►
From YouTube: DevoWorm (2023, Meeting #33): DevoSAM, 2D and 3D Gastruloids, Models of Differentiation
Description
DevoSAM with prompts and masks. MultiGPU cell segmentation. Overview of 2-D and 3-D Gastruloid biology and suitable computational modeling. Artificial models of cell differentiation, symmetry-breaking, and cellular decision-making. Cyanobacterial filaments as models of differential single cell morphogenesis. Attendees: Sushmanth Reddy Mereddy, Luke Knoble, Bradly Alicea, Jesse Parent, Susan Crawford-Young, and Richard Gordon.
B
Are
you
fine,
Bradley,
ordered
in
last
week
meeting
and
actually
I
was
going
through
some
College
full-time
work,
so
I
couldn't
manage
to
join
okay
yeah,
that's
the
thing
going
on
and
right
now
this
week,
I
couldn't
work
on,
because
I
have
exams
going
on
right
now
my
college
exam,
so
I
stopped
a
little
bit
for
the
preparation
for
one
week.
Maybe
from
next
week
I
will
resume
my
work
and
whatever
the
update
I
have
given
in
the
slack.
That's
the
only
thing.
B
I
just
talked
with
Mayu
had
a
meeting
with
him.
I
explained
about
the.
What
are
the
right
now
situations
going
on
right
now,
I
was
working
on
the
actually
Sam
model,
Bradley
it's
on
a
large
model,
so
the
ram
axis
was
around
80
GPU
80
gb,
not
8,
GBP,
80GB
and
right
now,
a100
has
only
support
of
48
GB.
So
I
was
thinking
to
add
a
multi
GPU
support
for
this
thing,
which
could
work
fine,
so
right
now,
I
was
working
on
that
feature.
Multi
GPU
support
and
I
was
integrating
W
and
biases
Library.
B
B
Yeah
like
this,
like
how
the
neural
network
is
working
on
everything
like
graphs,
Etc,
how
the
weights
are
updated
here,
all
things
will
be
mentioned
over
here.
So
I
was
trying
to
add
the
support
in
our
code.
It's
like
pretty
messy
I.
Could
there
are
a
lot
of
tutorials
but
I
couldn't
integrate
with
the
Sam
model
right
now,
I
was
working
on
this
and
I
was
playing.
I
was
working
on
multi-gpu
Support
also
I
have
thought
about
of
the
further
improvements.
B
In
this
thing
we
can
add
fine
tuning
and
we
can
try
to
freeze
to
other
models
to
other
encoders
and
to
try
to
train
only
on
one
encoder.
So
that's
the
thing
in
my
mind:
I'll
try
to
create
a
documentation
for
all
this
for
further
improvements
and
Etc
right
now,
I
was
working
on
this
block
to
add
this
support
in
wnb
support
in
Sam
model
and
parallel
I
was
working
on
Multi
GPU
Support.
B
Also,
so
our
model
can
be
trained,
not
only
one
GPU,
it
could
be
trained
on
multiple
gpus
cover
can
be
distributed
like
around
this
all
gpus.
If
you
have
five
gpus,
the
work
will
be
split
on
all
these
5gpl,
so
it
can
be
easily
trained.
And
after
this,
when
we
try
to
add
this
multi
GPU
support,
we
can
train
our
model
on
more
data,
actually
not
only
see
any
against,
but
also.
B
There
are
some
I
am
not
getting
the
mechanical,
yes,
sorry
that
model
and
other
models.
Whatever
the
data
we
have
on
Cell
tracking
challenge
data
set,
we
can
directly
take
them
and
try
in
our
model
on
that,
so
it
would
work
so
fine
and
it
could
be
used
to
any.
You
could
use
it
to
try
to
segment
any
kind
of
cell,
not
only
Clans,
but
also
another
thing
also,
but
before
that
we
need
to
add
this
multi-gpu
support
right
now
is
working
on
this.
B
Only
till
now,
I
never
worked
on
the
multi
GPU
support,
so
I
was
trying
to
learn
it
first
and
I
will
Implement
on
the
Sam
mode.
These
are
my
update.
Traveling
I
think
I
have
nothing
to
show
right
now,
but
maybe
by
next
couple
of
weeks,
I
will
show
it
to
you
and
that's
the
results
which
I
have
shared
in
the
group.
Actually,
okay,
let
me
share
flow
to
here.
These
people
also,
okay,.
B
This
is
the
model
I
have
trained.
This
is
how
it's
working.
We
can
segment
based
on
the
prompts
if
any
cell
is
not
working.
This
is
a
general.
It
has
segmented
it,
but
if
we
provide
the
problems
over
here
sandbox,
if
we
give
a
box
around
any
cell,
it
will
try
to
segment
it
as
a
single
cell.
This
is
the
beauty
of
instant
segmentation,
so
that's
what
it
I
will
try
to
work
more
on
this
and
try
to
increase
the
model.
B
A
B
B
It
is
morally
used
for
the
visualization
of
predictions.
If
you
see
here
is
the
data
set
of
images?
Actually,
you
can
see
in
this
picture.
There
are
multiple
images
of
it
for
this
guess
and
the
truth
Etc
will
be
given.
Actually,
according
to
that,
we
will
get
different
type
of
all
things
of
tracking
how
the
our
experiment
is
going
on
how
accuracy
is
going
on
loss
functions.
You
mentioned
me
like
adding
more
loss,
functions,
graphs,
Etc,
so
I
am
using
this
library
to
track
all
these
experiments.
B
A
B
This
could
be
used
for
any
ml
models.
You
can
track
our
experiment
just
by
giving
input
to
it
so
I'm,
trying
to
add
this
support
to
Sam
model
and
multi
GPU
Support.
Also
multi
GPU
support
means
Bradley
as
Sam
as
375
million
parameters.
It's
large
huge
model
right
now.
Devoland
was
90
million
parameters.
It
has
no
problem
with
me,
but
it
is
when
it
has
increased
to
375
million
parameters
need
more
RAM,
actually
I
know
80
gb
Ram
right
now.
Whatever
GPU
we
are
using
is
a100
GPU.
B
B
D
A
B
This
is
a
fluorescence
image
of
a
different
Model
A
different
organism-
I
just
it
has
so
many
cells,
so
it
has
more
than
sense
of
Cl
gas,
so
I
thought
of
taking
this
as
a
input
image
and
try
to
segment
it
so
I
have
given
the
weights
file
and
try
to
segment
it
as
a
tool.
I
use
a
napari,
another
Library
actually
to
segment
it.
So
when
I
tried
this,
this
is
these
are
the
results
I
was
getting.
B
B
Cells
are
missing,
I
mean
when
we
are
trying
to
segment.
If
some
cells
are
missing,
we
could
just
give
a
bonding
box
and
it
will
segment
that
part
also
missing
parts
also
yeah,
and
this
could
be
useful
for
data
annotation
tool
for
cells
and
we
can
directly
segment
cells
and
track
the
live
cell.
Also
right
now.
There
is
no
support
for
tracking
live
cells
in
segmenting
them.
But
if
we
try
to
train
the
model
on
the
video
data,
it
could
be
useful
actually
for
tracking
live
cells
and
segmenting
them.
B
A
B
Yeah
we
need
to
create
a
bonding
box
around
the
cell;
it
will
directly
segment
it
okay
and
we
can
not
only
give
prompt
as
bonding
blocks
but
also
points
I
mean
if
you
just
keep
a
point
on
this
cell.
