►
From YouTube: DevoWorm (2022, Meeting 7): Assorted Topics, Different Ways to GRN, Genomes and Embryo Phenotypes
Description
Assorted topics (Brain Gyrification and Evo-devo, Computational Embeddings, Towhee open-source, Special Issue on Developmental Network Structures). Different ways to model GRNs (computational, biological, inferential). Discussion about genomic datasets -- ENCODE and model organisms. Integration of genomic datasets and 4-D embryo datasets. Attendees: Karan Lohaan and Bradly Alicea.
A
Hello,
welcome
to
the
meeting
we
haven't
had
a
recording
in
about
two
weeks,
so
I'm
giving
an
update.
Our
original
meeting
today
was
pretty
disjoint.
People
were
joining
a
different
time,
so
I
didn't
want
to
record
from
the
original
meeting
so
I'll
be
giving
an
update
on
some
things
that
are
going
on
in
the
group
and
then
at
the
end.
Karan
is
going
to
talk
a
little
bit
about
some
of
the
work
he's
been
doing
recently.
Some
really
interesting
things
with
genomics
next
generation
sequencing
and
deep
learning
techniques.
A
So
those
are
two
areas
that
are
pretty
intense
and
so
joining.
Those
together
is,
you
know,
kind
of
a
hard
problem
he's
also
going
to
talk
a
little
bit
about
the
actual
model
data
set
that
we've
talked
about
in
the
meetings
before
and
maybe
finding
some
appropriate
genomic
data
to
integrate
with
that.
A
So
I
wanted
to
talk
about
a
couple
things
that
are
going
on
in
the
group.
The
first
is
that
I
prepared
a
blog
post
for
darwin
day
and
I'll
talk
a
little
bit
about
that,
we'll
go
through
that.
The
second
is
to
go
through
the
gsoc
google
summer
of
code
projects.
A
We've
had
a
couple
of
inquiries
already
for
the
graph
neural
networks
project.
I
wanted
to
go
over
that
a
little
bit,
so
I
want
to
talk
about
a
special
journal
issue,
we're
working
on
we're
trying
to
propose.
A
I
also
want
to
talk
about
this.
We
have
a
new
partner
that
we're
going
to
be
involved
with
perhaps
our
google
summer
of
code,
it's
toki,
it's
an
open
source
project
and
krishna
katyal
who's
been
one
of
our
contributors,
he's
actually
working
with
them
now
and
he's
going.
You
know:
they've
proposed
a
project
and
they're
applying
for
our
master
organization
that
we're
applying
to
to
be
a
part
of
google
summer
of
code.
A
A
So
the
first
thing
I
want
to
talk
about
today
is
this
blog
post.
For
darwin
day
it's
on
purification
and
we
talked
about
verification-
I
think
in
the
past
two
weeks,
so
this
is
from
a
blog
room
called
synthetic
daisies,
and
I've
run
this
blog
since
the
end
of
2008
and
there's
a
lot
of
content
on
here,
like
almost
400
posts
over
that
time
and
there's
a
lot
of
content
here.
That's
I've
worked
on,
you
know,
they're,
just
blog
posts,
so
they're
actually
not
dismis,
not
to
be
dismissive,
but
you
know
they're.
A
They
can
be
pretty
long.
You
get
to
like
4
000
words,
so
there's
a
lot
of
good
content
on
this
blog,
and
so
if
people
want
to
go
through
this
blog
and
there's
a
lot
of
content
on
development
on
evolution
on,
you
know
computing
on
general
theoretical
biology,
a
lot
of
different
topics.
A
So
you
know
if
you
go
through
this
blog
and
you
find
a
post
that
your
interest
that
you
find
of
interest,
you
know,
maybe
you
want
to
expand
it
and
work
with
me
on
that.
Please
contact
me
and
let
me
know,
because
we
can
expand
this
out
into
a
paper-
there's
also
a
a
github
repository
or
a
github
site
for
synthetic
daisies,
which
is
like
a
recap
of
the
greatest
hits.
So
I
I
can
talk
about
that
later
after
I
talk
about
this
post
so
anyways.
A
This
is
a
post
for
darwin
day
and
I've
been
posting
for
darwin
day
since,
like
2014.
darwin
day
is
february
12th.
I
believe-
and
this
was
a
little
bit
late,
but
that's
okay.
So
this
is
this
title
is
called
verification
of
the
tree
of
mammals.
This
is
a
picture
of
old
darwin.
Here
you
usually
see
young,
german
and
old
german
old
darwin
has
the
beard
and
young
darwin
is
like.
You
know
younger,
and
I
guess
before
he
went
on
the
beagle
versus
afternoon
on
the
beagle.
A
So
that's
the
idea,
but
basically
it's
an
opportunity
for
people
to
communicate
different
topics
in
evolution.
So
this
year,
I'm
going
to
focus
on
the
evolutionary
origins
and
developmental
emergence
of
gerification
in
the
mammalian
brain.
So
gerification
is
when
the
neocortex
switches,
this
outer
layer
of
the
brain,
becomes
wrinkled,
and
so
it
often
becomes
wrinkled
due
to
physical
processes.
A
Well,
more
often
due
to
the
growth
of
this
outer
layer
in
terms
of
its
size
and
it
grows
and
grows,
and
it
can't
really
fit
into
the
skull
anymore
as
a
smooth
surface.
So
you
get
these
folds
and
wrinkles
what
they
call
girai
and
sulci,
which
are
these
areas
that
are
like
sort
of
still
on
the
surface
and
areas
that
are
folded,
inward
and
so
those
you
know
that
basically
changes
the
topology
of
that
sheet
and
it
has
effects
on
function.
It
has
effects
on
its
on
different
aspects
of
the.
A
If
you
look
at
the
different
layers
of
cortex,
you
see
that
you
know.
Sometimes,
if
you
have
a
developmental
disorder
like
microcephaly,
you
have
a
disruption
in
how
those
sheets
are
developed,
and
so
you
end
up
with
a
smoother
brain.
So
all
these
things
are
tied
together,
the
development
of
the
cortex
and
folding,
and
all
that
so
I
went
through
this.
I
went
through
a
bunch
of
different
data
on
this
topic,
differentiated
between
gurification
and
listencephaly,
which
is
where
brains
have
a
smooth
surface.
A
A
So
in
fact
it's
not
even
really
that
much
developed
it's
just
very
smooth
and
that
part
of
the
brain
is
smooth.
It's
the
frontal
part
of
the
brain.
So
in
vertebrates
you
have
a
three
lobe
brain.
Essentially,
you
have
the
brain
stem
and
you
have
a
middle
section
and
then
you
have
the
frontal
section.
Middle
section
is
what
in
human
brains,
we
call
the
medial
temporal
cortex,
which
is
where
we
have
things
like
the
amygdala
and
the
hippocampus
things
like
that.
A
The
the
front
part
of
the
brain
is
this:
is
this
neocortex,
which
is
six
layer,
has
six
layers
and
there's
a
very
specific
processing
that
goes
on
through
what
they
call
columns
across
these
layers?
And
so
we
have
that
in
rabbits
and
rodents-
and
this
is
what
we
call
out
group
and
phylogenetics.
A
You
see
very
small
brains,
very
small
forebrains,
with
an
eliminate
amount
of
folding.
And
then,
when
you
get
into
some
of
the
monkeys
and
apes,
then
you
start
to
see
larger
brains
where
the
cortex
is
folded,
so
I'm
adorable,
monkeys
and
gibbons
and
then
in
great
apes
and
humans.
It's
a
lot
bigger
and
I'll
have
more
folded.
And
so
this
is
all
you
know,
this
has
genetic
factors
involved.
A
This
has
morpho
morphogenetic
factors
involved,
so
you
know
cell
proliferation
and
things
like
that,
and
then
you
have
these
physical
factors,
which
are,
if
you
expand
a
sheet
in
a
space-
and
I
think
we
talked
about
this
with
crumpled
paper
a
couple
weeks
ago.
