►
From YouTube: DevoWorm (2023, Meeting #31): DevoLearn and Docs, Temporal Synthetic Bio, Collective Cell Mechanics
Description
DevoLearn updates. How to create effective open-source documentation. CompuCell 3D workshop. Synthetic Biology circuits in cells. Temporal transcriptomics and transcriptional dynamics during the process of differentiation. How to infer cell mechanics parameters from computational models. How to infer symmetry from droplets. Attendees: Sushmanth Reddy Mereddy, Bradly Alicea, Susan Crawford-Young, Lukas, and Richard Gordon.
A
B
A
A
B
A
B
I
have
trained
it
up
to
10
epochs,
only
whatever
the
last
function
I
am
using,
it
is
Dice
loss,
so
dice
loss
generally
lies
the
loss
value
between
0
and
1..
If
it
gives
near
to
0
it's
the
worst
model,
I
mean
it's
not
predicting
the
correctly,
but
here
we
can
see
it
is
giving
0.99.
Which
is
approximately
near
2
1.,
the
modal
accuracy
was
good
and
it
is
predicting
perfectly
perfect
in
next
week,
meeting
I
will
add
the
visualization
of
all
this
model.
Train
model.
B
Issues
but
I
cleared
them
out
and
I
have
trained
it
up
to
10
epochs
bad
size
was
4
and
model
was
able
to
learn.
This
can
I
mean
yeah.
The
model
was
completely
learned
here.
Okay,
here
we
can
see
the
all
the
data
how
much
the
final
thing
is.
I
took
around
three
days
to
train
it
on
batches,
so
yeah
and
I
was
working
on
the
paper.
Actually
I
was
reading
some
papers.
I
was
writing
paper
mainly
for.
B
Writing
the
cell
segmentation
part
so
I
was
reading
some
research
past
research
papers
about
CL
gas
when
they
are
in
between
him
when
they
are
going
through
embryogenesis
right
and
in
that
domain
we
are
using
deep
learning
to
segment
these
cells.
That's
the
whole
paper
report
and
creating
like
comparisons.
First
I
need
to
understand
like
why
it
is
hard
to
segment
C
elegans
and
all
these
things.
B
And
so
you
see
any
one
another
paper,
C
shaper
was
another
paper
which
you
have
shared
and
I
am
trying
to
implement
them.
I
mean
not
Implement
them.
I
am
trying
to
understand
the.
What
is
the
main
problem
of
segmenting,
C,
elegance
and
those
matter.
I
will
use
in
our
paper,
but
in
mechanism
and
all
this
stuff
and
I
mean
these
are
the
reasons
actually
I
need
to
show
it
to
mayuk
actually,
but
before
my
I
have
showed
it
to
you.
Maybe
next
by
next
week,
meeting
I'll
show
it
to
my.
B
You
can
confirm
it,
but
the
model
accuracy
was
good.
I
mean
the
loss.
Function
was
very
high,
I
didn't
expect
it
like
it
would
come
around
0.99
I,
expected
it
around
0.80
or
something
but
started
starting
only
started
working.
Fine
I,
don't
know
really
whether
I
am
wrong
or
right,
but
I
will
discuss
with
my
health
and
I
will
let
you
know.
But
finally,
the
model
was
able
to
learn
and
need
to
create
some
visualizations
for
papers
like
pictures,
graphs,
Etc
right.
A
Yeah
I
think
the
that's
that's
true,
so
that
those
are
good
those
results.
And
then
you
know
the
the
kind
of
thing
you
usually
put
in
a
paper.
You
report
I
think
in
this
case,
you're
going
to
have
your
loss
function
versus
your
epics,
so
it
goes
from
like
zero
to
ten
and
then
watch.
You
know
if
you
I,
don't
know
if
you
have
the
loss
function
for
each
epic,
but
it
should.
B
B
B
A
A
You
know
we
should
probably
plan
out
the
different
graphs
that
we
want
to
put
in.
You
know
there
obviously
be
graphs
like
the
boss
function
graph.
There
will
be
other
types
of
graphs
that
we'll
want
to
make,
and
then
you
know
we'll
we'll
be
able
to
make
a
report
on
it.
We'll
probably
also
have
some
other
performance
metrics.
A
Sometimes
people
will
like
train
a
model
on
some,
like
you
know,
trivial
data
set
just
to
show
what
the
model
can
do.
So
you
know
it's
like
something:
that's
really
easy
to
segment
like
a
bunch
of
circles
that
are
just
very
basically
really
easy.
There's
no
background.
The
background
is
very
obvious,
and
then
all
of
that
would
be
like
trained,
and
it
would
be
really
quick,
but
to
show
like
that,
this
is
actually
doing
something.
Instead
of
like
just
picking
up
noise,
yeah.
B
Okay,
I
will
try
to
create
some
resolution
for
it
like
graphs
pictures,
Etc
I'll,
try
to
host
it
like
on
hugging
face
I
will
try,
it
I
mean
Sam
model
was
not
supporting
to
hold
as
a
invest
spaces.
Maybe
I
can
use
some
Napoli
some
segment
and
if
you
have
like
I
provide
some
other
things,
I'll
try
to
go
through
into
it
and
I
will
try.
I'd,
say
library.
I
will
show
it
to
you
next
week.
Meeting
in
this
life,
maybe
yeah.
C
B
I
was
just
thinking,
of
course.
You
can
ask
it
because
this
this
what
should
be
in
sessions-
it's
not
anymore,
not
as
a
paper,
but
we
need
to
more
develop
this
thing,
not
only
for
cl
against.
We
need
to
adapt
it
to
other
cells,
also,
not
only
CL
games
like
more
cells
Etc.
We
need
to
train
that
due
to
GPU
issues,
I
haven't
trained
on
all
of
them,
but
I
will
try
to
train
them
on
ourselves.
Also,
if
I
get
a
chance
all.
B
A
B
Hanging
face
spaces.
The
model
was
around
375
million
parameters.
It's
not
a
small
model
and
hugging
phase
spaces
provides
only
up
to
some
extent
level
to
host
it.
Not
like
I
mean
it's
a
big
model
right
yeah
on
free
version.
We
can't
host
it
and
modern
will
not
also
work
fine
as
we
as
right
now.
We
have
three
modems
over
there,
which
are
working
fine,
because
the
I
mean
parameters
are
very
small,
but
whatever
the
Sam
model,
it
is
there.
It
has
so
many
parameters
and
the
free
version
type
will
not
satisfy
that.
A
A
B
A
Okay,
yeah,
that
sounds
good
yeah.
A
lot
of
the
large
language
models
are
sort
of
having
that
same
problem
where
you
want
to
run
them
on
your
local
machine,
but
you
want
to
be
able
to
like
also
host
them
in
different
places
and
they're.
Pretty
big
models
but,
like
you
know,
working
with
them
is
a
little
challenging,
but
still
you
know,
but
that's
that
sounds
good.
I
mean
that'll,
be
good
to
have
you
know
in
place.
A
I
think
this
is
a
pretty
big
improvement
from
what
we've
had
before
so
and
again,
we'll
we'll
have
to
plan
out
the
figures
and
the
tables
and
things
there
are
certain
things
you'll
want
to
report.
So
you
know
when,
when
I
know,
you've
probably
read
a
lot
of
papers,
but
about
you
know
in
in
that
area,
but
to
look
at
some
of
the
papers
and
how
people
report
things.
C
A
You
know
yeah,
you
want
to
have
like
these
control
conditions
like
very
simple
things,
or
sometimes
people
use
gaussian
noise.
To,
like
you
know,
just
kind
of
say.
This
is
like
this
shouldn't
segment,
or
this
should
segment
very
easily
and
if
it
does
or
it
doesn't,
then
you
know
that,
like
it's,
not
picking
up
on
it's
picking
up
on
the
right
kinds
of
things,
so
you
know
you'll
see
the
the
loss
function.
B
A
A
A
C
A
So
yeah
I've
been
working
on
the
paper
I'm
almost
ready
to
send
it
out
to
you
guys
so
I
yeah
it'll
happen
this
probably
today
later
on
today,
yeah
and
yeah
Lucas
finished
up
with
his
analysis.
So
thank
you,
Lucas
and
then
yeah
I've
been
working
on
some
analysis
and
that
Lucas
is
there
anything
you
wanted
to
say.
C
A
A
B
A
So
that's
good
all
right
yeah,
so
schmante
was
talking
about
documentation
and
the
importance
of
documentation
and
I
know
that
Mayo
said
that
he
would
fail
such
month
if
he
didn't
do
proper
documentation.
So
but
it
we
it's
very
important
and
it's
something
that
we
talk
about
a
lot
in
open
source
and
I
know
hamanchi's
not
here
yet,
but
I'll
kind
of
use.
This,
hopefully
he'll,
come
in
when
we're
as
we're
talking
about
this.
A
So
there
are
a
lot
of
examples
of
documentation
that
you
want
to
use
and
I
have
some
other
gsoc
students
I
have
an
open.
We
have
an
open
source
meeting
on
Fridays
where
I
actually
went
over
this.
Some
of
this
stuff
went
over
four
of
these
documents
and
then
one
of
the
people
from
our
lab
group.
He
actually
gave
some
more
links
and
more
examples,
so
we're
going
to
go
through
some
documentation
examples
for
open
source
software.
A
So
let
me
share
my
screen
here,
so
you
know
there's
a
lot
of
there
are
things
that
go
into
good
documentation.
