Hello.
Today
I'm
glad
to
present
you
scA2SIGN
method
to
define
the
signatures
for
the
convolution
starting
from
single-
cell
data.
We
will
make
also
a
short
demonstration
of
it.
This
work
was
done
in
collaboration
with
Paul Fogel
and
Alexey Fedorov.
Before
to
do
the
deep
dive
about
our
method,
let's
do
the
short
introduction
in
order
to
explain
you
what
is
the
cell
deconvolution.
The
cell
type
deconvolution
is
an
silica
method
that
allows
to
identify
the
cell
proportions
from
bulk
RNA
sequencing
data.
This
method
requires
to
have
specific
signature
matrixes
containing
the
marker
genes
for
each
type
of
the
interest.
Our
method
proposed
to
identify
this
signature
matrix
and
start
from
the
single-cell data
because
in
single-cell data,
as
you
can
see
here,
we
do
have
the
transcriptomic
profiles
for
each
single
cell
from
the
data
present
in
the
tissue
in
the
samples.
Each
dot
here
corresponds
to
the
cell.
Cells
with
the
closed
transcriptomic
profiles,
they
are
grouped
together
and
form
clusters.
We
can
go
to
the
next
slide,
please.
Here
you
have
a
brief
overview
of
our
method.
Our
method
is
called
scA2SIGN,
which
stands
for
single
cell
agnostic
algorithm
for
the
convolution.
As
was
mentioned
previously,
it
starts
from
the
single-cell data
ideally
coming
from
several
subjects,
individuals,
in
order
to
cover
biological
variability
that
can
be
present
between
individuals.
The
single-cell data
need
to
be
preprocessed
and
assembled
in
one
data
set.
Here
the
cells
with
a
close
transcriptomic
profile
so
they
are
assembled
into
the
clusters.
This
representation
in
two
dimensions
is
called
UMAP.
From
this
single-cell data,
we
need
to
retrieve
the
matrices
which
are
going
for
each
individual.
They
are
going
to
contain
dreams
on
the
rows
and
clusters
on
the
columns.
It
means
we
need
to
aggregate
the
data
and
several
matrices
corresponding
to
each
subject
represents
a
tensor.
If
we
try
to
identify
the
signatures
right
away
at
this
step,
it's
going
to
be
difficult
because
some
clusters
of
single-cell data,
they
close
from
transcriptomic
al point
of
view.
It
can
raise
a
problem
of
multicollinearity.
In
order
to
deal
with
this,
we
run
several
rounds
of
non-
negative
tensor
factorization
in
order
to
get
the
reduced
tensor
and
assemble
close
clusters
together
in
order
to
define
the
cell
types
with
distant
transcriptomic
profiles
present
in
the
sample.
Once
again,
the
structure
of
the
data
is
going
to
be
the
same.
We
are
going
to
show
it
in
Jump
application
with
dreams
on
rows,
cell
types
on
columns,
and
several
matrices
corresponding
to
several
subjects.
From
this,
our
method
proposed
to
two
simple
criteria
to
define
the
marker
genes.
This
part
is
going
to
be
covered
by
Alexey.
Last
but
not
least,
the
same
genes
can
be
used
in
order
to
label
cell
types
are
defined
by
our
method
based
on
groups
of
the
clusters.
Last
but
not
least,
we
can
use
a
non-
negative
regression
in
order
to
perform
the
convolution
on
bulk
RNA
sequencing
data.
With
this
we
are
going
to
talk
a
little
bit
more
about
non-
negative
tensor factorization
in
order
to
explain
what
is
it
exactly.
Non-negative
tensor factorization
is
a
dimensionality
reduction
technique.
As
it
was
mentioned,
the
advantage
of
this
method
is
that
it
allows
to
take
into
account
the
several
matrices
coming
from
several
subjects
and
to
do
it
at
one
step,
comparing
to
other
dimensionality
reduction
techniques
like
ICA,
when
you
need
to
run
several
rounds
of
ICA
for
each
matrice
and
next
to
find
a
consensus
in
the
data.
As
output
is
an
input
of
the
NTF,
we
use
the
matrices
previously
described
as
the
output
of
the
NTF.
We
get
the
components
.
We
get
the
fingerprints
for
clusters
here
represented
by
W,
the
fingerprints
for
the
subject
represented
by
Q.
In
this
fingerprint,
they
have
the
genes or
the
signature.
Fingerprints
means
a
set
of
genes
with
lower
[inaudible 00:05:19]
defining
by
H.
