►
From YouTube: DevoWorm (2023, Meeting #25): DevoNet, TDA/GNNs, Open Biophysics Problems, and Embryoid Models
Description
GSoC Updates: Devonet and model selection, topological cell segmentation. Discussion on Topological Data Analysis (TDA) and GNNs. Large Language Models for genomics and HyenaDNA. Discussion on biomedical image sampling problem and creating biophysics graphs. Embryoids III (more on human embryoids). Attendees: Sushmanth Reddy Mereddy, Jiahang Li, Bradly Alicea, Himanshu Chougule, Susan Crawford-Young, Richard Gordon, and Jesse Parent.
B
A
A
C
C
Yeah
there
are
two
photo
library:
four
folders
here:
one
is
Library
folder,
where
whole
code
is
their
model
architecture,
how
we
will
take
the
exit
and
all
last
week,
I
worked
on
the
model
and
the
loss
function
here.
I
have
took
two
modems
for
nuclear
segmentation
Network
one
modem
is
there
new
phase
detection
Network
one
model
is
and
yeah
it's
completely
written
in
by
torch.
C
I
made
some
changes
according
to
what
I
needed
and
the
last
function
I
thought
of
using
the
direct
line,
but
I
thought
a
single
line
from
pipe,
but
I
thought
it
wouldn't
be
that
appropriate.
So
I
wrote
my
class
for
own
dice
class
function
from
net
resources
and
I'm,
giving
I'm
working
on
that.
Only
and
last
week
I
wrote
last
four
week
blocks
all
group
yeah.
A
It
doesn't
look
like
it.
You
gotta,
reshare,
Maybe,.
C
Yeah
yeah
I
first
I
wrote
the
whole
code
here
and
below
that
explanation
of
it
is
that
if
someone
comes
up
and
sees
our
development
code,
some
may
may
not
understand
new
beginners,
so
I
thought
of
writing
a
Blog
about
it.
Every
quote:
I
wrote
so
whatever
I
am
writing
those
I
am
just
explaining
you
to
make
clear
way
in
the
points
whatever
each
word
means
in
that.
So
that's
what
I
am
doing
if
some
I
was
trying
to
comment
the
whole
code
also,
but
right
now.
C
This
is
my
process
like
in
this
week
for
I
worked
on
the
data.
So
it's
fine
3D
data
sets
right,
I
wrote
in
every
explanation
of
it
and
then
I
mean
this
last
function.
Also
I
don't
report
it,
and
this
week
I
completely
worked
on
the
model
architecture,
so
my
main
motive
would
be
writing
code.
Explainable
code
of
model
architecture,
yeah
I,
have
done
these
things
only
this
week
and
I
tried
to
work
on
Sam
also
but
I
couldn't
handle
both
at
the
same
time.
C
This
week,
I
will
try
to
manage
model
architecture
with
sound
model.
Also
by
next
week,
yeah
I'm
just
working
on
the
model
right
now
there
are
no
fine
errors.
I
am
using
ciprit
library
to
image,
processing,
so
image
processing,
because
Mayo
once
told
me
that
he
wants
to
use
circuit
image
library
for
develop,
but
he
couldn't
use
it
of
last
time,
so
I
started
to
use
a
scientific
image
library
for
the
level.
C
Naturally,
it
is
mentioned
in
development
issues
phase
also
that
psychic
image
like
update
to
psychic
image
Library,
so
whatever
I
was
working
on
I,
am
writing.
In
psychic
image,
only
I
am
using
a
special
method
called
Watershed
segmentation.
It
generally
helps
like
when
two
cells
are
overlapped
to
each
other.
It
doesn't
serve
it
as
a
single
cell.
It
takes
us
two
different
cells,
so
it
would
be
more
clear
to
us
for
finding
the
centrants
of
each
cell
rather
than
overlap.
C
C
C
C
C
Yeah,
this
part
is
for
NSM
I,
usually
mainly
convolutional
3D,
because
whatever
data
set,
we
have
this
3D
data
set.
So
I
used
a
3D
convolutional
layers
and
tried
to
implement
it,
and
this
part
is
of
unit
I
just
want
to
compare
it
with
our
Network
and
unit
Network,
so
I'm
just
trying
to
implement
it.
Mostly
I
will
remove
this
part
of
Port,
because
my
proposal
was
made
upon
NSN
and
Indian.
Only
if
I
can
compare
these
results
of
unit
NSN
NDM
I
will
try
to
implement
it
and
see.
C
First
I
was
working
on
this.
No,
no!
It's
not
similar
to
unit
in
3D
I
was
just
comparing
the
segmentation
of
all
these
three
networks,
Sam
model
and
whatever
I
proposed
in
the
proposal
and
unit
model,
which
are
gives
the
best
results.
That
would
be
useful
to
us.
I'm
just
checking
on
which
segmentation
network
works.
Fine,
okay,.
C
A
Do
you
could
you
describe
the
what
what
a
unit
model
is.
C
Unit
model
is
mainly
used
for
medical
image
processing.
It
has
nine
convolutional
layers
in
a
u-shape.
Let
me
show
you
with
an
example.
C
Yeah,
this
is
how
it
looks
in
the
shape
of
I
mean
there
are
nine
convolutional
layers.
First,
continuation
will
happen
after
that.
Deconvolution
will
happen.
You're.
C
I'm
not
sharing
your
screen,
that's
okay,
sorry
yeah!
These
are
nine
convolutional
layers
totally
first
convolutional
convolutional
will
happen
until
here,
then
decodational
will
happen,
and
this
is
mainly
used
for
medical
image
processing.
See
here.
You
can
see
biographically
image
segmentation
yeah,
and
this
is
the
old
model
I'm
just
con,
seeing
which
gives
a
better
result
either
Sam
model
or
demonet
model
or
unit
I
will
try
to
extract
I
will
take
two
I
will
try
to
compare
it
with
the
rcmn
model,
which
was
implemented
by
himanshu.
C
Also
I
will
just
compare
the
results
and
provide
the
graph
which
model
outperforms,
which
on
which
one
performs
well
I,
will
take
it
as
for
better
one.
That's
what
my
idea
is
just
not
giving
up
the
model.
I
am
telling
it
looks
good,
but
with
the
perfect
explanation
which
model
out
corresponds
with
and
according
to
situation,
we
can
use
the
model.
C
Using
cell
tracking
challenge
and
with
that
I
will
I
will
just
take
a
common
I
will
eliminate
some
part.
Some
images
from
the
cell
tracking
and
I'll
just
check
the
test
results
of
all
models
on
that
and
after
that,
I
will
produce
the
loss,
function
and
I
I
O,
U
score
and
I'm
just
trying
to
compare
them.
If
the
io
user
and
the
loss
we
can
just
find
it
out
which
model
performs
well.
A
C
A
C
Yeah
I
was
gonna,
I
was
thinking
to
work
on
the
training,
the
model
because
the
model
is
not
completed
yet
I
tried
to
write
the
model
architecture,
so
many
errors
I
have
encountered,
but
at
last
week
is
working
fine.
This
week,
I
will
write
the
trainer
code
for
it
and
yeah
after
that,
I
will
try
to
train
the
models.
Actually
I
was
having
GPU
issues.
All
kaggle
gpus
are
completed,
which
I
have
so
I
was
I
took
a
I
buyed
some
of
gpus
from
online
from
Google
collab
of.
A
Good
yeah,
that's
always
a
good
resource.
School
lab
for
doing
that
sort
of
stuff.
C
Yeah,
actually,
the
resources
are
emptied
because
collab
provides
up
till
like
around
I,
don't
know
50
GPU,
compute
units
I,
think
so
I,
don't
remember,
I
completely
used
them
off
for
this
month
and
I
use
a
kaggle
also,
and
the
work
has
been
stopped
since
two
days,
so
I
took
a
another
resources:
I
bought
it
from
I,
bought
it
100
compute
units
I'll,
try
I
will
try
to
try
in
the
model.
That's
it.
E
Is
it
visible?
Yes,.
E
So
I
was
working
on
the
model
data
I
initially
proposed
about
the
mask
CNN
model
and
I
was
able
to
like
figure
out
how
how
to
solve
all
the
errors
and
basically
create
my
own
custom
class
for
the
segmentation
data
set
in
which
I
could
incorporate
like
the
image
mask,
as
well
as
the
bounding
boxes
for
it.
E
And
after
like
a
few
error,
solving
methods
I
was
able
to
like
finish
this
class
of
the
data
set
which
just
Returns
the
transform
damage,
as
well
as
the
Target
and
the
target
has
all
the
information
which
is
needed
for
like
segmentation.
E
Let
us
the
boxes,
the
masks
and
the
area
of
the
Boxes
Etc,
so
I
was
able
to
use
this
and
like
successfully
get
the
data
set
Bible
for
training
and
then
I
basically
created
the
model
for
like
instant
segmentation
as
well
as
I
was
able
to
like
use
the
bounding,
Boxes
Etc
and
also
create
visualizations
for
it.
E
So
this
is
how
it's
it's
it
looks
and
for
the
data
set
that
I've
created
so
and
now,
I
also
trained
it
for
like
just
I
used
like
the
basic
optimizers
and
schedulers,
and
I
tried
to
train
it
for
like
two
ebooks
and
like
each
Epoch
took
like
around
35
to
40
minutes
because,
like
the
size
of
the
data
set,
is
a
lot
large
and
and
then
I
just
saved
that
model,
and
basically
since
I,
just
since
I
was
just
checking
if
I
could
train
it
and
if
the
model
fits
perfectly
so
I
just
standard
for
two
ebooks
just
to
see
that
and
I
was
able
to
like
this
is
like
the
bounding
boxes
and
for
the
prediction
so,
but,
like
the
actual
masks,
have
only
got
like
only
one
mask
as
you
can
see
over
here,
because
it
was
just
straightforter
epochs.
E
So
what
my
plan
right
now
is
to
do
like
some
hyper
parameter
tuning,
as
well
as,
like,
probably
like
make
some
changes
in
the
data
set
but
like
if
I'm
creates,
making
some
changes.