Obviously,
you
want
to
have
like
be
precise
and
be
clear
about
what's
in
the
repository,
but
you
also
want
to
have
different
types
of
documentation
and
I
think
we
talked
about
this
earlier
about.
A
You
know
it
might
be
a
little
hard
for
like
someone
who
just
wants
to
segments
themselves
to
run
the
program
versus
someone
who
wants
to
Benchmark
the
software.
You
know
you
just
have
different
skill
levels
and
they
want
different
things
out
of
the
software,
so
it's
important
to
cater
to
those
different
groups
of
people
and
you
kind
of
have
to
think
about
who
you
want
to
get
to
use
your
software
who's
more
most
likely
to
use
your
software
when
you
write
the
documentation.
A
So
you
know
you
might
write
like
one
piece
of
documentation
for
people
who
maybe
just
want
to
use
it.
Maybe
just
a
down
and
dirty
sort
of
segmentation
and
they
don't
care
so
much
about
the
performance,
but
other
people
might
want
to
you
know
they
carry
greatly
about
the
performance
they're.
Basically,
you
know
professional
computer
scientists,
and
so
they
want
to
know
about
that.
So
you
offer
documentation
both
those
areas.
A
So
the
first
example
I'm
going
to
give
is
there's
this
article.
We
went
over
in
our
open
source
meetings,
the
anatomy
of
a
great
open
source
documentation
how
to
document
your
projects
on
GitHub,
and
this
kind
of
goes
over
some
of
the
points
that
they
talk.
You
know
that
are
important
for
good
documentation.
A
You
know
with
documentation,
it's
usually
something
that
you
write
up.
You
maybe
give
examples
of
how
to
use
the
software.
You
describe
what
it
is
what's
in
the
package
and
the
idea
is
that
your
code
becomes
more
understandable
and
reusable
people
can
understand
how
it
works.
They
can
use
it
for
their
own
purposes.
A
This
will
help
with
recruiting
people
for
open
source
participation.
So
you
know
I
I,
think
with
what
we
have
with
Devoe
learn
people
enough
people
had
seen
like
the
project
and
they
were
able
to
basically
understand
it.
But
you
know
there
was
a
lot.
There
was
a
lot
missing,
of
course,
but
you
want
to
get
it
to
the
point
where
people
can
identify
what
what
they
can
contribute
to,
what
are
the
missing
pieces
or
what
can
be
improved,
and
some
of
that
involves
documentation
and
making
sure
people
understand
how
things
work.
A
You
know
we
also
Empower
people
who
maybe
either
want
to
understand
the
code
base
or
they
want
to
use
the
project
and
Empower
them
so
that
they'll
be
able
to
use
the
software,
and
so
that's
just
basically
instructions
on
how
to
use
it
or
cases
of
how
to
use
it
or
things
like
that,
and
so
then,
but
you
know,
documentation
can
also
help
build
a
reputation
of
the
project.
So
there
are
a
couple
of
examples
of
GitHub
documentation
that
are
really
well
done.
A
This
one
is
a
game,
a
Space,
Invaders
game
that
someone
made
this
one
actually
is
a
I
guess
it's
a
class
project
that
someone
made.
So
this
is
an
example
here,
where
they
just
have
this
project
navigate
the
Mars
rover.
That's
the
first
project
help
the
Mars
Curiosity
Rover
find
the
shortest
path
between
two
points,
while
avoiding
obstacles
on
the
way.
A
So
they
have
this
software
here
that
you
can
run,
and
you
have
these
different
projects
that
you
can
do
or
these
different
like
sort
of
demos,
and
so
this
basically
shows
how
to
do
this.
It
gives
you
an
animation,
and
so
it
shows
you
kind
of
how
to
do
the
task,
how
to
use
the
software.
What
kinds
of
things
you
want
to
do
with
it
and
then
so.
It
gives
a
lot
of
Graphics
here.
A
A
Not
ever
you
know,
so
you
might
suggest
that
people
run
in
a
container
or
that
they
run
it
from
the
command
line,
or
they
run
it
from
like.
You
know,
they're
different
ways
that
people
want
to
start
the
software
and
run
it.
So
these
are
you
know,
including
these
in
the
documentation
is
important.
Sometimes
you
know
what
you
can
do
is
if
people
want
to
use
it
for
different
things,
you
might
have
different
tabs,
so
you
might
have
different
markdown
files
where
you
have.
A
This
is
for
people
who
want
to
do
want
to
get
into
the
program
and
work
with
the
code.
There's
another
piece
of
documentation
for
people
who
just
want
to
run
the
software
and
get
some
statistics
out
of
the
images.
In
other
words,
they
don't
care
so
much
about
you
know
the
software
details.
They
just
want
to
run
it,
and
so
those
are
things
you
can
do,
and
so
then
you
know
you
list
the
things
that
are
in
your
library.
You
talk
about
like
the
different
classes
in
this
program.
A
So
this
is
an
object,
oriented,
programming,
Paradigm.
You
know
you
have
different
classes
that
are
defined
in
the
software
and
you
can
document
those
as
well.
Then
there's
this
Health
Bridge,
which
is
a
project
Health
Bridge.
This
is
producing
the
gap
between
patients
and
doctors.
You'll
have
a
link
to
the
website.
You
can
view
the
demo.
You
can
report
bugs
right
from
the
documentation
or
request
to
feature,
and
usually
those
kinds
of
things
are
like.
A
You
have
a
lot
of
like
kind
of
main
points.
What
the
project
does.
What
is
it?
You
know
what?
How
is
it
built?
How
do
you
install
it
and
so
forth,
and
then
there's
this
third
project,
which
is
resonance,
which
is
a
content-based
music
recommendation
service
again
they
have
visiting
the
website
viewing
the
demo
reporting
bugs
requesting
features,
a
table
of
contents,
a
project
description
and
so
on.
So
this
is,
you
know
how
you
build.
You
know
open
source
documentation.
A
I
can
provide
these
links
in
this
in
the
slack.
So
that's
fine,
but
you
know
you
just
have
to
think
about
this
as
like
an
exercise
and
how
do
you
communicate
to
people
what
your
software
does,
how
to
use
it
and
so
forth?
So
you
start
with
a
readme
file,
usually
in
the
repository.
So
that's
usually
the
convention
is
to
provide
a
readme
file,
which
is
this
file
here,
and
you
know
you
can
have
other
documentation,
but
it
usually
starts
there
because
that's
how
people
expect
to
see
it.
A
A
And
you
know
it'll
help
you
people
navigate
your
Repository,
so
then
you
know
there
are
these
other
style
guides
that
people
have
made
for
documentation.
Basically,
there's
this
idea
of
minimal,
viable
documentation,
which
is
where
you
have
the
small
set
of
fresh
and
accurate
docks,
which
are
better
than
a
sprawling
loose
assembly
of
docks,
and
the
thing
you
have
to
remember
with
documentation
is
when
you
create
documentation,
someone's
always
going
to
have
to
update
it.
So
as
the
software
evolves,
people
have
to
update
the
documentation.
A
You
want
to,
you
know,
don't
be
afraid
to
delete
that
documentation.
I
know
we
have
documentation
in
there,
which
you
probably
need
some
trimming.
We
need
to
like
change
things,
but
we'll
maybe
also
need
to
get
rid
of
old
materials
and
that's
true
I
think
in
any
project.
But
you
know
when
we're
improving
upon
a
project,
especially
you
know.
The
new
version.
Isn't
there
are
things
in
the
old
version
is
not
really
relevant
to
the
new
version
and
we'll
have
to
do
that
with
pre-print.
A
That
exists
now
as
well,
and
you
know
those
sorts
of
things.
So
that's
that's
what
we
need
to
do
and
then
you
know
we
don't
want
to
like
over
obsess
about
documentation,
just
put
something
out:
that's
readable
or
start
with
something
that's
readable,
and
then
we
can,
you
know,
modify
and
edit
from
there.
So
don't
make.
A
Don't
think
you
need
to
make
this
perfect
document
right
off
the
bat
it
just
needs
to
be
like
all
the
relevant
information
needs
to
be
there,
and
then
we
can
worry
about
the
design
of
it
and
style
of
it
later
or
you
know
in
different
drafts
and
then
so
you're
telling
a
story
about
your
code.
You
know
you're
you're,
basically
taught
you're
walking
people
through
it.
A
What
it's
about
you
know
what
the
components
are
and
those
sorts
of
things:
that's
that's
documentation,
best
practices,
then.
Finally,
there's
this
idea
of
a
document
management
system.
So
a
lot
of
the
you
know.
The
whole
idea
of
documentation
comes
from
this
document
management
system
approach.
So
when
you
have
documentation
that's
more
than
like
a
some
single
readme
file,
which
is
important
for
a
lot
of
bigger
projects,
you
want
to
have
this
sort
of
document
management
system
approach.
A
Take
this
document
management
system
approach,
which
means
you
have
to
manage
the
documents
along
with
the
code
and
it's
important,
because
you
need
to
update
the
docs
and
everything.
But
it's
also
important,
because
you
know
you
want
to
keep
track
of
all
the
docs
and
make
sure
that,
like
everything
is
linked
to
everything,
is
updated
and
so
forth.
Now
GitHub
takes
care
of
a
lot
of
this.
A
This
is
just
kind
of
a
generic
stub
on
document
Management
systems,
but
you're,
basically
working
from
like
the
set
of
documents
and
you're,
also
integrating
your
code
into
it,
because
you're
going
to
have
things
like
metadata.
So
if
we
have
a
data
set
that
we
analyze
or
use
as
a
benchmark,
we
want
to
have
data
about
that
data
set.