Actually
it
will
segment
that
cell
directly
and
you
need
to
eliminate
the
negative
region
over
here
the
background,
so
it
will
try
to
segment
only
that
single
cell,
for
example.
At
this
point
of
time,
you
could
see
he
will
keep
a
dog
I
mean
I
have
kept
a
DOT
over
here
see
it
has
been
segmented
singlely.
B
A
Or
different
like
embryos,
but
also
you
have
you
know
different
types
of
data
like
what's
inside
the
cell,
so,
for
example,
if
you
want
to
say
this,
these
cells
Express,
this
Gene,
you
can,
you
know,
maybe
find
like
a
gfp
signal
which
is
like
an
intensity,
and
you
can
train
the
model
on
that
and
then
match
it
to
these
cell
identities
and
then
use
that
in
another
like
if
I
were
to
take
another
sample
or
another
type
of
image.
I.
B
Analysis
we
can
track
them
yeah.
We
could
add
that
support,
but
it
requires
a
lot
of
computation.
Actually
yeah
I
mean
we
don't
have
the
gpus.
This
gpus
only
I
took
from
my
University
and
tried
to
train
the
model.
If
you
have
the
better
gpus
support,
you
could
definitely
add
that
support
it's
easy
only,
but
we
need
to
train
the
model
as
a
model
is
so
big,
I
couldn't
print.
B
B
I
have
updated
the
whole
thing
to
make.
He
was
not
actually
we
had
a
meeting
yesterday.
There
is
a
bug
I
was
caught
with.
He
helped
put
me
out
with
that,
and
the
bug
was
like
to
access
this
GPU
issues.
Only
the
model
is
so
big,
so
I
was
having
the
problem,
so
he
came
and
tried
to
split
the
work
and
there
are
a
lot
of
things
he
cleared
out
when
I
never
left
the
meeting.
Yeah
yeah,
that's
the
complete
my
update.
Thank
you
agree.
A
A
A
A
microscopy
data
from
Susan
and
I
had
Susan
sent
me
the
data
that
she
has
produced
recently
that
she
sent
me
in
the
mail
and
I
put
that
on
like
a
private
sync
box,
so
he's
using
that
and
he
actually
requested
something.
I
don't
know:
Susan
got
the
email,
it
was
about
like
I,
guess
a
coordinate
map
or
a
reference
map
of
the
different
I.
Don't
know
what
what
he
means
by
that.
A
A
C
D
C
D
B
D
D
C
B
C
Well,
I
I
do
have
one
figure
from
him
or,
and
it's
like
at
the
actin
ring,
is
and
Mega
pascals
like
it's
very
stiff.
Oh,
it
makes
sense
for
a
tense
agree
to
have
a
stiff
streaks.
It
does
make
sense.
Okay,
good,
but
I
need
like
oh
well.
I
I'll,
try
looking
it
up.
It's
just
like,
sometimes
like
finding
a
needle
and
a
haystack
yeah.
C
C
Yeah
yeah
I
hope
they
just
they
just.
Let
me
do
this
since
I've
already
started
it
and
I'll
just
I
need
to
write
it
up
like
I've
done
a
lot
and
I
haven't
properly
written
it
up,
so
I'm
going
to
write
them
up
a
couple
of
things
about
it,
and
so
they
they
know
where
I'm
coming
from
with
it
and
yeah,
especially
the
dentist.
Maybe
you'll,
have
some
things
to
add,
add
to
it
a
bit
of
critique
yeah.
A
All
right
Jesse,
how
are
you
I,
don't
know
if
you're
available
to
talk
and
it's
fine
yeah?
So
it's
let
me
share
my
screen
so
found
a
couple
of
interesting
papers
this
week.
The
first
one
is
on
and
I
know
that
Morgan's
not
here.
He
likes
this
sort
of
thing,
especially,
but
this
might
be
very
interesting
to
people
a
lot
of
the
stuff
we're
doing
with
or
we're
talking
about,
embryoids
and
gastroides,
and
these
these
different
types
of
cell
models,
where
they're
trying
to
differentiate
cells
in
a.
A
Culture
or
two-dimensional
culture,
and
so
we
have
let
me
start
with
stem
cells.
We
get
a
bunch
of
differentiated
cells
that
resembles
something
in
like
an
organoid
or
a
gastroloid
or
an
embryoid,
and
so
this
is
a
nice
review
that
came
out
recently
in
current
opinion
in
genetics
and
development.
So
this
is
on
the
ever-growing
world
of
gastroloid.
So
this
is
one
of
the
subsets
of
organoids
Auto
genius
models
of
mammalian
embryogenesis.
So
this
is
something
this
is
an
ember.
This
is
mammalian
tissues,
mammalian
development.
A
So
that's
where
a
lot
of
this
activity
is
going
on
in
this
area,
I'm,
not
sure.
If
they've
built
an
embryoid
of
C,
elegans
I,
don't
think
it's
I,
don't
think
they
really
need
to,
because
I
think
we
know
kind
of
how
that
how
those
you
know
we
have
that
sort
of
deterministic
development,
but
in
mammals
you
know
we
get
a
lot
of
chemical
signals
and
different
ways
that
cells
differentiate.
So
this
is
this
is
why
this
is
so
interesting.
A
So
that's
why
they
call
these
gastroides,
because
they're
trying
to
approximate
gastrulation
and
not
necessarily
just
general
embryogenesis
like
we-
have,
for
example,
neural
organoids,
which
approximate
things
in
the
brain.
You
know
you
want
to
differentiate
tissues
in
the
brain
or
other
types
of
things
like
that.
Advances
in
stem
cell
biology
have
resulted
in
the
generation
of
stem
cell
based
in
vitro
models.
So
they're
doing
this
in
a
culture,
and
so
this
is
where
we
started
with
stem
cells.
A
A
Gastroloids
and
subsequent
gastrolyte-based
models
are
tractable,
scalable
and
more
accessible
than
mammalian
embryos.
So
you
can
use
these
kind
of
models
to
sort
of
approximate
a
mammalian
embryo,
but
they're
also
much
more
easy
to
work
with
for
medical
applications.
So
a
lot
of
applications
of
of
organoids
and
gastrolides
in
particular,
have
been
used
to
look
at
early
development,
or
you
know
different.
A
The
precursors
of
different
diseases
and
things
like
that
as
such
they've
opened
an
unprecedented
Avenue
for
modeling
in
vitro
self-organization,
which
is
of
course,
the
self-organization
or
self-assembly
that
we
see
in
the
formation
of
tissues,
patterning
and
fate
specification.
This
review
focuses
on
discussing
the
recent
advances
of
this
rapidly
moving
research
area
and
so
basically.
A
Hierarchy
within
the
tissues
themselves
or
within
the
cells,
so
you
know
we
can
understand
that
with
these
models
because
we
can
go,
we
can
measure
things
in
the
model.
You
know
once
as
it's
forming
or
after
it's
formed.
Sometimes
you
have
to
do
it
after
the
fact,
which
is
not
as
easy
to
extract
data
from,
but
sometimes
you
can
do
it
sort
of
dynamically.
So
you
can
measure
things
in
as
the
cells
are
in
the
process
of
differentiating.
A
B
A
Summary
of
the
main
difference
between
2D
and
3D
gastroloids,
so
I
said
that
they
we
do
things
with
3D
gastroids,
but
we
can
also
do
things
with
2D
gastroloids,
and
the
difference,
of
course,
is
that
you
have
cell
culture,
two-dimensional
cell
culture
being
like
a
flat
surface.