I'll
like
talk
with
sushma
as
well,
because
he's
also
talking
figuring
out
how
how
to
deal
with
the
big
data
set
and
all
and
similar
to
him.
E
I,
also
why
I
have
zero
complete
units
available
and
which
is
why
I
was
thinking
of
doing
like
some
kind
of
parameter
tuning
on
the
CPU
as
well,
but
like
after
done
making
some
small
small
changes
in
the
model.
I'll
just
get
some
Computer
Resources
and
just
finish.
The
training
like
on
one
night
session
like
just
keep
my
laptop
running
for
like
10
hours,
I
guess,
yeah,
yeah
and
yeah
I've
not
used
tackled
until
now.
E
So
I
can
just
do
the
hyper
parameter
tuning
on
toggle
and
then
I'll
drive
it
back
to
collab
and,
along
with
this,
I
started
like
working
on
the
demo.
This
demograph
part
of
the
project
like
to
derive
or
find
a
better
ways
to
derive
graph
embeddings
and
right
now,
I'm,
just
in
the
learning
stage
of
it.
So
I've
been
I.
E
Also
watch
the
like
the
recording,
which
was
given
of
the
video
like
in
which
Zhang
like
explained
like
what
is
the
way
of
doing
the
preferential
attachment
and
also
I'm
doing
that
simultaneously
as
well.
E
So
the
data
set
I've
been
working
on
is
also
the
cell
tracking
Challenge,
and
after
this
month
after,
like
customize
pipeline
is
created,
I
can
work
on
the
two
over
the
different
data
sets
for
the
membrane
segmentation
and
after
that,
I
can
just
work
on
the
data
set,
which
I
had
proposed
like
two
weeks
back,
which
was
like
for
the
longest
agent,
which
was
like
kind
of
elongated
like
this,
so
just
to
see
how
the
model
kind
of,
if
it
works
on
that
or
some
changes
have
to
be
made
etcetera.
A
C
E
E
E
Yeah
the
meeting
was
great
and
he
basically
like
kind
of
taught
told
me
about
the
day.
He
wanted
the
things
to
go
like
to
use
Mesa,
instead
of
like
the
which
we
initially
did
for
asian-based
modeling,
that
is,
to
use
net
logo
and
now
and
I.
Think
that's
the
right
direction
to
go
because,
like
using
Mesa,
is
different
from
net
logo,
because
that
logo
has
its
own
architecture,
and
you
have
to
follow
some
rules
that
that
logo
has
for
it.
E
So
creating
the
agent
based
models
for
that
for
that
environment
itself
is
different
than
what
we
can
do
in
Mesa.
E
So,
like
I,
came
from
a
basic
walkthrough
of
what
he
should
be
doing
and
what
should
be
the
next
steps
like
to
get
some
kind
of
a
small
algorithm
of
agent
learning
algorithm
ready
for
like
some
for
the
reinforcement
learning
part
of
it
so
like
you
can
just
do
a
basic
file
iteration
and
then
move
on
to
Q
learning,
just
for
a
basic
implement
creation
of
it
and
figure
out
a
way
to
just
connect
those
both
things
like
the
Asian
Taste
algorithm,
as
well
as
the
environmental
Creator.
A
Yeah,
thank
you.
We
were
discussing
this
Friday
about
kind
of
the
plan
forward
and
he's
also
talking
to
ancient
Grover
who
works
in
the
lab,
so
he's
acting.
Kid
is
actually
really
interested
in
RL.
So
you
know
that's
that's
going
to
be
fun
to
kind
of
develop
the
RL
capabilities
there
yeah.
So
that's
great!
E
So
right
now
I
saw
a
nuclear
segmentation
graph
as
well
and
in
the
meantime,
I'll
also
look
into
what
the
open
sustainability
project
is
going.
So
I'll
try
to
attend
the
meat.
On
fact.
Oh.
A
Yeah
that'd
be
great
yeah,
so
you
know
sushma
has
been
at
the
Friday
meeting,
so
we
run
a
Friday
meeting
it's
at
noon,
eastern
time
in
North
America.
So
we
do.
We
talk
about
the
two
projects
in
the
orthogonal
lab,
but
we
also,
you
know,
can
talk
about
the
diva
one
projects
if
necessary
and
especially
when,
if
you
can't
make
this
meeting
or
you
know
for
whatever
reason,
but
we
also
have
some
Crossover
with
the
projects,
so
yeah
that'd
be
good.
A
So
I,
don't
know
if
we
had
Dia
hung
a
couple
weeks
ago
gave
an
update
on
his
some
things.
He
was
thinking
about
and
it
was
mentioned
by
himanshu.
So
I
don't
know
if
Tia
hung
had
a
follow-up
to
that
or
not.
If
he
wanted
to
talk
anymore
about
that
or.
F
Great
a
long
time
to
see
how
yeah,
yeah,
hi
yeah,
actually
I'm
kind
of
big,
because
I'm
busy
as
preparing
for
my
integrated
confirmation
reports,
and
it
would
be
like
after
one
month
and
I,
have
to
present
to
some
professors
for
my
degree
confirmation.
So
it
could
be
a
little
bit
challenged.
So
yeah
I
spent
some
time
on
this,
so
maybe
not
much
updates,
because
I
didn't
of
commencement
votes
on
those
on
the
research
project
that
I
mentioned
two
weeks
ago
and
I
see
someone
mentioned
that
in
this
meeting
today.
F
So
so,
firstly,
like
like
what
I
mentioned
in
two
weeks
ago,
I
said:
I
wanna
apply
the
topological
data
analysis
to
the
analysis
of
of
ASO
temporal
topology
within
the
cell
development
process,
and
it
is
a
general
like
General
topic
and
I
I
haven't
like
thought
about
any
like
model
design
or
some
more
specific
details
about
them
about
this.
Talk
this
topic
two
weeks
ago
and
for
now
I
think
there
could
be
some
some
approaches
to
solve
this
task.
F
For
example,
the
first
of
all
is
compared
with
other
study
static
system
for
the
cell
development
process.
We
have
a
topology
which
is
dynamic.
It
is
time
depending
that
means
we
have.
We
can
explore
the
topology
of
like
the
cell
dependence,
Network
or
something
else.
F
For
example,
if
we
have
a
microscope
video-
and
we
have
this
video
recording
the
cell
development
process,
we
can
sample
a
lot
of
friends
and
for
each
frame
we
can
sample
a
lot
of
cells
like
using
some
sales
segmentation
method
and
and
then
we
have
a
point
count,
and
we
have
some
ways
to.
Maybe
we
don't
have
to
build
a
network,
we
can
apply
the
position
of
homology,
which
is
to
explore
the
topological
information
underline
this.
F
This
points
Cloud
it
could
be
a
it
could
be
approach,
but
if
we
deliveries
this
approach,
that
means
we
have
a
static
topology
for
each
friend.
But,
however,
this
topology
is
dynamic.
F
So
that
means
we
need
to
find
like
some
ways
to
to
build
this
Dynamic
topology
like
because,
in
my
in
my
opinion,
I
think
if
we
just
start,
if
we
just
consider
the
topology
of
each
friend
and
we
do
not
consider
the
temporal
connection
between
the
tolerance,
the
topology
across
different
Trends,
it
seems
like
we
are
not
considering
the
Dynamics
of
this
topology.
So
this
will,
if
we
do
not
consider
these
dynamics,
that
means
well.
F
We
are
regarding
the
whole
stuff,
as
just
investigating
the
static
topology
of
of
a
network
that
could
be
Trevor,
like
maybe
I,
didn't
make
myself
career.
Like
then,
give
your
example,
for
example,
we
have
another.
We
have
seen
another
paper
in
maybe
one
one
year
ago,
which
is
a
topological
graph
in
your
network
and
I'm,
not
sure.
F
If
you
also
remember
that
paper-
and
in
that
paper,
the
authors
proposed
graph,
your
network
to
explore
the
topology
within
static
Network,
for
example-
you
you
have
a
graph
which
is
not
Dynamic
and
you
apply
a
topological
graph,
your
network,
to
explore
the
topology
within
this
network,
so
I
think
in.
In
our
case
it.
The
the
approach
that
I
mentioned
before
is
very
similar
to
this
topological
problem
Network.
F
So,
in
my
opinion,
I
think
so.
First
of
all,
I
can
take
some
experiment
to
test
how
this
topological
graphing
network
performed
in
our
case,
and
then
I
would
like
to
try
some
the
answers
to
export
the
Dynamics
of
this
topology.
F
What
I
think
this
Dynamic
topology
is
very
very
important,
because
you
need
to
remember
an
example,
which
is
the
global
rotation
formed
by
the
collective
cell,
migrations
that
I
mentioned
in
two
weeks
ago,
and
actually,
if
you,
if
we
want
to
like
if
we
want
to
see
this
type
of
topology,
which
is
a
global
rotation,
most
specifically
a
circle.
If
you
want
to
see
such
real
topology,
we
actually
need
to
look
at
the
spatial
topology
as
well
as
a
temporal
topology.
F
Yes,
so
and
I
think
for
most
of
multicellular
systems,
then
Dynamics
is
very
important.
So
that
means
we
need
to
find
a
way
to
investigate
the
dynamic
topology,
but
first
of
all,
I
I.
Think
I
would
like
try
the
technological
graph
in
the
network
that
the
simple
approach
is
for
the
static
topology
yeah
they
were
about.
They
would
be
my
first
step
and
for
the
next
step,
I
would
like
to
investigate
the
dynamic
topology
like
there
could
be
a
lot
of
matters
to
consider
the
dynamic
topology.
F
The
first
approach
is
like
you
can
think
of
that
you
for
each
frame.
You
have
a
state
in
topology,
and
maybe
you
would
use
some
kind
of
embedding
or
some
Vector
to
represent
this
static
topology
and
then
you
can.
We
can
like,
like
just
like
other
recurring,
that's
work
or
other
words
which
focus
on
the
some
serious
data.
Maybe
we
can
use
some
recurrent
neural
network
or
we
can
use
some
Transformer
to
aggregate
this
to
logical
information
across
different
fans.
