,
but
we
didn't
get
to
it.
So
we'll
probably
do
a
release
this
at
the
end
of
this
year.
At
the
end
of
this
g-soc,
and
hopefully,
it'll
have
Devo
graph
in
it
as
well,
but
also
the
updated
capabilities
for
Diva
learn
the
release
just
means
that
we
have
a
formal
release.
C
C
Then
we
have
the
Devo
graph,
which
is,
of
course,
this
project,
or
this
repository.
This
is
from
last
year's
gsoc,
and
such
month
has
been
working
on.
This
galthem
has
been
working
on
this
a
couple.
Other
people
are
interested
are
in
the
process
of
forking
this,
so
we
have
six
forks
right
now
so
we're
you
know
we're
working
on
these
different
aspects
of
graph,
neural
networks
and
building
a
pipeline
for
this.
But
you
know
we
have
to
sort
of
make
sure
that
Devo
Eva
learn.
C
The
original
package
that
you
showed
is
working
well
for
all
different
types
of
applications.
We
want
to
use
that's
kind
of
like
the
core
of
this
whole
thing,
so
devograph
is
basically
taking
those
data
that
we've
collected
and
Evo
learn
and
turning
it
into
some
sort
of
graph
embedding
so
that
we
can
do
things
like
build
trees
and
build
Network
topologies,
and
things
like
that.
So,
if
you're
interested
again,
we
have
no
open
pull
request.
We
had
a
couple
social
month,
I
think
and
Gotham
have
pushed
some
things
and
they've
been
accepted.
C
So
that's
great
also
wanted
to
bring
people's
attention
to
the
slack
we
have.
Not
only
have
you
join
the
open
worm
slack
to
join
the
discussion
about
gsoc,
you
have
the
diva
worm
Channel,
which
is
of
course
the
main
Channel.
We
have
a
lot
of
people
coming
in
I
see.
You
know,
I
think.
In
the
last
couple
weeks
we've
had
about
10
people
here
join
again.
C
C
So
you
know
there's
a
paper
on
C
elegans
locomotion,
Python
tutorial.
You
know
this
shape
of
things
to
come,
which
is
actually
a
nice
review,
article
expressivity
of
persistent
homology,
so
in
graph
neural
networks.
This
is
there's
this
idea
of
expressivity,
or
what
can
the
algorithm
do
or
what
can
it
represent
and
that's
a
common?
That's
a
actually
a
pretty
popular
Topic
in
graph
neural
networks.
You
have
this
idea
of
expressivity
and
you
know
if
you
write
an
algorithm
or
if
you
build
a
representation,
how
expressive
is
it?
C
C
C
Jiahang
lee
is
he
worked
on
the
graphical
networks
last
year,
so
he's
actually
he's
been
in
and
out
of
the
slack
recently,
but
he's
he
posted
this
last
week
he
posted
a
couple
of
I
think
he
posted
his
proposal
from
last
year.
So,
if
you're
interested
in
reading
a
proposal
that
got
accepted,
you
can
read
his
proposal
and
that
has
the
proper
structure
for
a
gsoc
proposal.
So
that's
that's
the
place
you
want
to
go
for
that.
C
He
also
has
a
couple
readings
here
on
graph
neural
network
for
cell
tracking
and
microscopy
videos,
topological
graph,
neural
networks.
This
is
you
know,
merging
topological
data
analysis
and
graph
neural
networks
and
then
the
Deep
graph
Library,
which
is
another
tool
you
might
be
interested
in.
If
you
needed,
because
graph
neural
networks
have
a
very
high
barrier
to
entry,
it's
just.
You
have
to
learn
a
lot
before
you
can
really
do
a
lot
of
work
on
it.
C
So
we
that's
why
we've
been
struggling
with
this,
for
you
know,
last
year,
and
then
this
year
it's
not,
you
know,
really
easy.
It's
it's
hard.
It's
a
new
area,
there's
a
it's
kind
of
like
a
wild
west.
There
are
a
lot
of
types,
different
types
of
approaches.
People
are
using,
and
so
you
know
whatever
you
can
come
up
with
is
probably
you
know
it's
probably
a
good
contribution,
I'm,
not
saying
that
you
can
just
do
anything.
You
want
it's
just
you
know
it's.
C
It's
ultimately,
we
want
to
apply
these
graph
neural
networks
to
biological
systems
to
biological
data
sets,
and
so
we
we've
done
a
lot
of
work
with
like
networks
with
trees,
and
this
is
going
to
fit
into
that
that
work
and,
if
you're
interested
in
that
work,
you
know
we
have
our
website
Evo
worm
at
weebly.com,
and
that
has
a
Publications
page.
You
can
check
out
the
Publications
there.
C
Okay,
so
I'm
gonna
talk
actually
about
there's
a
new
paper
on
graph
neural
networks
from
deepmind,
and
this
has
just
come
out
in
the
last
month.
This
one's
called
graph
schemas
is
abstractions
for
transfer,
learning,
inference
and
planning,
and
this
is
a
set
of
authors
from
deepmind.
They
leave
George's,
pretty
famous
you
know,
person
in
this
area
he's
done
a
lot
of
really
interesting
stuff,
with
Ai
and
machine
learning
and
deep
learning.
C
So
if
we
look
at
the
abstract
here
so
this
paper
is,
we
propose
schemas
as
a
model
for
abstractions
that
can
be
used
for
Rapid
transfer,
learning,
inference
and
planning,
common
structured
representations
of
Concepts
and
behaviors,
or
what
they
call
schemas
have
been
proposed.
It's
a
powerful
way
to
encode
abstractions.
C
So
basically
they
have
these
abstractions,
which
are
these
I
guess
they're
models
for
different
things.
You
want
to
do
their.
You
know,
abstractions
mean
you
take
a
process
and
you
extract
like
the
essential
aspects
of
it.
So
it's
kind
of
like
model
building
you
take
a
process,
You
observe
it.
You
say
these
are
the
most
important
things
we
want
to
take
and
put
into
our
model
and
that's
your
abstraction,
so
you
can
have
different
levels
of
abstraction.
You
can
have
a
high
level
abstraction
or
low
level
abstraction
if
we
take
like
the
embryo.
C
One
of
the
things
we're
doing
in
Diva
learn
is
we're
taking
the
centroids
or
taking
the
volume
of
the
cells
and
we're
using
that
as
a
proxy
for
embryogenesis
we're
not
talking
about
genes.
We're
not
talking
about
physics
necessarily,
but
those
are
that's
because
that's
the
model
we
chose,
that's
the
abstraction
we
chose,
and
so,
if
we
wanted
to,
we
could
build
an
abstraction
with
many
many
different
details,
but
we
don't
do
that.