You
know
you
might
have
a
couple
of
layers
of
cells
where
the
cells
expand
on
a
two-dimensional
plate
and
they
can
kind
of
stack
on
top
of
one
another,
but
they
don't
really
have
any
depth.
A
B
A
They
give
you
geometry
in
two
Dimensions,
but
not
three
dimensions.
So
in
a
two-dimensional
gastroid
the
shape
is
fixed,
but
the
sizes
can
vary
in
3D
gastroides.
The
sizes
can
vary,
but
the
shape
is
variable,
so
you
can
have,
for
example,
in
2D
gastroloids.
It
kind
of
expands
from
the
center
of
the
plate
out
to
the
edge
of
the
plate,
and
you
can
have
different
geometries
around
that
fixed
point.
A
D
A
Shape
can
be
quite
variable.
They
have
some
references
here
for
the
shape
and
geometry
of
these
esteroids
here
so
I.
Can
you
know
I'll
make
this
paper
available.
You
can
look
these
references
up
notice
that
a
lot
of
these
dates
are
past
post
2010.
So
this
is
a
young
field.
This
is
not
a
very
they've,
not
been
doing
this
for
very
long,
so
maybe
15
years
at
best,
maybe
not
even
that
long.
So
this
is
some.
This
is
an
emerging
field.
It's
not
something
that
a
lot
of
you
know.
A
The
second
thing
are
cell
types
one,
so
we
can
form
different
cell
types
from
our
gastroloid,
we're
not
guaranteed
any
cell
types
at
all.
So
we
have
to
have
the
right
culture
conditions,
actually
the
right
like
surface
conditions
in
the
culture,
but
also
the
medium.
Now
the
the
surface
can
be
patterned,
it
can
be.
There
could
be
a
scaffold
like
in
the
3D
gastro
light
example.
A
You
might
have
like
a
circular
scaffold
that
you
might
have
the
cells
you
know
plated
on
and
that
can
take
the
shape
of
some
sort
of
like
Bowl
structure
or
a
spherical
structure.
So
if
you're
trying
to
engineer
a
tissue,
you
might
try
to
build
it
around
the
spherical
structure
and
then
the
cells
will
differentiate
and
form
this
gastroloid
and
then
eventually,
maybe
form
the
precursor
of
some
sort
of
ladder
or
something
you
know.
A
So
that's
the
idea,
so
you
can
actually
get
different
cell
types
just
by
introducing
geometry
so
in
but
in
2D
gasteroids.
It's
interesting
that
you
get
both
embryonic
cells,
so
cells
that
are
known
to
be
specifically
embryonic
and
you
get
cells
that
are
extra
embryonic,
which
means
that
they're
beyond
sort
of
what
we
expect
out
of
an
embryonic
culture
or
an
embryonic
sample
in
3D
gastrolytes.
We
can
be
more
specific.
A
We
only
get
embryonic
cells
and
I,
don't
know
the
reason
for
that,
necessarily
it
could
be
because
you
have
that
geometry,
a
3D
geometry.
It's
not
clear!
The
references
are
here:
initial
cell
number
in
2D
gastrolides,
you
get
several
thousand
and
that's
because
you
have
that
flat
surface
where
you
might
have
a
two
or
three
cell
layer,
one
one
on
top
of
the
other.
A
So
you
don't
you're
limited
in
in
terms
of
the
size
that
you
could
have
basically
by
your
plate,
diameter
and
your
you
know,
maybe
two
or
three
layers
of
cells
in
that
space
and
that's
again
that
geometry
is
responsible
for
that.
Then
we
go
to
Signal
ingredients
and
TD
gastroloids.
We
have
external
and
internal
signaling
sources.
A
A
Sometimes
if
you,
if,
for
example,
stem
cells
in
the
center
and
stem
cells
up
towards
the
edge
of
those
two
populations
can't
ever
contact
one
another
and
so
they're
vulnerable
to
external
and
internal
signaling
sources,
whereas
if
you
have
you
know
a
a
structure
that
can
maybe
migrate
towards
itself
towards
a
like
cell
type,
you
end
up
with
these
structures
that
are
self-sustaining.
I,
don't
know,
I
just
came
up
with
that
explanation.
These
are
the
references
for
that.
A
B
A
Is
400
and
not
hours,
as
I
said
before,
as
I
misspoke,
there
are
many
more
oh
I
see
two
to
three
hundred
cells
for
a
mouse
400
for
a
human,
so
you
don't
get
as
many
cells
in
the
3D
gastroid.
You
actually
get
fewer
like
if
there's
a
difference
in
the
cells
for
Mouse
and
human
a
little
bit,
but
it's
less
than
the
2D
asteroid,
so
yeah.
So
scalability
is
medium
to
high
in
the
TD
gastroy
you
have
high
reproducibility
3D
gas
droids.
A
A
A
You
know
they're
kind
of
just
kind
of
working
on
starting
to
work
on
them
now
trying
to
get
things
you
know
trying
to
get
them
to
maybe
14
days
into
development
and
they're,
not
necessarily
embryo
real
embryos
are
just
things
that
resemble
our
burritos.
So
as
a
result,
you
know
you
don't
really
have
the
reproducibility
that
you
might
have
in
Mouse,
where
we've
been
doing
this
a
little
bit
longer.
A
A
2D
gastroids
do
provide
the
spatial
organization,
they
provide
lineage
information
and
3D
gastroyed
self-pattern
and
organize
to
generate
embryonic
cell
type,
so
3D
in
3D
gastroides.
We
get
this
I
guess,
maybe
more
controlled
cell
type
and
in
2D
gasteroids
we
get
this.
Maybe
greater
spatial
organization
and
so
I
I,
don't
know
if
there's
a
lot
of
data
available
for
us,
but
it
might
be
interesting
to
look
at
some
of
these
Trends
in
2D
versus
3D
gas
droids.
A
A
lot
of
the
data
that's
been
collected
on
these
are
rna-seq,
so
it's
a
sec,
next-gen
sequencing
data,
but
I
don't
know
if
there's
like
a
lot
of
position,
data
or
things
like
that,
but
it'd
be
interesting
to
see
because
a
lot
of
times
what
they'll
do
is
they'll
take
a
a
sample
and
they'll
have
to
sacrifice
a
sample
at
the
end.
So
you
can't
measure
it.
As
I
said
dynamically.
B
A
Can
just
measure
it
at
the
end
and
find
like
Mark.
You
know
RNA
seek
markers,
but
you
know
being
able
to
track
some
there's.
Some
data
I
know
that
where
they
track
this
with
a
camera,
so
they're,
actually
looking
at
the
cells
and
they're
able
to
track
the
cells
like
we're
doing
in
Embry
real
embryos.
A
A
So
in
2D
gastroloids
you
get
stem
cell
culture,
you
dissociate
the
cells,
you
seed,
your
micro
pattern
or
your
plate,
and
then
you
get
patterning
in
3D
gastroides,
you
dissociate
the
cells,
you
aggregate
them
into
a
3D
culture,
you
get
symmetry
breaking,
and
then
you
get
this
elongation
and
patterning.
So
you
get
the
you
know
the
shape
here,
so
you're
able
to
get
elongation
or
readily
than
you
have
in
a
two-dimensional
one,
but
you
get
patterning
in
both.
A
A
I'm
not
sure
they
may
have
I
know
that
they've
people
have
done
the
time
lapse,
kind
of
analysis,
I've
talked
to
several
people
about
modeling,
yeah
embryoids
as
well,
and
we've
been
trying
to
do
this
for
a
while
conversation
keeps
coming
up
and
the
the
thing
is
like
you
know
you
can.