We
want
to
integrate
the
docs
with
the
code
with
the
comments
and
the
code
and
with
the
data
sets,
and
you
know
we
want
to
have
all
these.
A
We
want
to
be
able
to
store
them
in
a
place.
That's
good!
So
that's
all
I'm
going
to
say
about
documentation,
I
think!
That's!
You
know.
I
I
presented
this
to
the
open
source
meeting
and
they
got
a
taste
of
it.
So
I
wanted
to
present
it
here
as
well
in
limited
form,
and
thank
you
to
Ankit
Grover
for
pointing
out
those
repository
examples.
A
D
A
A
C
A
A
A
D
A
A
Well,
they
might
show
the
derivation,
but
they
also
might
show
what
the
components
do
the
different
parts
of
the
equation
and
then
show
what
it's
related
to,
and
you
can
do
all
that
in
a
documentation
setting,
whereas
it's
not
really
acceptable.
Do
that
in
a
paper.
It
kind
of
detracts
from
the
paper
itself.
So.
D
A
All
right,
okay,
so
that's
that's
all
I
had
to
say
about
that.
Now.
I
wanted
to
present
a
little
bit
on
so
there's
this.
We've
worked
on
copy
cell
3D,
I,
remember
very
early
in
the
history
of
Diva
worm.
We
were
interested
in
using
this
program
for
modeling
cells
and
modeling,
so
it's
an
asian-based
modeling
platform
for
cells
itself,
it's
cell
Centric!
So
it's
you
can
model
tissues.
A
You
can
model
different
types
of
biological
systems
at
the
level
of
the
cell,
so
I
actually
visited
their
headquarters
a
long
time
ago
and
gave
a
talk
there,
but
they
you
know
it's
it's
a
software
that
I've
tried
to
get
people
to
use
several
times
in
the
group
and
each
time
we've
tried
to
use
it.
It's
been.
No
one's
really
been
able
to
advance
anything
on
it
and
they
I.
Guess
it's
not
very
easy
software
to
use
it's
very
hard.
So
they
have
these
workshops
that
are
like
kind
of
help.
A
A
You
can
use
it,
but
you
know
it's
very
hard
to
figure
out
kind
of
how
to
implement
your
model
system
of
interest,
and
so
this
Workshop
was
recently
done
where
they
had
all
these
different
topics:
Network
modeling
compartmental
cells
links
cell
crawling
conversion,
extension
tissue,
folding
Etc,
and
you
can
Implement
all
these
in
copy
cell
3D
now
Morgan,
who
is
not
here
today,
pointed
this
out
to
me
that
this
was
going
on
and
we
had
a
discussion
about
it.
So
there
yeah
there's
a
software
copy
cell
3D.
A
A
So
there's
a
lot
of
physics.
There's
a
lot
of
different
types
of
you
know:
cell
interactions,
you
can
model,
you
can
specify
a
lot
of
parameters
like
that,
and
so
you
know
some
of
the
projects
they
have
in
the
software.
They
have
different
modules
that
you
can
download.
One
of
them
is
I,
think
the
chick
embryo,
something
about
I,
think
conversion,
extension
and
chick
embryo.
A
There
are
some
other
examples
from
like
the
liver
and
other
things
or
they're,
actually
modeling
tissues,
they're
modeling,
the
cells
and
the
tissues,
and
because
it's
so
it's
like
a
agent-based
model.
It's
like
a
cellular
automata.
You
have
these
cells
that
interact
on
a
grid,
and
in
this,
in
this
case
it's
a
3D
grid
and
so
you're
able
to
get
a
lot
of
you
know
you're
able
to
do
model
capture
a
lot
of
things
in
that
way.
So
this
is
something
that
I
think
people
would
be
interested
in.
A
They
also
have
a
YouTube
channel,
but
their
YouTube,
like
the
last
videos
that
they
had
are
a
couple
years
ago.
So
I
didn't
know
if
that
was
I
mean
it's
probably
relevant
to
today,
because
I
don't
think
the
software
changes
that
much.
But
this
is
something
that
people
might
be
interested
in
again.
You
know
we
try.
We've
tried
to
do
a
couple
things
with
this,
but
it's
just
really
kind
of
hard.
There's
a
high
learning
curve
on
it.
So
it's
it's.
A
You
know
people
have
to
be
really
invested
in
it
to
really
kind
of
get
something
out
of
it,
but
I'd
like
to
do
something
with
it.
If
people
are
interested
if
anyone's
interested
in
working
with
copy
cell
3D,
if
anyone
has
any
specific
questions,
they'd
like
to
ask
with
the
software,
especially
I,
think
because
you
know
you
can
learn
the
software
and
just
play
around
with
it,
but
it's
you
need
to
really
have
some
reason
why
you're
using
it-
and
so
there
are
a
lot
of
questions
that
we
have,
that
are
really
interesting.
A
A
Those
are
kinds
of
things
you
can
answer
with
this
software
and
there
are
a
lot
of
you
know
different
plugins
and
things
you
can
do
to
really
kind
of
refine
your
model.
You
can
also
you
know,
train
it
with
a
model
or
data
from
different
sources.
Physiological
data
Etc
I'm,
not
as
familiar
with
how
to
do
that.
But
you
know
that's
part
of
the
fun
of
setting
up
these
kind
of
models,
so
it
might
be
useful.
We
might
be
able
to
do
something
with
that.
D
A
A
So
it's
hard
to
say,
but
yeah
I,
like
the
where
you
have
these
flexible
modeling
platforms,
where
you
can
apply
just
about
any
question
to
it,
but
you
have
to
figure
out
for
your
specific
question
what
you
want
to
actually
ask,
and
then
you
have
to
figure
out
how
you
want
to
represent
it.
So
it's
like
I
have
a
cell
I
have
cells
that
are
interacting.
A
A
You
know
what
I
guess
I
want
to
talk
about
these
papers
here,
so
this
is
getting
into
again.
Like
the
synthetic
biology
field,
where
people
are
creating
different
Gene
circuits
and
they're,
making
things
with
them,
so
they're
able
to
take
like
a
gene
circuit-
that's
well
known
and
they're
able
to
sort
of
design
new
circuits
because
they
know
that
they
have
a
promoter
and
they
have
a
some
coding
DNA
and
they
can
make
Gene
products.
A
And
you
know
they
can
do
things
like
make
enzymes,
but
they
can
also
design
these
things
because
they
know
the
sequence
and
make
new
things
that
we've
not
seen
before.
So
people
have
made
different
circuits
where,
like
it,
it
expresses
like
letters
and
shapes
and
things
like
that,
but
you're
kind
of
fun.
Little
toy
models
to
work
with
this
is
an
interesting
paper
where
they
talk
about
regulating
synthetic
Gene
networks
and
3D
materials.
A
A
So
so
the
abstract
reads:
combining
synthetic
biology,
Material
Science,
will
enable
more
advanced
studies
of
cellular
regulatory
processes,
in
addition
to
facilitating
therapeutic
applications
of
engineered
Gene
networks,
one
approaches
the
couple
genetic
inducers
into
biomaterials.
So
this
means
that
we
have
these
Gene
circuits
that
we're
making-
and
we
can
you
know-
they're
regulated,
they're
self-regulating
here,
inducing
it
with
certain
things
like.
So
you
have
this
inducer,
which
is
where
you
stimulate
it
and
it
induces
a
biochemical
Cascade
and
it
produces
what
you
want.
A
A
A
In
addition,
systemic
applications
of
iptg
were
used
to
induce
genetic
circuits
and
cells
encapsulated
in
material
into
materials
and
implanted
in
Vivo,
so
they
were
actually
able
to
induce
a
circuit
in
a
Cell.
They
were
able
to
put
the
cells
in
a
material
and
they
were
able
to
build
a
system
around
that.
A
The
flexibility
of
modifying
biomaterials
with
genetic
inducers
have
allowed
for
pattern.
Placement
of
these
inducers
that
can
be
used
to
generate
distinct
patterns
of
gene
expression,
so
they're,
basically
trying
to
embed
cells
and
materials,
and
you
know
they
had
they're
inducing
some
gene
expression
that
allows,
for
you
know,
maybe
to
Aid
in
pattern
formation.
A
So
one
of
one
of
the
things
we
want
to
use
this
for
are
different
types
of
controlling
diseases.
You
know
treating
cancer,
promoting
different
types
of
biofilm
degradation,
and
things
like
that.
So,
if
you
were
thinking
about
like
if
you're
using
this
for
therapy
we're
doing
plant
like
one
of
these
systems
into
a
patient
you'd
be
able
to
affect,
like
you,
could
implant
in
your
tumor
and
reduce
the
size
of
the
tumor?
Are
you
able
to
implant
it
in
like
some
sort
of
gene
therapy
context
and
help
that
process
along?
A
So
that's
where
they're
going
with
this
and
then
what
they
want
to
do?
Is
they
want
to
integrate
it
with
biomaterials?
Because
a
lot
of
these
medical
contacts
they
have
biomaterials
that
need
to
be
in
place.
You
know
to
to
serve
as
a
scaffold
or
to
you
know,
serve
as
sort
of
you
know
if
you're
implanting
some
sort
of
mechanical
device.
You
know
you
want
to
have
be
able
to
interact
with
that
without
it
without
the
body
rejecting
it.
So.
A
They
want
to
do
this.
The
flex
flexibility,
modifying
biomaterials
with
genetic
inducers
allows
for
the
pattern
placement
of
these
inducers.