Each
component
corresponds
a
strong
biological
signal
present
in
the
data.
If
we
make
the
sum
of
these
components
of
these
strong
biological
signals
present
in
the
data
plus
a
small
error,
we
can
approximate
the
data
which
is
represented
by
this
equation.
With
this
we
can
go
to
the
next
step
and
to
see
exactly
in
the
application
how
the
data
looks
like.
Right
now,
we
are
in
the
Jump
application
and
you
can
see
the
input
data
required
for
our
method.
On
the
left,
you
do
see
the
eleven
matrices
coming
from
different
donors.
The
matrices
have
different
names
and
we
used
the
idea
of
the
donors
as
their
names.
In
the
tables,
we
have
17
columns
corresponding
to
17
clusters
identifying
in
single-cell data.
On
the
rows
we
do
have
the
gene
names
preceded
by
the
idea
of
the
individual
in
order
to
distinguish
different
transcriptomic
data
coming
from
different
donors.
Once
we
loaded
the
data
into
the
system,
we
can
launch
our
add-in
and
we
are
going
to
see
how
it
looks
like.
It
looks
like
a
window
with
several
parameters.
These
parameters
are
written
in
Python
because
the
code
behind
use
Python
language.
But
you
do
not
need
to
be
afraid
because
we
are
going
to
change
only
one
important
parameter,
which
is
number
of
components
in
here.
We
suggest
to
start
with
the
number
of
components
equal
to
number
of
clusters
present
in
the
data
and
it's
going
to
be
equal
to
17.
Thank
you.
With
this,
we
can
simply
click
on
the
submit
button
and
the
application
will
run
the
algorithm
behind.
Next
with
this,
I'm
going
to
leave
the
floor
to
Paul
in
order
to
explain
the
output
of
NTF
and
to
guide
you
following
the
next
step
of
our
method.
Thank
you
for
your
attention.
Thank
you,
Garria.
The
output
of
the
a
dd-in consists
of
three
tables
with
NTF
components
along
each
dimension
of
the
tensor:
cell
clusters,
subjects,
and
genes.
First,
let
us
look
at
the
cell
cluster
component
stage.
I
open
the
table
and
each
row
on
this
table
corresponds
to
a
cell
cluster
in
each
column
to
a
component.
There
are
17
rows
and
17
components
as
specified
in
the
add-in
input
parameter
by
Galina.
Each
entry
in
the
table
corresponds
to
the
loading
of
a
cell
cluster
onto
a
component.
Now,
scrolling
through
the
columns
of
the
table,
we
see
that
some
of
the
components
contain
many
zeros
like
components
16,
which
is
good
because
it
indicates
that
these
components
approximate
the
expression
data
coming
from
the
few
cell
clusters
with
nonzero
loadings.
Here
we
have
exactly
one
cell
cluster
with
a
one,
and
all
of
us
are
zeros,
or
almost
zero.
The
fact
that
cell
clusters
have
nonzero
loadings
on
the
same
component
indicates
a
high
degree
of
collinearity, so
regrouping
of
these
clusters
is
advisable.
We
have
an
example
here
with
component
12,
where
three
actually
cell
clusters
have
nonzero
loadings.
Now,
let
us
look
at
the
subject
components
table.
I
open
it
and
scrolling
through
the
columns
of
the
tables,
we
see
now
that
unlike
the
site
clusters
components
table,
the
columns
here
tend
to
be
dense,
with
only
a
few
zeros
here
and
there.
This
is
good
because
it
shows
that
most
subjects
contribute
to
each
component.
In
other
words,
the
expression
data
approximated
by
each
component,
or
we
call
it
fingerprint,
applies
to
most
subjects.
In
a
sense,
the
fingerprint
is
universal.
Now,
Jump
has
an
excellent
feature
that
allows
you
to
view
the
histograms
of
the
columns
at
the
top
of
the
table.
Let's
click
on
this
icon
here,
and
we
now
see
all
the
histograms.
We
see
that
the
columns
are
more
or
less
dense
anyway,
and
it
will
be
good
to
be
able
to
quantify
the
degree
of
sparsities
of
each
column.
How
do
we
do
that?
To
quantify
the
degree
of
falsity
involves
the
cell
cluster
and
subject
components.
We
use
the
index
represented
by
this
formula.
This
formula
was
introduced
by
Herfindahl and
Hirschmann
to
evaluate
the
degree
of
concentration
achieved
by
fields.