Probably
it
would
be
a
good
like
the
3D
gastroids
I
think
especially,
would
be
a
good
model
for
looking
at
waves.
But
I
don't
know
if
the
data
set
exists.
A
C
Yeah,
so
it
makes
sense
because
you've
got
on
the
surface.
It's
actually
the
connection
point
between
cells,
it's
an
active
ice
and
complex
yeah.
So
if
you
have
a
calcium
wave,
you
probably
have
a
contraction
wave
to
go
with
it
at
least
most
times
yeah,
so
I
actually
have
found
an
article
about
contract,
calcium
waves,
I've
been
downloading
too
many
things
lately
and
kind
of
panic.
So
I
don't
know
where
I
put
that.
C
A
A
Kinds
of
structures
you
get
in
3D
and
2D
gas
Droid
models,
so
if
we
have
they're
different
types
of
gastroides,
these
are
human
and
mouse
examples.
So
in
both
human
and
mouse
gastroids.
In
the
3D
case,
we
have
we've
been
able
to
reproduce
mesoderm
versus
A
Primitive
like
streak
in
the
structure
in
2D
gastroloid
models,
we've
had
we
get
sort
of
this
neuro
ectoderm
in
the
middle
amnion
outside
and
then
mesoderm
and
endoderm
kind
of
in
this
in
the
in
between.
A
We
also
have
human
and
mouse
gastroids.
We
have
anteriorized
and
cautilized,
which
means
that
they're,
you
know,
there's
polarity
there
between
the
front
part
and
the
back
part,
and
we
get
again
neural
cell
types
in
A
Primitive
like
streak,
so
you
get
these
differentiations
that
are
kind
of
weird
from
a
developmental
perspective,
because
they're
not
exactly
like
what
you
would
see
in
an
embryo,
but
you
get
this
organization,
you
get
the
neural
plate,
the
neural
caress,
the
play
code
and
the
non-neuro
ectoderm
in
these
type
of
anteriorized
gastrointuroloids.
A
So
this
is
where
it's
towards
the
front,
so
it's
forming
the
brain
basically
or
something
that
resembles
a
brain.
So
it's
just
the
front
part
of
the
human
embryo.
You
know-
and
it's
sort
of
this
analogy
of
that
and
it's
producing
neural
tissues.
So
you
can
do
things
like
that.
You
can
get
like
different
poles
of
the
embryo
and
you
can
it
again.
A
You
do
neural
cell
types
on
one
end
in
mesoderm,
endoderm
on
the
other,
and
then
you
get
these
neural
tube,
like
structures
and
human
embryoids
or
human-like
embryoids,
which
have
this
sort
of
neural
tube
where
I
guess
yeah,
it's
a
neural
tube
here
with
non-neuroactoderms
surrounding
it.
So
you
get
these
different
types
of.
You
know,
structures
that
resemble
things
you'd
see
in
development.
But
you
know
it's
not
really
like
development,
because
it
doesn't
always
I
mean
it.
A
You
know
we
could
say
it's
like
development,
because
you
know
it's
very
similar
to
it,
but
we
can't
know
for
sure,
because
we
don't
really
know
the
underlying
rules
of
it
so
yeah.
So
there
are
all
sorts
of
structures
you
can
form,
it
doesn't.
I.
Imagine
you
know
they're,
showing
the
sort
of
the
Royal
Road
here
they're,
showing
the
best
examples
of
what
you
can
achieve.
A
I'm
sure
there
are
a
lot
of
cases
where
you
could
do
this
and
you
could
just
get
like
random
cell
types,
but
most
of
the
time
you
get
these
sort
of
you
know
once
you
get
a
cell
type
because
you're
in
this
spatial
context
you
can
get
clusters
of
similar
cell
types
and
in
fact
you
might
expect
that.
But
if
they
form
you
know
whether
they
form
a
coherent.
B
B
A
Important
just
in
terms
of
studying
signaling
hierarchy,
they
have
people
have
looked
at
different
morphogenetic
markers
like
emp4
and
wind
and
nodal
signaling.
So
you
know
we
get
that
sort
of.
We
recapitulate
some
of
that
sort
of
stuff
they're
different
protocols.
You
can
follow
to
get
some
of
these
results.
You
know
so
people
are
always
experimenting
with
protocols
trying
to
get
different
types
of
structures
and.
A
C
A
A
A
So
this
is
actually
from
the
a
life
conference
from
last
year
or
this
last
year,
23.-
and
it's
really
interesting,
they're
doing
this
in
silicomorphogenetic
engineering
with
differentiable
programming,
and
so
this
is
a
paper
where
they're
actually
trying
to
build
an
embryo
using
I,
guess,
neural
networks
and
other
types
of
computational
tools.
So
you
can
see
here
in
figure
one.
This
is
a
schematic
representation
of
the
building
blocks
of
the
model
described
in
the
main
text.
So
you
have
this.
A
The
middle
of
the
cell,
you
have
these
different
aspects
of
development.
You
have
diffusion
cell
division
and
mechanical
relaxation,
so
you
have
these
chemical
signals,
you
have
these
sort
of
cell
behaviors
and
you
have
these
mechanical
factors
and
they're
modeling
all
this
in
this
and
so
they're
trying
to
figure
out
sort
of
what
are
the
maybe
the
causal
factors
in
in
a
developing
system.
So
let's
read
the
abstract
and
go
through
they're
just
interested
in
this
area.
A
Morphogenetic
engineering
and,
of
course,
we've
run
into
that
from
Mike
Levin
in
the
work
that
he's
doing
with
people
on
like
using
embryos
as
a
basis
for
looking
at
biological
engineering
and
even
like
bio
robots,
where
they're
building
little
robots
from
like
embryo
inspired
robots.
So
this
is
a
this
is
an
actual
field.
Everything
is
differentiable
programming
paradigm.
A
So
let's
see
how
they
do
this.
A
notoriously
difficult
challenge
in
biology
is
to
understand
how
cells
can
be
directed
to
grow
and
spontaneously
arrange
themselves
in
a
despite
desired
spatial
pattern
like,
as
we
saw
in
the
other
paper,
there's
a
spatial
patterning
we
want
to
achieve.
You
know
how
do
we
achieve
it?
We
can
put
cells
in
culture,
we
can
put
different
media
in
that
culture.
We
can
let
them
grow
on
a
substrate
and
they
can
grow
out
we're
in
a
three-dimensional
substrate
on
a
scaffold
or
whatever,
and
they
form
these
patterns
and.
A
We
can
control
it,
but
it's
hard
to
know.
You
know
how
to
get
precision
and
that
control.
In
this
study
we
leverage
recent
advances
an
automatic
differentiation
ingredient
based
optimization.
So
these
are
things
from
computational
science.
These
are
not
things
from
biological
science,
there's
automatic
differentiation
and
gradient-based,
optimization,
of
course,
grading
based,
optimization
being
you're,
finding
like
the
lowest
or
highest,
energetic
State
or
the
lowest
or
highest
mechanical
state,
or
whatever
and
you're
optimizing
to
that
point.
A
So
that's
that's
what
they're
trying
to
do
here
so
they're
trying
to
use
that
to
discover
local
interaction,
rules
that
yield
some
desired
emergent
system
level
characteristics
in
a
complex
biology,
inspired
model.