I
mean
this
could
be
an
approach
and
another
approach
is
I.
F
More
fancy
is
because
actually
I,
think
persistent.
Homology
is
a
very
cool,
very
Grace
tour
applied
to
the
points
cloud,
data
and
I
think
if
we
have
some
ways
to
build
such
a
points
cloud,
a
close
friends
like
a
single
point,
Cloud
improve
like
covering
several
friends.
Maybe
we
can
find
some
ways
to
use
the
positional
homology
to
like
investigate
the
spatial
temporal
topology
I'm,
not
sure
if
I
make
myself
career
the
the
challenge
here.
F
Is
you
have
you
reading
distance,
which
is
the
spatial
distance
and
you
maybe
you
can
Define
some
other
types
of
distance.
There
are
other
space
resistance,
but
if
we
want
to
connect
cells
across
different
different
frames,
According
to
some
some
kind
of
reasons,
that
means
we
need
to
consider
the
the
the
10
protists
like
the
distance
of
course,
time
Dimension.
It
is
not
very
intuitive,
like
the
spatial
distance
like
that,
you
can
eat
the
distance.
So
this
could
be
one
trendy
assume.
F
We
need
to
consider
how
to
define
such
a
temporal
distance,
and
the
next
challenge
is
how
to
unify
the
temporal
distance
and
the
spatial
distance.
So
we
can
build
a
point
Cloud.
So
we
can
apply
the
position
of
allows
you
to
a
certain
point.
Colleges.
Yes,
I
think
this
this
if
they
are
reasonable
or
if
they
are
meaningful,
but
maybe
I
would
take
a
try
after
I
finish,
my
information,
yeah
I,
think.
A
A
Yeah,
it's
great
yeah!
Thank
you
for
that.
Yeah
there
we
go
there,
we
go
thumbs
up.
Yeah
I
mean
that's.
That
I
talked
about
some
of
that
two
weeks
ago.
The
persistent
homology
thing,
I,
think,
is
really
interesting
from
the
standpoint
of
kind
of
taking
these
Point
clouds
and
breaking
them
down
into
you
know
a
network,
but
also
into
some
of
these
higher
order,
shapes
and
things
that
might
describe
the
data.
So
you
know
with
with
segmentation.
A
We
have
these
Point
clouds
and
midpoint
clouds
in
time,
but
the
point
clouds
in
time
are
interesting
because
they're
connected
so
like
you
have
a
point
Cloud
at
time:
zero
and
another
Point
Cloud
at
a
time.
One
point
cloud
of
time:
two
and
you
have
cells
dividing,
so
you
have
sort
of
this
regeneration
of
the
point
Cloud
as
this
maybe
there's
more
increasingly
dense
structure
with
things
in
different
positions,
and
then
the
idea
would
be
that
you
want
to
link
those
things
and
we
oftentimes.
A
We
use
lineage
trees
to
do
that,
but
in
a
point
Cloud,
it's
kind
of
hard
to
draw
a
lineage
tree
and
then.
E
F
E
A
We've
done
some
of
that
with
well
I
mean
we've
done
some
things
with
different
techniques.
We
don't
have
a
really
good
technique
for
it,
though
that's
the
thing,
and
so
that
might
be
a
good
way
to
kind
of
build
a
nice
model
of
that
that
change,
but
also
the
changes
in
position
and
all
that
and
then
characterize
it
as
like
this
change
in
spatial
orientation,
or
something
like
that.
A
A
You
know
this
is
like
kind
of
a
lesson
of
people
working
on
Cell
segmentation.
That's
where
that
go.
That's
where
the
data
goes.
Upstream
in
the
pipeline,
you
go
from
like
segmenting
the
cells
to
like
analyzing.
You
know
sort
of
a
first
order,
analysis,
which
is
the
exploration
of
the
point
cloud.
You
can
you
know,
label
the
point,
Cloud
points
you
know
as
cells
and
then
you
want
to
find
the
relations
between
the
cells,
and
this
is
one
of
the
things
that
we're
trying.
A
This
is
why
we
kind
of
come
up
with
these
techniques
to
kind
of
put
together
a
model
of
the
embryo,
or
at
least
the
cells,
and
how
they're
sort
of
related
to
one
another.
So
that's
that's
where
this
is
going,
and
you
know
I
like
these
kind
of
topological
approaches,
because
you
know
I,
think
people
have.
A
You
know
people
kind
of
know,
at
least
in
C,
elegans
sort
of
what
it
looks
like
you
know,
in
terms
of
you
know
where
the
cells
are
in
space
and
where
they're
differentiating,
they
would
kind
of
know
how
those
things
kind
of
unfold
from
the
data.
But
we
don't
really
have
a
higher
level
description
of
like
what
kinds
of
structures
are,
maybe
precursors
to
tissues
or
what
kinds
of
you
know.
Migration
oriented
things
are
emerging
in
the
embryo.
A
A
C
F
Yeah
yeah
I
couldn't
even
this
my
my
motivation,
but
first
of
all
my
motivation
is
actually
I,
think
toological.
This
analysis
have
been
widely
applied
to
the
biological
stuff,
but
I
think
one
thing:
one
interesting
thing
is.
F
They
have
also
been
applied
to
the
biological
analysis
as
well,
but
like
there
is
a
it's
a
gap
between
the
TDA
and
depending
approaches,
especially
the
proper
natural
I.
Think
government.
That's
one
interesting
points
about
graphic
natural.
Is
this
approach?
We
consider
the
data
as
a
set,
and
that
means
from
this
kind
of
perspective
you
will.
You
can
consider
the
geometry
properties
and
the
topological
properties
of
the
data.
F
Yes,
so
it
could
be
different
from
the
like
conversion
or
natural
which
consider
data
from
the
perspective
of
visual
information
from
the
from
the
images.
So
this
could
be
different,
so
so
I'm.
The
motivation
for
me
for
this
project
is
I
want
to
see
if
there
can
be
some
like
interesting
stuff
happen.
If,
when
we
like
integrates,
go
to
a
natural
way,
the
topological
analysis,
yes
on
the
biological
stuff-
oh
maybe
I
can
make
myself
more
clear,
like
we
all
know,
that's
the
multicellular
system,
for
example
the
cell
development
process.
F
They
have
Rich
topological
information
and
we
know
that
we
have
a
lot
of
approaches
such
as
the
TVA,
which
are
very
good
ads.
Exploring
the
topological
information
from
data
like
from
the
microscopic
video
and
we
have
graphene
Network,
which
are
very
useful
for
like
capture
information
from
the
big
data
from
a
lot
of
videos
when
the
data
is
very
large
data
science
is
very
large,
so
I
mean
they
have
their
own
like
advantages
for
the
downstream
task,
but
actually
I'm
a
graphic.
F
Network
guy
and
I'm
more
familiar
with
graphing,
Network
and
I,
see
that
in
the
field
of
crafting
network
actually
I
would
not
say
the
crafting.
Network
have
been
widely
applied
to
the
cell
development
analysis.
I
will
not
say
that
because
I
didn't
see
too
many
examples.
F
Looking
at
some
intersection
between
these
two
films
like
and
how
you
would
perform
on
the
this
multicellular
system
and
from
another
perspective,
I
think
is
I
think
when
I
said
what
I
want
to
say
is
I
wanna
I
wanna
like
explore
any
kind
of
explanation
of
the
different
technique
from
the
geometry
and
the
topological
perspective.
Yes,
like
so
now,
there
are
some
a
lot
of
photos
working
on
the
geometry
or
the
topological
techniques
within
the
context
of
keep
learning
and
they
thought
that's.
F
A
I
was
thinking
actually,
as
you
were
talking
about,
like
a
lot
of
the
things
that
people
do
in
3D
modeling
and
that's
highly.
You
know
topological
and
geometrical.
One
of
the
things
that
you
do
is
you
create
an
object,
and
then
you
extrude
it
in
space.
So
you
pull
it
up.
You
know
you
pull
it
out
and
you
give
it
some
depth.
A
A
We
change
the
network,
Point
Cloud
as
we
get
cells,
dividing
and
they're
changing
their
orientation,
and
so
you,
basically,
if
you
have
like
cyclical
complexes
or
just
like
shapes
that
form
in
the
network,
so
it
could
be
a
triangle
or
a
diamond,
and
then
that
diamond
would
sort
of
track
with
the
cells
dividing
and
that
Network
expanding
so
that
you
might
end
up
with,
like
maybe
a
hexagon
that
shifts
its
orientation,
maybe
even
45
degrees,
and
then
it
maybe
curves
a
little
bit
on
the
bottom
and
you
keep
following
that
shape
out
as
it's
moving
around
extruding
and
you
know
being
warped
in
different
ways.
A
So
that's
you
know.
That's
the
kind
of
thing
too
that
we
could
think
about.
Like
you
know,
what
is
it
we
start
with,
maybe
like
a
very
simple
shape,
but
a
very
Sim,
maybe
an
eight
cell
embryo,
maybe
like
a
a
diamond
of
four
cells
that
kind
of
are
in
that
part
of
the
embryo
or
maybe
like
the
anterior
end
of
the
embryo,
and
then,
as
the
number
of
cells
increases
as
you
get
daughter
cells
and
they
change
their
position,
they
migrate,
they
sort
of
form
these
structures
that
start
to
you
know
tissues.
A
Those
will
change
the
shape
and
you
can
actually
track
that
shape.
As
like
an
object,
that's
you
know,
it
has
properties
just
like
a
three-dimensional
object
in
a
in
something
like
blender
or
Unity,
or
something
like
that.
A
D
F
A
So
that's
that's
great
yeah.
Thank
you.
Jihang
I
think
there
are
a
lot
of
things
that
like
following
up
from
our
discussion
two
weeks
ago
and
I,
just
yeah
I
wanted
to
get
the
people
doing.
You
know
I
want
to
get
people
like
camancho
and
sushma
thinking
about
especially
himanshu
about,
like
you
know
what
what?
Where
are
we
going
with
this
stuff?
A
So
we're
going
somewhere,
that's
very
kind
of
conceptually
hard,
and
so
we
want
to
have
one
I
think
a
good
sort
of
a
good
sense
of
where
that
is
and
what's
possible.