C
You
know
certain
abstractions
are
good
for
certain
things.
What
they're
saying
is
they
want
to
build
a
schema
as
a
model
for
abstractions,
and
so
you
have
this,
so
you
can
encode
abstractions
in
different
ways.
You
have
this
these
schemas,
that
they
propose
that
can
be
used
to
build
use
as
a
model
for
that
late.
Let
latent
graph
learning
is
emerging
as
a
new
computational
model
of
the
hippocampus
to
explain,
map,
learning
and
transitive
inference.
So
in
this
case
the
hippocampus
is
a
part
of
the
brain.
C
C
In
learning
we
build
on
this
work
to
show
that
learned
latent
graphs
and
these
models
have
a
slot
structure
right.
They
go.
Schemas
that'll
offer
Quick
Knowledge
transfer
across
environments
in
a
new
environment.
An
agent
can
rapidly
learn
new
bindings
between
the
sensory
stream
to
multiple
latent
schemas
and
select
the
best
fitting
one
to
guide
Behavior.
C
To
evaluate
these
graph
schemas.
We
used
two
previously
publish
challenging
tasks,
the
memory
and
planning
game
and
the
One-Shot
Street
learn.
They
are
designed
to
test
Rapids
task
solving
in
novel
environments.
Graph
schemas
can
be
learned
in
far
fewer
episodes
than
previous
bass
lines
and
can
model
and
plan
in
a
few
steps
in
novel
variations
of
these
tasks.
C
C
They
have
this,
this
abstraction
of
the
important
parts
of
what
the
hippocampus
is
doing
and
then
they
build
the
schema,
which
is
a
graph
of
that
so
they're
able
to
build
a
relational
database
essentially,
and
that
will
allow
you
to
build
a
better
model
of
an
AI,
for
you
know,
navigation
and
say
like
a
typical
standard,
neural
network,
and
so
this
is
the
reason
why
we
do
a
lot
of
this
as
well.
We
do
graph
neural
networks
because
we
can
build
these
graph
representations,
so
sometimes
things
in
the
world
have
a
Graphic
representation.
C
So
you
know
there's
a
lot
of.
There
are
a
lot
of
things
in
the
world
that
have
sets
of
relations
between
them.
They
could
be
Maps,
they
could
be.
You
know,
structural
networks,
they
could
be
just
you
know,
networks
of
interactions
and
the
idea
is
we
want
to
build
these
graphs.
That
will
tell
us
basically
the
essential
aspects
of
those
relationships,
and
then
we
can
build.
You
know
we
can
use
that
for
AIS.
We
can
use
that
for
other
types
of
data
analysis.
C
So
it's
really
useful
to
do
and
that's
kind
of
what
people
are
doing
in
graph
neural
networks,
they're
going
Beyond
sort
of
the
typical
layered
Network.
That's
you
know
has
weights
that
you
know
you
have
to
figure
out
what
the
weights
are.
The
idea
is,
you
have
some
structure
in
there
and
you
try
to
recover
the
graph
structure,
the
real
graph
structure,
but
along
the
way
you
still
have
the
structured
representation.
C
We
further
demonstrate
learning,
matching
and
reusing
graph,
schemas
and
navigation
tasks
and
more
challenging
environments
with
aliased
observations
and
size
variations
and
then
show
a
different
schemas
can
be
composed
to
model
larger
two
and
3D
environments.
So
this
is
just
kind
of
their
work
on
this.
This
is
a
quite
a
long
paper,
but
it
kind
of
talks
about
how
they've
observed.
What's
going
on
in
the
hippocampus
that
you
have
these
different.
C
C
So
you
know
there
are
different
ways
that
you
can
do
this.
We
build
these
schemas
using
clone
structured,
cognitive
graphs,
a
computational
model
of
cognitive
graphs
using
high
performance
computing.
In
particular.
This
work
showed
that
latent
high
order
graphs
can
be
learned
in
highly
Alias
settings
using
a
smooth,
probabilistic
parameterization
of
the
graph
learning
problem.
So
this
is,
you
know
there
are
a
lot
of
problems
like
this
comes
from
graph
Theory
and
there
are
a
lot
of
problems
in
graph
theory
that
involve
finding
the
optimal
graph
for
a
given
set
of
relationships.
C
So
that's
that's
one
of
the
things
you're
trying
to
do,
and
you
know
they've
built
these
kind
of
graph,
these
these
graph
schemas
or
these
representations
that
can
generate
different
graph
Relic.
You
know
graphical
relationships
and
then
they
try
to
find
the
best
one
to
fit
the
problem.
So
you
know,
but
we
still
have
that
graph
relationship.
So
we
don't
have
we're
not
working
blindly
we're
working
from
like
a
set
of
relationships.
C
We
just
need
to
find
the
optimal
set
of
relationships,
and
then
you
know
this
is
work
that
people
have
done
in
cognitive
modeling.
So
you
know
if
we
apply
this
to
development,
you
know
there
are
going
to
be
different
types
of
tools
we
may
need
to
draw
from.
Unfortunately,
there
isn't
a
lot
of
work
at
all
on
this
in
in
the
brain
in
Neuroscience.
C
So
there
are
all
sorts
of
problems
you
can
apply
to
this.
You
know
for
solving
simplified
relational
tasks,
they're
different
benchmarks
and
other
types
of
work
that
you
can
draw
from.
So
this
is
a
diagram
here
of
using
graph
schemas
in
the
CS
CG
model.
So
this
is
your
schema
matching
with
binding.
You
have
these
experiences
and
you
have
these
schemas
that
extract
from
it.
So
what
you
do
is
You
observe
the
hippocampus.
There
are
these
different
experiences.
There's
this
network
of
neurons
that
get
activated,
and
then
you
have
these
schemas.
C
You
can
extract
which
tell
you
something
about
the
function
now
in
development.
You
know
we
have
kind
of
like
you
know.
We
have
real
lineage
trees
that
we
can
draw
from,
or
we
have
real
differentiation
trees
that
tell
us
something
about,
maybe
like
the
experience
of
a
certain
embryo.
How
does
development
unfold
which
cells
divide
and
turn
into
what
tissues
or
what
cells
divide
and
proliferate
at
certain
times?
So
we
have
these
analogs
in
development.
C
E
C
Know
passing
them
on
to
I,
don't
know
some.
Some
application
will
build
in
the
future.
Maybe
they'll
perform
really
well
as
an
algorithm.
Maybe
they
won't.
Then
they
have
these
cloned
hidden,
Markov
models.
So
you
have,
you
know,
observations
and
clones,
so
you
have
these
latent
States.