We
consider
a
model
where
cell
to
cell
interactions
are
mediated
by
physical
processes
such
as
morphogen,
diffusion
cell,
adhesion
and
mechanical
stress.
C
A
A
Well,
these
are
pretty
complex
processes,
so
diffusion
chemical
diffusion
itself
is
complex
and
mechanical
aspects
are
complex
as
we've
seen
and
then
the
division
probability
is
a
bit.
You
know
I
guess
that's
probably
the
easiest
part,
but
that's
something
that
you
you
know
it's
not
always
the
case
that
it's
a
constant,
so
yeah
anyways.
This
is
just
a
model
that
they're
introducing.
A
We
consider
a
model
where
cell
cell
interactions
are
mediated
cells,
take
internal
decisions
such
as
whether
to
divide
or
not.
So
that's
this
division
probability.
This
is
considered
a
decision
based
on
their
local
environment,
so
they
sample
their
local
environment.
They
take
this
decision,
they
say,
should
I
divide
it
or
not,
if
I,
if
the
local
environment
is
suitable,
if
I'm
in
the
right
mechanical
position,
my
neighbors
are
close
or
far
apart.
If
the
chemical
signals
are
right,
they'll
decide
to
divide
and
then
that's
my
division.
A
Probability
that's
generated
by
these
conditions
based
on
their
local
environment
with
learnable
policies,
meaning
that
they're
actually
using
something
akin
to
reinforcement.
Learning
where
you
have
a
policy
that
says
this
set
of
division.
Probabilities
is
best.
These
are
the
best
responses
to
these
conditions.
So
that's
what
they're,
trying.
D
C
A
Parameterized
with
feed
forward
neural
networks,
so
they're
using
these
neural
networks
as
a
way
to
sort
of
you
know,
find
the
best
probabilities
for
division,
the
best
conditions
for
dividing
and
so
forth,
and
then
so
that's
how
they're
doing
that
we
present
here.
Some
preliminary
results
that
showcase
how
this
approach
can
Discover
cell
interactions
that
breaks
symmetry
in
a
growing
cluster.
A
So
what
they're
doing
is
they're
trying
to
figure
out
these
interactions,
which
interactions
are
optimal
for
cells
to
divide
and
then
ultimately
break
symmetry,
so
they
want
to
create
emergent
chemical
gradients
and
homogenized
cluster
neural
biomechanical
stress
response.
So
that's
what
they're
trying
to
do
here
so
there
have
been
attempts
to
do
this
through
hand.
Devised
rules
are
optimized
by
genetic
algorithms.
A
That's
been
like
kind
of
the
standard
thing
for
artificial
life
only
in
recent
years,
due
to
our
ability
to
use
gpus
and
tpus
and
other
type
of
hardware
and
software
like
efficient
automatic
differentiation,
which
is
actually
different
from
cell
differentiation.
It's
a
computational
technique
has
gradient-based
optimization,
become
a
viable
option.
So
we.
B
A
Tools,
but
now
we
do
and
at
least
when
you
have
predium-based
optimization
to
work
from
now,
whether
that's
the
best
novel
model
or
not,
is
you
know,
that's
a
question,
but
you
know
so
one
one
example
of
this
actually
has
been
in
neural
cellular
automata,
which
we've
talked
about
before
this
is
where
we're
using
cellular
automata
combined
with
neural
Nets
and
we're
kind
of
getting.
You
know
producing
patterns
of
activation
cell
activations
in
cellular
automata,
using
a
neural
network.
A
So
we're
able
to
do
that
and
actually
produce
some
pretty
coherent
pattern
formation,
and
but
people
have
used
this
in
other
fields
like
in
the
design
of
self-assembling
materials
and
the
learning
of
data-driven
discretizations
for
partial
differential
equations.
So
people
have
used
this
for
looking
at
self
self
assembly
and
Material
Science
and
in
sort
of
you
know,
differential
equations
more
generally,
so.
A
Know,
there's
some
really
interesting
directions
for
some
of
these
things,
what
they
do
here.
They
want
to
take
a
step,
Beyond,
neural,
cellular
or
automata,
and
build
these
kind
of
models
of
cell
division
and
cell
differentiation
and
so
they're
building
these
different
these
cell
models,
which
are
soft
spheres
in
a
two-dimensional
space.
So
they
have
this
two-dimensional
space
like
the
two-dimensional
cultures
that
we
sell
in
the
last
paper
and
they're
using
different
potentials
for
the
mechanics
and
for
the
chemical
interactions.
A
So
they
have
this
sort
of
Assumption
of
symmetry
with
respect
to
divisions,
and
they
can
also
sense
mechanical
stress.
So
they
have
these
sense,
sensory
things
that
they
can
have
as
inputs,
and
then
they
have
a
decision-making
circuit.
So
this
is
where
they're
learning
to
break
symmetry
in
figure
two:
okay,
there
we
go.
This
is
where
they
have
so
in
panel
a
here
they
have.
A
This
is
I
guess
these
are
division
rates,
so
the
higher
the
division
rate,
I,
guess
the
more
app
the
environment
is
to
giving
it
to
allowing
for
divisions.
So
basically
they
have
two
conditions,
one
with
random
parameters.
Where
you
know
they're
basically
doesn't
know
what
it's
doing.
It's
just
dividing
and
the
other
is
where
they're
learning
the
parameters
from
the
environment
from
the
sort
of
the
local
conditions.
A
A
Some
effective
learning,
where
it's
learning
these
parameters
over
time
and
it's
figuring
out
how
to
make
so
it's
kind
of
making
more
of
a
pattern
here
in
B,
you
have
a
little
bit
different
situation
so
in
a
you're
trying
to
learn
to
break
symmetry,
so
you're
trying
to
figure
out
how
to
break
Symmetry
and
form
like
a
pattern,
or
you
know
a
differential
sort
of
shape
in
this
actually
down
at
the
bottom
of
a
this.
These
are
the
Learned
chemical
gradients,
which
are
actually
the
average
concentration
in
arbitrary
units
versus
Space.
A
So
you
can
see
that
in
in
We
have
basically
for
chemical
two,
we
have
a
peak
on
the
right
side
and
then
for
chemical,
one,
there's
a
peak
on
the
left
side
and
they
sort
of
intersect
at
around
zero.
So
you
can
see
that
there's
chemical
one
in
chemical,
two
for
the
one
parameter
condition
that
you
have
this
distinction
in
terms
of
which
cells
are
learning
which
chemical
gradient.
A
In
b,
we
have
random
parameters
and
learned
parameters
for
homogeneous
growth,
regulation
via
mechanical
stress,
so
in
a
you're
trying
to
learn
to
break
Symmetry
and
B
we're
looking
more
at
sort
of
growth
regulation,
that's
homogeneous
throughout
the
population
of
cells,
and
this
is
using
mechanical
stress
as
an
input.
So
this
again
on
the
left,
you
have
random
parameters
on
the
right.
You
have
learned
parameters
now.
A
It's
interesting
that
with
random
parameters,
the
division
rates
are
higher
in
the
middle
and
then
fall
off
towards
the
edge,
and
we
might
expect
that
if
we're
just
dealing
with
mechanical
parameters,
because
these
cells
in
the
middle
were
packed
in
more
tightly
and
so
they're,
you
know
experiencing
more
mechanical
stress,
so
they
may
be
dividing
more,
which
is
you
know
it's
it's
fine,
but
that's.
We
expect
that
from
the
random
condition,
but
from
the
learn
parameters.