So.
F
Yeah
yeah,
yeah,
yeah,
actually
I
think
so
again,
let's
just
because
actually
for
this
project,
I'm
working
on
graphing
network,
topological
data
analysis
and
less
good
ads,
the
like
the
segmentation
or
relating
stuff,
so
so,
first
of
all
for
This
research
projects.
If,
like
anyone
interested
in
this
and
I'm,
very
welcome
for
everybody
which
reach
out
to
me
and
another
point
is
during
this
whole
jigsaw
period.
If
you
have
any
related
questions
to
the
topological
analysis
and
The
Graduate
Network
I'm
also
welcome
for
everyone
like
reach
out
to.
A
A
A
Complex
that
you
find
often
in
Networks,
so
let's
consider
a
simple
Network
here,
where
we
have
connectivity
across
five
nodes.
Okay.
So
this
is
a
network
topology
where
you
have
this
triangle
and
these
two
lines
that
come
out
now
in
a
normal
network
analysis
we
might
be
interested
in
calculating
like
the
centrality
or
the
clustering
coefficient
or
things
like
that.
But
in
a
topological
network
analysis
you
might
be
interested
in
finding
shapes
in
the
network,
so
what
kinds
of
things
are
enclosed
in
the
network?
A
D
E
A
And
these
are
fine.
You
know
people
are
able
to
discover
these,
and
these
are
kind
of
ways
of
flipping.
The
typical
network
analysis
on
its
head
and
saying,
instead
of
maybe
finding
communities
or
finding
centrality
or
finding
a
connectivity
regime,
I
may
be
interested
in
finding
shapes
that
are
formed
by
the
network.
So
this
is
important
in
sort
of
the
spatial
networks
that
we're
dealing
with
in
the
embryo.
At
this
point,
Cloud
that
we
talked
about
earlier,
which
is
in
three-dimensional
space,
they're,
actually
a
four-dimensional
space.
A
We
have
this
point
cloud
of
points
centroids.
These
centroids
are
kind
of
worked
out
in
this
way,
so
that
you
have,
you
know
they're
always
changing.
We
talked
about
this
where
this
point
cloud
is
always
changing:
we're
getting
new
cell
centroids
we're
removing
old
cell
centroids,
and
so
this
topology
is
always
actively
evolving.
A
A
So
this
point
cloud
is
constantly
evolving
and,
as
we
said
in
the
meeting
in
the
main
meeting,
you
know
we
have
these
two
views.
We
have
the
static,
View
and
the
dynamic
View,
and
so,
if
we
look
at
the
actually
the
point
cloud
or
anything
from
those
two
points
of
view.
Well,
let's
think
about
the
network.
A
A
So
we
have
something
like
this
so
This
node
here
is
here:
it's
now
removed
and
you
have
two
daughter
nodes
in
its
place.
That
means
the
connectivity
changes
so
instead
of
a
triangle,
you
end
up
with
this
sort
of
rectangle
with
two
means
here,
and
maybe
this
node
actually
dies
and
you
end
up
with
daughter
cells
here,
where
you
end
up
with
maybe
a
not
a
triangle,
but
like
a
fan
motif,
let's
get
this
fan
and
then
this
might
be
connected
later
into
two
more
triangles.
A
A
So
we
can
take
two
static
views
of
the
network
on
a
different
two
of
different
points
in
time
and
end
up
with
this
Dynamic
shift
from
a
triangle
to
a
rectangle
or
rhomboid
I
guess
would
be
a
better
way
to
describe
it,
and
so
these
two
shapes
this
transformation
of
shape
is
actually
what
we're
dealing
with
in
development,
because
we
have
our
embryo
and
it's
changing
its
shape
constantly,
as
those
cells
are
dividing,
so
we've
removed
from
the
point
Cloud
to
these
kind
of
structures.
A
A
A
A
Shape
that
now
has
sort
of
it's
extruded
out
in
this
in
this
plane
and
then
also
in
this
plane.
So
it's
kind
of
hard
to
see.
I've
also
made
a
mistake
in
my
drawing,
but
basically
you
have
this
shape.
That's
changing
its
form,
and
so
this
is
where,
where
I
was
kind
of
going
with
this,
this
is
in
2D,
and
this
is
a
static
View
all
right,
and
then
this
is
where
you're
going
into
three
dimensions.
Where
there's
Extrusion.
A
Or
some
process
like
gastrulation.
A
And
so
we
can
actually
take
these
geometric
Transformations
map,
the
developmental
Transformations,
and
this
actually
goes
back
to
Darcy
Thompson's
work
way
back
in
1917,
and
you
know,
the
sort
of
the
idea
was
that
you
had
like
you
could
take
a
fish
and
you
could
put
it
on
a
grid
where
you
could.
Each
Landmark
would
have
like
a
certain
point
on
the
grid,
and
that
tie
point
then,
would
would
allow
you
to
transform
the
Grid
or
transform
the
phenotype
and
you'd
have
a
corresponding
mathematical
structure
or
you'd
have
a
corresponding
morphological
structure.
A
A
A
Basically,
where
the
landmarks
are
so
the
script
goes
underneath
the
morphology
and
there's
a
lot.
You
know
this
is
a
pretty
simple
representation,
and
so
you
might
have
type
minus
the
eye.
You
might
have
tied
points
like
this
under
winding
the
fish,
and
so
each
of
these
tie
points
corresponds
to
a
certain
Landmark
on
the
fish's
morphology
is.
E
A
A
From
a
linear
to
a
non-linear
sort
of
way-
and
this
is
of
course
changes
in
species-
but
it
could
be
changes
in
time-
it
could
be
a
develop,
it
could
be
a
developmental
phenotype
like
an
egg
or
a
some
sort
of
in
a
learnable
stage
organism.
A
So
we
can
do
this
sort
of
warping
with
developmental
forms,
and
then
we
can
look
at
this
relationship
between
embryo
networks
or
some
sort
of
network
topology
and
these
shapes
and
how
they
change
their
form
over
time.
A
Comments
about
that
or
questions
I,
don't
know
if
how
dick
is
doing,
how
are
you
doing
dick.
D
I'm
doing
better
today,
good,
that's
good
yeah!
If
you
want
they'll
present
that
Strange
New
Image,
processing
problem,
okay,
yeah
that'll
be
good.
Actually,
okay,
I
came
up
and
came
across
it
because
there
are
now
some
x-ray
sources
which
will
produce
a
narrow
beam
of
light
or
R
of
x-rays.
D
D
Okay,
entire
screen
and
back
up
here:
okay,
yes,
yeah,
okay,
okay,
now
I'll
switch
to
Auto
draw
to
show
you
what
the
problem
is
start
crying.
Okay,
take
any
picture
whatsoever
which
I'll
represent
as
a
square
and
let's
draw
a
line
across
it.
Okay,
suppose
that
we
sample
all
the
pixels
along
the
line.
D
D
And
we
could
start
to
do
in
what's
called
in
in
painting
or
in
filling
okay
interpolating
between
the
values
on
the
lines
this
one
that
produce
a
picture
of
some
quality,
and
if
we
have
enough
lines,
I
mean
it's
clear
here.
Let
me
do
another
one
here,
it's
clear
from
the
old-fashioned
CRT
televisions,
which,
if
you
have
enough
lines,
you
can
get
a
good
representation
of
a
picture.
That's
what
television
was
based
on
originally
until
the
flat
screens
have
shown
up
right,
and
so
those
are
all
parallel
lines.
D
But
I'm,
saying
let's
take
a
random
set
of
lines.
It
was
an
in-painting
algorithm
and
that
produces
a
picture,
and
the
question,
then,
is:
what
is
the
quality
of
the
image
versus
the
number
of
lines?
D
Okay,
now
the
lines
can
be
random
or
you
can
do
it
more
cleverly
and
say
we're
going
to
add
a
line,
but
it
depends
on
the
quality
and
the
image
and
where
we
put
it,
so
you
could
then
have
image
dependent
addition
and
lines
which
will
then
do
it
within
painting
and
then
you'll
be
further
Improvement
in
the
picture.
D
D
Okay
and
it's
it's
a
it's
a
it's
a
bit
strange
because
you're
taking
these
lines-
and
you
know,
function
along
each
line
which
the
value
of
the
pixels
and
and
then
here's
the
infilling
algorithm
to
fill
in
the
picture.
D
Well,
this
there
was
an
old,
an
old
problem
of
compressing
images
for
transmission,
so
they
went
much
faster.
It's
not
a
problem
nowadays,
but
this
would
also
your
particular
solution
to
any
progression
and
I
can't
find
anything
in
the
literature
except
you
know,
International
TV
screens,
which
do
anything
like
this.
D
D
This
is
abstract
it,
but
it
came
out
of
the
X-ray
sources
which
generate
lines
across
objects
if
these
will
be
x-ray
lines
instead
of
the
visible
lifelines,
but
backing
up.
It's
obvious
that
the
general
problem,
this
combination
of
lines
Plus
in
painting,
is
a
form
of
representing
an
image
with
a
lot
less
data
than
the
whole
image.
C
D
A
D
I'm
looking
for
a
partner
in
this
who
wants
to
it's
much
simpler
image
processing
than
these
segmentation
problems?
Yeah,
okay,
but
it
seems
to
be
a
brand
new
topic
for
image
processing.
A
For
one
thing,
yeah
one
area
I
could
imagine
immediately
is
imagine
if
you
had
a
an
image
of
a
lot
of
noise
in
it
and
you
couldn't
unmasks
yeah.
So
you
had
things
like
a
bunch
of
noise
as
a
mask
over
an
image,
and
you
couldn't
tell
what
the
signal
and
noise
like
you
couldn't
decouple
them.
So
you
know
you'd
have
to
do
that.
You
could
do
this
kind
of
sampling
where
you
go
over
basically
you'd
end
up
sampling
things
at
random.
A
You
just
kind
of
pick
a
pixel
and
you
okay,
flip
it
over
and
say:
what's
the
content,
information
content
of
this
pixel
or
that
pixel
you
did
it
this
way,
you
could
like,
like
kind
of
divide
up
the
image,
so
you
could
send
a
signal
across.