So
you
have
these
sets
of
observations,
and
then
you
have
these
latent
states
where
they're
not
observed
but
they're.
You
know
hidden
Markov
model.
C
Are
these
transitions,
these
different
states
that
are
hidden
from
view
you're,
not
observing
them,
but
you
know
they
have
to
exist
because
the
system
has
a
certain
structure
that
suggests
that
they're
there.
So
you
can
infer
latent
states
that
you
may
not
observe
you
know.
If
you
watch
an
animal
behave
in
space
and
navigate
certain
parts
of
the
hippocampus
will
be
active,
maybe
if
they
behaved
in
a
different
way
in
another
environment.
C
Other
parts
of
the
hippocampus
would
be
active,
but
you
can't
observe
that
but
they're
late,
what
we'll
call
latent
States,
and
so
we
can
build
a
model
of
those
late
States
and
then
we
can
build
a
model
of
continuous
activation.
This
is
like
a
continuous
Behavior,
where
the
organism
is
planning
to
some
goal.
State,
and
so
you
know
this
is
actually
relevant
to
embryos,
because
you
know
you
have
a
process
that
unfolds
to
say
the
adult
phenotype.
So
the
question
is,
you
know:
can
we
model?
C
Can
we
build
a
model
that
you
know
goes
from
like
an
egg
one
cell
to
an
adult
phenotype,
and
then
what
paths
are
possible?
Are
all
paths
equally
possible
or
we
know
from
development
that
that's
really
not
true,
that
there
are
certain
paths
that
are
viable
and
certain
paths
that
are
inviable
so
I
can
see
elegans
you
have
you
know
the
the
lineage
tree
always
unfolds
the
same
way
except
when
you
have
a
defined
mutant,
and
there
are
differences
in
that
lineage
tree
in
something
like
a
mouse
where
development
is
not
as
deterministic.
C
B
B
Okay,
it's
developing
into
adults,
but
it
doesn't
have
the
adult
as
the
goal
you
know,
I
can't
find
the
quote,
but
I,
remember
I.
Remember
somebody
saying
if
you
have
an
average
difficult,
it
goes
towards
the
goal
of
being
an
average
adult,
no
matter
what
you
do,
okay,
that
in
other
words
it
doesn't
have
the
normal
adult
as
its
goal
right.
Maybe
there
is
no
problem.
B
Yeah,
the
the
in
between
compromise
on
this
is
being
janus-based
approach,
where
the
differentiation
tree
consists
of
a
set
of
small
holes,
it's
logos,
for
example,
or
how
far
does
the
differentiation
wave
go
and
what
stops
it?
Okay?
So
it
stops
at
a
given
point
and
that's
sort
of
like
a
relay
race
right.
If
you
imagine
a
relay
race
under
open
field
which
could
go
in
all
different
directions,
do
you
have
any
control
over
what
happens
right?
Okay,.
C
B
B
D
B
B
Okay
in
various
fields-
and
it
was
quite
an
inspiring
paper
for
the
the
case-
the
problem
is
when
you
have
something
with
the
slightest
deviation,
can
cause
a
huge
error
because,
for
instance,
you
exponentially
increasing
the
number
of
cells.
How
on
Earth?
Do
you
keep
the
adult
goal
in
mind
without
having
a
mind?
Okay,.
C
B
C
C
This
is
basically,
you
know.
The
the
key
here,
though,
is
that
you're
trying
to
extract
these
graph
structures.
That
will
tell
you
something
about
you
know
different
contexts,
different
systems-
and
you
know,
it'll,
have
as
long
as
it
has
a
sort
of
graphical
structure
to
it,
which
means
it
has
a
set
of
associations
or
it
has
this.
B
C
Right,
yeah
yeah,
it's
not
yeah,
so
yeah
I
mean
a
lot
of
the
extra
a
lot
of
the
you
know,
the
a
lot
of
the
machine,
learning,
literature
and
deep
learning
that
you
know
it's
all
about
like
minimizing
energy
or
optimizing
the
result
and
of
course,
in
algorithms.
You
know
that's
kind
of
what
you
want
to
do,
but
at
the
same
time
that
can
you
know
if
you
have
like
a
generative
algorithm
or
a
generative
process?
That's
not
really
an
algorithm.
It's
not
algorithmic!
C
Yeah,
so
you
know
there,
there
are
different
things
like
in
that
context.
You
know
they're
different
ways
that
people
have
approached
this
in
the
artificial
life
Community,
for
example,
they've
used
open-ended
models
and
open-ended
algorithms
to
sort
of
approach
this,
because,
if
you're
modeling
like
an
evolutionary
system,
you
know
evolution
is
producing
outcomes
that
are,
you
know
very
variable,
so
you're
not
out
you're,
not
optimizing
things.
B
C
Yeah,
so
that's
so
that's
yeah,
that's
the
paper
and
they
go
through
a
bunch
of
benchmarks.
Here
they
talk
about
their
schemas
and
how
they
approach
these
different
tests.
So
they
have
the
memory
and
planning
game
and
they
have
Street
learn.
Basically,
you
can
see
that
they
have
a
sample
environment
and
from
that
you
know
from
both
of
those
you
know
you
have
like
different
places
that
are
visited.
You
get
this
learned
graph
and
this
learn
graph
has
structure
in
it.
C
So
in
different
ways,
the
street
learn
is
a
little
bit
different
from
memory
and
planning
and
you
can
see
they're
optimizing
it
the
average
reward.
So
this
is
sort
of
a
reinforcement,
learning
aspect
to
it
where
they're
built,
you
know,
they're
optimizing
for
the
average
reward
over
a
series
of
training
steps.
You
can
see
that
it
converges
at
a
large
number
of
training
steps,
and
so
this
is
just
two
benchmarks,
so
these
are
have
nothing
to
do
with
development,
but
they're.
C
You
know
you
can
have
multiple
plans
like
you
know,
multiple
networks,
and
you
know
you
look
at
like
depending
on
the
room
size,
you
have
more
plans
and
you
know
you
can
make
a
difference
between
like
different
schemas
or
plans,
and
you
could
think
of
that.
As,
like
you
know,
in
the
embryo,
you
know
there
isn't
necessarily
a
plan,
but
they
have
these
hypothesized
things
called
developmental
programs
and
you
can
look
at
those
in
terms
of
like
evaluating
different
alternative
developmental
programs.
B
Maybe
if
you
look
at,
if
you
buy
seeds,
you
know
that
seeds
have
a
are
often
rated
by
a
viability
which
may
go
down
with
time,
yeah,
okay
and
we
for
animals
that
have
multiple
eggs.
You
also
have
that
problem
with
liability.