For
example,
to
avoid
monopoly
situations
that
can
result
from
mergers.
Originally,
the
inverse
formula
was
published,
but
for
simplicity,
we
will
continue
to
refer
to
this
index
as
the
HHI index.
Similar
indexes
have
been
proposed
more
recently.
They
are
mathematically
equivalent,
but
this
index
has
the
advantage
of
being
very
interpretable
as
it
gives
the
number
of
non-negligible
entries
in
the
vector
which
are
highlighted
here.
In
the
example,
we
have
a
number
of
values,
but
only
the
highlighted
values
are
non-negligible.
The
HHI index
of
this
vector
is
five.
Returning
to
our
workflow,
we
see
how
this
index
allows
us
to
identify
components
that
are
sparse along
the
cell
cluster
dimension
and
dense
along
the
subject
dimension.
We
call
these
components
sparse and robust
.
They
are red
crosses
here
with
high
subject
HHI and
low
cluster
side
cluster
HHI on
the
top
of
this
field.
Based
on
these
four
components,
we
can
make
the
necessary
groupings
of
sight
clusters
to
eliminate
collinearities
and
the
remaining
cell
clusters
here... Sorry for this interruption.
[inaudible 00:13:22]. I just lost
[inaudible 00:13:22], I'll put it here.
The
blue
crosses here
correspond
to
the
remaining
cell
clusters
that
will
be
reintroduced
in
a
second
iteration
of
the
analysis.
Note
well
that
this
addresses
with
the
strongest
signal,
after
being
identified
by
the
NTF,
components
are
removed
from
the
data
set
and
the
next
NTF
starts
over
with
the
reduced
data
set.
To
illustrate
this
process...
I'm
sorry
but
I
have
a
problem
with
here
with
the...
It keeps
reappearing.
To
illustrate
this
process,
let's
now
look
at
how
these
iterations
of
the
recursive
NTF
play
out.
We
see
here
the
HHI table
from
the
first
iteration
which
will
help
us
decide
which
NTF
components
to
keep
and
which
to
discover.
As
mentioned
earlier,
good
components
are
those
that
are
specific
of
only
a
few
cell clus
tures
which
translates
in
the
low
cell cluster
HHI index.
These
are
labeled
rows
here
at
the
bottom
of
the
table.
We
also
run
one
high
degree
of
robustness
to
subject
diversity
and
this
translates
into
a
high
HHI index.
How
do
we
decide
whether
an
HHI value
is
low
or
high?
We
can
use
that
the
Jump
[inaudible 00:15:47]
platform
which
is
available
in
the
specialist
modeling
platform.
We
run
this
platform
and
we
apply
a
normal
niche
model
to
distinguish
between
low
and
high
HHI values
and
stop
the
resulting
classification.
When
we
do
that,
we
see
here
the
red
crosses
corresponding
to
the
components
that
are
selected
because
they
have
a
high
HHI subject
index
and
a
low
HHI
cell cluster
index
and
they
are
automatically
identified
as
such
by
the
Jump
Normal
platform.
Let
us
look
now
at
the
loadings
of
the
cell clusters
on
these
particular
components
.
We
see
on
component
15
and
component
16
that
cell cluster 15
and
cell cluster 17
has
exclusively
associated
these
components
so
they
can
be
well
approximated
by
these
components.
In
contrast
to
these
cell clusters, on
component
13,
we
see
three-
side
clusters
regrouped
along
this
component.
This
means
that
they
are
correlated
somehow,
it
will
be
hard
to
distinguish
them,
so
it's
better
to
regroup
them.
Similarly
on
component
12,
with
the
cell clusters
5, 2, and 14.
We
can
now
look
at
a
second
iteration
of
the
Recursive
NTF.
Similarly
we
look
at
the
components
HHI on
the
subjects
and
on
the
cell clusters
and
we
see
now
that
only
one
component,
component
four,
satisfies
these
retrainments
requirements
for
being
retained
during
this
iteration.
If
we
look
now
at
the
components
audience,
we
see
that
cell cluster
seven
is
exclusively
associated
with
this
component.
Good.
Now
that
we
have
done
that,
we
are
not
going
to
show
you
all
of
the
iterations.
Actually,
there
are
four
iterations
and
we
end
up
now
with
these
groupings
that
we
see
here.
We
have
actually
nine
cell
clusters
after
regrouping.
Some
of
them
are
groupings
of
three
or
four
original
cell clusters.