You
could
send
a
signal
across
here
and.
C
A
D
D
Then
you
get
lines
between
the
two
shafts
and
you
can
reconstruct
the
Earth
between
the
two
Jets
from
the
set
of
lines:
yeah.
Okay,
that's
another
example:
okay,
so
there
might
be
some
a
couple
of
practical
applications:
yeah,
that's
just
the
old
comparison,
Village
compression
right.
A
D
You
don't
bother
with
it,
okay,
but
it's
still
curious,
It's,
a
curious
problem,
reconstructing
an
image
from
Lion's
glove.
A
Okay,
all
right!
Well,
thank
you!
That's
great,
okay,
sure,
all
right.
Anyone
else
wanna
have
any
updates
or
Susan
or
Jesse.
A
B
My
my
Imaging
problem
is
a
bit
of
a
strange
one:
I
have
X
Y
and
Z
axis
and
I
have
let's
see
this
is
this
is
the
z-axis
and
this
is
the
x-axis
and
then
I
have
the
y-axis
going
this
way
here.
I,
don't
know
if
you
can
see
that
yeah
yeah
and
then
so.
My
diagrams
are
90
degrees
flipped
from
the
paper
I'm
trying
to
follow
and
I
suspect
that
they're
they're
y-axis
is
going
out
of
the
page
instead
of
into
the
page,
and
so
things
are
90
degree
flipped.
A
Anyways
yeah
I,
don't
know
what
what
program
are
you
using.
A
D
A
B
A
A
B
So
Amy
could
I
show
you
show
you.
B
A
B
Okay,
yeah
now
so
I
think
this
graph
that
looks
like
a
mountain
here.
A
pink
mountain
I
was
trying
to
duplicate
this
and
I
get
graphs
that
are
like
this.
B
On
the
left,
hand,
side
and
I
took
this
blue
line
here
this
it's
at
0.4
stream
and
I,
moved
it
around
the
axis
and
just
changed
the
negative
of
positive
and
I.
Still
I
don't
get
I.
It's
got
a
curve
going
up,
doesn't
doesn't
have
this
curve
going
down
like
this;
in
other
words,
it's
kind
of
90
degrees
flecked.
It
needs
to
be
turned
over
right
and
then
I
I
looked
at
this,
and
this
is
rubber
and
it
looks
kind
of
like,
like
my
curve
and
the
pot,
the
first
quadrant
anyways.
B
You
know
trying
to
find
the
other
one.
Oh
and
then
I
have
this,
and
this
is
for
a
tall
prism
instead
of
a
short
one.
Anyway.
I
got
this
to
work
for
me
here,
because
I
had
my
numbering
system
backwards.
B
A
B
Memory
system
is
bad
guys
here
here,
I,
don't
know
if
you
can
see
it
in
the
corners,
the
X,
X,
Y
and
Z,
and
the
Y
is
going
into
the
plane.
Okay,.
B
C
A
B
A
A
B
It
needs
90
degrees
slipping
anyway.
It's
a
really
stupid
problem,
but
it's
consistent
I
get
it
this
this
one
for
a
really
tall
triangle
prism.
This
is
sort
of
a
medium
tall
one,
and
this
is
for
a
shorter
one.
Yeah.
You
can
see.
There's
a
theme
here:
okay,
not
it
doesn't
follow
this
pill.
It's
not
a
hill,
it's
more
of
a
swirled
right
right
and
that's
my
problem,
so
I,
just
it's
stupid
and
I'm
hung
up
because
I'm
supposed
to
duplicate
this
paper
or
explain
why
it
doesn't
work.
A
Yeah
I
think
that's
some
interesting
data
and
it's
always
hard
to
like
get
a
to
get
the
graphs
to
plot
correctly
because
they
use
their
own
conventions.
Like
you
know,
each.
A
For
the
XYZ,
though,
couldn't
you
just
transform
the
y-axis
to
something
it's
like.
A
B
A
A
A
simpler
case
of
like
one
over
X,
and
that
would
give
you
like
a
like
a
decimal,
but
it
would
be
like
transformed
y,
so
it
would
like
turn
it
around,
but
it
would
be,
and
it
would
be
a
weird
number
like
the
weird
interval,
but
it
would
be
transformed
y,
so
it'd
basically
give
you
that
orientation
and
you
just
well.
This
is
transformed.
Why
and
it's
going
in
a
different
direction.
So
I
don't
know
if
that
would
work,
but
something.
B
A
B
That's
why
I've
been
having
problems?
Yeah,
yeah,
that's
exactly
why
I've
been
having
problems.
A
A
It's
all
right,
I
had
a
few
things
to
share
her
myself.
So
let
me
share
my
screen
or
at
the
top
of
the
hour,
but
it
shouldn't
take
that
long
to
talk
about
so
one
thing
I
found
this
week
was
this
interesting.
This
is
having
to
do
with
using
genomic
data
as
sort
of
like
an
input
for
what
they
call
Foundation
models.
A
So
this
is
I,
think
from
Stanford
research
group
at
Stanford
and
what
they're
doing
is
they're
doing
they're
working
with
hyena
DNA.
A
What
they're
trying
to
do
is
build
what
they
call
long-range
genomic
Foundation
model.
So
this
is
a
model
where
you
can
train
a
model
on
DNA
like
genomic.
E
A
A
bunch
of
nucleotides
and
it
can
predict
or
make
predictions
for
different
types
of
DNA
stretches.
So
one
of
the
problems
that
we've
had
in
in
genomics
is
DNA
alignments,
dino,
weinmans
and
then
discovering
stretches
of
DNA.
So
we've
in
the
past,
we've
used
things
like
blast.
We
use
in
all
sorts
of
mapping
software
where
we're
mapping
between
a
reference
genome
and
a
genome
that
we're
sequencing
and
to
get
those
aligned
to
get
them
to
know
where
they
are
in
the
in
the
reference.
So
this
is
a
a
little
post
on
this.
A
This
is
we're
excited
to
introduce
hyena
DNA,
a
long-range
genomic
Foundation
model,
with
context
links
up
to
1
million
tokens
that
single
nucleotide
resolution.
So
this
is
just
a
term.
That's
you
know.
Tokens
are
like
what
these
in
this
model
to
you
know,
provide
this
context
thing
hyper
hyena
DNA
is
pre-trained
on
the
human
reference
genome
and
sets
new
SAT
a
on
23
Downstream
tasks.
A
This
is
a
I
think
state
of
the
art,
including
predicting
regulatory
elements,
chromatin
profiles
in
species
classification,
so
they
can
take
a
like.
They
have
a
reference
genome
that
they
train
the
model.
A
On
and
then
it's
it's
kind
of
like
a
large
language
model
where
you're
predicting
on
a
bunch
of
letters,
you're
finding
features
like
regulatory
sequences,
chromatin
profiles,
if
you
want
to
find
a
chromatin
different
sites
for
chromatin
activity
and
species
classification,
so
you
can
detect
like
a
certain
stretch
of
DNA,
and
usually
we
do
this
in
Blast,
where
we
find
potential
matches
for
some
piece
of
DNA
that
we
put
into
a
blast
search.
A
So
this
is
you
know
this
will
give
you
those
predictions
I,
don't
know
how
it
compares
the
blast,
they
don't
really
say,
but
basically
that's
what
they're
trying
to
do.
It's
like
kind
of
a
Next,
Generation
blast
and
kind
of
using
a
large
language
model
to
or
a
large
language
model
techniques
to
sort
of
get
at
this
problem.
A
A
Paper
here
this
is
hyena
DNA.
They
talk
about
the
paper.
This
is
what
they
call
it
a
foundation
model.
A
So,
let's
see
they
do
actually
have
some
interesting
things
with
respect
to
other
types
of
machine
learning
models,
so
leveraging
hyenas
new
long-range
capabilities
represent
hyena
DNA.
They
talk
about
hyenas.
E
A
Here,
which
is
an
approach
that
they've
developed,
hyena
large
language
model
based
on
implicit
convolutions,
was
shown
to
match
attention,
which
is
a
method
that
they
use
in
machine
learning
and
deep
learning
in
quality,
while
allowing
longer
context
lengths
and
lower
time
complexity.
So
it's
a
large
language
model
that
they're
using
they're
fine-tuning
it
with
respect
to
the
data
here,
leveraging
Hyena's
new
long-range
capabilities,
we
present
hyena
DNA.
So
this
is
a
subset
of
hyena
hyenas,
which
is
the
model
that
they
have
here.
A
A
genomic
Foundation
model
pre-trained
on
the
human
reference
genome
with
context
links
up
to
1
million
tokens
at
the
single
nucleotide
level.
So
they
train
at
a
single
nucleotides
and
up
to
500
500
fold
increase
over
previous
density
tension-based
models.
So
people
were
using
attention-based
models
for
some
of
these
genomic
problems
earlier
in
the
so,
people
have
been
applying
machine
learning
and
deep
learning
to
this
to
sort
of
I
guess:
sort
of
get
improvements
over
black
traditional
blast,
searches
and
they're,
actually
improving
over
these
attention-based
models.
A
Hyena
DNA
scale,
Sub
quadratically
in
sequence,
length,
training
up
to
160
times
faster
than
Transformer
use
a
single
nucleotide
tokens
and
has
full
Global
context
in
each
layer.
We
explore
what
longer
context
enables
so
what
that
means
is.
It
includes
first
use
of
income
tax
learning
in
genomics
for
a
simple
adaptation
and
novel
tasks
without
updating
retrained
model
needs,
so
basically
they
can
find
you
know
very
specific
sequences,
they
can
find
novel
sequences
and
it
doesn't
need
a
lot
of
updating
in
terms
of
training.
A
A
A
You
know
getting
this
kind
of
get
going
from
like
blast
searches
which
were
kind
of
so
you
know,
2000s
to
you,
know
different
machine
learning
techniques
to
now
this
type
of
model,
so
they
were
talking
about
long
context,
models
which
are
state-of-the-art
thing.