If
you
expand
a
sheet
in
size,
if
you
crumple
it
up,
you
can
restrict
that
size
down,
but
you
still
have
the
same
amount
of
space
there.
You
still
have
the
same
amount
of
surface
area.
A
So
that's
why
you
get
a
lot
of
this
folding
and
so
in
this
tree
you
see
that
there
are,
if
we
trace
these
traits,
what
they
do
in
a
phylogeny
is
usually
these
trees
are
built
on
on
discrete
traits,
and
so
you
can
actually
look.
Every
branch
represents
a
certain
number
of
traits
that
distinguish
that
branch
from
another
branch.
So
we
have
in
primates.
We
have
these
traits
here:
neocortical
enlargement,
visual
motor
and
somatosensory
system,
specializations
dorsal
outer
prefrontal
cortex.
A
There
they
don't
exist
and
an
out
group
is
just
a
comparison
from
somewhere
else
in
the
tree
of
life,
so
people
use
l
groups
to
root
their
tree
to
figure
out
which
traits
are
ancestral
which
traits
are
derived,
and
so
the
the
l
group
is
assumed
to
have
some
of
these
ancestral
traits.
They
also
have
derived
traits,
but
they
shouldn't
be
anything
similar
to
what
you
would
see
in
the
in-group
or
the
primate
group
in
this
case.
A
So
you
have
your
different
primate
groups.
You
have
another
set
of
traits
with
arthropoid
anthropoids,
brain
enlargement,
reduction
of
olfactory,
receptors
and
trichromatic
vision.
These
all
appear
at
this
branch
and
then,
as
you
go
on,
you
get
the
hominoidaea
bring
enlargement
reduction
of
olfactory
receptors
and
so
and
then
they
place
the
folding
of
the
neocortex
in
this
context.
A
Okay,
so
then
the
second
topic
is
to
talk
about.
You
know
how
purification
may
be
a
product
of
convergent
evolution,
not
just
simply
as
a
matter
of
this
sort
of
specialized
evolution,
where
there's
not
what
they
call
monophyle,
which
is
where
it's
restricted
to
a
single
lineage
or
a
single
part
of
the
tree.
You
see
cortical
enlargement
throughout
actually
mammals.
You
see
it
in
pilot
whales.
You
see
it
in
elephants,
you
see
it
in
humans.
A
See
humans
are
in
the
primate
tree
in
a
primate
part
of
the
tree
of
mammals.
Elephants
are
in
another
part,
and
whales
are
in
another
part,
so
those
expansions
of
cortex
occur
throughout
that
the
tree
of
mammals,
and
you
know
you
have
smaller
brains
also
in
between.
So
it's
what
we
call
polyphonic
or
it's
evolved
multiple
times,
maybe
from
the
same
precursors.
A
Maybe
not
so
we
you
know
we
can
look
at
different
genes.
We
can
look
at
other
conditions
that
are
exist,
but
the
idea
is
that
basically
it's
a
product
of
this
convergent
evolution
or
what
they
call
polyphenol,
okay,
so
but
there
also.
Yes,
there
are
a
lot
of
physical
constraints
that
can
be
involved
in
this.
Another
thing
that
we
look
at
in
evolution
as
we
look
at
elemetric
scaling.
A
So
in
this
case
we
have
a
bunch
of
species
that
are
all
lined
up
and
they
have
their
brains.
We
have
rats
lemurs,
wolves
and
humans,
and
we
can
look
at
those
brains
to
look
at
the
relative
size,
relative
size
of
the
brain
relative
size
of
the
cortical
sheet
and
this
metric
gi,
which
is
the
verification
index
which
marks.
A
You
know
how
much
folding
there
is
in
in
the
sheet,
and
so,
if
we
look
here,
we
see
that
rats,
lemurs,
wolves
and
humans
are
on
this
sort
of
continuum,
and
if
you
plot
all
these
brains
out
with
their
measurements,
they
form
this
scaling
factor,
which
is
where
you
compare
brain
size,
which
is
a
certain
scaling
of
the
size
versus
this
tangential
expansion,
which
is
basically
an
expansion
of
size.
And
so
you
see
that
this
verification
index
increases
with
the
expansion
of
the
the
sheet
and
with
brain
size,
and
you
get
this
characteristic.
A
Regression
function,
that's
curvilinear,
so
you
can
actually
characterize
in
terms
of
a
mathematical
formula
and
other
brains
should
fit
on
that
on
that
trajectory.
So
this
is
actually
an
interesting
point,
because
this
means
that
it's
not
necessarily
a
matter
of
this
is
this
is
a
little
bit
different
than
the
phylogenetic
analysis,
because
it's
basically
a
matter
of
the
size
of
the
brain,
the
expansion
of
that
cortical
sheet
and
you
we
see
this
increased
scarification.
A
So
it's
a
mixture
of
physical
factors
and
genes
that
are
controlling
this
process,
and
we
can
see
that
that
isn't
necessarily
you
know
part
of
any
one
lineage,
it's
part
of
the
mammalian
sort
of
developmental
milieu.
If
you
want,
so
that's
that's
basically
it.
So
we
see
in
general
that
larger
brains
also
have
a
larger
gi
value.
A
A
A
These
are
just
images
from
mri
data
sets,
so
it's
just
basically
the
anatomy
and
what's
going
on
on
the
anatomy
and
then
modeling
that
with
a
computer
and
modeling
that
with
a
physics
engine
to
be
more
specific,
so
the
physics
engine
is
modeling,
this
folding,
and
so
this
is
a
purely
physical
process,
and
so
we
find
that
like.
If
you
do
this,
you
can
actually
replicate
what
what
the
brain
looks
like
during
this
process.
A
What
this
neocortical
sheet
looks
like
you,
get
the
same
type
of
folds,
you
get
the
same
type
of
behavior,
and
so
this
is.
This
is
maybe
evidence
for
this
physical
hypothesis,
but
we
do
know
that
genes
control
this.
We
know
that
there's
a
whole
family
of
mcph
genes
or
loci
that
lead.
You
know.
If
you
have
mutations
in
those
genes,
you
have
developmental
mutations
that
affect
the
size
of
this
cortical
sheet
and
thus
the
folding.
We
know
that
mutations
in
this
family
of
loci
leads
to
microcephaly.
A
I
misspelled
that
there,
where
this,
where
this
mature
brain,
remains
small
and
likes
your
urification.
So
if
you
look
at
like
people
with
this
developmental
condition,
they
reach
adult
and
they're,
very
small
brains
and
those
brains
are
smooth.
They
have
some
mental.
You
know
they're
mentally
disabled
as
well,
so
you
know
the
the
sides
of
that
sheet.
Is
you
know
as
it
scales
up
it?
A
You
know
it.
It
has
to
scale
up
in
size
if
it
doesn't
scale
up,
the
someone's
cognitive
abilities
are
impaired.
So
this
is
some
evidence
that
genes
are
controlling
this
process
and
that
it's
not
just
a
physical
process.
A
A
There's
a
there's,
a
no
matter
what
the
size
of
your
cortical
sheet,
those
folds,
always
come
out
at
a
consistent
frequency,
meaning
perhaps
that
there
are
different
places
where
there's
a
propensity
for
folding
or
that
the
folds
just
happen,
because
the
sheet
is
of
a
certain
material
is
of
a
certain
type
that
it
just
folds
in
that
way
or
what?
If
we
go
back
to
our
example
of
a
flat
piece
of
paper,
and
we
crumple.
B
A
A
That
crumpling
is
consistent
over
a
crumple,
so
if
we
just
take
a
bunch
of
sheets
of
paper
and
crumple
them
up
and
we
measure
the
folds
and
the
full
wavelength,
I
don't
know
what
the
answer
would
be
on
that.
I
don't
think
anyone's
ever
done
that
and
that's
kind
of
one
of
those
ignoble
ignoble
experiments
that
you
could
do
where
you,
like.