Together,
these
genetically
interactive
materials
can
be
used
to
characterize
genetic
circuits
and
environments
that
more
closely
mimic
cells,
natural
3D
settings.
A
A
You
know,
and
that's
fine,
but
that
doesn't
really
have
a
lot
of
applicability
to
Medicine,
because
in
the
body
the
cells
are
in.
This
three-dimensional
space,
they're
interacting
in
three
dimensions
and
they're,
interacting
with
you
know,
materials
in
three
dimensions.
So
you
know
you're
different
tissues
are
interacting
muscle
with
bone
and
so
forth.
A
D
A
Is
a
nice
paper
where
they
talk
about
putting
this
in
into
these
biomaterials?
So
this
is
an
example
here
of
sort
of
they're
doing
this
in
a
dish
they're
able
to
they
always
do
this
validation,
where
they
look
at
the
markers
that
they're.
You
know,
they're
looking
for
the
expression
of
iptg
in
these
cells,
so
they
put
it,
but
the
circuit
and
the
cell
when
they
see
that
it's
able
to
express
this
marker
and
that's
how
they
can
validate
that
the
circuit
works,
then
they're
able
to
do
these.
A
They
use
these
electrospun
PCL
fibers
and
they
embed
them
in
these
fibers.
And
you
know
it's
able
to
sort
of
survive
in
that
kind
of
environment,
and
then
this
eye
is
where
they
have
rtpcr
expression
of
egfp,
which
means
that
they're
getting
this
upregulation
of
G
the
gfp
signal
in
these
different
contexts.
So
they
have
the
two-dimensional
context
of
the
dish
they
have
it
embedded
in
fibers.
A
B
A
A
where
they're,
where
they're,
showing
it
in
the
context
of
a
mouse,
so
this
is
where
they're
talking
about
the
therapeutic
applications.
This
is
an
example
of
local
induction
at
the
top
with
this
little
mouse
here
is
like
supposed
to
be
inside
the
mouse,
this
little
Cube
plga
sponges
with
and
within
iptg.
So
this
is
where
they
put
in
like
some
sort
of
sponge
inside
the
body
of
the
mouse
containing
Cho
cells.
So
they
have
this
sponge.
It's
embedded
with
these
cells.
It's
stably
transfected
with
this
dfp.
A
So
that
means
they
put
this
marker
into
the
cells,
and
then
this
was
implanted
subcutaneously
in
these
mice.
This
shows
this
B
B,
Prime,
C
and
D.
They
basically
show
examples
of
what
what
happened
here
so
B
is
the
bright
Field
image
of
the
sponge
removed
13
days
after
implantation.
So
that's
basically
they
put
the
sponge
in.
They
wait
13
days
they
take
it
out,
and
so
this
is
the
egfp
expression
here.
A
So
I
guess
what
that
means.
Is
it's
stable?
It's
kind
of
like
the
sponge
acts
as
like
a
source
of
these
cells.
So
that's
both
stable.
The
cells
are
able
to
maintain
themselves
in
the
material
and
they
can
migrate
out
if
they
need
to
so.
The
the
point
is
is
that
those
cells
are
still
alive
in
that
material
and
you
can
tell
with
the
markers.
A
So
the
marker
is
always
the
thing
we
want
to
see
in
the
in
the
image,
and
so
that's
just
saying
that
those
cells
are
there
and
that
they're
not
like
invaded
by
other
types
of
cells
and
E.
You
have
this
inducer,
so
this
is
the
sort
of
a
local
induction,
and
then
this
is
the
inducer,
and
so
in
E
you
have
a
systemic
induction
of
genetic
circuits
and
Vivo
sponges
and
hydrogels
each
contain
Cho
cells,
stably
transfected,
where
implanted
subcutaneously
or
in
the
injured
peritoneum
of
the
rhythmic
or
thymic
mice.
A
So
this
is
where
they
put
it
in
the
mice
again,
then
they
have
this.
This
inducer,
which
is
in
the
drinking
water
24
hours
after
the
surgery
where
they
put
this
in
they
add
this
iptg
added
in
the
drinking
water,
so
they're,
basically
using
the
iptg
to
induce
these
cells,
and
so
that's
what
they're
doing
in
this
example.
So
you
know
you
can
imagine
like
you
both
want
to
have
some
sort
of
material
that
you
can
implant
in
the
organism.
A
Have
it
expressed
or
have
it
be
stable,
but
you
may
also
want
to
induce
it
in
some
cases
with
a
drug
that's
administered
through
drinking
water,
or
something
like
that.
So
you
have
like
a
pill
or
something
you
would
take
and
induce
the
expression
of
this,
and
so
this
is
just
an
example.
This
is
where
they
validate
these
results.
So
that's
all
they're
doing
here,
so
they
kind
of
go
through
and
they
say
that
this
is
a
very
versatile
approach
for
attaching
the
genetic
inducer
to
several
different
classes
of
biomaterials.
A
This
is
important
because
different
biomaterials
promote
distinct,
morphological
and
lineage
characteristics
that
has
significant
application,
dependent,
experimental
and
therapeutic
relevance
by
including
an
inducible
circuit.
These
unique
cellular
differentiation
outcomes
be
further
enhanced
or
tightly
regulated
by
modulating
the
expression
of
Downstream
differentiation
factors,
so
you
can
hook
these
circuits
into
other
targets
within
the
the
Genome
of
the
cells
and
get
them
to
express
those
things
as
well.
A
So
there
are
a
lot
of
applications
that
one
can
use
and
they
get
into
this
part
about
using
different
types
of
logic
gates
in
these
in
these
circuits.
So
it's
not
just
that
you
can
build
something
that
expresses
a
gene
and
you
can
modify
that
if
you
want,
but
you
can
set
up
these
logic
gates,
so
you
can
set
up
an
and
gate
which
is
a
you
know
where
you
have
this
logical
operation,
where
you
take
two
things
and
they're
activated
together.
A
The
result
could
be
genetically
interactive
Matrix,
in
which
successive
cellular
differentiation
events
could
be
triggered
in
a
spatial
order
and
or
temporal
manner.
So
you
can
do
a
lot
of
things
with
respect
to
sort
of
differentiation
and
space
and
time,
and
this
is
of
course
important
for
a
lot
of
processes
that
we're
interested
in.
A
So
that's
that
paper
and
then
there's
this
other
paper
where
they're
actually
looking
at
temp
what
they
call
temporal
transcriptomics-
and
this
is
more
of
a
modeling
paper
where
they
do
this
deep
dynamical
modeling
of
Developmental
trajectories.
So
I
think
this
is
a
yeah.
This
is
a
pretty
recent
paper.
This
is
actually
from
the
bio
archive.
A
So
this
is
actually
where
they're
they're
developing
this
labeling
method,
but
they're
actually
doing
this
they're
investigating
the
temporal
transcriptomics
and
so
they'll
Define.
What
that
is,
but
why
don't
we
walk
through
the
abstract,
so
developmental
sulfate
decisions
are
Dynamic
processes
driven
by
the
complex
behavior
of
Gene
regulatory
networks.
So
we
talked
about
that
in
the
last
paper,
not
exclusively
in
development,
but
in
this
context
of
these
different
processes,
a
challenge
in
studying
these
processes
using
single
cell
genomics.
A
But
the
the
one
of
the
challenges
of
this
method
there
are
a
lot
of
single
cell
methods
out
there
for
looking
at
gene
expression
is
that
the
data
provides
only
a
static
snapshot
with
no
detail
of
Dynamics.
So
a
lot
of
times.
The
data
sets
that
we
get
are
like.
We
can
take
a
single
cell,
which
is
a
very
tiny
spatial
context
and
we
can
actually
get
a
very
rich,
very
large
amount
of
data.
A
A
You
have
different
phases
of
cell
cycle
where
the
cell
divides,
and
but
you
know,
if
you
can
only
capture
one
part
of
that
process,
you're
really
not
getting
the
full
picture,
and
so
that's
a
problem
with
a
lot
of
molecular
data
as
I
think
we
found
in
our
analyzes
that
we've
been
doing
recently.
Is
that
it's
not
like
you
know,
it
doesn't
represent
everything
going
on
in
the
organism.
A
It's
just
like
a
very
bulk
measurement
of
it.
So
it's
like!
Maybe
one
point
in
time
or
you
know
you
sometimes
you
don't
even
synchronize
your
sample
and
you're
just
getting
like
a
bunch
of
cells,
which
is
why
the
single
cell
context
is
so
useful.
You
take
a
bunch
of
cells
and
they're,
maybe
in
different
parts
of
cell
cycle,
or
they
represent
a
kind
of
a
mean
behavior
of
the
process.
And
so
it's
it's.
You
know
it's
hard
to
get
that
precise
sort
of
information
with
respect
to
time.
A
So
here
we
present
an
experimental
and
computational
methods
that
overcome
these
limitations
to
a
lot
of
dynamical
model
and
your
gene
expression
from
a
single
cell
data.
So
they
need
to
get
you
need
to
get
the
data
in
place
from
different
points
in
time.
In
order
to
get
this
expression
and
you
could
go,
you
know,
collect
the
data
from
different
points
in
time.
That's
fine,
but
you
have
to
remember
that
it's
you
know
it's!
It's
sort
of!
A
You
have
to
have
the
be
able
to
label
these
kinds
of
pro
things
that
are
involved
in
these
processes.
So
you
know
it's
not
enough
just
to
get
that
at
different
points
in
time.
Oftentimes
you
have
to
label
things
so
that
they
can
be
identified
and
normalized.