It's
very
trendy
opening
I
stated
a
goal
to
reach
1
million
tokens.
These
are
these
long
context
models,
while
magic
announced
they've
reached
5
million
tokens
for
code
world
for
it,
but
well
most
of
the
focused
on
natural
language
and
code.
A
So
this
is
where
a
lot
of
large
language
models
focus
on
like
using
English
and
predicting
text
in
English
they're,
using
something
in
biology,
they're
using
DNA
sequences,
and
so
we
have
of
course,
four
nucleotides
that
get
rearranged
from
different
ways,
and
you
have
these
long
stretches
of
DNA
and
there
are
actually
quite
a
few
possible
combinations
that
it
could
be.
So
it's
a
hard
problem
to
solve
and
you
have,
like
you
know
things
within
that
those
stretches
of
DNA,
so
you
have
different
patterns.
A
You
know
you
have
repeats,
you
have
of
different
lengths.
You
have
different
types
of
combinations
of
codons
and
things
like
that.
You
have
stopped
codons,
you
have
Stark
codons,
and
so
you
have
a
lot
of
things
in
the
structure
of
DNA
sequences,
which
need
to
be
predicted.
A
Of
like
people
have
made
the
analogy
between
DNA
and
and
language
before,
but
this
is
where
you're
actually
using
this
to
find
to
to
improve
the
search
process
so
to
give
a
sense
of
the
scale.
The
human
genome
is
3.2
billion
nucleotides
and
can
be
seen
as
characters
in
the
sequence
so
they're
able
to.
A
They
have
some
blog
posts
on
this
as
well,
so
they
they
train
this
hyena
model,
which
is
a
convolutional
large
language
model,
and
so
they've
actually
been
able
to
do
this
with,
or
you
know
more
generally,
but
they've
applied
it
now
to
genomics
where
they
want
to
find.
You
know
they
want
to
really
kind
of
put
this
to
the
test,
so
here's
the
sample
from
the
so
they
have
a
reference
Gene.
You
know,
usually
you
do
this
some
sort
of
reference
genome
that
you
build
so
the
first
team
in
genome.
A
You
know:
first,
human
draft
of
the
human
genome
is
kind
of
like
a
reference
genome
and
they
do
this
for
different
species
where
they
kind
of
sequence,
a
bunch
of
organisms,
and
they
they
find
a
consensus
for
the
sequence.
And
then
they
they
put
a
draft
out.
A
The
the
sort
of
the
reference
genome
and
then,
if
you
just
go
to
sample
a
genome
from
your
own
sample,
that's
your
sample
and
then
you
try
to
map
it
to
the
reference
genome,
so
you
can
actually
align
things
and
make
sure
that
they're
in
the
right
place
in
the
genome.
So
that's
what
they're
trying
to
do,
but
they're
trying
to
pre-train
the
model
with
a
reference
genome
and
then
throw
data
at
it
from
different
samples
and
see
if
they
can
make
the
predictions.
So
this
is
how
it
works.
Basically,
they
have
this
reference.
A
Gene
now
I'm
going
to
use
that
as
the
pre-training
step.
Then
they
have
these
predictions,
these
token
predictions
that
they
they
propose.
They
match
it
up
with
this
technique
here
and
then
they
have
these
Downstream
tasks
where
they
try
to
classify
things
structures
within
the
DNA
sequences
like
finding
regulatory
Elements,
which
are
these
you
know
sometimes
they're
repeats,
sometimes
they're,
very
specific
sequences
domains
in
within
the
DNA.
A
Then
they
have
chromatin
profiles.
So
you
know
you
might
have
a
certain
two
or
three
DNA
sequence
or
nucleotide
sequence,
Motif
within
your
DNA
sequence,
that
you
can
discover,
and
these
are
potential
sites
for
chromatin
binding.
Then
you
have
in
context
learning
which
are
you
know,
other
features
that
you
want
to
learn
about,
and
that's
really,
you
know
sort
of
the
the
heart
of
genomics.
A
That's
why
we
we
do
genomics
is
because
we
want
to
find
these
patterns
that
we
want
to
say
like
this
is
useful,
for
you
know
some
biological
function
or
you
know
it's,
maybe
a
new
gene,
or
something
like
that.
So
you
know
this
is
kind
of
they're,
really
kind
of
moving
this
area
forward
and
so
they're
trying
to
learn
all
sorts
of
things
they
want
to
learn,
maybe
to
Define,
like
coding,
regions
from
non-coding
regions.
They
want
to
find
out
more
about
gene
expression.
A
They
want
to
find
out
about
other
types
of
structures
in
the
DNA
and
they
ultimately
want
to
go
to
building.
You
know
going
from
DNA
to
proteins.
So
you
know
this
is
where
you
have
a
sequence
of
nucleotides
that
gets
transcribed
into
RNA,
which
has
you
know
usually
there's
some
changes
in
the
RNA
that
derive
from
that
transcriptional
process
than
that
RNA
gets
built
into
proteins
through
another
process
of
translation,
and
so
you
know
this
is
the
first
step
in
predicting
what
these
proteins
will
look
like.
A
If
you
can
predict
patterns
in
the
DNA,
you
can
then
take
samples
of
RNA
predict
patterns
and
you
can
use
something
like
rna-seq
or
something
like
that
predict
patterns
in
the
RNA
and
then
link
those
two
things
and
then
link
that
to
the
proteins.
Because
now
you
know
we
can
of
course
model
in
sequence,
proteins.
We
can
actually
make
predictions
on
that
and
I
think
people
are
applying
larger
language
models
to
proteins
as
well
I'm,
not
familiar
with
how
that's
you
know
with
the
state
of
that
field
is.
A
But
so
you
know
you
have
these
different
DNA
motifs
promoter
silencers
enhancers,
and
you
want
to
be
able
to
discover
those
in
the
sequencing
context.
A
So
I'll
provide
this
paper
where
this
is
in
the
slack
and
if
I
can
send
it
out
an
email
as
well.
People
want
to
read
it.
A
Let's
see.
So
let
me
check
the
chat
here.
Okay,
so
Jia
hung
had
to
leave
and
if
you're
interested
in
what
jiahang
had
to
say
about
networks
and
topology,
don't
hesitate
to
reach
out.
He
put
his
email
in
the
chat
if
you're
interested
in
TDA
and
gnns
and
yeah.
A
A
Have
a
good
week,
thank
you.
Bye!
All
right
now,
I'd
like
to
talk
about
sort
of
a
third
installment
in
the
discussion
we've
been
having
about
different
types
of
embryoids.
A
So
two
weeks
ago
we
talked
about
human
embryoids
and
some
of
the
work
that's
been
happening
there
and
then
last
week
we
talked
about
monkey
embryoids
and
non-human
primates,
so
we
have
monkeys
and
humans,
and
so
we've
been
able
to
demonstrate
this
in
two
model:
organisms
in
in
primates
this
week,
I'm
going
to
talk
more
about
human
embryoids
and
some
of
the
gastroloid
work.
A
That's
been
done
so
you
know
this
is
like
I
I
showed
a
diagram
two
weeks
ago,
showing
sort
of
this
progression
from
organoids
to
gastroides,
to
embryoids
and
just
kind
of
how
those
models
are.
You
know
they're
kind
of
refining
those
models
to
give
us
answers
to
different
things.
We
could
look
at
different
things
in
development
and
get
certain
answers
depending
on
the
type
of
model.
We
build.
There's
a
lot
of
control
here.
It's
either
transgene
control,
transgenic
control,
which
means
you
put
a
gene
into
the
cells.
You
maybe
transform
them
into
an
induced
portal.
A
You
can
also
do
things
with
defined
media
which
we'll
talk
about
today,
so
the
Hana
paper
from
the
first
week
that
talked
about
defined
media
and
kind
of
getting
the
perfect
formula
for
your
medium
and
then
using
that.
Then
this
is
something
that
goes
back
to
and
do
support
you
know
making
into
supporting
potent
cells
where
the
some
people
use
transgenes.
Some
people
use
a
defined
medium
of
a
certain
type
and
I
can
tell
you
that
there's
a
lot
of
sort
of
technical
gamesmanship
there.
A
So
you
know
finding
the
right
medium
is
great,
but
some
people
like
there's
a
lot
of
skepticism
surrounding
some
of
the
results
for
both
transgenes
and
and
define
media.
So
you
know
when
you
read
these
results,
it's
amazing
stuff,
but
you
can
also
ask
the
question
you
know:
are
we
really
getting
the
thing?
We
think
we
are
we're
getting
you
know,
differentiated
cells,
we're
getting
tissues,
we're
getting
things
like
that.
A
To
say,
like
a
whole
embryo
setting,
where
you'd
actually
be
able
to
make
some
priming
or
maybe
clone
a
primate-
or
you
know,
creative
primate
embryo
that
could
be
implanted,
which
we
talked
about
in
the
first
week
as
well.
A
So
with
that
in
mind,
let's
go
to
the
papers,
so
this
is
actually
an
article
in
I.
Think
it's
in
the
Observer
It's
a
UK
publication.
This
is
by
Philip
ball.
This
is
a
feature
that
they
had
on
this.
This
human
embryo
models
were
so
this
is
kind
of
considering
some
of
what's
been
going
on.
They
showed
an
example
of
how
these
embryo
models
are
made.
So
these
are
embryo
models,
they're,
not
necessarily
embryos,
they're,
technically
embryoids
or
they're,
technically
something
to
some
of
the
embryoids.
A
They
could
be
a
gastroyed
or
even
just
the
standard
organoid.
Although
what
they're
trying
to
do
is
build
this
embryo
with
the
tissues
and
with
the
structure
of
an
embryo
as
opposed
to
just
creating,
maybe
like
Joe
Myers,
but
they
are
creating
germ
layers
as
well
as
we
talked
about
in
the
first
and
second
weeks.
So
you
know
they're
creating
these
embryo
models
out
of
stem
cells,
they're
transforming
them
to
these
different
Fates
building.