Just
crumple,
a
bunch
of
pieces
of
paper,
spread
them
back
out
and
measure
the
fold
frequency.
A
But
if
you
do
that,
I
mean
I
can't
imagine
that
it
would
be
systematic
like
this,
so
that
might
suggest
that
there's
another
rule
for
genetics
here
so
yeah.
So
this-
and
this
again
shows
that
you
have
this
mammalian
ancestor.
You
have
some
transitional
form
and
then
you
have
in
extent
species.
You
have
all
this
variation
and
there
are
a
lot
of
different
things
going
on
here.
So
it's
you
know.
You
have
increases
and
decreases
in
brain
size
across
the
tree
of
mammals.
A
You
have
increases
in
decreases
in
this
verification
index
and
there
are
a
number
of
factors
that
lead
to
that.
There
there's
one
study
that
suggests
that
there's
different
differences
in
cell
proliferation
that
affect
this.
So,
for
example,
when
you
have
an
increase
in
brain
size
and
an
increase
in
purification
index
from
this
mammalian
ancestor,
which
is
like
what
we
see
in
humans,
we
have
an
increase
in
something
called
brg
and
an
increase
in
tap
where
taps
are
present.
So
what
are
brg
and
temps?
So
brg?
A
Is
this
basal
radial
radial
ganglion
precursor
cell
type,
so
these
precursor
cells
are
increased
in
their
proliferation
during
this
process.
So
from
the
mammalian
ancestor,
we
have
an
increase
in
this
certain
type
of
precursor
cell
that
proliferates
in
that
cortical
sheet
and
that
drives
the
size
increase
and
that
also
drives
folding.
A
But
we
also
have
this
these
taps
and
so
taps
are
trans
transit,
amplifying
progenitor
cells,
and
so
these
are
another
class
of
cell,
that
is
when
it's
present
they're
able
to
proliferate
quickly
and
convert
to
different
cell
types,
and
so
this
might
have
something
to
do
with
larger
brains
being
maybe
smarter
or
having,
like
you
know,
enhanced
cognitive
capacity
as
opposed
to
species
of
smaller
brains
like
rabbits,
and
so
you
know,
you
have
different
cell
types,
like
you
know,
there's
special,
very
specialized
cell
types
in
the
cortex
aside
from
the
typical
cell
types,
and
so
you
know
some
of
those
cells
might
actually
play
a
role
in,
in
other
things
that
are
going
on
these
bigger
brains.
A
So
that's
the
kind
of
thing
we
have
just
increases
and
decreases
in
those
different
factors.
We
also
have
to
account
for
changes
in
cell
cycle.
So
when
cell
cycle
decreases
in
its
frequency,
you
get
smaller
brains.
When
cell
cycle
increases
in
its
frequency,
you
get
larger
brains,
it's
just
how
many
cell
cycles
can
you
have
during
development?
A
If
you
have
fewer,
you
have
fewer
cells.
If
you
have
more,
you
have
more
cells
and
that
drives
decreases
and
increases
in
this
sheet
this
cortical
sheet.
Similarly,
you
can
increase
the
brain
size,
a
decrease
in
purification
index
in
some
cases,
and
that
also
leads
to
a
different
kind
of
brain.
So
this
brain
on
the
bottom
is
is
where
you've
increased
the
brain
size
but
decreased
the
verification
index,
and
that
brain
is
going
to
look
very
different
from
a
brain
like
a
human
brain
with
a
lot
of
verification.
A
So
this
is,
I
guess
this
is
still
an
open
question.
I
have
some
references
here.
If
you're
interested
you
can
explore
this
further,
you
can
get
into
some
of
the
details.
It's
a
really
fascinating
area.
It's
a
tough
area
to
kind
of
understand,
but
you
know
once
you
get
into
it.
I
think
you'll
see
that
there
are
a
couple
of
leading
competing
hypotheses
here,
a
couple
leading
hypotheses-
and
you
know-
maybe
we
can
talk
about
this
later
on.
A
It
definitely
has
a
lot
of
relevance
to
development
and
developmental
differentiation
that
we
talk
about
often
in
this
group.
A
So
next
thing
I
want
to
talk
about
is
the
gsoc
for
this
year,
and
this
is
our
one
of
two
of
our
projects.
So
we
have
this
graph
neural
networks
project
and
we
also
have
the
project
of
axolotl
building
that
axolotl
atlas,
which
is
called
the
digital
microsphere.
So
it's
like.
We
have
this
specialized
microscope
that
takes
images
from
nine
different
angles
and
we're
going
to
have
data
from
these
different
angles
and
they
all
take
just
different
sides
of
a
an
embryo.
A
Some
sort
of
visualization,
but
this
one
is,
I
think
this
is
the
one
I've
had
inquiries
about.
This
is
the
graph
neural
networks
as
developmental
networks,
so
this
maybe
takes
a
little
bit
of
explanation
as
to
what
we're
looking
for
here.
A
So
here
it
goes.
We
have
we've
kind
of
played
around
with
networks.
In
this
group
we've
played
around
with
neural
connectives,
we've
played
around
with
gene
regulatory
networks,
what
we
call
interactive
networks
and
anatomical
networks.
More
broadly,
so
we've
thought
about
all
these,
like
processes
going
on
in
development,
and
you
know
we're
thinking
about
networks
where
a
bunch
of
different
components
are
connected
together,
and
so
these
could
be
gene
regulatory.
These
could
be
cells.
These
could
be
at
the
level
of
you
know,
tissues.
A
There
are
all
sorts
of
ways
you
can
look
at
these
networks.
We
get
the
data
for
these
networks
from
microscopy
images.
We
get
them
from
experiments,
do
genomic
experiments
where
people
are
looking
at
gene
expression,
we're
looking
at
microscopy
where
people
are
looking
at
cells
and
tissues
in
their
spatial
organization.
So
we
have
all
this
data
and
we
have
these
network
models,
and
so
there's
one
other
thing
we
do
in
this
group,
which
is
look
at
deep
learning
and
machine
learning.
A
We
also
look
at
modeling,
but
that's
not
part
of
this
and
the
deep
learning
and
machine
learning
work
have
been
largely
done
in
the
service
of
t
extracting
data
from
these
data
sources
and
and
sort
of
giving
us
a
better
sort
of
way
to
segment
those
data
and
define
salient
features
in
those
data.
A
They
should
resemble,
instead
of
resembling,
like
a
deep
alerting
network
where
it's
sort
of
this
connectionist
monster,
where
you
have
like
different
layers
and
fully
connected,
and
then
you
sort
it
out
through
waiting
the
the
the
different
connections
between
the
nodes,
you're
actually
building
these
embeddings
that
maybe
look
more
like
one
of
these
networks
that
we've
been
working
on
these
developmental
networks.
So
they're
not
like
typical.
A
You
know
all
the
all
connected
networks,
they
have
this
structure,
and
so
the
structure
is
actually
quite
important.
It
tells
us
a
lot
of
things
about.
You
know
centrality
about
modularity
about
things
that
you
know
we
talked
about
in
different
meetings
and
and
actually
you
care
about
when
you
want
to
understand
development,
so
we're
actually
interested
in
building
some
of
these
graph
embeddings.
So
the
first
part
of
this
project
will
be
defining
graph
embeddings,
something
that's
computable,
but
also
something
that
resembles
development.
A
So
we
want
to
use
you
know
biological
rules
and
we
want
to
use
ideas
from
like
differentiation
and
differentiation
trees,
which
we've
talked
about
in
the
group
and
other
other
sources
to
generate
different
types
of
say,
like
you
know,
things
like
movement,
output
or
developmental
output,
so
you
know,
maybe
even
like
a
model
of
cell
proliferation
would
be
acceptable.