So
there's
a
lot
of
you
know,
there's
a
lot
of
work
that
needs
to
be
done
to
take
these
data
sets
and
line
them
up
and
get
make
them
comparable.
A
So
this
is
why
this
is
such
a
big
challenge.
So
this
group
developed
a
platform
called
SciFi
2.,
so
they
love
these
acronyms
and
optimize
metabolic
labeling
method
that
substantially
increases
data
quality
and
profiled.
Approximately
45
000
embryonic
stem
cells,
different
Union,
differentiating
into
multiple
neural
tube
identities.
A
So
they
have
this
system
where
they
can
look
at
cells
that
are
started,
stem
cells
and
they're,
starting
to
express
the
markers
of
neural
cells,
and
this
is
a
very
complex
process,
because
you
have
a
lot
of
markers
that
are
potentially
useful,
but
you
have
to
label
them
so
that
you
can
figure
out.
You
know
how
they're
changing
what
where
they
are,
what
cells
they're
in,
where
you
know
what
kind
points
that
they
become
relevant,
these
sorts
of
things.
A
So
they
have
these
label
this
labeling
technology
they're
using
and
it's
able
to
label
different
different
genes-
and
you
know
so.
They're
able
to
I
did
sort
of
capture
this
process
of
going
from
a
stem
cell,
which
has
its
own
sort
of
transcriptional
profile
to
some
of
these
neural
tube
identities,
which
change
their
profile
in
some
way.
And
the
question
is
we
want
to
figure
out
which
markers
or
which
genes
are
changing
and
how
so
to
recover
the
Dynamics?
A
We
develop
velvet
a
deep
learning
framework
that
extends
Beyond,
instantaneous
velocity
estimation
that
modeling
gene
expression
Dynamics.
There
are
neural
stochastic
differential
equation
systems,
so
they're
using
a
lot
of
fancy.
Math
neurostochastic
differential
equation
system
within
a
variational,
encoder
autoencoder,
so
they're
using
a
deep
learning
framework
here,
where
they're
using
it's
sort.
C
A
This
variational
autoencoder
model
and
then
they're
using
these
stochastic
differential
equations
within
them.
So
this
is
basically
how
they're
modeling
gene
expression,
Dynamics,
basically
they're,
trying
to
discover
you
know
the
different
patterns
within
gene
expression
by
putting
by
training
this
model,
and
then
it's
basically
creating
this
sort
of
a
dynamical
system.
That's
not
really
my
classification,
but
it's
I,
don't
know
exactly
have
to
look
down
in
the
paper
and
figure
out
what
they're
doing
exactly,
but
actually
I.
Think
Sichuan
can
find
this
interesting.
A
I
know
he
had
to
leave,
but
velvet
outperforms
current
velocity
tools
across
quantitative
benchmarks
and
predicts
trajectory
distributions
and
accurately
recapitulate
underlying
data
set
distributions
while
conserving
known
biology.
So
that
means
that
they're
able
to
match
what
they're
measuring
with
some
of
the
things
that
they
know
about
this
process.
So
we
know
these
static
snapshots
of
what
the
genes
are
supposed
to
be
like
we
kind
of
know
what
the
markers
are,
but
we
don't
know
how
it
gets
from
point
A
to
point
B.
So
that's
what
they're
trying
to
capture
with
this
model?
A
That's
why
we
need
to
use
all
this
all
these
deep
learning
techniques-
and
you
know
some
of
these
other
things-
is
to
capture
that
process
using
velvet
to
provide
a
dynamical
description
of
in
Vivo,
neural
patterning,
so
they're,
actually
not
just
looking
at
the
sulfate
but
they're.
Looking
at
this
process
of
neural
patterning,
because
a
lot
of
times
a
lot
of
like
in
especially
in
like
the
neural
tube,
you
know
differentiation
doesn't
just
happen
in
a
vacuum.
A
It's
happening
in
the
tissue,
so
it's
like
you're
getting
these
cells
they're
becoming
neural
tube
they're,
also
forming
a
neural
tube,
and
so
you
know
in
C
elegans.
We
don't
see
this
as
much,
but
in
like
mammals,
for
example,
or
other
vertebrates.
We
see
this
where
it's
in
the
tissue
and
it's
like
forming
the
tissue,
and
you
know
if
you've
separated
the
cells
or
the
tissue.
You
might
not
see
those
patterns
of
gene
expression.
Now,
that
being
said,
I
don't
think
they
account
for
white
geometry.
A
So
there's
this
whole
process
that
can
happen
and
that's
what
they're
trying
to
model
as
well
together,
these
experimental
and
computational
methods,
recast
single
cell
analyzes
from
descriptions
of
observed
data
distributions,
and
so
that's
where
we
have
these
static
snapshots
of
things
in
time.
So
you
might
take
the
snapshot
of
a
stem
cell,
a
snapshot
of
a
neural
precursor
cell,
a
snapshot
of
another
neural
precursor
cell.
That's
Farther,
Along
down
that
set
of
Fades,
and
so
you
get
that's,
usually
the
kind
of
data
we
have.
A
But
now
we
can
actually
move
towards
models
of
the
Dynamics
that
generate
them
describing
some
of
these
distributions.
So
we
build
a
distribution
of
what
these
responses
should
look
like
to
this
process
and
then
model
these
Dynamics
using
a
fancy
technique.
And
then
this
provides
a
new
framework
for
investigating
developmental
Gene
regulation
and
sulfate
decisions.
A
So
that's
that's
what
they're
getting
in
this
paper.
So
this
kind
of
shows
a
little
bit
of
what
they're
doing
here.
They
have
cyphate
and
siphate
too.
So
this
is
like
two
different
versions
of
this:
they
label
the
cells
they
fix
them.
They
do
all
this
other
stuff
and
then
they
actually
we
do
the
r
SCI
rna-seq,
which
is
the
Single
Cell
method
for
getting
gene
expression
data.
A
Basically,
the
chop
The
genome
up
and
fragments
they
measure
the
number
of
copies
in
the
cell,
and
then
they
map
it
to
a
draft
genome.
So
they
can
figure
out
where
it
came
from
what
the
transcript
is
and
then
they
do.
These
dimensionality
reduction
plots
just
a
low
dimensional,
visualization
I.
Guess
that's
a
PCA,
so
it's
just
a
normal
PCA.
A
That
shows
basically
these
in
different
categories.
So
you
can
take
the
all
the
different
transcripts
from
different
cells,
44
713
cells-
and
you
can
put
them
in
these
get
these
in
these
in
these
categories
from
these
principal
components.
So
you
have
mesoderm
here:
early
neural
cells,
neural
cells,
nmp
cells
and
so
forth.
So
if
this
shows
that,
like
their
method,
is
sort
of
validating
their
method,
that
they're
able
to
measure
what
they
say,
they're
able
to
measure
that
their
markers
are
consistent
with
like
different
cell
types
and
they
cluster
together.
A
So
the
idea
would
be
like,
if
I
think
it's
a
neural
cell,
it's
not
enough
just
to
say
it's.
A
neural
cell
I
need
to
show
that
it's
a
neural
cell,
that
the
markers
you
know,
other
cells,
the
same
markers
cluster
in
the
same
places
and
then
I
can
compare
it
with
other
types
of
Fates.
Other
types
of
gene
expression
profiles
and
you
get
a
different.
You
know
you
get
these
different
clusters
that
come
out
of
it.
A
A
So
they
have
this.
So
the
model
gene
expression
Dynamics.
We
focus
on
a
generalized
idea,
velocity
that
is
not
specific
to
splicing
Dynamics,
so
they
apply
a
biophysical
model
of
RNA
transcription
labeling
and
degradation.
We
Define
a
cell's
velocity
in
terms
of
three
observables
labeled
reads:
total
reads
and
labeling
time,
and
these
are
technical
details,
but
they
base,
and
then
they
have
a
free
parameter
and
then
they
have
something.
A
This
represents
the
genome-specific
degradation
rate,
so
they're
able
to
look
at
the
velocity
of
like
a
RNA
transcription
they're,
looking
at
like
the
number
of
transcripts
and
then
sort
of
their
degradation,
and
that
should
tell
them
something
about
like
you
know
how
long
a
transcript
will
be
around
if
it
degrades
and
then
there's
another
Gene
that
gets
expressed
associated
with
another
profile.
If
those
get
expressed,
those
copies
will
be
around
for
a
certain
amount
of
time
and
then
they'll
degrade,
and
the
idea,
basically,
is
that
you
have
these
pools
of
RNA.
A
They
get
produced
when
a
cell
is
in
a
certain
state.
So
a
certain
marker
say
for
a
cell
stem
cell
marker
it'll
be
expressed
at
a
certain
amount
or
produce
a
certain
number
of
copies
of
the
RNA,
and
then
those
copies
will
degrade
over
time.
But
if
the
cell
is
still
a
stem
cell,
then
it
will
retain
the
sort
of
stem
cell
fate
and
it'll
like
keep
producing
copies.
A
So
even
though
it
degrades
you
still
get
copies,
and
you
still
have
that
upregulation
of
the
stem
cell
marker,
but
on
the
other
hand,
when
the
stem
cell
transitions
into
a
neural
cell,
some
of
these
stem
cell
fat
genes
stop.
You
know,
producing
copies
of
the
RNA
so
then
that
RNA
will
be
around,
but
only
for
a
while
until
it
degrades
then
once
it
degrades.