However,
five
cell clusters
did
not
need
to
be
regrouped.
What
is
the
biological
meaning
of
these
groupings?
By
matching
the
most
highly
expressed
genes
in
each
cell
cluster
with
bioinformatic
databases,
it
is
possible
to
associate
the
cell clusters
with
known biological
cell
types.
For
example,
the
NTF
groupings
2-
5-14
was
identified
as
a
natural killer
CD8
cell
type.
We
found
that
there
were
several
groupings
that
actually
pointed
to
the
same
major
key
cell
type.
Therefore,
we
decided
to
merge
them
and
came
up
with
five
main
cell types.
However,
it
is
still
possible
to
extract
signatures
for
[inaudible 00:20:12]
cell typ
es
that
can
be
used
in
a
second
stage
of
the
convolution.
But
we
will
come
back
to
this
point
at
the
end
of
this
presentation.
Let
us
now
look
at
the
resulting
reduced
tensor.
I'm
just
opening
the
table.
You
see
here
the
table,
and
as
expected,
there
are
now
five
rows
corresponding
to
the
regrouped
cell
clusters.
Now
that
we
have
introduced
cluster
grouping
to
minimize
collinearity,
I
will
have
Alexey
explain
how
we
proceed
to
create
signature
genes
that
are
most
effective
for
decomposition.
Once
the
cell
clusters
are
grouped,
we
do
still
need
to
create
gene
signatures
for
each
of
the
groupings
to
ensure
accurate
convolutions
of
the
mixed
samples.
So far,
we
have
considered
the
cell
and
subject
component
vectors.
Let
us
now
consider
the
gene
component
vector.
The
table
we
see
here
is
not
really
informative.
We
can
only
see
zeros
everywhere
due
to
the
fact
that
only
small
minority
of
genes
have
actually
nonzero
loadings
on
the
NTF
components.
If
you
can
go
to
the
next
iteration,
please,
Paul.
Let
us
now
examine
the
loadings
of
the
genes
on
this
representation
of
the
components
which
we've
called
sparse
and
robust.
These
were
selected
during
the
first
iteration
of
the
recursive
NTF.
We
see
here
in
red
that
some
of
the
distinctive
genes
are
shared
by
all
components.
We
know
well
that
each
component
is
specific
to
a
particular
cell
cluster
or
a
group
of
closed
cell clusters.
This
means
that
some
of
the
genes
are
actually
strongly
expressed
in
several
clusters
with
some
degree
of
linearity
between
the
cell
clusters.
This
is
exactly
how
niche
works.
Biological
parts
are
not
necessarily
orthogonal
to
each
other.
Therefore,
we
cannot
rely
solely
on
NTF
gene
components
to
identify
gene
markers
that
are
unique
to
a
cell
cluster
or
a
group
of
this,
we
must
return
to
the
actual
expression
data.
We
are
getting
back
to
expression
data
and
we
will
introduce
two
criteria.
The
first
one
for
a
gene
to
be
considered
the
marker
gene
of
a
particular
cell cluster,
we
say
that
it
should
always
be
more
expressed
in
the
cluster
than
in
other
clusters,
regardless
of
which
subject
is
considered.
This
is
shown
in
figure
bottom
left,
where
we
have
depicted
two
genes.
With
some
[inaudible 00:23:20],
we
see
that
these
points
on
the
left
correspond
to
a
unique
marker
gene
for
B
cells.
We
can
see
only
B
cells
everywhere.
On
the
right,
a
non-marker
gene
is
highly
expressed
in
both
B
cells,
non
classical
monocytess
and
natural
killer
cells.
The
second
criteria
will
be
that
the
associated
cell
cluster
should
not
only
be
at
the
top
of
the
other
clusters.
When
we
look
at
the
expression
of
the
market
gene
over
[inaudible 00:23:55]
subjects
and
cell
types,
but
also
it
should
be
well
ahead
of
others.
For
this,
we
use
this
score
with
a
formal
issue
on
the
right
that
were
proposed
by
[inaudible 00:24:09]
in
their
original
paper.
This
score
reflects
the
specificity
of
a
gene
marker.
This
score
is
based
solely
on
the
distribution
of
gene
expression
across
subjects
and
cell
types.
Thus
applying
this
to
criteria,
we
obtained
a
list
of
173
marker
genes
that
can
subsequently
be
used
for
the
convolution.
In
this
table,
they
are
listed
together
with
the
values
that
meet
the
two
criteria
I've
talked
about.