A
The
goal,
though,
is
to
create
a
library
of
these
graph
neural
network
embeddings
that
we
can
use
to
simulate
using
a
set
of
tools,
so
we
can
use
pi
torch,
we
can
use
tensorflow
and
then
I
think,
oh
well,
next
I'll
talk
about
tohi
and
then
you
can
see
that
you
can
also
use
something
like
that
as
well.
So
you
you
know
you
can
use
a
tool
like
that.
You
can
create
these
embeddings.
You
can
then
simulate
those
graph
neural
networks
using
some
algorithm,
and
then
you
end
up
with
this
really
nice
model.
A
You
know
we
work
with
simulation
models
in
the
group
we
work
with
cellular
automata
and
other
types
of
simulation,
so
those
are
pure
simulations.
Sometimes
they
don't
connect
very
well
to
the
actual
phenomena
or
the
data
like,
for
example,
with
a
cellular
automata.
We
can
simulate
pattern
pattern
formation,
but
it
may
or
may
not
resemble
what's
actually
going
on
in
the
embryo.
A
It
just
looks
like
it,
and
so
we
want
to
actually
have
something
that
is
a
little
bit
more
connected
to
the
data
that
looks
like
it's
something:
that's
developmental,
it
could
form
patterns,
it
could
move,
it
could
proliferate,
it
could
divide.
You
know
the
visualizations
are
created
from
the
from
the
simulation
and
the
same.
Those
visualizations
have
different
properties.
They
have
to
look
sort
of
like
they're
developing,
but
that's
not
enough,
because
we
want
statistics
that
go
along
with
it
that
show
that
they
have
these
sort
of
properties
of
assist
underlying
system.
A
That's
doing
these
sorts
of
developmental
things.
So
then
you
know
you
will
join
our
group
you'll
attend
our
weekly
meetings.
You'll
give
us
reports
on
your
progress
and
so
forth.
We
have
other.
You
know
things
that
we've
worked
on
in
the
past,
so
as
the
successful
candidate
for
this
position
will
contribute
to
the
diva
learn
platform,
which
is
this
platform,
where
we're
doing
a
lot
of
pre-trained
deep
learning,
divorm
ai,
where
we
have
other
types
of
algorithms
that
one
might
use
they're,
not
pre-trained
models
but
they're
segmentation
models
and
other
things.
A
That
might
you
know
we
will
add
to
that
library.
You'll
be
a
part
of
that
library,
but
you
can
also
you
know
utilize.
You
can
also
make
connections
between
these
graph
neural
networks
and
some
of
those
earlier
models.
These
are
your
standard.
Some
of
these
are
standard.
You
know,
machine
learning
and
deep
learning
models.
A
Some
are
pre-trained
models.
The
diva
learn
platform
is
a
pre-trained
model
for
c
elegans
embryogenesis.
So
there
are
a
lot
of
opportunities
here.
We
can
also
tie
this
into
neural
cellular
automata,
which
is
another
thing
that
we've
been
working
on
in
the
group.
So
these
are.
These
are
a
lot
of.
There
are
a
lot
of
possibilities.
A
We
would
like
for
you
to
have
some
basic
knowledge
of
biology
and
complex
networks.
Theory.
Now,
knowing
that
not
everyone
has
that
you
know,
that's
not
a
requirement.
You
can
learn
that
in
the
group,
but
you
definitely
need
to
be
able
to
program
these
kind
of
models
and
build.
A
A
He
did
a
lot
of
stuff
he's
been
less
active
recently,
but
he's
working
now
with
tohi,
and
this
is
a
group
that
is
trying
to
build
machine
learning
pipelines,
for
you
know,
people
who
want
to
automate
their
machine
learning
process
and
so
they're
looking
at
this
x2vec,
which
is
an
approach
that
originally
comes
out
of
text-to-speech
or
linguistic
generation,
so
it's
generating
language,
it's
a
language
model
type
thing,
but
there's
this
idea
of
anything
devex,
so
tohi
allows
you
to
automate
this
process
of
taking
anything
and
turning
it
into
a
vectorization
which
then
allows
you.
A
Build
a
machine
learning
model
from
it,
so
they're
doing
this
they're
helping
you
encode
unstructured
data
into
embeddings,
and
so
this
would
actually
be
quite
useful
for
the
graph
neural
networks
project.
If
one
decides
to
do
that,
there's
documentation
here,
there's
a
list
of
pipelines.
So
this
is
there
on
github.
You
can
start
the
repository
here.
A
We
have
out
of
the
box
pipelines.
So
there
are
a
number
of
out
of
the
box
pipelines.
One
can
use
for
different
things.
They're,
not
all
you
know
one
area
they're,
not
all
like
nlp
or
they're,
not
all
like
image
processing.
They
have
different
things
going
on
here,
so
yeah.
These
are.
These
are
a
lot
of
embeddings
or
pipelines
to
choose
from
there's
documentation.
A
A
You
can
build
a
top
pytorch
scikit
learn
opencv
libraries,
so
this
works
with
other
libraries
that
you
might
want
to
use
and
so
yeah.
This
is
great.
So
this
is
something
that
they
they're
they're
applying
for
g-shock
students
as
well.
A
Through
our
organizations,
our
organization
is
the
orthogonal
web,
and
so,
if
you're
familiar
with
that,
we
have
a
list
of
projects
through
there,
and
you
know
it
includes
this
year-
we're
applying
to
be
an
organization
including
projects
from
orthogonal
lab,
which
is
a
ai
ethics
project,
we're
applying
through
diva
worm,
which
has
the
two
deborah
projects
we're
applying
to
rockwire
community,
which
is
another
group
that
does
a
lot
of
smart
communities
and
mobile
app
development.
A
There
are
two
projects
there
and
then
there's
this
toy
project
so
definitely
check
that
out
if
you're
interested
the
difference
between
the
toki
project
and
the
gnns
project
is
that
the
tohi
project
is
more
focused
on
these
pipelines,
building
pipelines
and
quality
control.
Things
like
that,
whereas
the
gnns
project
is
diva,
worm,
oriented,
you're,
going
to
be
working
with
biological
data,
but
you
might
use
things
like
tohi
or
some
other
tool
set
or
some
other
libraries
to
accomplish
that.
A
So
that's
that's
all
we
have
there.
I
wanted
to
talk
about
this
blog
post
because
it
links
into
what
what
what
they're
doing
at
tohi,
and
so
this
is
transform
anything
into
a
vector.
So
I
talked
about
anything
to
vec.
What
does
that
mean?
There's
this
approach
called
entity
to
back
or
anything
to
vac.
This
is
using
cooperative
learning
approaches
to
generate
entity
vectors,
and
so
this
is
by
this
author
here
works
at
ibm,
watson.
A
A
The
challenge
comes
from
trying
to
represent
these
entities
in
a
meaningful
and
compact
way
to
feed
them
into
a
machine
learning
classifier.
So
that's
what
they're
doing
when
they
create
these
embeddings
they're,
creating
basically,
data
structures
that
allow
you
to
represent
different
entities.
So
in
developmental
biology
we
have
a
lot
of
entities
at
the
cellular
level
at
the
molecular
level.
A
We're
probably
interested
more
on
the
cellular
level
in
our
in
these
gnns
in
these
embeddings
that
we're
asking
for
in
the
gsoc
project,
but
also
you
know,
combining
the
two
scales
genetic
genomics
and
cell
biology
is
also
you
know.
It's
just
a
matter
of
finding
the
right
data
that
matches
up
and
the
mining
data
is,
is
very
challenging.
So
you
know,
maybe
maybe
we
don't
go
there.
First,
maybe.
B
A
Save
that
for
later,
but
so
this
this
goes
back
to
natural
language
processing,
there's
a
there
was
a
team
of
researchers
at
google
that
created
something
called
word
defect
and
word
devec,
and
this
is
the
preprint
here,
which
is
a
technique
that
represents
words
as
continuous
vectors
called
embeddings.