A
You
know
there
is
a
building,
differentiated
tissues,
they're,
building,
a
structure
that
looks
like
an
embryo
and
so
that
that's
where
we
are
and
they're
a
host
of
ethical
challenges
surrounding
this
I,
don't
want
to
Discount
those.
This
article
I
just
talked
about
these
rare
to
see
like
a
print
version
of
an
article
anymore,
and
you
can't
really
read
the
text
but
basically
goes
over
the
papers
we
talked
about
in
the
first
week
and
and
some
of
the
so
you
can
find
that
article
online.
You
know
that
would
be
weird
about
it.
A
So
the
first
paper
I
want
to
talk
about
is
this
paper.
This
is
a
bunch
of
authors
here,
just
a
lot
that
all
and
the
title
was
modeling
human
post-implantation
development
by
extra
embryonic
Nation
engineering.
So
this
is
the
work
on
defined
media
and
the
idea
behind
defined
media
is
that
you
have
the
cells
grow
on
this
medium
and
you
have
to
find
the
right
recipe.
So
it's
a
chemical
inducement
of
this
to
create
these
underneath
holes.
A
A
All
sorts
of
there
are
all
sorts
of
chemicals
that
keep
the
pH
at
the
right
level,
and
things
like
that,
so
they're
all
sorts
of
things
that
we
can
modify
to
make
optimize
that
environment
and
that's
what
they're
doing
in
this
paper
they're
trying
to
kind
of
they're
talking
about
this
in
terms
of
extra
embryonic
Niche
engineering,
so
extra
embryonic
just
means
out
in
the
environment
of
the
cell
or
of
the
embryo,
and
so
when
you're
out
in
the
environment,.
A
Or
something
like
that
now,
the
environment
is,
you
know
we
talked
about
environment
before
in
the
group
and
that
is
to
say
that
it's
hard
to
Define,
you
have
to
really
Define
what
you
mean
by
environment,
which
is
of
course
a
concern
here,
because
this
is
a
chemicals
formula
for
the
environment.
Basically,
so
this
is.
Where
kind
of
you
know,
this
is
one
approach
to
making
these.
A
So
the
abstract
reads:
implantation
of
the
human
embryo
commences
a
critical
developmental
stage
that
comprises
profound
morphogenetic,
alteration
of
embryonic
and
extreme
embryonic
tissues,
access
formation
and
gastric
relation
events,
so
you're
not
only
creating
the
tissues
you're,
creating
the
axial
polarity,
so
things
are
towards
the
header
towards
the
tail.
In
that
sense,
you
know
so
they
have
to
be
in
the
right
order.
A
Our
mechanistic
knowledge
of
this
window
of
human
life
remains
limited
due
to
restricted
access
to.
In
Vivo
samples,
well
Technical
and
ethical
reasons,
so
we
can't
necessarily
create,
like
you
know
something
in
the
womb
that's
frowned
upon,
but
we
can
create
it
in
a
three-dimensional
culture.
This.
A
In
every
models,
because
we
can't
experiment
on
fetuses
basically-
and
you
know-
there's
a
lot
of
you
know
that's.
This
is
actually
why
we
have
abuse
policy
cells,
because
there
were
a
lot
of
ethical
concerns
and
people
have
moral
objections.
Sometimes
the
stem
cell
models
actually
harvesting
them
from
the
embryo,
and
so
that's
where
a
lot
of
stem.
E
A
Thing,
that's
been,
you
know
at
the
floor
of
this
field
for
a
long
time,
so
this
gives
us
a
model
where
we
can
work
on
these
problems
without
some
of
these
ethical
concerns.
A
Additionally,
human
stem
cell
models
of
early
post-implantation
development
with
those
embryonic
and
extra
embryonic
tissue
morphogenesis
are
lacking.
So
we
don't
really
know
anything
like
I
said
in
the
first
week
about
posting
Plantation
development.
We
know
a
lot
about
pre-implantation
development
in
terms
of
like
working
on
it
and
culture
and
models,
but
this
posting
Plantation
world
is
quite
different
here.
A
So
it's
a
small
eye
discoid
and
it
sounds
like
something
Apple
would
make,
but
it's
actually
a
biological
thing
here
we
present
I
discoid
produced
from
human
inducible
and
stem
cells
by
an
engineered
synthetic
Gene
circuit,
so
they're
actually
using
a
synthetic
Gene
circuit
to
sort
of
helpless.
You
know
help
the
embryoid
model
respond
to
its
environment,
the
optimal
environmental
conditions.
So
you
have
this
you're,
basically
creating
an
inducibility
stem
cell
you're
using
an
engineered,
a
synthetic
Gene
circuit
you're
using
that
to
sort
of
control.
A
The
fate
of
the
cell
I
discoides,
exhibit
reciprocal
co-development
of
human
embryonic
tissue,
an
engineered
extreme,
organic
niche.
Remember
all
these
chemicals
in
a
model
of
human
post-implantation,
the
end
they
exhibit
unanticipated,
self-organization
and
tissue
boundary
formation
that
recapitulates
the
oak
Sac
like
tissue
specification.
A
So
this
is
at
a
certain
point
in
development
where
we
get
this
yolk
Sac,
we
get
the
tissue
specification,
the
boundaries
of
tissue
start
to
form,
and
this
is
all
sort
of
self-organized.
There's.
No,
you
know
they're
chemical
signals
within
the
embryo,
but
particularly
their
chemical
signals
outside
the
embryo
that
are
sort
of
enforcing
this.
A
A
They
were
looking
at
how
the
blood
the
the
circulatory
system
develops
in
the
formation
of
blood.
So
hematopoetic
characteristics
are
the
formation
of
bilamina
dislike
embryonic
morphology,
which
is
you
know
your
red
and
white
blood
cells.
I.
Think
the
red
blood
cells
are
what
they
focused
on
in
the
papers.
In
the
first
week,
the
development
of
the
amniotic
like
cavity
and
acquisition
of
an
anterior
like
hypoglass
pole
and
posterior,
like
axis
idisk,
leads
offered
easy
to
use
high
throughput
reproducible
and
scalable
platform
to
promote
multifaceted
aspects
of
human
early
posting
implantation
development.
A
So
we
can
do
a
lot
of
things
with
this.
We
can
do
drug
testing,
developmental
toxicology
and
modeling
diseases
that
occur
in
utero.
So
here
are
some
images
of
what
they're
doing
here.
This
is
where
they
have
zero
percent
I
get
a
six
highest
supplementation.
This
is
okay,
so
this
shows
some
of
the
features
in
here.
Here's
the
hemoglobin
starting
to
form
here.
This
is
another
example
of
hemoglobin.
A
Signal
here
you
have
some
again
six
High
supplementation
25,
so
it's
a
zero
percent.
This
is
25
percent.
This
is
showing
some
of
these
hemoglobin
in
context
here.
So
here's
some
of
the
structures
here
in
the
hemoglobin
surrounding
it.
A
Here's
some
more
pictures
of
where
you
have
kind
of
a
you're
going
through
the
structure
from
the
mesoderm
to
the
endothelial,
to
the
endoderm,
so
from
the
bottom
to
the
top
and
here's
some
images
here
where
you
have
stained
for
Fox
A2,
cd34
Desmond,
which
is
a
muscle
thing,
an
effect
in
which
is
another
muscle
thing.
So
you
see
a
Desmond
is
in
here
being
expressed.
Fox
A2
is
in
the
blue,
see
some
of
this
in
the
endothelial,
and
especially
in
the
endoderm
and.
A
You
have
a
lot
of
Desmond,
which
decreases
as
you
go
towards
the
endoderm
and
then
cd34,
which
is
another
marker,
which
decreases
as
you
go
towards
the
endoderber,
the
top.
So
you
get
these
three
germ
layers,
these
three
types
of
precursors
to
the
tissues
and
you
get
different
things
expressed.
A
So
you
get
things
expressed
with
respect
to
muscle
and
some
other
things
that
are
in
so
you
can
see
that
they
differ
where
you
are
in
the
sculpture
same
thing
here
you
have
Desmond
and
white,
you
have
cd43
and
yellow
and
then
cd31
and
red,
and
you
see
that
as
you
move
through
from
the
mesoderm
to
end
of
the
oil
hematopoietic
cells,
that
it's
changing,
it's
composition.
A
So
this
is
just
an
example
of
how
these
things
form
in
time.
This
is
actually
an
example
here
of
what's
happening
in
these
embryoid
models.
This
is
a
figure
where
they
show
the
first
symmetry
breaking
here,
and
then
you
see
confinement
and
migration,
so
cell
migration
is
happening,
and
then
you
have
this
bilometer
boundary,
so
this
is
corresponding
to
differential
gene
expression,
so
you
have
so.
First
of
all,
you
start
off
with
these
two
cell
lines.
The
I
got
a
six
which
are
uninduced
in
the
wild
type.
A
Then
in
the
I-discoid
cell
line
you
have
the
combination
of
them,
so
you're
inducing
the
I
get
a
six,
but
it's
also
mixed
with
mild
type
cells.
You
have
the
Symmetry
breaking
between
the
I
get
a
six
and
the
wild
type.
We
have
this
confinement
of
migration
of
different
I
get
a
six
cells.
You
have
this
bilaminar
boundary
between
the
Gata
6
being
expressed
here
and
then
the
wild
type
you
get
nanog
expression,
which
is
these
are.
A
Of
course,
an
important
marker
of
differentiation
and
stemness.
A
Is
where
you
get
more,
this
sort.
A
A
So
this
is
like
the
amniotic
SAC
and
the
disc
associated
with
it
the
posterior
formation.
Here,
then,
this
is
the
model
of
the
human
yolk
sac
and
you're
modeling
hematopoiesis
here
from
the
music.
So
these
are
all
things
where
you're
starting
to
get
structure
you're
starting
to
get
different
structures.
You
recognize
as
embryonic
especially
human
embryonic,
and
it
leads
to
this
model.
A
So
this
is
from
nature,
and
this
is
an
accelerated
article
preview,
so
they're
going
to
have
a
different
version
of
this
out
soon,
where
it's
the
final
version.
This
is
weird
the
way
they
do
this
in
this
part
of
science,
where
it's
very
competitive.