Only
a
certain
fraction
of
the
producer
of
an
adult
organs,
people
don't
study
the
ones
who
don't
understand.
Okay,.
D
B
D
D
B
B
A
B
B
A
C
All
right,
I
think
that's
all
I'm
going
to
talk
about
for
this
paper
again.
I
can
make
this
available.
I
wanted
to
go
over
this
to
kind
of
show,
the
sort
of
the
way
that
we're
thinking
about
graph
neural
networks
and
graphs
and
their
relationship
to
some
of
the
data
that
we've
been
talking
about
in
the
earlier
part
of
the
meeting.
So
again,
this
is
just
kind
of
what
how
they
do
this
with
you
know,
looking
at
the
wiring
and
optimizing
the
wiring
for
the
problem
that
you're
trying
to
solve
and
an
embryos.
C
E
E
Oh
yeah,
no
problem,
no
problem:
what's
the
first
intellectual
contact
with
them,
I
know:
okay,
I
think
he
I
think
he
actually
I
think
he
might
have
a
couple
papers
if
you're
it's
a
it's.
A
massive
literature
sounds
like
it's
20
pages
with,
like
300
references,
yeah,
I,
I
I've
been
following
him
for
decades,
but
I
can't
convince
him
to
combine
his
work
with
the
differentiation
tree
work:
okay,
okay,
I.
B
B
B
B
B
B
C
A
D
D
E
E
B
B
C
C
We
identified
a
unique
property
of
undifferentiated
mesenchyme,
isolated
from
the
distal
early
autopod
to
autonomously
reassemble,
forming
multiple
autopod
structures,
including
digits
interdigital
tissues,
joints,
muscles
and
tendons.
So
you
have
these
forming
digits
and
I.
Think
we've
talked
about
an
earlier
meeting
that
you
have
like.
You
have
programs
cell
death
that
occurs
that
reveals
digits
from
it's,
basically
like
a
paddle
or
a
pad
like
a
4lm
pad,
and
then
you
get
programmed
cell
definitely
digits
that
form
from
that.
C
So
this
reveal
distinct
cell
clusters
that
Express
canonical
markers
of
distal
limb
development,
including
coal,
2a1
called
10a1
and
sp7.
This
is
for
families
formation
for
perichondrium.
You
have
a
thbs2
and
coal
1a1
for
the
joint
enterzone
of
gdf5,
wind,
5A
and
June,
and
then
interdigital
tissues
are
aldh1a2
and
Ms
X1
and
then
for
muscle
progenitors.
You
have
my
o
B1
and
then
verticular
perichondrium,
articular
cartilage.
You
have
prg4
and
then
for
scx
tnmd,
which
are
the
genes
for
tendons.
C
So
you
can
see
that
they've
defined
all
these
different
tissue
types
or
precursor
tissue
types
and
they've
Associated
cells
expressing
these
genes.
That
will
go
on
to
form
these
these
different
tissues.
So,
for
example,
if
you
have
tissue
or
if
you
have
cells
that
are
going
to
form
muscle,
progenitors
or
muscle,
you
have
those
progenitor
cells
expressing
high
levels
of
myod1.
C
So
you
see
that
you
have
this
muscle
pain.
Let
me
go
to
our
jamboard
here
and
I'll
draw
this
out.
Basically,
you
start
with
this
pedal,
or
this
it's
almost
like
a
a
fin
because
that's
its
Origins,
invertebrates
and
then
this
fin
has
you
know
all
these
These
are
stem
cells
within
this,
then
you
have
programmed
cell
death
in
between
what
will
become
the
digits.
C
So
your
program
cell
death
in
these
regions
and
programs
Zelda
through
apoptosis,
is
just
simply
where
the
cells
die
with.
You
know
a
certain
number
of
Divisions
so
that
you
know
this
is
something
that's
encoded
in
the
genotype,
so
that
you
know
the
the
cells
can
be
removed
in
a
timely
manner.
Sometimes
people
get
webbed
fingers
and
webbed
toes.
C
If
this
doesn't
this
process
doesn't
work
properly,
but
once
you
get
this
process
completed,
then
you
get
something
that
resembles
a
limb
with
digits,
and
so
then
the
digits
have
to
differentiate
within
their
own
sort
of
each
digit.
So
you
have
cells,
of
course,
that
make
up
the
digit
and
these
cells
are
going
to
differentiate
into
different
tissue
types,
different
structures
and
so
forth.
C
You
know
we
don't
I,
don't
want
to
be
too
descriptive
about
it
here.
The
joints,
of
course,
are
expressing
different
genes
and
so
forth.
So
we
can
identify
them
at
this
stage
and
then
they'll
go
on
to
form
those
tissues
later
foreign,
so
the
system
actually
is
good.
For
kind
of
you
know
you
have
to
first
of
all
characterize
the
genes
that
are
being
expressed
in
these
cells
and
then
the
cells
will
go
on
to
form
these
tissues.
C
C
C
This
Innovative
system
will
provide
access
to
the
developing
limb
tissues,
facilitating
studies
to
discern
on
digit
and
articular.
Joint
formation
is
initiated
and
how
undifferentiated
mesenchyme
is
pattern
to
establish
individual
digit
morphologies,
so
we
have
mesenchymal
stem
cells
with
these
little
circles
that
I
drew
here.
Hermesanimal
stem
cells
you're
plating
these
in
a
dish.
So
it's
not
actually
on
the
digit
itself.
C
C
So
this
is
this:
is
a
nice
system
for
kind
of
recapitulating
development
in
a
dish,
the
in
vitra
digit
system
also
provides
a
platform
to
rapidly
evaluate
treatments
and
that's
stimulating
the
repair
regeneration
of
mammalian
digits
impact
by
con
General
malformation,
injury
or
disease.
So
this
is
a
way
me
perhaps
that
we
can
stimulate
some
sort
of
repair
regeneration
mechanisms
that
exist
in
in
the
digits.
We
can't
do
that
if
we
do
this
in
Vivo,
because
we
can't
necessarily
change
the
fate
of
cells.
C
So
this
is
where
you
get
limbo,
dissection
and
dissociation,
so
they
take
like
what
will
become
the
digits
from
a
developing
Mouse.
They
basically
take
the
end
of
the
limb,
bud
the
four
limb
Bud
they
put
it.
They
dissociate
the
cells.
Then
they
grow
this
tissue
culture
dishes
they
do.
The
single
RNA
seek
analysis
to
see
what
genes
are
in
up
or
down
regulated
in
each
cell.