So
what
they
did
in
word
divec
is
they
created
words
that
were
continuous
vectors,
these
embeddings
and
then
they
would
embed
them
in
this
dimensional
space
and
they'd
find
the
distance
between
these
words.
A
So,
for
example,
you
have
a
mandel
woman
akin
to
a
queen
walking
to
walk
and
swimming
to
swam.
So
you
have
these
distinctions
between
male
and
female.
You
have
these
distinctions
between
verb
tense
and
then
you
have
this
distance.
Basically,
in
this,
in
this
three-dimensional
space
of
you
know
you,
you
can't
do
that.
You
can
do
it.
I
guess
by
looking
at
its
levenshtein
distance,
which
is
the
distance
between
the
characters
and
the
word
so
walking
and
walked.
A
You
have
a
certain
number
of
letters
that
are
different
in
those
two,
and
so
that
gives
you
a
distance,
but
there
are
other
ways
to
calculate
distance
based
on
context
based
on
other
things,
and
so
I
think
they
use
11
stream
distance
in
the
in
the
word
defect,
which
is
just
like
a
simple:
it's
not
a
hamming
distance.
It's
like
a
hamming
distance,
it's
a
it's
a
specific
type
of
sort
of
hamming
or
or
diff
type
distance,
and
then
you
also
have
correspondences.
A
So
you
correspond
a
country
to
its
capital
and
you
can
draw
a
line
between
them.
So
these
are.
These
are
how
embeddings
work
in
the
linguistic
example.
You
have
all
these
things
that
are
going
on.
You
figure
out
which
terms
or
entities
are
closer
to
other
entities
and
you
and
you
build
a
network
from
that
or
you.
They
just
show
the
linear
relationship,
but
you
can
actually
build
networks
from
this.
So
it's
kind
of
an
interesting
way
to
approach
it.
A
So
word
devec
has
been
a
staple
of
nlp
or
natural
language
processing,
but
now
we
want
to
take
that
outside
of
the
linguistic
context
and
make
it
available
to
other
types
of
data.
So
they
talk
about
business
in
this
post.
I
thought
it
was
nice
because
they
kind
of
talk
about
it.
Very
simply,
but
you
know
we're
kind
of
interested
in
in
biology
and
biology
is
quite
a
bit
different
because
you
don't
really
have
I
mean
you
have
like
you.
Don't
really
have
words,
you
need
words,
but
you
want.
A
Your
objects
of
analysis
are
maybe
more
complex
than
words
or
like
customer
databases,
they're
often
like
cells
or
identities,
or
you
know,
gene
sequences,
but
they
don't
exist
in
isolation.
You
have
to
express
them,
so
there
are
all
sorts
of
like
different
entities
and
associations
we
can
make,
and
the
key
is
to
get
in
there's
great
complexity.
There,
like
there
isn't
language.
The
key
is
to
find
you
know
a
subset
of
those
and
build
these
representations.
A
So
here's
an
embedding
for
words.
You
have
source
text,
you
have
training
samples,
you
have.
You
know
they
kind
of
go
through
this,
where
they
take
each
word
in
this
sentence
and
they
train.
You
know
an
association
of
brown
with
different
words,
so
you
know
how
would
you
associate
wrong
with
in
this
sentence?
A
So
this
is
the
way
you
build
these
kinds
of
things
you
build
it
from
a
corpus
of
data,
so
in
this
case
they
had
100
billion
words
from
google
news
and
building
word
devec
in
the
case
of
our
our
data
sets
are
maybe
smaller
than
that,
the
ones
at
least
we're
going
to
work
with
in
the
project,
but
there's
a
lot
of
biological
data
and
potentially
even
more
than
this,
depending
on
how
you
define
the
entities
and-
and
you
know
how
much
data
you
can
actually
transform
into
something
that
you
can
get
associations
from.
A
So
that's
why
I'm
connecting
with
all
this
other
work
we've
been
doing
on
extracting
data
sets
from
microscopy
data.
You
know,
they're
possible,
you
know
genomic
information.
We
can
use
so
there's
a
lot
of
potential
and
there
seems
to
be
a
lot
of
you
know
these
seem
to
be
pretty
large
models.
These
embeddings,
but
you
know
our
embeddings-
don't
have
to
be
that
big.
They
can
be
rather
small,
and
you
know
we
don't
know
what
the
performance
is
going
to
look
like.
A
A
You
have
an
architecture,
these
embeddings
you're,
a
person
embedding
a
single,
dense
layer
and
an
output.
So
you
have
this
person
embedding
which
is
like
you
give
it
a
person.
You
have
a
database
of
people
and
you
have
it
try
to
understand
like
so.
This
is
a
picture
of
barack
obama,
which
embedding
fits
barack
obama,
and
you
have
the
labels
here,
but
the
machine
doesn't
know
what
these
labels
are.
A
Just
knows
that
they're
different
embeddings,
and
so
it
just
tries
to
match
up
this
image
with
the
embedding
or
like
the
residual
network
that
comes
from
training
and
on
images
of
barack
obama
or
hillary
clinton
or
bill
gates
or
whatever,
and
so
you
would
just
you'd
be
able
to
differentiate
between
those
images
by
its
statistical
properties,
and
so
that's
what
you're
trying
to
do
you're
trying
to
take
any
one
sample,
fit
it
into
an
embedding
and
then
see
which
embedding
matches
the
most
closely
and
then
that's
the
embedding.
A
You
would
use
that's
how
it
recognizes
these
things,
so
that's
the
dense
layer
and
then
the
output,
so
they're
different
methods
in
in
linguistics.
They
use
entity
to
vec
these
pre-trained
word.
Defects.
Bag
of
words
is
most
frequent.
These
last
two
are
older
techniques
and
they're,
not
as
good.
These
are
the
newer
techniques
and
they're
much
better,
but
you
know
these
are,
of
course,
reliant
on
a
lot
of
data.
A
So
this
is
something
to
keep
in
mind,
so
you
can
do
this
with
the
yelp
data
set
as
well,
where
you
look
at
different
recommendations
and
you
try
to
fit
it
to
a
you
know.
You
look
at
different
retail
establishments
and
then
you
fit
them
to
this
output.
So
yeah
you
have
to
it's
another
classification
task.
A
A
Where
you're
looking
at
physical
forces,
they
could
also
be
other
networks
that
are,
you
know,
like
networks
where
you
have
divergent
function.
Things
like
that.
So
there
are
a
lot
of
different.
I
have
to
think
I
haven't
written
up
the
description
of
what
this
is
going
to
look
like,
but
this
is
what
I'm
thinking,
and
so
I
think
this
will
be
a
nice
special
issue.
Hopefully,
people
want
to
help
with
this.
You're
definitely
welcome
to
join
in.
A
Let
me
know-
and
this
is
something
longer
term-
I
don't
know-
there's
a
call
for
a
special
issue
from
this
journal,
but
we
might
not
go
with
this
journal.
We
might
go
somewhere
else
with
this,
but
this
is
definitely
a
nice
topic
and
we'll
have
to
work
on
that.
A
So
last
point
I
want
to
make
today
is:
I
want
to
talk
about
readings
for
today
and
we
have
a
lot
of
things
in
our
queue
here.
So
why
don't?
I
talk
about
gene
regulatory
networks
in
different
ways,
and
so
what
I
mean
by
this
is
that
there
are
different
ways
to
represent
a
gene
regulatory
network.
A
You
know
people
are
modeling
these
things
in
terms
of
what
they're
observing
from
the
natural
system.
So
they
look
at
things
that
they
think
are
important
and
they
generate
a
model.
It
could
be
a
computational
model
or
a
phenomenological
model
where
they
look
just
what's
going
on
and
then
they
see
how
it
predicts
the
data.
A
A
A
This
just
kind
of
goes
over
what
gene
net
regulatory
networks
and
how
they've
sort
of
evolved
in
in
the
tree
of
life,
but
also
how
they,
how
play
a
role
in
development,
and
so
there
are
a
lot
of
things
we
you
know,
consider
about
like
genome
size
and
different
modular
structures
for
these
networks.