A
We
see
that
there's
not
only
no
difference
across
the
entire
population
in
terms
of
division
rate,
but
the
edges
are
kind
of
like
kind
of
sticking
out
and
so
they're
actually
learning.
Maybe
the
you
know
what
the
chemical
stresses
should
be
and
what
this
you
know,
you're
kind
of
maybe
getting
a
little
bit
more
structure
or
you're
getting
exploration.
A
Basically,
the
difference
is
that
the
the
mechanical
stresses
are
being
compensated
in
hereby
learning
what
they
should
be
or
what
they
are
and
then,
with
with
respect
to
stresses
we
have.
This
is
in
pascals
or
I.
Don't
know
what
the
actual
units
are
here
for
stress,
but
we
have
0
to
negative
250
and
so
again
in
the
middle.
You
have
this
negative,
High,
negative
stress
and
at
the
outer
edge
you
have
a
neutral
stress
and
it's
basically
the
same
for
random
parameters
versus
alarm
parameters.
A
Although
learn
parameters,
you
have
a
little
bit
more
homogeneous
homogeneity
in
terms
of
stress,
so
you
know
again
it's
learning
about
what
the
stresses
are
about.
What
different
things
are
mechanical
parameters
and
it's
being
reflected
in
the
division
rates,
which
vary
quite
a
bit
between
the
Learned
random
case,
so
they're
able
to
like
train
this
internal
model
for
each
cell.
It
learns
about
the
environment,
it
learns
about
the
conditions
under
which
it's
living
and
then
it
divides
accordingly,
and
you
get
eventually
responses
to
chemical
gradients
and
maybe
pattern
formation
and
symmetry
break.
A
So
this
is
kind
of
an
interesting
paper
from
that
respect,
because
you
can
actually,
oh
that's
it,
that
you
can
have
this
sort
of
intrinsic
stochasticity,
which
is
sort
of
the
you
know
the
things
that
are
internal
to
the
cell
that
are
stochastic.
So
you
know
the
the
internal
model
has
to
learn
how
to
deal
with
those
things
you
have
extrinsic
stochasticity,
which
are
the
effects
of
the
environment,
and
these
things
are,
you
know,
sort
of
randomly
distributed
in
time.
A
So
if
you
have
a
learning
algorithm,
it
has
to
pick
up
all
these
fluctuations
and
things
like
forces
and
and
chemical
signaling
and
put
them
together
into
some
sort
of
model
of
prediction.
So
they
use
a
lot
of
different
tools
here.
They
use
reinforcement,
learning
scoreboard
based
method
from
Sutton
Umberto,
these
Gumbel
soft
Max,
which
is
a
gradient
estimator,
and
then
they
use
optimization
different,
optimization
protocols.
So
they
use
a
lot
of
different
computational
methods.
A
A
So
any
questions
I
see,
Jesse
has
a
question
unrelated
question:
we've
covered
green,
lighter
already
here
in
Devon,
happened
upon
a
copy
of
mastered
as
Emissary
by
mcgilchrist,
but
I.
Don't
know
it's
too,
outside
the
scope
we
well.
We've
talked
about
symmetry
breaking,
we
haven't
already
talked
about
brain
laterality.
A
To
say:
we've
talked
about
some
of
the
differences
in
like
different
hemispheres
like
if
we're
dealing
with
like
C
elegans.
You
know,
there's
the
left
hand,
side
and
the
right
hand
side
which
are
usually
symmetrical,
but
sometimes
you
get
this
fluctuating
asymmetry
and
we
talked
about
that
in
the
Royal
Society
paper
that
we
just
put
out,
but
that
would
be
interesting
to
follow
up
on
because
I
think
there's
some
interesting
questions
about
function
in
that
so
yeah
yeah,
the
Symmetry
breaking
is
really
interesting.
Yeah,
you
know
so
yeah.
C
No
no
I'm
just
trying
to
follow
along
here,
I,
don't
know
how
much
I
need
to
get
into
the
background
of
other
modeling
for
what
I'm
doing
yeah,
but
this
does
look
like
another
modeling
set,
so
my
advisor
is
very
into
math
modeling.
So
if
I
mentioned
it
to
him,
he
might
want
you
to
do
a
whole
thing
on
it.
I'm
not
sure.
A
Yeah,
the
the
like
stuff,
like
this,
like
reinforcement,
learning
and
some
of
these
other
optimization
techniques
there
a
whole
different
can
of
worms
re.
You
know
optimization,
of
course,
is
General
useful
in
engineering,
so
that
might
be
something
to
focus
on.
You
know
just
because
they're
different
optimization
methods,
you
could
use
for
say
like
tensegrity
and.
D
Have
anything
yet
but
they're,
you
know
cell
types,
classically,
there's
the
vegetative
cells,
the
header
assists
akinates
and
sometimes
terminal
cells
yeah,
because
they
you
know
these.
These
grow
and
chains
very
often
yeah.
And
apparently,
if
you
look
at
Google,
Images
cyanobacteria
differentiation,
apparently
there's
some
other
cell
types
like
homogonia.
A
D
D
A
Yeah
yeah,
it
sounds
like
something
I'd
be
interested
in
is
even
yeah.
So
he's
been
talking
about
that
yeah
we're
supposed
to
do
proposal.
He
wants
to
do
the
open,
worm,
studentship
and
he's
going
to
produce
a
outline
of
what
he
wants
to
do.
I
don't
know
if
that
would
be
suitable
for
that,
but
it
would
be
interesting.
A
A
A
reference
on
restricted,
cellular
differentiation
in
cyanobacterial
filaments.
This
is
something
that
dick
was
talking
about
in
the
meeting.
This
is
a
paper
from
pnas
by
Enrique
Flores
from
2012..
It
talks
about
differentiation,
model
of
differentiation
and
restricted
differentiation
in
cyanobacteria.
A
A
Species
of
cyanobacteria
pump
to
form
means
it's
punctate
or
it
has
like
Teeth.
They
look
like
a
little
teeth
lined
up
so
these
are
photosynthetic
vegetative
cells.
So
these
are
plants
like
cells,
they
have
they
perform
photosynthesis
and
they're
lined
up
from
end
to
end.
So
we
can
build
a.
We
can
look
at
this
under
a
microscope.
We
can
also
build
a
model
here,
and
so
the
model
is
one
where
we're
assembling
things
from
single
to
cells
that
divide.
So
we
have
this
chain
here
we
start
at
the
beginning
of
our
sequence.
A
With
this
chain
we
have
cell
a
and
cell
b,
so
a
and
so
B
come
into
contact
here,
so
a
and
cell
b
divide
so
that
you
have
a
prime,
an
a
prime
prime
B
divide
so
that
you
have
B
Prime
and
B
prime
prime.
This
is
cell
division.
Number
one
cell
division
number.
One
then
continues
down
in
time
and
then
cell
division
number
one
serves
as
the
basis
for
cell
division
number
two,
which
is
where
a
prime
and
a
prime
prime
divide
into
a
prime
1
and
a
prime
2.
A
A
So
as
they
divide,
they
expand
widthwise
and
starts
to
form
this
chain,
so
the
chain
elongates
as
cells
divide.
So
if
we
go
back
up
to
the
top
of
the
paper,
we
see
why
this
is
important.
So
when
a
cyanobacterial
filament,
the
heterosis
differentiate
from
some
vegetative
cells
in
response
to
nitrogen
deficiency
in
cyanobacteria
of
the
Genera
and
a
bienna
or
nostock
and
I
think
we're
dealing
with
nostal
care,
heterosis
form
intercalary
along
the
filipin,
producing
a
non-random
pattern
of
one
heterosis
to
approximately
10
to
15
vegetative
cells.