C
So
you
get
a
single
cell
resolution,
so
it's
actually
useful
for
that,
although
you
don't
have
any
geometric
provenance
like
you,
do
if
you're
doing
this
in
in
Vivo-
and
you
also
have
RNA
fish
and
IHC
Imaging,
so
you
can
image
the
cells
in
the
dish.
You
don't
have
to
harvest
them
and
harvest
the
RNA,
so
this
is
actually
quite
useful
for
building
a
you
know,
a
proof
of
concept
system
that
will
tell
you
something
about
what
the
cells
are
doing,
but
not
necessarily
in
in
an
anatomical
context.
C
C
So
you
see
that
the
formation
of
digits
not
only
digits
but
carpals,
which
are
Knuckles,
and
so
this
is
all
sort
of
the
precursors
of
this,
probably
from
Full
digits
I,
don't
know
how
far
along
the
pathway,
you
get
a
digit
development
in
here,
but
you
definitely
get
cells
differentiating
in
different
ways.
So
this
isn't
really
an
organoid
study,
but
it
doesn't.
You
know
it's
it's
kind
of
a
cross
between
a
typical
cell
differentiation
protocol
and
maybe
like
an
organoid
because
I
don't
think
they
do
the
same
three-dimensional
culture.
C
We
see
that
they
call
them
autopods,
distinct,
autopod
muscles
musculoskeletal
tissues
and
structure
present
at
day,
seven
in
did
culture,
so
this
did
system
Days
by
day
seven.
You
start
to
get
some
of
these
tissues
and
structures
that
are
present
in
the
musculoskeletal
system
of
the
digits,
and
so
this
is
an
example
of
umap
analysis
of
gene
expression.
So
this
is
just
showing
the
sort
of
the
you
know.
The
the
variation
of
the
different
cell
types
and
how
they
segregate
out
so
we
can
identify
the
cell
types
by
different
gene
expression
profiles.
C
C
This
next
set
of
papers
is
on
something
called
curvatures
and
condensates,
and
so
these
are
actually
from
soft
materials,
but
we're
going
to
now
talk
about
them
in
the
context
of
development,
so
the
first
paper
is
spatially
non-uniform
condensates
emerge
from
dynamically
arrested,
phase
separation,
and
so
you
know
the
formation
of
biomolecular
condensates
through
phase
separation.
Basically,
where
you
get
these
condensates,
that
are
things
that
precipitate
out
of
liquids.
This
happens
through
phase
separation.
C
This
is
the
formation
of
these
from
proteins
and
nucleic
acids
is
emerging
as
a
spatial
organizational
principle
used
broadly
by
living
cells.
So,
what's
going
on
here,
is
you
have
these
biomolecular
condensates
that
segregate
out
in
terms
of
spatial
organization?
So
if
you're
looking
at
things
like
different
parts
of
the
cytoplasm
or
different
parts
of
the
cell,
this
is
how
these
things
sell
from
self-organize
and
self-assemble.
So
this
is
an
organizational
principle
used
broadly
by
living
cells.
C
Many
such
condensates
they're,
not,
however,
homogeneous
fluids,
but
possess
an
internal
structure
consisting
of
distinct
subcompartments
with
different
compositions,
so
I
think
in
an
earlier
meeting.
We
talked
about
this
where
we
have
these
different
aspects
of
like
cell
condensates,
and
so,
if
we
look
at
like
a
cell,
we're
talking
I
think
talking
about
a
minimal
cell
or
some
type
of
model
of
a
cell
where
we
have
this.
This
is
the
cell.
This
is
the
nucleus,
and
then
this
area
here
is
the
cytoplasm
and
there's
some
organelles
in
in
inside
this.
C
D
C
So
that
term
means
that
you
have
various
things
that
form
in
here
and
there's
a
spatial
organization
within
the
cytoplasm
and
that
this
fluid
is
heterogeneous.
So
when
we
go
to
model
this,
we
can't
model
this
as
a
homogeneous
substance.
We
have
to
model
what
is
a
heterogeneous
substance
with
its
own
Dynamics
and
I,
can't
remember
which
paper
that
was,
but
we
did
talk
about
this
and
to
some
extent
several
weeks
ago,
it
was
recent,
notably
condensates,
can
contain
compartments
that
are
depleted
in
the
biopolymers
that
make
up
the
condensate.
C
So
this
is
again
the
self-organization
of
you
know
these
different
biomolecules
and
how
they
form
rings
and
how
they
form
clusters.
And
so
we
have
to
understand
that
you
know
some
of
these
compartments
can
are
depleted
and
some
aren't.
So
if
you
go
back
to
our
model,
we
have
the
cell
and
there's
a
heterogeneity
to
the
cytoplasm.
If
we
go
to
a
model
with
a
cell
or
we
maybe
just
have
a
nucleus,
and
then
we
have
several
compartments
within
the
cytoplasm
within
which
may
be
different
types
of
metabolism.
C
C
Some
of
them
may
be
packed
with
molecules.
Some
of
them
may
have
no
molecules,
so
we
have
to
think
about,
like
you
know,
if
what
the
consequences
of
our
setting
up
our
model
are
on
the
future
of
this
material,
and
so
we
have
to
consider
during
self-organization
that
things
are
heterogeneous
and
not
homogeneous
and
that
different
compartments
of
different
properties.
We
can
set
those
up
in
a
model
and
see
the
outcome.
C
Here
we
show
that
such
double
Emulsion
condensates
emerged
by
a
dynamically
arrested
phase
transitions.
The
combination
of
a
change
in
composition,
coupled
with
a
slow
response
to
this
change
and
lead
to
the
nucleation
of
biopolymer
poor
droplets
within
the
polymer-rich
condensate
phase.
So
the
way
they're
compartmentalizing
this
through
this
process
of
condensate
formation
is
that
you
have
droplets
that
form,
and
so
the
droplets
have
sort
of
a
sampling
of
the
the
soup.
C
That
is,
this
biopolymer
within
the
condensate,
it's
sort
of
like
what
we
were
talking
about
with
simulating
early
life
and
simulating
this
molecular
soup,
and
you
know
samples
vesicles
that
sample
the
molecular
soup.
This
is
basically
the
same
thing.
The
droplets
are
represent.
Basically
these
these
vesicles
and
you
know
they're,
making
the
point
that
you
can
that
just
because
it's
a
droplet
doesn't
mean
that
it's
always
the
same
makeup.
C
The
the
greater
soup
can
be
sampled
differentially
so
that
some
droplets
have
very
are
very
rich
in
biopolymerism
are
very
formed
biopolymers,
and
so
this
is
an
interesting
point
in
light
of
some
of
the
work
you've
been
doing
on
simulating
early
life
because
it
suggests
that
there
is
this
sort
of
sampling
bias
that
happens,
and
this
is
not
something
that
happens
through
any
sort
of
conscious
intent.