So
in
this
case
we
have
a
network
where
you
have
two
different
modules
that
are
being
turned
on
and
off.
A
You
have
different
features,
so
these
modules
consist
of
spatial
repressors,
meaning
like
where,
in
space,
some
gene
is
going
to
be
regulated,
upregulated
or
down
regulated.
So
it's
going
to
be
turned
up
or
down
they're
different
signals
from
adjacent
cells.
So
if
you
have
another,
this
is
in
a
single
cell.
A
If
you
have
another
cell,
as
your
neighbor,
it's
going
to
send
a
signal
to
tell
you
to
turn
down
your
gene
expression
or
turn
up
your
gene
expression
cell
cycle
control,
so
like,
as
I
said
earlier,
cell
cycle
has
a
certain
frequency
and
if
it's
higher
there
are
more
cells
produced
in
terms
of
proliferation.
If
it's
a
lower
fewer
cells
produced
and
then
different
lineages,
so
the
cell
might
change
its
lineage,
it
might
branch
off
onto
a
new
lineage
or
be
a
lineage
identifier.
A
You
have
two
modules
in
this
example,
and
these
two
modules
actually
can
be
controlled
by
an
or
gate.
So
you
can
have
two
modules
that
have
settings
they
might
be.
You
know
deterministic
settings
it
might
be,
they
might
fluctuate.
You
can
then
have
this.
A
You
can
also
have
an
and
gate
or
an
exclusive,
or
so
you
can
have
these
logic
gates
here,
where
it
turns
module
one
and
module
two
on
and
off,
or
combines
them
or
something
like
that,
and
then
this
results
in
an
output
which
is
rna.
So
your
dna
is
transcribed
into
rna.
These
regulators
move
across
the
dna.
A
The
putative
dna
there's
no
dna
in
this
model,
but
this
is
a
cis-regulatory
element
which
sits
right
directly
upstream
of
the
gene
or
the
coding
region,
and
then
it
basically
decodes
that
dna
into
rna
using
these
different
factors.
So
it
might
produce
a
lot
of
rna.
It
might
produce
a
little
bit
of
rna,
it
might
produce.
A
A
So
in
this
paper
they
talk
about
gene
regulatory
networks,
encode
the
complex
molecular
interactions
that
govern
cell
identity.
They
propose,
in
this
paper
a
method
called
deep
sem,
which
is
a
deep
generative
model
that
can
jointly
infer
grn's
gene
regulatory
networks
and
biological.
Meaningful
representation
of
single
cell
rna
sequencing,
so
this
data
is
scr
and
ac.
This
is
a
next-gen
sequencing
technique.
A
You
know
number
a
copy
number
so
again,
if
it's
turned
up
you're
going
to
have
a
larger
number
of
copies,
if
it's
turned
on
there
in
a
smaller
number
of
copies,
they're,
all
going
to
be
more
or
less
the
same
sequence,
the
thing
about
transcription
is:
it
produces
transcripts
of
different
length
and
different
sequence,
because
it's
taking
you
know
different
parts
of
that
coding,
region
and
transcribing
it
in
rna.
So
it's
going
to
look.
A
A
In
particular,
we
developed
a
neural
network
version
of
the
structural
equation
model
to
explicitly
model
the
regulatory
relationship,
ships
among
genes.
So
a
structural
equation
model
is
a
statistical
model
where
they
take
a
different
number
of
different
variables
and
they
look
at
their.
You
know
their
relationship
and
they
use
some
statistical
tests
to
assess.
So
it's
like
almost
like
a
network
of
different
variables
and
how
they're
all
related
to
one
another.
A
A
So
this
helps
us
regulatory
relationships
because
we
want
to
know
what
came
first
and
what
came
second
in
in
terms
of
genes
just
looking
at
genes,
we
don't
know
which
genes
activate
other
genes.
So
a
lot
of
these
gene
regulatory
networks
are
like
one
gene
that
is
transcribed
and
then
that
rna
transcribes
another
gene
by
activating
its
regulatory
machinery
and
then
that
in
turn,
activates
some
other
genes.
A
So
you
actually
have
these
hierarchies
in
some
cells,
where
you
have
some
function,
like
muscle,
you
know,
if
you
want
to
go
from
like
one
type
of
cell
to
another,
I'm
going
to
go
to
my
board
here,
you
might
have
a
network
of
genes,
so
you
have
like
say
a
single
gene
here,
that's
an
activator
and
it
might
activate
a
bunch
of
genes
down
here
so
like
if
you
activate
this
cell
here
or
this
gene,
it
activates
these.
It
turns
on
these
other
genes
by
virtue
of
its
rna
production
or
it's
mrna
production.
A
I
should
say
it's
a
messenger
rna
that
is
activates
the
transcriptional
machinery
here.
The
regulatory
machinery
turns
these
genes
on
and
then,
when
they're
turned
on
their
mrnas
in
the
cytoplasm,
it's
signaling,
but
it's
also
turning
on
these
other
genes
down
here.
So
you
can
see
that
like.
If
you
want
to
create
a
muscle
cell
from
a
stem
cell,
you
can
use
a
factor
called
myod.
A
You
can
introduce
it
into
the
cell
and
I
will
turn
on
this
myodgin
it'll
activate
it
activate
its
its
regulatory
machinery.
It
produces
a
lot
of
mrna
and
this
mrna
then,
will
activate
these
genes
down
here.
It'll
activate
the
regulatory
machinery
and
then
the
gene
it'll
produce
a
lot
of
mrna
copies
here
and
it'll
go
on
and
on
and
activate
more
genes,
and
they
call
this
a
gene
activation
cascade.
A
So
this
is
a
a
nice
example.
That's
not
too
complicated,
but
suffice
it
to
say
it.
Isn't
that
simple,
there's
a
lot
of
a
lot
of
cases
where
it's
not
the
simple
it's
not
hierarchical.
For
example,
it
can
be
very
the
causality.
A
To
follow-
and
you
know
this
is
something
we've
talked
about
in
this
group
in
my
other
group-
where
we
have
there-
we
go
where
we
have
these
reciprocal
causality
situations
and
all
these
other
things,
so
don't
think
that
it's
always
hierarchical
like
this.
In
this
case
it
is,
but
in
other
cases
it's
not
so
they
have
to
use
these
different
models-
structural
equation,
modeling
and
deep
learning
to
kind
of
get
at
some
of
this.
A
So
these
the
deep
generative
model
to
generate
to
generate
these
conditions,
these
grns
and
some
of
their
products,
and
then
they
try
to
use
the
sem
to
figure
out
which
genes
are
causing.
What
so
benchmark
results
show
that
deep,
sem
or
achieves
a
comparable
or
better
performance
on
a
variety
of
single
cell
computational
tasks
such
as
grn
inference
scr
and
ac
data,
visualization
clustering
and
simulation
compared
with
the
state-of-the-art
methods.
A
In
addition,
the
gene
regulations
predicted
by
deep
sem
on
cell
type
marker
genes
and
the
mouse
cortex
can
be
validated
by
epigenetic
data,
which
further
demonstrates
the
accuracy
and
efficiency
of
our
method.
Deep
sem
can
provide
a
useful
and
powerful
tool
to
analyze.
Scrna
seek
data
in
in
for
a
grn,
so
again,
you're
using
this
deep
learning
component
to
generate
these
models
or
of
grns
and
then
using
the
sem
part
to
analyze
them
and
figure
out
how
they're
structured.
A
So
it's
a
nice
paper.
It's
very
you
know
it's
again
like
this
is
very
hard
to
understand,
because
they're
using
you
know,
not
only
are
they
using
a
lot
of
the
genomic
terminology,
they're,
also
using
deep
learning
and
statistical,
so
you
can
see
that
you
know
you
have
these
transcription
factors
here
that
serve
as
hubs.