Let
us
now
look
at
the
table
of
signature
genes
for
each
of
the
subject
and
major
cell
types.
Here
we
see
the
list
of
genes and
the
columns
correspond
to
each
other's
subjects
and
the
same
list
of
cell
types
for
each
of
the
subject.
In
contrast
to
the
table
of
NTF
loadings
that
we
showed
previously,
this
table
is
not
sparse,
so
it
is
limited
only
to
173
marker
genes
that
we
selected
according
to
our
criteria
with
[inaudible 00:25:26]
gene.
As
I
said,
the
columns
are
for
subject
cell
types.
We
will
now
visualize
how
the
signature
genes
are
expressed
across
subjects.
For
that,
we
will
use
the
second
plug-in
we
developed
which
is
called
adrubix.
This
will
allow
us
to
easily
assess
the
robustness
of
the
gene
signatures
we
got.
We
open
a
small
window
with
a
list
of
parameters
in
Python
syntax
as
for
the
first
plugin.
The
default
values
have
been
already
filled
in
and
we
will
now
only
fill
in
for
demonstration,
the
value
for
the
parameter
which
is
called
scale
along.
Since
the
expression
ranges
differ
significantly
between
genes,
we
will
set
this
parameter
to
zero
to
have
all
the
columns
scaled
and
we
submit.
Just
a
few
seconds
and
we
see
a
picture
with
distinct
signatures
for
each
major
cell
type.
However,
we
see
that
this
heat
map
is
glued
together
in
between
subjects
and
we
can
actually
, with
this
plug-in,
very
easily
add
separators
in
between
the
columns
for
each
of
the
subjects.
For
this,
we
change
the
value
of
the
parameter
metadata
cols sep
to
subject
and
we
resubmit.
Now
we
get
the
separate
heat
maps
for
each
of
the
subjects.
This
is
much
clearer
and
as
you've
seen,
this
small
window
contains
numerous
parameters.
You
can
refer
to
the
documentation
for
the
underlying
Python
package
which
is
available
clicking
on
this
link.
Conclude,
we
see
here
actually
that
the
signature
genes
are
very
robust.
We
see
very
little
difference
in
different
subjects.
Like
that
we
can
visually
very
easily
confirm
the
robustness
of
[inaudible 00:28:09]
.
Now
we
are
nearing
the
end
of
our
presentation,
I
will
let
Galina
present
some
deconvolution
results
that
were
obtained
with
the
gene
signature
[inaudible 00:28:22]
. Galina.
Yes,
thank
you.
Last
but
not
least,
we
would
like
to
show
you
the
performance
of
our
method.
To
evaluate
the
performance
of
our
method,
we
use
the
published
data
coming
from
White
paper
so
mentioned
in
hints
his
reference.
This
data
set
containing
21
different
cell
mixtures.
There
are
the
mixtures
of
different
T
cell
subtypes,
NK
cells,
monocytes
and
B
cells.
In
these
mixtures,
we
have
also
three
different
cell
types
not
covered
by
our
signatures
corresponding
to
colorector
or
breast
cancer
cells
and
two
cell
lines
like
fibroblaster
and
the
tail
cells.
We
decided
to
use
two
criteria
evaluates
overall
performance
of
our
method
Pearson
correlation
coefficient
or
[inaudible 00:29:18]
a
concordance
coefficient
because
we
can
get
very
significant
correlation
between
estimated
and
ground
truth
proportions,
but
we
can
have
the
change
in
the
scale
identifying
the
quantity
of
cells
present
in
the
sample.
That
is
why
we
use
two
criteria.
Overall,
we
obtain
good
results
for
cell
types
covered
by
our
signatures.
The
results
are
slightly
not
so
beautiful
for
T
cells
because
we
don't
have
exact
correspondence
between
T
cell
subset
present
in
the
mixtures
and
T
cells
group
presented
in
our
signatures.
We
have
also
effecto
r
cells
and
in
the
mixtures,
this
cell
type
was
not
added.
Last
but
not
least,
we
would
like
also
to
mention
with
our
method
you
can
go
beyond
and
identify
the
subtypes
of
T
cells,
the
monocytes
running
recursive
NTF
and
defining
the
marker
genes
for
each
big
group
of
immune
cells.
Also
that
these
add-ins
that
have
been
presented
and
demonstrated
today,
they
are
available
in
Jump
application.
With
this,
thank
you
for
your
attention.