A
A
Looking
at
this
velocity
of
the
marker
velocity
of
the
RNA
copies
to
see,
you
know
how
long
it
sticks
around
you
know,
and
eventually,
as
the
cell
State
changes
it'll,
you
know
it'll,
the
profile
will
change,
but
it's
not
instantaneous
like
we
might
think
from
something:
that's
the
static
measurements,
so
it
doesn't
just
switch
very
abruptly.
You
have
these
Dynamics,
where
you
get
things
that
are
up
regulated
and
then
they
kind
of
follow
a
number
that
decay
and
then
there's
another
marker
that
comes
up,
and
then
there
are
these
interactions
that
happen.
A
As
you
know,
one
marker
is
shutting
off
and
another
marker
is
turning
on.
That
can
be
interesting,
but
that's
what
they're
trying
to
model
so
to
get
a
little
bit
more
context
about
the
temporal
transcriptomics
paper,
I'd
like
to
do
a
little
bit
of
drawing
on
the
board
to
show
what
these
kind
of
experiments
look
like
over
time.
A
A
Different
transcripts
that
are
collected
you
get
these
fragments
of
transcripts
and
then
you
get
a
summary,
usually
transcripts
at
different
locations.
So
this
is
our
copy.
Number
I
will
just
say
that,
like
you
know,
we
have
like
three
examples
here
where
the
copy
number
is
variable
over
time.
We're.
A
Just
in
one
snapshot
of
time,
so
this
is
the
gene
expression
is
the
number
of
copies
that
you
get
for
a
certain
fragment
or
a
certain
site,
and
so
then
you
can
look
at
those
over
time
in
these
sort
of
static
summaries
that
we
usually
get
where
you
sample
cells
at
different
times
and
for
these
different
locations.
You
get
different
changes
in
how
they're
upregulated,
so
that's,
usually
the
typical
sort
of
snapshot
that
we
get.
A
A
A
So
if
we
look
at
a
dynamical
representation
here-
and
we
have
our
time
access
and
we
have
our
copy
number
on
the
left-
this
copy
number
gets
produced
at
a
certain
point
in
time
and
that
reflects
the
activity
of
that
Promoter
on
that
stretch
of
DNA
producing
RNA.
But
what
happens
if
you
stop
producing
this?
A
So,
for
example,
if
you
have
a
stem
cell
that
transitions
into
a
neural
precursor
that
stem
cell
marker
shuts
down
its
production
of
RNA,
and
so
this
marker
then
begins
to
decay
the
copy
number
Falls
over
time
as
it
stops
being
produced,
or
at
least
at
the
level
that
it
was
originally
producing
it.
So
you
get
these
time
points
where
you
get
this
reduction
in
the
amount
of
transcript
and
they
call
these
RNA
pools
they're
experiments.
You
can
do
where
you
actually
either.
A
You
know,
change
the
fade
of
the
cell,
or
sometimes
there
are
experiments
where
you
can
shut
down
the
transcriptional
Machinery
in
the
cell
and
then
all
transcription
stops.
So
you
can
watch
the
Decay
profile
of
a
certain
Gene,
so
you
can
measure
the
transcript
for
that
Gene.
In
terms
of
the
RNA.
You
can
look
at
it
as
the
cells
decaying,
so
you
shut
off
the
mechanism
and
then
you
sample
after
a
certain
amount
of
time.
Usually
it
takes
48
to
72
hours
for
Decay
to
be
complete.
A
So
after
that
time
you
end
up
with
a
negligible
amount
of
the
transcript,
and
so
that's
and
that
that
basically
happens
in
this
process.
So
you
end
up
with
a
very
low
amount
after
that
time
if
it
stops
producing,
but
of
course,
for
most
cases
your
marker
will
continue
to
produce
RNA.
Some
of
it
will
naturally
Decay,
because
there
are
copies
that
are
older,
but
they're
also
copies
that
are
younger,
so
they'll
replace
them
in
this
pool.
So
this
pool
gets
replenished,
and
so
this
time
Point
here
this
one
remains
stable
over
time.
A
So
this
profile
is
much
less
decay-like,
and
so,
if
we
actually
think
about
like
what
happens
when
these,
when,
like
one
marker
shuts
off
and
another
turns
on,
we
can
represent
one
in
circles
and
one
in
line.
So
this
is
the
stem
cell
marker
say
it
gets
shut
off
because
there's
a
transition
to
a
neural
fate.
A
A
The
neural
marker
is
subject
to
the
same.
Dynamics
is
what
we
saw
in
the
last
example
where
you
have
production
of
the
transcript,
but
you
also
have
Decay
natural
Decay
and
eventually
you
end
up
with
this
profile,
so
you
transition
from
one
marker
to
another
marker.
You
have
this
transition
of
state,
that's
how
they're
measuring
that.
So
there
are
other
ways
you
can
measure
the
transition
of
State.
You
can
look
at
the
morphology
of
the
cell,
but
this
is
the
transcriptomic
profile.
A
D
D
A
A
Yeah
well
some
people
criticize
it
and
say
that,
like
those
maps
that
I
showed
you
of
the
PCA
and
of
their
other
types
of
maps,
you
can
build
like
umap
and
and
other
things
that
they
criticize
them
for
being
kind
of.
Like
can
you
distinguish
these
things
from
random
patterns
because
we
kind
of
think
well,
we
throw
these
data
into
a
dimensionality
reduction.
Algorithm
I
get
a
pretty
picture
and.
A
So
that's
like
something
to
keep
in
mind
with
a
lot
of
the
data
analysis
that
you
might
see
that
it's
you
know
it's
really.
Our
methods
are
really
early
in
the
sort
of
scope
of
things,
especially
when
the
technology
moves
really
fast.
You
get
you
say
well,
we
can
use
this
method.
This
is
a
good
method
for
doing
this
and,
of
course,
you
know
you
want
to
really
have
multiple
methods
for
analyzing
these,
and
maybe
some
tried
and
true
ones,
but
there
isn't
really
anything
like
that.
Yet
so.
A
D
D
A
A
This
title
was
inferring
cell
Junction
tension
and
pressure
from
cell
geometry.
So
this
is
inference
by
the
experimenter
and
they're
trying
to
figure
out
how
to
measure
these
variables
in
they're.
Getting
this
information
from
cell
geometry,
so
they
abstract
rates
of
recognizing
The,
crucial
role
of
mechanical
regulation
and
forces
and
tissue
development
and
homeostasis
has
started
demand
for
insight
to
measurement
of
forces
and
stresses
so.
A
This
approach
is
non-destructive
and
fast
and
statistically
validated
based
on
comparisons
with
other
techniques.
However,
it's
qualitative
and
quantitative
limitations
in
theory,
as
well
as
in
practice,
should
be
examined
with
care.
This
is
a
primer
that
they've
designed
for
people
in
developmental
biology.
To
summarize,
the
underlying
principles
and
assumptions
behind
stress,
inference,
discussing
its
validity
criteria
and
providing
guidance
to
help
beginners
make
the
appropriate
choice
of
its
variance.
A
We
did
extend
our
discussion
from
two-dimensional
stress
inference
to
three-dimensional
so
we
talked
about
three-dimensional
tissue
engineering
earlier
in
the
meeting.
This
is
where
we
go
from
the
two-dimensional
case,
which
is
a
like
a
culture
dish
to
three-dimensional
case
which
is
like
the
actual
tissue,
and
so
this
is.
These
are
two
very
different
environments.
A
A
I,
don't
know
if
Engineers
I
guess
there
may
be
more
receptive
to
the
idea
of
computational
modeling
in
general,
but
this
is
a
you
know,
one
of
these
things
about
the
audience,
as
I
said
before
they
they're
asking
different
questions,
they're
trying
to
understand
things
for
different
ends,
and
so
it
kind
of
means
that
you
have
to.
Sometimes
you
don't
run
into
certain
methods
that
other
fields
are
sort
of
come
as
second
nature
to
another
field.
A
So
so
the
role
of
mechanical
interactions
during
morphogenesis
goes
way
back
to
you
know
the
turn
of
the
last
century.
Darcy
Thompson
of
course
mentions
it
in
his
work
on
the
mathematics
of
phenotype
and
so
forces
are
generated
by
molecular
Motors,
notably
by
actin
myosin
networks
and
they're
transmitted
by
a
cell
cytoskeletal
elements
through
cell
cell
adhesive
complexes.
A
So
we've
talked
about
these
networks
before
they're,
within
the
cell
they're,
in
the
cell
membrane,
these
actin
myosin
networks,
which
act
similar
to
muscle
in
a
lot
of
ways,
they're
flexible
and
they
produce
forces
and
they're
able
to
they
have
to
interact
with
other
cells.
So
the
cell
can
change
its
shape,
but
it
can
also
generate
forces
for
migration
or
for
sort
of
maintaining
its
own
space.
You
know
whatever
it's
used
for
they're.
These
adhesive
complexes
that
mediate
the
connections
between
cells
and
those
are
important
for
a
number
of
reasons.
A
A
A
To
things
like
Integrity,
because
you
generate
these
Dynamic
situations
where
you
have
Dynamic
tension,
stresses,
are
part
of
that
forces
and
stresses
contribute
to
the
determination
of
static
cell
shapes
and
packing,
as
well
as
Dynamic
changes
in
cell
size,
shape,
number
and
position
and
gene
expression.
So
we
have
all
these
other
changes.
You
have
cell
size.
So
as
this
tissue
grows
in
size,
the
forces
change
and
everything
else
you
have
shape.