It's
just
a
product
of
the
sampling.
It's
it's
an
outcome
of
the
sampling
process.
C
Just
because
it's
heterogeneous-
and
you
know
we
don't
know
spatially-
there
isn't
necessarily
any
distinct
spatial
segregation,
but
when
we
draw
a
sample
out
of
it
either
by
locking
it
up
in
a
vesicle
or
getting
droplets
out
in
some
way.
You
know-
and
in
this
case
it's
their
condensation-
we
can
get
the
sort
of
we
can
get
these
these
differences
in
number
of
biopolymers.
C
So
they
talk
about
this
dynamically
arrested
phase
transition.
So
we
have
this
change
in
composition
and
a
slow
response
to
change.
So
this
is
a
combination
that
I
think
they
may
talk
about
hopefully
more
later
in
the
paper,
but
that's
I
think
key.
Maybe
to
some
of
this.
Our
findings
demonstrate
the
condensates
of
the
complex
internal
architecture
can
arise
from
kinetic
rather
than
purely
thermodynamic
driving
forces,
and
so
this
is
where
you
have
connect
kinetic
activity
going
on,
that
biopolymers
are
being
made
and
that
can
actually
impact
this
heterogeneity.
C
It's
not
just
thermodynamic
driving
forces,
it's
not
just
energy
minimization
or
just
you
know,
diffusion
processes,
the
I
guess
diffusion
processes
would
be
partially
kinetic,
but
I
think
in
one
of
the
papers
we
talked
about
where
there
were
diffusion
processes
that
were
well.
We've
talked
about
this
in
the
past,
where
you
have
diffusion
processes
that
are
non.
Not
you
know
like
a
random
distribution.
They're
sort
of
super
super
normal
or
super
gaussian
in
in
their
nature.
C
So
that's
that's
something
to
think
about
and
then
provide
more
generally
an
Avenue
to
understand
and
control
the
internal
structure
of
contents.
That's
both
in
vitro
and
in
viewer,
so
they
kind
of
go
over.
So
what
they're
interested
in
here,
these
spatially
structured
condensates?
They
arise
in
minimal
mixtures
and
the
kind
of
biomolecules
are
interested.
Are
these
peptide
RNA
condensates?
C
So
you
can
get
vacuole
formation
within
them?
Vacuoles
are
just
like
different
compartments
within
a
Condon
set,
but
also
it
could
be
within
a
cell.
The
the
this
core
shell
structure.
You
can
see
this
in
micro
gels,
for
example,
in
living
cells.
You
see
vacuoles
in
a
lot
of
places
and
they
can
be
filled
with
different
types
of
biomolecules
nuclear
and
cytoplasmic
germ
granules,
with
Hollow
centers
and
poly
ra
and
RNA
nuclear
condenses
of
low
density
space.
C
So
one
of
the
results
that
they
have
is
a
doubly
Emulsion
structure
can
form
reversibly.
So
they
want
to
go
beyond
the
formation
of
this
internal
structure,
with
a
model
system
containing
single-stranded,
poly,
ra,
RNA
and
Peg.
So
these
are
cocundances,
so
they
form
in
the
same
condensate.
They
don't
significantly
solidify
over
time
and
change.
Compositional
temperature.
C
We
vary.
The
temperature
determine
the
effect
of
charging
the
condensate
composition
on
the
formation
trap.
Droplets
liquids
pour
in
these
in
these
biomolecules
inside
the
condensate
we've
observed
these
droplets
incontinence,
it's
at
20
degrees,
C,
but
not
at
55
degrees,
C.
So
there's
a
an
effective
temperature
there's
also
an
effect
of
what
they
call
annealing,
which
is
where
you
heat
the
condensate,
and
then
you
cool
it,
and
you
do
this
multiple
times
and
then
this
is
something
it
they
observe
this
effect
in
Anne
old
condenses
as
well.
C
We
conclude
that
trap.
Droplet
formation
is
reversible.
It
takes
place
during
a
composition,
change,
droplet
formation
could
be
a
kinetic
process
or
a
thermodynamic
process.
If
the
formation
of
enclosed
droplets
is
a
kinetic
process,
we
would
expect
droplets
to
form
a
fast
composition
changes,
whereas
droplets
would
also
form
a
slow
composition
changes
if
if
it
is
a
thermodynamic
process,
so.
D
C
A
condensate
with
a
radius
of
24
microns
will
contain
only
one
trip,
droplet
after
cooling,
slowly
to
one
degree
C
per
minute,
but
24
droplets
after
cooling,
faster
at
20
degrees,
C
per
minute,
and
so
basically
you
get
differences
in
the
number
of
droplets
or
the
amount
of
cooling.
Thus,
depending
on
the
rate
of
composition,
change,
we
obtain
a
kinetic
product,
a
condensate
with
many
trap
drop.
What's
inside
a
more
thermodynamically
stable
product
of
Condon
set
with
less
enclosed
droplets.
C
This
observation
implies
that
these
droplets
are
formed
when
the
condensate
is
unable
to
reach
the
thermodynamic
equilibrium,
in
which
case
we
would
expect
the
number
of
trap.
Droplets
also
depends
on
the
size
of
the
condenses,
so
the
the
size
of
the
continent
sits
in
the
rate
of
change
in
temperature,
aren't
necessarily
equivalent.
So
there's
no
Trend
between
those
two
things
smaller
condenses
might
equilibrate
faster
with
the
surrounding
dilute
phase,
as
opposed
to
large
condensates.
C
C
C
To
get
to
the
phase
transition
part
but
I
think
it's
really
about
thinking
about
some
of
these
things
with
droplet
formation
and
kinetics
versus
thermodynamics.
So
this
is
the
dynamically
arrested
phase
transition.
Combining
your
observations
allows
us
to
understand
the
mechanism
behind
the
formation
of
this
double
Emulsion,
condensing.
C
Consider
the
binodal
of
a
phase
diagram
for
the
poly
ra
Peg
system,
which
is
the
equilibrium
concentrations
in
the
dilute
and
the
dense
phase.
This
is
figure
five
b,
so
this
is
way
down
when
the
change,
when
the
change
in
the
environmental
conditions
occur,
the
composition
of
that
dilute
and
dense
phase
will
change
to
their
new
equilibrium
positions
on
the
binodal
we
have
observed.
Multiple
systems
that
have
composition,
change
occurs
relatively
quickly.
Droplets
of
dilute
phase
are
formed
in
the
condensate.