They
have
these
a
borrowed
papers.
They
have
accelerated
papers,
they
try
to
bring
them
on
all
at
the
same
time.
It's
just.
This
is
part
of
the
culture.
A
A
So
we've
talked
about
how
these
embryoids
provide
an
alternative
to
accessing
experimental
models
that
we
couldn't
exist
excess
before
here
we
show
the
human
stem
cells
can
be
triggered
to
self-organize
into
three-dimensional
structures
that
recapitulate
some
key
spatio
temporal
events
of
the
human
proposed
implantation
of
organic
development.
So
this
is
where
we're
focusing
in
these
spatio
temporal
events,
this
organization
and
spaces
you
saw
in
the
previous
paper
and
then
over
time.
A
A
Guess
in
terms
of
their
spatial
distribution,
and
then
you
get
the
Symmetry
breaking,
which
is
where
they
differ,
so,
like
a
left
and
right,
symmetry
braking
would
be
where
the
left
side
differs
from
the
right
side,
there's
an
event
that
triggers
a
difference,
and
these
are
common
in
development.
This
is
involving.
You
know
these
signaling
hubs,
where
you
get
secretions
of
different
chemicals
out
of
the
environment,
into
the
embryonic
environment.
So.
A
The
medium,
the
extreme
moronic
environment,
this
is
within
the
embryo,
the
environment,
and
you
get
these
secreted
modulators
that
get
pumped
out
into
the
local
environment
and
it
tells
other
cells
kind
of
what
to
do
so.
You
want
to
establish
these
key
signaling
hubs.
It
helps
to
sort
of
organize
the
model,
so
self-organization,
maybe
is
a
little
fuzzy
in
this
context,
because
self-organization
just
means
that
it's
not
involving
some
sort
of
supervised
process,
there's
no
top-down
process,
it's
just
kind
of
that.
These
signaling
UPS
start
to
form
you
get
the
Supreme.
A
This
is
a
secret
tone
that
sort
of
exists
in
the
model
and
then
the
secretion,
those
secreted
molecules
actually
help
to
you-
know
signal
other
cells
and
tell
them
where
to
go.
What
to
do
so.
These
single
scale
cell
transcriptomics
to
confirm
differentiation
in
diverse
cell
States,
so
we
were
using
a
lot
of
labeling
in
the
last
paper,
a
lot
of
fluorescent
labeling
in
this
case
they're
using
single
cell
transcript.
So
it
makes
this
confirms
a
lot
of
these
sort
of
sulfates
that
they're
actually
forming
different
types
of
cells.
A
E
A
A
Of
what
they
have
to
offer
here
in
terms
of
their
evidence,
you
have
a
couple
things.
First
of
all,
I
want
to
highlight
the
epiblast
and
hypoblast
layers
that
that
form
first
and
then
you
end
up
with
these
more
detailed
structures.
So
here's
an
example
here,
A
Carnegie
stage
for
you
of
this
implantation
stage
where
you
have
the
Inner
Cell
Mass
in
CS5,
you
have
this
epiblast
and
hyperblast
form
around
within
a
trophoblast,
and
then
you
have
the
amnion
forming
and
you
have
this
more
detailed
structure.
A
So
you
can
see
that
there's
this
sort
of
layering
and
pattern
of
formation
that
accompanies
this
protocol,
then
they
show
some
stains
here,
go
to
the
second
figure,
where
they
actually
do
some
Gene
Expressions.
So
they're
doing
this
gene
expression,
profiling
and
they're,
plotting
it
out
in
a
umat
space
here,
so
they're,
actually
showing
some
of
these
cell
types
in
a
three-dimensional
three
components
of
a
umap
analysis,
so
they're,
showing
that
their
differences
in
the
expression
profiles
that
there's
some
overlap.
A
But
it
largely
is,
you
know,
clustered
in
the
space,
so
we'll
call
it
clustering
per
se
and
emap.
But
this
shows
another
analysis
on
two
umap
components
where
they
show
some
of
these
cell
morphologies,
and
this
is
actually
shows
like
some
of
these
differences
where
you
have
mesoderm
here.
You
have
extra
embryonic
mesoderm
here
the
blood
progenitors,
as
we
saw
in
the
last
paper
trophoblasts.
A
Epiblasts
and
hyperblasts,
so
these
are
all
separated
out
in
this
umap
space.
So
this
is
dimensionality
reduction,
you're,
basically
trying
to
find
how
they
you
know
if
they're
separate
or
if
they're
sort
of
commingled
with
each
other
and
that's
how
you
can
tell
generally
how
you
know
different
the
different
label
cell
types
are,
so
they
may
identify
the
cell
types
from
some
other
Criterion
like
a
staining
for
markers,
and
then
they
look
at
this
gene
expression
profile,
which
includes
email,
thousands
of
transcripts,
and
they
can
actually
do
this
sort
of
analysis.
So.
E
A
Good
that
we
have
the
next
gen
sequencing
and
the
dimensionality
reduction
here,
and
then
this
just
shows
more
examples
of
some
of
the
structures
that
are
forming
in
the
embryoid.
You
have
these
different
states
for
effect
and
gappy
and
AFP.
So
you
can
look
at
like
how
some
of
these
things
are
expressed
in
the
structure
and.
A
A
Then
they're
self-sorting
into
these
layers,
they're
self-patterning,
where
the
layers
start
to
pattern,
and
then
you
get
this
these
secret,
these
secretion
centers
here
and
here
and
then
these
secreted
inducers
are
pumping
Out,
secret,
home
and
they're
inducing
other
cells
to
form
other
things.
That's
basically
the
process
here.
So
you
can
see
as
you
get
this
aggregation
you
get
patterning,
then
you
get
further
patterning
from
sort
of
these
hubs
of
secretum,
and
so
this
shows
again
what
this
looks
like.
A
You
have
EPA
blast,
like
cells
hypoglass,
like
cells,
they
get
transformed
through
fgf
and
middle
actions
in
this
sort
of
aggregation
step
here,
and
then
you
get
the
Sorting
of
cells.
So
once
you
get
an
aggregation
of
these
cells,
you
know
they're
sort
of
from
Mosaic.
They
get
sorted
into
these
layers
by
gene
expression
profile
or
by
how
much
fgf
and
nodal
they're
producing,
and
then
they
start
to
form
this
patterning,
which
needs
to
be
Senators
for
me.
A
So
the
hypoglass
like
cells,
either
go
to
this
anterior
visceral
endoderm,
like
cell
fate,
or
this
amnion
like
fate
and
then
there's
a
it
goes
to
a
mesoderm
like
thing.
If
it's
given
enough
bmp4,
there's
feedback
bmp4
for
amnion,
like
cells
and
then
emp2
from
hyperblast
amnion
like
so
and
so
they're,
you
know,
bmp2
is
playing
a
big
role
here
in
some
of
these
transitions.
A
E
A
A
And
then
the
final
paper
is
this
paper
on
gastroloids:
this
is
in
stem
cell
reviews
and
reports.
This
is
on
these
gastroloid
models,
so
I
said
that
they're
embryoids
and
then
their
gastroloids,
which
are
a
little
bit
more
specific.
So
these
are
novel
systems
for
disease,
modeling
and
drug
testing.
So.
E
A
A
So
by
virtue
of
inaccessible
nature,
mammalian
implantation
stage
development,
which
is
what
we've
been
talking
about,
has
remain
one
of
the
most
enigmatic
and
hard
to
investigate
periods
of
riogenesis
derived
from
pluribose
stem
cells.
Gastroise
recapitulate
key
aspects
of
gastro
with
Stage
embryos
and
have
emerged
as
a
powerful
in
vitro
tool
to
study
the
architectural
features
of
early
post-implantation
and
versions.
A
So
they're,
focusing
on
the
stage
with
these
gastroides-
and
you
know
it's
just
really
kind
of
recapitulating
this
process,
which
is
actually
very
important
to
understand,
and
you
can
see
that
there's
a
lot
of
self-organization
and
there's
a
lot
of
things
going
on.
So
it's
important
to
have
a
very
specific
biological
model
that
we
can
use
for
this.
A
While
the
majority
of
the
work
in
the
emerging
field
is
focused
on
the
use
of
gastro
lights,
to
model,
embryogenesis
retractable,
nature
and
suitability
for
high
throughput,
scaling
is
presented
in
unprecedented
opportunity
to
investigate
both
Developmental
and
environmental
aberrations
to
the
embryo
as
they
occur
in
vitro.
So
we
want
to
look
at
some
of
these
aberrations
and
development.
Maybe,
like
things
you
know,
dysregulation
of
some
of
the
factors
that
we
talked
about
like
emp4
or
something
like
that.
Then
you
turn
it
off.
A
If
there's
a
mutation
that
turns
it
off
or
whether
there's
a
circumstance
that
turns
it
down,
what
is
the
effect
on
the
asteroid
or
the
embryoid
or
if
the
environment's
different,
if
the
environment
is
not
as
it
should
be,
we
talked
about
the
within
the
embryo
environment,
which
is
this
secretion
of
molecules
and
signaling
molecules
that
guide
cells
to
their
faith
or
the
external
environment.
So
if
the
medium
is
not
as
it
should
be,
and.
A
A
standing
for
the
the
environment
in
the
womb,
because
we
don't
you
know
we
can't
necessarily
know
exactly
what
that's
like,
but
we
can
try
to
approximate
it.
So
we
have
these
two
different
types
of
levels
of
environment
and
if
they,
you
know,
have
aberrations
for
some
reason,
then
that
can
affect
you.
We
want
to
know
what
that
looks
like
what
those
effects.
Look
like
this
review
summarizes
the
recent
developments
in
the
use
of
gastroites
to
model
congenital
abnormalities.
A
This
is
the
review
I'm
not
going
to
go
through
the
whole
thing,
I'm
just
going
to
say
that
basically,
this
period
of
development
has
been
kind
of
like
a
black
box
and
now
we're
kind
of
making
this
gray
box
as
the
same
cybernetics,
meaning
that
we
know
something
about
right
now,
but
we
don't
know
everything,
and
so
that's
work
to
come.