They
target
different
genes
in
their
activation,
and
this
is
basically
how
a
gene
network
is
structured.
There
is
typically
this
sort
of
semi-hierarchical.
A
At
least
I
saw
my
hierarchical
nature
to
it.
You
have
these
hubs
that
activate
a
bunch
of
downstream
genes.
They
can
be
strictly
hierarchical,
like
in
the
myod
example
where
they
can
be
like
this,
and
this
is
more
kind
of
like
a
a
scale-free
network
with
hubs,
so
this
makes
it
look.
You
know
a
little
bit
more
like
it's
a
little
more
decentralized,
but
then
again
it's
also
there's
a
lot
of
you
know
direct
control
here.
A
So
another
example:
I'm
going
to
use
here
or
draw
from
is
this
paper
that
came
out.
I
think
recently
very
recently,
it's
a
gene
regulatory
network
balances,
neural
and
mesoderm
specific
specification
during
vertebrate
trunk
development.
So
this
is
a
nice
very
long
paper,
it's
nice,
but
I
want
to
go
through
some
of
the
highlights.
A
So
again
they
use
a
single
cell
rna,
seq
technology.
It
reveals
a
signature
of
neural
mesodermal
progenitors,
so
these
are
types
of
cells
that
are
sort
of
in
that
mesodermal
layer
of
the
embryo,
but
they're
sort
of
this
neural
precursor
property.
These
are
progenitor
cells.
They're
able
to
find
this
with
the
single
cell
already
seek.
A
They
look
at
these
in
vitro
and
nmps
resemble
and
differentiate.
These
are
the
neuromas
and
normal
progenitors
resemble
and
differentiate
similar
similarly
to
their
in
vivo
counterparts.
So
they're
doing
this
in
a
I
guess
in.
A
In
I
guess
it's
an
in
vitro
model
so
so
neuro
mesodermal
progenitors
generate
cells
of
the
spinal
cord
and
somites,
so
they're
in
the
embryo,
they're
not
quite
forming
any
tissue.
Yet
but
they're
going
to
go
on
to
form
cells
in
these
areas
and
they're
doing
this
in
vitro
so
they're
trying
to
match
the
viewer
results,
and
so
this
is
yeah.
B
A
So
there's
also
a
dual
role
for
retinoic
acid
signaling
in
nmp,
induction
and
neural
differentiation
and
a
transcriptional
network
regulates
neural
versus
mesodermal
allocation.
So
this
is
about
this
is
in
a
biological
system.
This
is
where
they
look
at
the
network
and
they
see
what
it's
doing
during
this
process
of
differentiation.
A
A
Fate,
they
typically
live
a
certain
number
of
divisions
and
they
die
they
they
apoptose,
and
so
this
is
an
important
sort
of
switch
that
gets
turned
on
and
off
in
these
cells,
and
so
they
kind
of
talk
about
the
cell
type,
and
then
they
try
to
discover
the
gene
regulatory
network.
That's
involved
in
this
I'm
trying
to
see
if
they
have
a
figure
where
they
show
this
network.
A
And
they
actually
show
a
little
bit
of
the
signaling
cascade
and
maybe
some
of
this
network,
so
this
is
a
route
from
this
precursor
cell
lineage
to
these
different
other
types
of
cells
is
neural,
lineage
cells,
so
the
psm
is
what
that
is.
Npc
is
the
neural
precursor
cell.
So
this
is
a
neural
precursor.
It's
outside
of
this
germ
layer,
and
it's
now
you
know
on
its
way
to
becoming
some
sort
of
neural
cell.
A
A
A
In
and
try
to
find
a
figure
of
this
just
like.
Suffice
it
to
say
that
this
is
another
method
of
looking
at
gene
regulatory
networks,
you
can
actually
look
at
the
concentration
of
the
products
during
different
periods
of
development.
So
that's
a
nice
way
to
look
at
the
activation
of
these
networks
over
time.
A
Final
paper
I'm
going
to
talk
about
is
this
evolving
differentiable
gene
regulatory
networks-
and
this
is
doug
wilson
and
sylvain
planck
and
kyle
harrington.
So
these
people
been
working
in
the
artificial
life
community
and
on
gene
regulatory
networks
and
development,
so
over
the
past
20
years,
artificial
gene
regulatory
networks,
so
these
are
artificial.
A
Computational
models
of
these
type
of
networks,
they've
shown
their
capacity
to
solve
real-world
problems
in
various
domains,
such
as
agent
control
signal
processing
and
artificial
life.
Experiments
they've
also
benefited
from
evolutionary
approaches
and
improvements
to
dynamic
which
have
increased
their
optimization
efficiency.
A
A
That
differentiate,
and
so
they
actually
can
be
optimized
using
stochastic
gradient
descent,
so
they
use
a
standard
machine
learning,
benchmark
data
set
they
evaluate
these
networks
and
how
they
can
be
optimized,
so
in
biology,
they're,
not
always
optimized,
and
the
difference
here
is
that
you
can
create
these
networks
that
are
inspired
by
gene
regulatory
networks,
and
I
think,
in
the
last
paper
you
saw
how
messy
that
can
be
in
a
computational
model
that
can
be
optimized
in
terms
of
gradient
descent
and
that
sort
of
thing
so
they're
not
using
any
sort
of
genetic
algorithm
here.
A
They're
using
radiant
descent
and
so
artificial
grns
were
first
proposed
using
a
binary
encoding
of
proteins
with
a
specific
start
and
stop
codon
similar
to
biological
genetic
encoding,
which
was
like
about
20
years
ago.
Grms
have
since
been
used
in
a
number
of
domains,
including
robot
control,
signal
processing
and
wind
farm
design
and
reinforcement
learning.
A
So
this
is
a
biologically
inspired
model.
That's
been
used
for
a
lot
of
other
applications
and
they've
been
used
to
investigate
a
number
of
questions
in
the
context
of
evolution.
So
one
such
question
is
the
relationship
between
the
robustness,
against
noise
and
against
genetic
material.
Deletions
redundancy
of
genetic
material
was
proven
to
enhance
evolvability
up
to
a
point,
and
it
was
shown
that
modular
genotypes
can
emerge
when
grns
are
subject
to
dynamic
fitness
landscapes.
A
So
these
networks
that
we
saw
in
the
first
example
in
the
davidson
book,
they've
sort
of
formalized
these
a
little
bit
and
then
they're
using
them
to
simulate
these
types
of
things,
and
they
have
these.
They
can
generate
these
types
of
behaviors
that
look
biological
but
they're.
Actually,
not
you
know,
they're,
not
biological
in
the
organic
sense
they're
they
look
biological,
their
outputs
too.
A
A
You
can
use
modern
platforms
like
tensorflow
and
gr
gpus,
and
that
should
be
able
to
help
you
make
these
networks
all
the
more
powerful,
and
so
that's
they're
trying
to
do
here
and
so
the
first
set
of
grn
models.
Weren't
necessarily
very
good,
but
these
are
better
that's
what
their
point
is.
They're
able
to
actually
show
some
really
interesting
effects
like
the
baldwin
effect,
which
is
where
learned
adaptive
learned
and
adaptive
behaviors
have
a
fitness
advantage
that
can
facilitate
the
genetic
assimilation
of
equivalent
behaviors
by
subsequent
generations.
A
They
also
can
model
plasticity
as
an
adaptive
response
so
yeah.
They
just
show
this
differentiable
grn
model
and
they
walk
through
it,
and
I
don't
know
if
they
have
any
figure
or
any
nice
figures
here.
This
is
a
nice
figure.
A
grn
layer
model
is
a
composition
of
other
layers,
so
here
you
have
inputs,
you
have
outputs
and
you
have
regulatory
units,
and
then
you
use
the
soft
max
function.
A
This
output
mask
gets
fully
connected
and
that's
how
they
do
this
so
they're
able
to
get
like
these
they're
roughly,
like
you
have
these,
which
you
might
consider
to
be
genes.