A
So
we
have
a
heterosis
that
divides
into
these
vegetative
cells.
However,
can
any
someone
a
filament
differentiate
into
a
heterocyst
in
another
pnas
paper
reserto,
which
is
this
reference?
Five
present
evidence
consistent
with
the
idea
that,
at
a
given
time,
only
some
cells
in
the
filament
have
the
capacity
to
differentiate,
so
not
every
cell
in
the
filament
that
we
saw
there
can
differentiate
only
some
of
them
and
so
investigating
this
species.
A
Non-Stock
pointiform
a
filament
to
cyanobacteria,
capable
of
producing
the
four
different
cell
types
described
earlier,
which
they
describe
above
so
there
are
different
cells
here
that
do
different
functions.
So
so,
in
this
model,
with
our
lineage
here
of
these
different
cells,
they
have
different
components,
so
they
have
different
Regulators
that
are
involved
in
different
functions.
A
So
we
get
this
head
R,
which
is
a
regular
heterosis
differentiation
regulator
that
appears
to
be
in
her
inscription
Factor
they
get
induced
after
nitrogen
deprivation.
This
produces
an
amplification
Loop
of
gene
expression.
So
you
get
this
factor
that
gets
expressed
in
some
cells
and
they
end
up
dividing
and
then
some
of
those
cells
are
inheriting
this
Factor.
A
So
there
these
two
genes,
the
pedest
and
head
engine.
These
include
polypeptides
containing
conserved
amino
acid
sequence
that
inhibits
heterosis
differentiation,
a
pedest
derived
peptide
produced
by
some
cells
early
in
differentiation
at
a
head
and
or
head
in
related
compound
produced
by
mature
heterocytes,
inhibit
the
differentiation
of
nearby
cells
in
a
process
that
can
be
described
as
lateral
inhibition.
A
So
this
is
something
that
we
know
from
neural
modeling,
which
is
where
cells
can
inhibit
their
neighbors
to
not
be
active,
and
it
refines,
in
that
case,
the
processing
of
sensory
information.
But
in
this
case
it's
actually
inhibiting
gene
expression
in
its
neighboring
cells.
So
this
this
process
of
lateral
inhibition
likely
involves
an
interaction
with
headar.
So
this
is
why
we
have
the
differentiation
of
some
cell
or
cells
from
some
cells
in
that
chain
over
others.
A
Now
I'd
like
to
talk
about
a
few
issues
that
came
up
in
our
discussion
of
gastroloids,
so
we
talked
about
gastroloids
and
specifically,
we
talked
about
gastroloids
in
2D
there
versus
3D,
which
is
interesting,
but
I
also
talked
about
how
we
might
model
this
computationally,
which
is
different
than
actually
looking
at
the
biology.
It
goes
beyond
that
to
sort
of
predict
or
to
characterize
what
these
different
types
of
models
offer.
So,
basically,
what
we
have
is
we
have
geometry.
A
Pluripotent
stem
cells
are
usually
pluripotent,
but
you
can
also
have
a
germ
cell,
which
is
totipotent,
so
we
could
do
Tony
potency
well,
we'll
just
do
a
pluripotent
cell,
and
so
pluripotent
cells
are
interesting
because
first
of
all
can
produce
any
cell
in
a
body.
So
if
you
have
a
human
Toady
potent
cell,
it
can
produce
any
cell
in
the
human
body,
but
pluripotent
cells
are
generally
restricted
to
a
certain
set
of
Fates.
So,
for
example,
if
you
have
a
neural
precursor
or
a
neural
stem
cell,
you
have
a
neural
precursor.
A
A
A
A
A
So
back
to
our
modeling,
we
start
with
a
set
of
stem
cells
here
that
are
pluripotent
and
then
those
cells
divide,
and
we
start
to
get
cells
coming
out
like
this.
So
we
have
a
sort
of
a
Center
Point
and
then
we
start
to
get
these
cells
that
emerge
and
we
can
model
this
in
two
Dimensions
or
three
dimensions.
A
Right
now,
I
mean
it
looks
like
I'm
in
two
Dimensions,
but
let's
suppose
that
I'm
in
three
dimensions.
So
let's
suppose
that
this
is
a
sphere
I'm
going
to
try
to
draw
a
sphere
here,
I'm
not
going
to
do
a
very
good
job,
but
I'm
going
to
draw
sort
of
like
a
volume
which
is
that's
pretty
good.
So
this
is
a
volume
which
is
three
dimensions.
A
It's
a
sphere,
and
the
idea
is
that
you're
going
to
get
this
outward
production
of
cells-
and
these
are
computational
objects.
So
you're,
going
from
like
a
Center
Point
centroid,
where
there
are
a
number
of
constituents
to
those
constituents
dividing
and
migrating
out
and
they'll,
eventually
migrate
out
to
a
certain
radius.
A
A
Is
migration
rate
and
so
migration
rate
might
be?
You
know,
determine
how
far
after
division
a
cell
will
travel
out.
So
if
the
cell
originates
at
the
centroid,
it
divides
and
migrates
out
this
cell,
then
that
results,
the
daughter
cell,
divides
and
results
in
two
cells
that
migrate
out
further,
and
so
they
keep
migrating
out
at
a
certain
rate.
A
A
A
A
So
that's
what
we're
trying
to
do.
The
other
factor
in
this,
of
course,
is
that
these
cells
change,
State
and
so
they're,
not
always
stem
cells.
They
eventually
become
differentiating
cells,
so
our
differentiated
cells
that
need
to
have
we
need
to
have
assign
a
state
to
each
cell,
so
the
cell,
can
you
know
we
can
think
of
this
as
like
a
zero
or
one.
A
A
So
like
you
know,
we
can
either
have
like
a
stem
cell
or
a
neural
cell
or
a
stem
cell
or
a
muscle
cell
there's
a
critical
point
in
which
it
switches
there's
an
internal
model
for
each
cell
that
we
might
describe
and
get
this
outcome.
We
could
also
have
a
gradient
from
zero
to
one.
So
we
could
actually
use
gradient-based
methods,
which
we
talked
about
in
the
meeting
to
decide
where
you
know
the
cell
is
kind
of
searching
for
an
optimal
State.
A
Given
its
position
in
space,
this
volume
can
be
heterogeneous
with
respect
to
chemical
and
mechanical
factors,
so
this
could
be
like
a
chemical
gradient
here.
This
could
be
a
chemical
gradient
here,
there's
a
boundary
here
and
then
the
cells
that
end
up
here
have
a
stronger
incentive
or
a
stronger
signal
in
the
zero
to
one
interval
to
differentiate
into
a
certain
type
or
any
cell
on
this
Zone
would
have
a
stronger
incentive
to
differentiating
into
a
certain
type.
So
we
might
have
multiple
radians
throughout
the
cell
and
they're
defined
by
their
position.
A
Also
have
a
label
which
just
basically
says
what
the
state
of
the
cell
is
at
any
given
observation.
So
in
time,
you'd
have
you
know
over
time.
You'd
have
a
number
of
observations
and
a
DJ
observation.
It
would
have
a
state
C
generate
a
sort
of
a
there
are
states
here,
knowing
that
this
is
not
reversible.
A
A
So
we
can
deal
with
differentiation.