C
We
conclude
that
during
the
composition,
change
as
the
system
moves
from
its
original
to
its
new
equilibrium
position
and
you'll
see
that
the
black
dots
the
system
deviates
from
the
binodal
during
this
change
black
lines.
So
if
we
go
down
to
five,
we
can
see
this
is
the
phase
transition.
So
you
have
you
have
this
initial
condition?
C
So
and
then
you
have
these
two
different
types
of
change
and
then
you
have
this
reversing
composition,
change
where
it
reverses
to
the
original
form.
So
for
soul,
composition,
changes,
it
takes
less
change,
less
energy,
perhaps
fast
compositions
and
changes
take
longer
and
involve
more
so
here
we
see
this
temperature,
which
is
Centigrade.
You
go
up
in
terms
of
temperature
and
then
you
go
down
and
you
have
these
different
phase
transitions
between
spinal
and
vinyl
binaural
and
spinaural
change.
So
you
have
this
phase
transition
up.
D
C
Towards
its
peak
temperature
and
then,
as
you
cool,
you
get
these
different
phase
transitions
or
these
different
states
of
of
composition,
change,
and
so
looking
at
the
diameter
of
the
droplet
in
microns
versus
the
cooling
rate,
you
see
that
you
get
no
droplets
formed
in
these
dark
green
circles.
Droplets
formed
in
the
transparent
circles,
and
you
see
that
there's
a
decrease
in
the
diameter
with
an
increase
in
the
cooling
rate,
but
that
the
diet,
larger
diameter
droplets
always
10.
C
You
know
that's
the
formation
and
droplets,
so
this
region
here
is
where
you
get
droplets
formed
within
the
condensate.
Here
you
get
no
droplets
formed,
so
the
cooling
rate
can
be
very
large
for
a
small
diameter
droplets.
You
don't
get
any
sort
of
droplet
formation
and
so
the
same
thing
here
we
get
this
condensate,
that's
maybe
at
a
low
cooling
or
low
changing
cooling
rate
or
a
low
cooling
rate,
and
if
the
diameter
is
big
enough,
you
end
up
with
these
droplets.
C
C
C
C
So
this
is
local
constriction,
it
doesn't
happen
across
the
embryosis
locally.
This
can
lead
to
long-range
flows
due
to
tissue
viscosity.
So
is
your
tissue
gets
deformed
there's
this
viscosity
and
it
can
result
in
long-range
flows.
The
tissues
are
all
connected
by
forces,
so
you
can
actually
get
this
effect
kind.
D
C
C
However,
we
showed
with
experiments
and
modeling
the
onset
of
polarized
tissue
flow
in
early
drosophila,
morphogenesis
drosophila
fruit
flies
in
there.
There
we'll
maybe
see
examples
of
the
embryos
and
how
this
works
occurs,
independent
of
adhesion
and
is
instead
driven
by
a
geometric
coupling
of
apical
actomyosin
contractility
tissue
curvature.
C
So
this
is
related
to
curvature
in
the
embryo
and
how
the
tissues
start
to
curve.
And
then
the
outcome
is
this
polarized
tissue
flow
and
this
so
yeah
we
get
these
tissue
Flows
In,
This
Way,
particularly
the
onset
of
polarized
flow,
is
driven
by
a
mismatch
between
the
position
of
apical,
myosin
activation
and
the
position
of
P
curvature
at
the
posterior
pole
of
the
embryo.
C
C
And
so
that's
one
example
of
tissue
flow.
This
emerges
from
apical
constriction
of
well-defined
groups
of
cells
such
as
the
mesoderm
or
endoderm,
and
C
elegans,
drosophila,
sea
urchin
or
acidians,
which
are
sea
squirts.
Similarly,
apical
constriction
has
been
shown
to
derive
cell
shape
changes
during
vertebrate
development.
So
when
they
say
apical
constriction,
they
mean
there's
a
constriction
at
one
end
of
this
cell
population-
and
you
know
it's
polarized,
the
cells
are
polarized,
so
there's
this
Collective
Behavior
going
on.
C
C
These
two
types
of
symmetric
flows
May
coexist,
which
is
conver
the
process
of
convergent
extension.
So
during
early
drosophila
morphogenesis,
a
single
layer
of
epithelial
cells
is
formed
by
simultaneous
cellularization
of
6000
nuclei,
composing,
the
centitial
blastoderm.
So
there's
this
large
sort
of
single
tissue
that
forms
it's
a
single
layer
of
epithelial
cells
that
results.
When
these
all
this,
this
sort
of
layer,
this
single
tissue
gets
cellularized
by
the
formation
of
many
nuclei
in
a
single
centation.
So
you
get
the
centitium,
which
is
a
big
structure
with
a
water
nuclei.
C
So
we
get
this
this
pattern
of
sort
of
cellularization
and
then
this
need
for
tissue
flow
to
form
some
of
the
structure
that
will
come
later.
We
don't
really
know
what
creates
this
polarized
flow.
It
could
be
due
to
other
processes
that
are
unknown
or
other
processes
that
are
known,
such
as
mesoderm
in
vagination
or
germ
band
extension,
but
it's
kind
of
a
mystery
at
this
point.
So
this
is
an
example
of
the
persophila
embryo.
This
is
the
membrane.
C
This
is
the
interior
of
the
centitium
and
then
this
whole
area
gets
cellularized
and
ends
up
forming
tissue.
So
this
is
the
sort
of
what
happens
here.
You
have
the
posterior
around
the
anterior
end.
This
is
the
eggshell
on
the
outside.
The
dorsal
side
is
on
top.
The
ventral
side
is
on
the
bottom.
What
you
see
is
that
there's
this
movement
that
results
in
this
exact
at
the
dorsal
side,
and
then
you
get
so
you
know
the
whole.
D
C
Ification
of
tissue
flow,
these
cartoons
show
kind
of
what's
going
on,
and
then
these
graphs
show
by
position
these
movements.
So
you
see
that
there's
that
the
tissue
flow
moves
over
time,
there's
this
distinction
between
symmetric
the
symmetric
stage
in
a
polarized
stage.
So
this
is
when
this
whole
thing
is
symmetric
versus
when
it's
polarized
by
these
different
structures,
especially
cephalic
Furrow,
which
moves
down
the
cell
or
down
the
embryo
and
differentiates
things
into
cells.
C
So
they've
performed
these
same
quantifications
on
embryo
mutant
embryos
for
different
types
of
defined
mutant,
which
have
blocked
mesoderm
in
vagination,
German
extension
encephalic,
Furrow
formation,
We
Still
Still,
saw
this
transition
of
polarized
flow.
The
timing
of
the
transition
was
even
slightly
earlier
in
the
mutant
ETS
embryos
than.