Of
course,
if
it's
packed
in
a
certain
way
there,
it's
a
sphere
versus
a
sheet
number
and.
A
In
gene
expression,
of
course,
gene
expression
is
something
we've
talked
about
as
affecting
sort
of
the
state
of
the
cell,
but
sometimes
gene
expression
can
be
moderated
in
different
cells
by
external
stimuli
like
the
chemical
microenvironment
or
something
like
that,
and
so
all
these
ultimately
contribute
to
tissue
morphogenesis.
So.
A
A
You
can
use
visual
sensors,
such
as
fret
sensors,
which
use
proteins
to
sort
of
as
a
Target
liquid
droplets
and
elastic
beads
in
non-mechanical
observation
techniques.
So
these
are
just
ways
you
can
do
imaging
that.
Allow
you
to
observe
what
you
want
to
observe,
and
so
these
are
all
ways
you
can
measure
these
stresses
and
and
forces.
A
A
Stressor
inference
is
a
very
promising
approach
to
link
mechanics
and
developmental
processes,
because
in
stress
inference
you
need
to
make
the
minimal
measurements
sometimes
they're,
non-invasive,
sometimes
they're
non-interfering
with
the
system,
and
then
you
just
get
the
minimal
measurements
you
plug
them
into
the
inferential
engine,
which
is
usually
a
simulation
or
some
sort
of
computational
technique,
and
you
can
model
the
result.
So
that's
basically
the
idea
here.
So
it's
it's
helpful.
A
It's
helpful
for
biologists,
because
it
doesn't
interfere
with
the
system
and
interferes
and
minimally,
and
especially
if
it's
in
Viva,
in
vitro
or
in
Vivo,
you
know
it
can
be
hard
to
actually
get
in
and
actually
do
manipulations.
So
this
is
always
a
good
way
to
go.
So
there's
this
glossary
of
different
terms.
They
actually
use
Bayesian
methods.
They
mentioned
Bayesian
methods
here
it's
interesting,
and
so
they
actually
talk
about
using
stress,
inference
and
study
development.
A
This
was
implemented,
though,
more
formally,
the
first
time
by
Stein
and
Gordon
in
1982..
So
this
is
a
froth
made
of
soap
bubbles.
This
is
a
well-known
example
of
a
pattern
with
equal
tensions.
So
you
have
this
fraud
of
soap
bubbles.
It's.
It
forms
a
pattern.
The
pattern
has
equal
tensions,
analyzing
a
five
bubbles,
cluster
Stein
and
Gordon
check
the
tension
values
they
inferred
were
equal
with
within
their
position.
A
It's
validating
their
approach
using
this
approach.
They
further
analyze
the
patch
of
11
cells
and
The
Superficial
cell
layer
of
a
frog
Gastro,
and
revealed
significant
differences
among
cell
Junction
tensions.
So
this
is
where
they
have
sort
of
a
soap
level
model
where
they
have
equal
tensions
that
form
a
pattern.
A
They
were
able
to
infer
that
these
were
equal,
certainly
were
able
to
validate
that
approach,
using
the
soap
bubbles,
and
then
they
went
to
a
model
of
the
Frog
The
Superficial,
cellular,
frog,
gastrula,
and
they
were
able
to
measure
or
or
see
that
there
were
significant
differences
among
cell
Junction
tensions,
since
this
example
using
a
sort
of
a
control
system
or
a
sort
of
a
at
rest
system,
a
random
system
or
whatever.
We
want
to
call
it
and
then
using
something
that
has
some
structure.
A
C
A
Biological,
the
one
that
is,
you
know
the
one:
that's
not
biological
is
uniform.
Now,
what's
the
effect
of
the
biology
or
the
structure
in
the
biology
in
the
2000s,
with
progress
in
computers
line,
Imaging
and
finite
element,
modeling,
broadland
and
co-workers,
we
visited
this
principle
to
infer
attentions
from
cell-shaped
images.
A
So
they're
able
to
use
these
different
features
of
the
image
to
infer
these
differences,
and
some
of
the
ratios
that
you
might
want
to
know
like.
If
you
want
to
look
at
different
parts
of
the
image
and
say
that
those
visual
features
change,
you
can
infer
that
those
are
due
to
these
different
effects
of
tensions
and
forces
and
pressures.
So
this
is
something
that
you
know
we
can
basically
take
from
an
image,
so
we
can
actually
infer
it
stress,
because
we
know
that
the
laws
of
physics
are,
you
know,
you
know
pretty
predictive.
A
We
can
just
use
equations
to
model
some
of
these
changes
and
assume
that
the
material
changes
in
a
certain
way
when
exposed
to
these
sorts
of
physical
parameters-
and
so
this
is
a
stress-
inference
basically
is
represented
by
its
current
variants.
Discussed
in
the
primer
were
first
introduced
by
ciao
in
2012..
A
It's
been
extensively
used
in
two
Dimensions
to
obtain
mechanical
information,
mostly
in
planar
epithelial
tissues,
meaning
that
it's
this
flat
surface
of
epithelial
tissues.
So
you
know
really.
You
need
to
have
like
The
Benchmark
for
this
works
best
when
you
have
a
way
to
say
this
is
what
the
profile
should
look
like
in
a
control
condition
where
it's
very
simple
and
then
we
can
make
these
inferences
and
they
should
fit
the
The
observed
value
should
match
the
predicted
values.
That's
how
you
do.
A
A
So
this
is
the
variable
to
match
this
with
just
parts
of
one
development
as
well
as
embryo
Ventra,
referral
formation,
germ
band
extension
or
yeah
turn
band
extension
in
German
retraction
stress
inference
has
also
been
applied
to
study
avian
development
revealing
tissue
scale,
pressure
gradients
in
the
quail,
primitive
ectoderm
and
mechanical
heterogeneity
associated
with
hair
cell
differentiation
during
avian
cochlea
formation.
So
this
is
again
this
idea
that
in
biology
these
mechanics
are
heterogeneous,
they're
not
homogeneous
like
in
a
soap
film
or
something
else,
there's
actually
the
structure,
that's
associated
with
these
processes.
A
So
this
is
the
mechanical
heterogeneity
here
with
hair
cell
differentiation,
and
there
are
many
examples
where
you
can
actually
take
images
and
like
basically
infer
with
the
what
the
physical
parameters
should
be,
so
they
they
talk
about
a
number
of
cases
here
this.
These
are
some
nice
images
here
where
they
show
examples
of
stress,
inference
and
development.
So
a
is
a
junction
tension,
distribution
and
the
drosophila
Egg
chamber.
So
you
can
see
this
is
where
you
have
this
heterogeneity
in
the
image
and
you
can
infer
forces
from
that.
A
Tissue
we've
seen
these
kind
of
maps
before
where
you
have
this
parameter,
that
has
a
value
and
it's
going
in
a
certain
direction.
This
is
like
this
Force
map,
where
the
forces
are
applied
in
different
directions,
and
you
can
see
this
is
where
you
get
this
sort
of
be
at
the
furrow
here
and
you
get
these
different
changes
in
the
direction
of
forces,
and
you
can
tell
from
looking
at
the
image
what
direction
the
forces
are
being
applied
because
sort
of
the
sort
of
the
layout
and
shape
of
the
cells
underlying
this.
A
So
these
are
the
two
examples.
These
are
the
stresses
here
these
circles.
These
are
the
inferred
tensions
as
a
as
a
heat
map.
So
you
can
see
that
the
tensions
are
higher
on
the
outside
and
lower
on
the
inside
area.
And
again,
if
you
look
at
the
images
you
can
tell
the
shape
of
the
cells
and
the
orientation
as
to
what
the
tensions
are.
It
just
translates
right
into
the
image,
and
then
e
is
a
map
of
inferred
pressures
of
the
Primitive
ectodermal
equival
embryo.
So
this
is
where
you
have
this
map
of.
A
The
central
hypothesis
underlying
stress
inference
is
that
cell
shapes
result
from
the
balance
of
cell
cell
interactions
through
Junctions
between
two
cells
and
advertices,
where
the
Junctions
meet.
So
you
can
see
these
cells
are
all
sort
of
these
geometric
shapes
and
they
all
have
vertices
kind
of
like
you
know
you
have.
A
You
know
you
have
a
corner,
let's
see
like
right
here
and
you
have
like
three
cells
meeting
at
that
corner
and
that
vert
that
corner
is
a
vertex
of
a
network.
It's
it's
a
mechanical
Network
where
forces
and
stresses
are
transmitted
across
them,
because
it's
just
kind
of
the
edge
of
the
cell
and
cell
cells
cells
are
meeting
at
that
edge,
and
so
the
edges
come
together
in
that
vertex
and
then
you
know
the
shape
of
that,
of
course,
is
varied
by
the
amount
of
forces
and
stresses
that
are
experienced
there.
A
So
you
can
see
there's
a
lot
of
heterogeneity
in
terms
of
the
sort
of
where
these
vertices
are.
With
respect
to
the
edges,
the
the
cell
shape
changes
and
you
can
see
in
the
color
coding
that
the
cell
shape
changes
as
this
value
changes
as
well.
So
you
know
there
are
a
lot
of
ways
you
can
infer
it
from
just
looking
at
the
image
it's
kind
of
like
when
you
can
look
down
on
an
area
and
look
at
the
geologic.
A
A
In
the
present,
so
this
is,
you
know,
basically
the
method.
So
again
you
don't
just
infer
things
from
the
images,
but
this
is
the
idea
behind
some
of
these
easier
sort
of
you
know,
taking
things
and
running
it
through
a
computational
model
and
getting
the
answer,
you
don't
have
to
do
a
lot
of
direct
manipulation
with
this.