We
can
deal
with
sort
of
parameters
of
that
space,
just
in
terms
of
its
shape
and
its
extent.
The
other
thing
we
want
to
model
is
electrochemical
interactions,
so
you
know
we
talk
about
like
excitable
cells,
we
talk,
you
know
that
that
involves
a
lot
of
biophysical
modeling
of
ion
channels
and
of
ion
balance,
and
so
it's
important
to
note
that
not
just
neurons
have
this,
but
stem
cells
have
this
as
well,
so
stem
cells
have
or
have
an
electrical
sort
of
profile.
A
A
A
You
need
to
have
a
model
of
ION
channel
balance
and
we
need
to
have
a
model
of
sort
of
an
internal
integration
model
and
that
could
be
a
gene,
regulatory
Network
that
could
be
sort
of
Flex
balance
analysis
or
whatever.
We
have
a
number
of
ways
of
doing
this.
Basically,
we
want
to
create
an
internal
model
that
describes
the
internal
state
of
the
cell.
We
want
to
describe
a
change
from
the
stem
cell
to
the
neuron
or,
if
intermediate
cells,
which
are
important,
because
the.
A
A
So
we
actually
have
to
build
a
tree
that
represents
where
do
they
sell
division
events?
So
it's
not
just
enough
to
have
them
in
this
volume
and
track
their
differentiation
across
space
like
this,
but
you
also
have
to
have
a
differentiation
for
your
lineage
tree.
That
shows
like
this
relationship,
but
also
a
differentiation
tree
which
shows
how
these
states
change
across
divisions.
Sometimes,
factors
are
inherited
by
one
cell,
but
not
another,
sometimes
their
size,
differences
between
two
daughter
cells.
A
Sometimes
these
daughter
cells
will
have
different
Fates
because
their
gene
expression
networks
are
are
Gene.
Regulatory
networks
are
triggered
in
different
ways.
Sometimes
that's
my
location,
as
we
saw
in
the
sphere.
Sometimes
that's
by
some
of
the
other
activity
in
the
cell.
Sometimes
it's
programmed
within
the
genome
itself,
and
so
I
can
see
elegans
we
have
this
deterministic
genotype
or
deterministic
cell
types,
so
those
are
turned
on
and
off
by
The
genome.
So
there
are
a
lot
of
different
ways
you
can
proceed,
and
so
those
are
that's,
that's
all
interesting.
D
A
So
we
can
get
these
different
cell
types
that
you
see
there
they're
not
really
like
a
real
embryo,
but
this
and
this
in
this
case
it's
a
gastroyed,
so
we're
interested
in
sort
of
the
gastrulation
aspect
of
it,
but
we're
interested
in
like
that
part
of
the
gastrulated
embryo
so
but
notice
that
we
have
these
different
cell
types
that
mimic
that
state.
But
we
don't
necessarily
have
identical
organization,
and
so
the
question
is:
do
we
have
things
like
with
calcium
waves
or
other
types
of
waves
that
serve
as
organizing
principles?
A
B
A
A
A
A
You
get
this
wave
of
activity,
it
goes
like
this.
Here's
what
they
call
a
Furrow,
which
is
a
a
line,
a
chemical
line
that
goes
across
the
imaginal
disc,
and
these
are
undifferentiated
cells
here
and
as
the
furrow
passes,
they
become
differentiated
and
ordered.
So
there's
this
sort
of
chemical
self-organization.
That
happens
where
the
furrow
moves
across
the
disc,
and
it
goes
for
the
wave
of
activity
that
goes
across
the
disc
and
it
changes
these
undifferentiated
cells
into
differentiated
cells
that
have
some
order.
A
So
you
end
up
with
these
only
tidia
in
the
eye
that
serve
as
the
structures
of
vision.
We
did
a
paper
on
this
several
years
ago,
where
we
looked
at
this.
We
looked
at
drawings
of
the
imaginal
disk
of
the
furrow
and
then
calculated
statistics
about
the
cells
on
either
side
of
the
frill,
so
these
leaves
have
an
effect
of
self-organization
in
an
embryo.
You
usually
have
things
like
calcium
waves
that
serve
as
organizing
courses.
A
Differentiation
waves
that
serve
to
structure
the
embryo,
so
you
get
these
contraction
waves
and
expansion
waves
that
occur
as
cells
divide
differentiate.
So
you
get
these
areas
of
restriction
in
areas
of
expansion.
So
this
is
that
would
explain
some
of
the
differences
in
the
size
of
these
tissue
types
or
these
regions
of
differentiation.
A
A
We
can
see
that
we
add
in
like
an
internal
model
for
each
cell
in
different
ways,
and
then
we
can
have
these
regions
of
cells,
the
cells,
each
having
an
internal
model
and
they're
responding
to
these
waves.
So
we
can
Implement
waves
across
the
geometry
to
sort
of
trigger
the
cells
in
terms
of
signaling.
A
So
imagine
you're
the
cell
here
and
you
are
responding.
You
have
an
internal
model
which
is
like
maybe
a
gene
regulatory
Network,
and
you
know
I'll
just
draw
something
really
simple
here
and
there's
an
output
there's
an
input
actually
in
an
output
It's.
Usually
the
in
well
there's
actually
like
an
endogenous
signal
here
which
might
be
the
activity
of
genes
or
the
actual
structure
of
genes.
The
activity
is
triggered
by
environmental
signals
and
then
the
output
is
the
self-enotype,
and
maybe
you
know
a
cell
migration
or
something
like
that.
A
The
input
could
be
chemical,
it
could
be
mechanical
if
you
have
chemical
waves
or
mechanical
waves,
they
serve
as
an
input
to
the
grn,
and
it
tells
it
what
to
do
what
to
turn
on
what
genes
to
turn
on
how
much
to
turn
them
on
and
so
forth
that
results
in
an
output
that
is
maybe
different
from
its
neighbors.
Maybe
the
same.
A
It
depends
on
the
rules
you
program
into
the
grn
or
the
you
know
the
way
that
the
inputs
work.
So
sometimes
you
can
have
things
like
whatever
one
ambition
which
silences
the
cells
next
to
it,
but
you
know
keeps
that
so
active,
so
the
grn
and
neighboring
cells
would
behave
differently
than
a
cell.
That's
active
proactively,
so
there
are
a
lot
of
ways
that
these
waves
can
be
implemented
at
a
crumb,
Rio,
the
output,
of
course,
being
the
phenotype.
A
And
so
we
expect
to
see
the
Sorting
over
time
where
we
get
these
regions,
that
behave
alike,
and
so
they
end
the
structure
in
the
embryo.
So
there
are
a
lot
of
parts
to
modeling.
These
These
are
sort
of
a
factors
of
I've
identified
a
number
of
conversations,
so
one
of
the
reasons
why
we
might
want
to
do
this
is
to
have
sort
of
a
more
non-descript
embryo.
So
when
we're
modeling
embryos,
we're
often
saying
which
organism
is,
it
is
C.
Elegans
is
a
drosophila.
A
A
Put
those
in
as
parameters
as
well,
but
for
now
I
think
this
is
just
a
good
model
of
sort
of
a
nondescript
embryo
nondescript
developmental
process,
and
hopefully
we
can,
you
know,
maybe
expand
this
out
to
maybe
like
a
model
of
an
organism
that
just
is
good
at,
like
you
know
single
cell
morphogenesis
and
plug
that
into
this.
In
a
more
you
know,
it's
sort
of
borrowing
from
some
of
the
things
that
we
see
at
the
embryoid
models.