You
have
an
input,
you
have
a
recurrent
state
which
is
a
feedback.
You
evaluate
using
a
soft
max
criterion
and
then
you
have
this
output,
which
is
the
output
mask,
and
then
that's
that
gives
you
your
answer,
and
so
you
can
use
this
on.
A
number
of
you
know
non-uh
biological
problems.
A
So
this
is
again
another
example.
A
gerund
layer
model
is
a
composition
of
other
layers
of
protein
concentration,
normalization
and
so
protein
concentration
is,
you
know
where
we
have
the
switch
that
goes
on
and
off,
and
you
know
if
you
use
a
soft
max.
I
guess
you
get
like
this
concentration
gradient
that
it's
normally
has
to
be
normalized
or
not.
So
you
end
up
that's
how
you
do
that.
A
And
so
you
know
they
have
some
nice
graphs
of
before
and
after
training,
so
they
clearly
learn
and
they
clearly
minimize
their
error,
and
so
that's
that,
so
that's
that's
our.
I
think.
That's
all
I'm
going
to
go
over
with
you
on
for
this
week.
I
just
wanted
to
show
you
the
different
versions
of
gene
regulatory
network
that
you
can
have.
I
know
that
we've
been
talking
about
this
in
past
meetings,
so
I
think
that
was
a
nice
overview
of
that.
A
So
next
up
is
quran
and
some
of
his
materials
and
after
he's
done
then
see
you
next
week.
B
B
Like
abnormalities
or
things
like
that
in
the
pictures
you
are
trying
to
find
a
connection
between
the
two
like,
if
you
have
a
gene
sequence-
and
you
have
a
picture
associated
with
that,
so
they
were
trying
to
find.
You
know
like
the
neural
net
was
being
trained
on
both
of
those
things,
so
that.
B
Like
that,
so
there
are
there
like
two
three
pre-trained
networks
that
are
already
existing,
so
I
don't
know
which
people
exactly
dealt
with
the
accuracy
and
all
those
things
so
transformers
were
again.
You
know
big
thing
that
we
use,
I
think
yeah.
A
So
this
is
from
deep
mind.
B
I
don't
think
this
is
the
exact
one,
but
it
was
something
similar.
So
they
were,
you
know
talking
about
the
accuracy
models
that
are
there
and
how
you
usually
for
mammals,
like
mostly
mice
and
the
most
studied
things,
so
even
they
had
a
lot
of
isolated
data.
That
was
there,
I
think
so
they
trained
it
on
that
and
I
think
it
was
kind
of
a
like
it
could
be
used
for
a
lot
of
things.
That's
why
I
was
you
know
trying
to
find
it.
D
D
B
So
also
regarding
gene
sequences,
the
epigenome
is
like
it
contains
the
actual
transcript
of
how
you
know
it
is.
It
doesn't
work
that
way
here.
The
epg
don't
think
like.
If
you
have
a
gene
sequence,
that's
there.
The
mpg
normally
tell
you
how
it's.
A
Yeah,
I
guess
it's
yeah,
so
it's
you
know
you
have
your
gene,
which
is
your
dna
sequence
and
then
there's
a
method
of
transcription,
so
there's
a
transcriptional
machinery
that
hooks
onto
the
sequence
and
and
and
transcribes
it
to
rna.
Okay.
So
it's
going
down
the
gene
and
there
are
a
lot
of
different
marks
on
top
of
that,
like
in
the
gene
and
the
promoter
regions,
what
they
call
it's
just
before
the
main
gene
that
tells
it
what
to
do
so.
A
You
have,
you
know
different
promoters
and
enhancers,
and
they
kind
of
you
know
tell
the
transcriptional
machinery
where
to
go
and
what
to
take
transcripts
of,
and
then
you
also
have
these
these
histone
markers,
and
these
are
sites
where
they
either
enhance
the
activity
of
a
certain
downstream
component
which
is
down
from
the
from
the
histone
mark
or
they
you
know,
turn
it
off.
So
there
are
ways
that
you
know
there
are
a
lot
of
different
ways
that
the
gene
gets
expressed
through
a
lot
of
the
epigenomic
stuff
and
transcription.
A
B
B
D
B
B
B
Informal,
I
think
it
depends
on
the
when
you're
implementing
the
transformer
mechanism.
You
know
for
something
that
is
very
custom
made
like
this
gene
expression
is
more
of
a.
I
think.
B
A
D
C
C
B
D
D
D
D
B
C
B
A
See
anything
there
I
mean
we
should
we
could,
and
I
mean
it
would
be
kind
of
hard
to.
You
know
it's
always
kind
of
hard
to
integrate
the
two
kinds
of
data,
but
it
would
be
nice
to
have
something
I
I
I
actually
think
they're
like
a
lot
of
spatial
expression
data
sets.
I
know
that
they
have
them
for
like
c
elegans
and
maybe
for
zebrafish.
I
don't
know,
I
don't
know
about
axolotl.
I
can
look
into
that
and
see
if
they
exist,
because
I
know
it's
a
common
model
organism.
A
Yeah,
I
don't
think
that's
not
it,
but
yeah.
They
have
like
drosophila
and
and
yeah,
yeah
and
and
c
elegans
mouse
human.
A
But
yeah
this
is
great.
This
is
the
they
have
the
combination
of
like
deep
learning
and
genomics,
and
that
can
break
a
lot
of
people
like
this.
A
Both
fields
are
very
intensely
like
technical
and
they
have
their
own
terms.
So.
B
Yeah
very
very
much,
but
you
know
I've
seen
transformers
here:
they're
mostly
used
in
applications
where
they're,
using
both
like
they
either
using
one
and
generating
a
prediction
and
they're
using
the
picture.
Maybe
to
you
know,
verify
or
really
confirm
that
that
prediction
is,
you
know,
giving
us
this
and
the
kind
of
going
kind
of
deal
with
that.
D
Yeah
aspect.
B
So
that
was
the
another
topic
that
I
was
going
through
was,
I
think
again,
a
hydrophobic
problem.
Molecular
it's
it's.
It
has
to
be
deeply
rooted
in
molecular
physics,
because
you
have
a
physics
model.
You
know
that's
checking
out
whether
how
those
hydrogen
bonds
are
actually
working
with
the
d
transformation,
yeah,
okay,.
B
B
So
I
have
like
pictures
from
one
shots.
Okay,
this
is,
I
had
to
find.
B
In
different
stages,
that
is
there.
So
if
I
could
have
like
a
really
count,
the
amount
of
times
that
thing
is
divided
and
along
with
that,
the
astrological
stage
that
is
going.
B
B
A
Yeah,
well,
that's
good
thanks
for
the
update.
B
A
Yeah,
if
you
find
anything
interesting,
you
know
you
can
put
it
in
the
slack
channel
or
you
know,
if
you
don't
do.
B
A
A
B
B
B
B
A
B
So
I
kind
of
take
it
like
stages.
I'll
say
this
so
like
this
was
maybe
probably
stage
six
okay.
D
B
B
A
big
drastic
change
in
the
pixel
number-
that
is
there
right
if
it's
going
like
this
traversing
this,
so
the
pixel
levels
will
be
like
kind
of
similar
to
that.
So
then
there'll
be
a
change
here.
Something
like
that.
Maybe
contrast
different
glass,
something
like
that.
A
B
To
at
least
find
out
how
many
times
it's
divided
and
what
one
can
we
expect
from
a
state
six
as
well,
maybe
give
a
prediction
also
a
plague.
It
should
have
these
many
divisions
by
this
day.
It's
kind
of,
let's
see.
D
B
D
B
B
A
Well,
yeah,
it's
nice
to
see
that,
and
you
know
I
think
it's
great-
that
kind
of
getting
it
together
and
putting
trying
different
techniques.
D
D
D
B
All
right
things
just
don't
relate
to
that
anything,
but
don't
play
you're
gonna
find.