A
So
this
is
an
example
here
of
some
applications
of
stress,
inference
and
living
tissues.
There's
this
paper
shoe
in
2018,
where
they
talk
about
C
elegans,
early
development,
and
this
is
about
tension
and
pressure,
inference,
sensitivity,
test
and
inference
scheme.
So
they
actually
do
this
in
C
elegans
that
they've
done
this
in
zebrafish,
Mouse,
avian
cochlea,
Quail,
primitive,
ectoderm
and
and,
of
course,
in
drosophila
stress
inference,
can't
really
give
us
the
answers
to
everything.
A
We
can
gain
information
from
stress
inference,
but
it's
not
all-encompassing.
It
doesn't
tell
us
everything
we
need
to
know
so,
there's
some
images
that
sometimes
deviate
from
mechanical
equilibrium.
So
one
of
the
main
ideas
here
is
that
there's
this
mechanical
equilibrium
and
when
the
tension
or
stresses
to
deviate
from
that,
you
can
see
the
effects
and
the
boundaries
between
the
cells
or
the
orientation
of
the
cells.
Now
you
know
there
are
different
cases
where
maybe
that
isn't
informative
or
maybe
it's
just
not.
You
know
that
useful.
A
It's
hard
to
really
kind
of
make
an
inference
from
that.
So
here's
some
images
here
that
are
sort
of
that
they've
identified,
as
maybe
not
representative,
of
where
we
want
to
go.
So
there
are
these
crenellated
edges
they're
these
s-shaped
edges.
Let's
see
the
S
there,
you
have
edges
of
the
constant
cyanic
curvature,
but
not
a
constant
value.
So
it's
not
a
circular
Arc.
It's
kind
of
a
compound
circular,
Arc
D
is
a
plant
tissue
with
a
very
irregular
cell
shape
says.
A
Plant
tissues
are
usually
like
square
or
sort
of
this
elongated
rectangle,
but
sometimes
you
have
these
irregular
shapes
in
plant
tissues.
So
you
don't
really
know
you
know
the
junction
is
wiggly
and
you
can't
really
infer
much
about
the
forces
from
that.
You
have
the
triskelion,
which
is
the
curvature
of
the
three
Junctions
that
meet
having
the
same
sign.
So
this
is
where
they're
all
sort
of
I
guess
positive,
and
so
there
are
a
number
of
things
you
can
look
at
the
Junctions
and.
A
A
The
edges
then
have
different
forces
they
can
have
or
pressures
they
can
have
a
positive
or
negative
sign
and
so
or
an
angle.
Even
and
so
then
you
know
we
can
measure
these,
but
you
know.
Sometimes
these
are
not
like
sometimes
they're,
all
the
same
sign,
sometimes
they're
negative,
so
the
idea
being
that,
like
sometimes
they
violate
the
physics.
The
second
paper
is
actually
more
research
paper
and
not
a
tutorial.
A
So
the
abstract
reads:
Symmetry
and
orderliness
of
two-dimensional.
Levitating
micro
droplet
clusters
are
Quantified
with
the
foreign
entropy
yeah,
we'll
sell
through
these
voronoi
packings,
and
you
can
see
that
they
have
these
sort
of
hexagons
that
are
packed
together
and
they
come
from
centroids
of
the
centroid.
You
know
you
go
up
from
the
centroid
and
the
edges.
Like
the
Midway
point
between
the
centroids
and
then
the
packing,
or
merges
from
that,
so
you
end
up
with
these
hexagons,
which
is
the
most
efficient
packing
here
on
each
centroid,
the
centroid
being
the
droplet.
A
A
The
time
evolution
in
both
ve
and
SCM
runs
investigate
to
compare
the
correlation
between
the
two
measures
of
orderliness.
The
Pearson
correlation
coefficient
is
used.
The
maximum
correlation
between
V
and
S
C
CSM
was
found
for
clusters
containing
the
number
of
droplets
enabling
the
formation
of
hexagons
so
droplet
cluster
Is
possessing
six-fold
symmetry.
A
So
what
they're
doing
here
is
they're
looking
at
these
droplets
they're
sort
of
making
ripples
against
one
another
as
they
kind
of
intersect
somewhere
and
they're
using
entropy,
and
this
continuous
symmetry
so
they're
saying
that
symmetry
is
important.
This
distance
between
droplets
is
important
and
with
these
measures
they
find
that
hexagons
are
the
best
fit,
so
this
is
actually
six-fold
symmetry.
So
you
have
these
six-sided
hexagons,
there's
the
six
volt
Symmetry
and
the
hexagon,
and
so
that's
the
best
fit
for
this.
A
For
what
we're
looking
at
another
case
is
the
maximum
minimum
of
the
V
and
CSM
are
not
only
well
correlated.
Moreover,
in
certain
cases,
maximum
of
the
CSM
May
correspond
to
the
minimum
of
the
vorinal
entropy
Symmetry
and
orderliness
of
2D
patterns
could
not
be
Quantified
with
a
single
mathematical
measure,
so.
A
So
there's
this
idea
of
symmetry
and
developmental
biology
as
well,
and
it's
basically
where
things
have
Pairs
and
they're
well,
ordered
or
whatever,
and
then
there's
this
asymmetry,
which
is
where
the
Symmetry
is
broken
and
there's
this
you
things
are
not
even
between
different
folds
of
symmetry,
so
you
can
have
like
in
a
bilaterally
symmetrical
organism.
You
can
have
two
sides
that
mirror
one
another
and
then,
when
that
symmetry
is
broken,
you
have
an
asymmetry
that
maybe
leads
to
a
change.
That
is,
you
know,
differentiated
tissue
or
something.
A
But
you
know
a
lot
of
times:
symmetry
is
this
sort
of
subjective
binary
feature?
You
need
to
have
a
mathematical
method
or
technique
to
understand.
You
know
what
that
looks
like,
and
so
what
they've
done
is
they've
they've
done
this.
With
this
droplets,
they've
put
the
droplets
down
they're,
fitting
it
with
a
model
of
symmetry
and
they're
able
to
get
this
sort
of
this
order.
So
what
they're
using
here
is
this
black
white
Paradigm?
This
is
the
continuous
symmetry
measure
that
they're
interested
in
such
an
approach
to
symmetry
work.
A
The
traditional
model
we
implement
this
approach
for
the
analysis
of
symmetry
of
2D
droplet
clusters,
levitating
above
heated,
water,
vapor
interfaces,
so
they're
able
to
look
at
this
large
clusters
form
hexagonally,
ordered
honeycomb
structure
similar
to
colloidal
crystals,
while
small
from
one
to
several
dozens
of
droplets
clusters
possess
special
symmetry
properties.
Small
clusters,
in
particular,
May,
demonstrate
four-fold
five-fold
and
seven-fold
symmetry,
which
is
absent
from
large
clusters
and
crystals.
A
Symmetry
properties
of
small
cluster
configurations
are
Universal,
that
is,
they
do
not
depend
on
the
size
of
the
droplets
and
the
details
of
the
interactions
between
them.
So
this
is
interesting.
This
is
not
only
this
sort
of
packing,
but
they
also
have
unique
properties
that
they
found
the
system.
A
And
so
this
is
a
typical
droplet
cluster
levitating
above
the
water
surface.
You
can
see
that
it
looks
like
a
it's
kind
of
a
sphere
with
a
bunch
of
droplets
sort
of
distributed
throughout,
but
you
want
to
use
a
vornoid
Tessellation
to
find
sort
of
the
boundaries
between
them
to
find
the
packing,
the
boronite
tessellation
or
the
diagram
of
a
2d
area
with
a
set
of
nodes
in
it
divides
the
Orient
into
a
polygonal
zone
of
cells,
with
each
cell
containing
points
that
are
closest
to
a
given
node.
A
Such
tessellation
represents
the
two-dimensional
special
case
of
the
vignor
shites
cell
tessellation.
They
cite
that
here
we
calculate
the
CSM
overnight
diagram,
a
droplet
cluster,
not
the
CSM
of
the
cluster
itself,
considering
the
theorem
stating
that
a
vornoy
cell
always
has
the
same
point.
Symmetry
group
as
the
underlying
had
a
two-dimensional
pattern,
so
they
Define
The
Continuous
symmetry
measure
here,
which
is
a
function
of
the
minimal
average
square
displacement
of
the
points
M
sub
I
of
the
shape,
those
that
have
you
have
to
be
undergone
to
acquire
the
prescribed,
G
symmetry.
A
So
basically,
these
droplets
have
to
be
displaced.
All
all
of
them,
at
the
same
time,
across
the
shape
to
undergo
they,
have
to
undergo
these
kind
of
transformations
to
acquire
a
certain
symmetry.
So
it's
basically
where
you
want
to
you
know
they
have
to
organize
themselves
so
that
the
Symmetry
can
be
achieved.
A
A
A
Is
an
example
of
the
foronoid
diagram
corresponding
to
the
droplet
cluster?
It's
basically,
where
you
find
the
sort
of
the
midpoint
between
the
two
droplet
any
two
droplets.
So
when
you
have
any
two
droplets
with
an
interface
that
forms
a
boundary,
so
there's
an
interface
between
this
droplet
and
this
droplet,
this
droplet
in
this
droplet.
This
travel
is
trouble
and
so
forth,
and
then
these
two
boundaries
kind
of
meet
that
we
can
actually
see
where
these
two
vornoy
spaces
feet.