C
Type
likely
to
the
lack
of
ventral
pulling
from
the
mesoderm
in
vagination,
as
evidenced
by
The
increased
flow
in
the
ventral
anterior
region
of
the
embryo.
This
clearly
demonstrates
that
the
polarized
fluid
is
not
dependent,
meiosin
polarization
in
The
Germ
band,
which
is
this
egg
Sac.
That
I
showed
you
wearing
geometric
constraints
imposed
by
the
cephalic
Furrow
as
previously
predicted.
This
leads
us
to
two
main
questions:
what
physical
mechanisms
Drive
tissue
full
in
the
early
to
soft
labrio
and.
D
C
E
C
D
C
Axis
you
have
a
ventral
side,
but
it
changes
it
can
change
as
as
things
happen.
So
basically,
the
embryo
isn't
locked
into
these
polarized
coordinates
it's
it's
free
to
move
anyway.
It
likes
as
long
as
it's
rolling
around
the
interior
posterior
axis
so
in
tool.
Vo
embryos,
the
posterior
tissue,
does
not
always
flow
dorsally,
but
instead
of
significant,
more
likely
than
wild
type
to
flow
towards
the
ventral
or
lateral
directions.
C
So
that's
interesting.
You
can
get
mutants
to
do
different.
You
know
have
different
things
emerge
through
these
flows
to
quantify
the
flow
in
the
Imaging
plane.
We
analyzed
only
embryos
of
flow
dorsally
tracking
the
pulse
cells
in
these
mutants
revealed
that
the
early
symmetric
flow
is
almost
completely
halted,
confirming
that
the
symmetric
flow
observed
is
driven
by
The
increased
activation
of
basal
myosin
on
the
dorsal
side
of
the
embryo.
C
This
coincides
with
the
previous
work
from
strikan
at
all,
which
are
the
twist
mutant.
Embryos
which
have
a
strong
reduction
in
DVA
symmetry
or
dorsal.
Ventral
asymmetry
show
a
strong
reduction
of
tissue
movement
towards
the
dorsal
side
of
the
embryo
during
early
stages
of
morphogenesis,
so
they're
able
to
characterize
some
of
these
movements,
some
of
these
tissue
flows
and
match
them
up
with
some
of
the
odd
results
that
people
have
been
getting
in
the
field.
C
C
So
this
is,
you
know
there
are
a
lot
of
processes
going
on
and
they're
able
to
quantify
the
Flow
by
pole
cell
tracking.
We
concluded
that
the
symmetric
phase
of
flow
requires
a
non-uniformity
of
basal
myosin,
and
the
polarized
floor
requires
apical
myosin,
since
polarized
tissue
flow
is
normal
and
ETS
means
and
which
mesoderm
in
vagination
and
germ
band
extension
are
blocked.
We
conclude
that
apical
myosin
is
required
strictly
in
the
posterior
region
of
the
embryo
and
not
in
other
adjacent
tissues.
C
So
this
kind
of
goes
back
to
the
last
paper,
with
condensates
and
just
kind
of
this
self
organization,
where
the
self-assembly
process,
where
things
aren't
meant
to
be
homogeneous,
they're
highly
heterogeneous,
and
sometimes
it's
a
matter
of
just
sampling
or
like
Regional
difference.
So
in
other
words,
if
you
have
some
process
going
and
you
just
take
a
sample
from
one
area
or
another
you'll
get
a
you
know,
you'll
get
some
sort
of
order
that
can
be
Amplified
later.
So
this
is
kind
of
an
interesting
effect.
C
D
D
D
C
V
bar
V
bar
a
minute
microns
per
minute:
I
guess
that's
the
velocity
and
then
the
position
PC
microns
in
microns,
and
you
can
see
the
results
here.
Basically,
it's
increasing
over
time.
In
this
case
you
get
this
difference
between
the
wild
type
and
the
mutant.
In
this
case,
the
mute
starts
higher
and
kind
of
gets
to
the
same
place
in
this
case.
I
think
the
mute
is
this
one.
D
C
D
C
Point
out
that
localized
friction
or
adhesion
is
not
responsible
for
polarized
flow,
so
this
idea
of
using
friction
or
adhesion
as
a
way
to
explain
it
is
not
exactly
what
we
want
as
an
explanatory
framework.
So
during
the
Polaris
phase
of
flow,
the
dorsal
posterior
activation
of
apical
myosin
leads
to
bending
of
the
epithelium.
This
causes
the
anterior
under
the
apical
myosin
domain
to
come
in
close
contact
with
the
egg
show,
which
could
create
a
localized
domain
with
higher
friction.
It.
C
Lead
to
adhesion
between
the
epithelium
and
the
eggshell
intuitively,
this
configuration
could
lead
to
asymmetric
contraction,
but
adhesion
at
the
interior
end
of
the
apical
myosin
domain
has
ever
even
been
shown
to
crucial,
be
crucial
to
the
anterior
wave
propagation
of
the
endoderm
in
vagination
slightly
later
in
development.
So
they
test
this
using
the
polarized
flow
we
observe
could
be
created
by
such
an
asymmetric
friction,
which
of
course,
is
what.
C
C
So
that's
just
the
cells
as
they
curve
and
the
interaction
of
that
gradient
with
an
active
moment,
which
is
from
the
theory
of
active
surfaces
and
so
active
moments
are
created
when
apical
and
basal
myosin
tensions
differ,
so
there's
a
difference
in
tension
and
that
can
create
what
they
call
an
active
moment
when
myosin
and
therefore
active
tension
is
present
at
similar
levels,
apically
and
basically
in
a
region
of
tissue.
This
acts
to
contract
this
region
which
in
turn
exerts
a
forces
on
the
surrounding
tissue.
E
C
So
when
the
levels
of
apical
and
basal
myosin
different
region
of
tissue,
an
active
moment
is
present
that
causes
this
region
to
observe
torques
on
the
neighboring
tissue,
which
then
in
turn
next
increase
or
decrease
the
curvature
of
this
region.
So
they
summarize
this
in
this
figure
active
tension
force
here
that
are
pushing
together
this
x
to
decrease
the
length
of
the
green
region,
which
is
this.
C
This
is
active
tension
pushing
in
a
bunch
of
cells
and
then
the
active
moment
here
is
where
you
get
a
moment,
which
is
where
you
get
torque
and
it's
acting
on
this
green
area
of
cells.
It's
like
an
A,
but
then,
instead
of
pushing
it
together,
it
pushes
it
down
and
it
decreases
curvature
on
the
green
region
and
that's
what
we
have
here
and
so
again
you
can
see
this
in
the
embryo
here
at
the
edge
you
get
this
green
region.