►
From YouTube: Accelerating Drug Discovery by Competitive Cooperation Through Ope... Bill Mulligan & Camille Marini
Description
Don’t miss out! Join us at our upcoming events: EnvoyCon Virtual on October 15 and KubeCon + CloudNativeCon North America 2020 Virtual from November 17-20. Learn more at https://kubecon.io. The conferences feature presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.
Accelerating Drug Discovery by Competitive Cooperation Through Open Source - Bill Mulligan, Kubermatic & Camille Marini, Owkin
Historically, pharmaceutical companies have kept their machine learning (ML) models and data strictly confidential. The EU/EFPIA IMI2 Joint Undertaking funded “MELLODDY” (Machine Learning Ledger Orchestration for Drug Discovery, Grant n° 831472) initiative, a 3 year 18.4M EUR effort, brings together 10 leading pharma companies and 7 tech partners to build a new collaborative ML platform that boosts drug discovery model development while addressing both security and privacy preservation concerns (including commercial confidentiality). Kubernetes provides the consistent computing infrastructure across companies ensuring that data owners maintain control while running the common ML software and sharing the resulting models. This project demonstrates the potential of Kubernetes to accelerate drug discovery and enable cooperative competition in IP-sensitive industries. https://www.melloddy.eu/ https://www.imi.europa.eu/projects-results/project-factsheets/melloddy
https://sched.co/ZetX
A
A
A
A
So
why
do
we
want
to
build
such
a
platform
of
federated
learning
for
drug
discovery
as
a
disclaimer?
I'm
not
a
chemist
but
I'll.
Try
to
explain
on
average,
bringing
one
clock
to
market
cost
a
bit
less
than
2
billion
euros
and
takes
13
years.
So
it's
a
lot
and,
and
one
of
the
steps
correspond
to
drug
discovery.
A
A
A
So,
but
in
order
to
do
this,
we
need
to
build
a
platform,
and
this
platform
is
based
on
on
one
hand,
kubernetes
and
on
the
other
hand,
we
can
connect.
So
I'm
going
to
start
with
the
presenting
hooking
connect,
so
it
corresponds
to
the
application
layer
of
the
of
the
platform
and,
and
it
makes
it
possible
to
do
federative
learning.
A
But
what
is
federated
learning
so,
as
just
mentioned,
each
farmer
has
its
own
liability
of
annotated
molecule
and
we
want
to
train
the
machine
learning
model
with
the
ensemble
of
data
sets
with
the
data
set
of
the
different
pharmaceutical
companies
so
to
train
a
model
on
multiple
data
sources.
The
standard
approach
would
be
to
pull
all
the
data
in
one
center
and
try
and
train
a
machine
learning
model
there.
A
B
A
A
A
So
this
is
the
overview
of
the
application
layer
once
it
is
deployed.
So
basically,
there
is
an
application
deployed
on
a
kubernetes
cluster
in
every
pharma
and
also
an
application
which
is
deployed
and
called
the
model
dispatcher
in
each
farmer.
We
store
data
and
the
data
are
going
to
be
remains
there,
they're
not
going
to
go
outside
of
of
the
application,
and
we
train
a
machine
learning
model
on
this
data
by
exchanging
the
model
updates
with
a
model
dispatcher
and
doing
the
orchestration
of
the
training.
B
A
A
So
the
training
task
occurs
in
a
pharma
application
and
the
aggregation
task
corresponds
to
aggregating
model
updates,
so
the
output
of
the
training
step,
which
is
done
on
the
model
dispatcher
and
it's
up
to
the
data
scientists
to
define
how
to
train
the
model
to
which
algorithm
is
used
and
which
aggregation
to
use
so
the
routine
connected
agnostic
to
this
type
of
algorithm.
So
it's
up
to
the
data
scientist
and
in
the
melody
project
is
done
by
your
partner,
which
is
k11.
A
So
before
letting
a
bill
present
kubernetes,
this
is
an
overview
of
the
architecture
of
connect.
So
it's
another
just
an
overview,
basically
in
every
pharma
center.
So
it's
called
node
here,
but
it's
not
a
kubernetes
node.
We
deploy
a
backend
called
subscribe
again,
which
is
responsible
for
managing
training
tasks
and
also
for
managing
the
storage
of
data,
and
this
back
end
interacts
with
a
non-sensitive
metadata
network
enables
through
hyperledger
fabric
in
order
to
exchange
non-sensitive
metadata
about
the
machine
learning
model
and
the
set
of
the
status
of
the
differ
of
the
compute
plan.
A
B
Cool
thanks
for
that
introduction,
yeah.
So
what
I'm
going
to
talk
about
now
is
how
we
use
the
kubernetes
kubernetes
platform
to
deploy
and
manage
the
kubernetes
clusters
that
are
supporting
all
the
federated
learning,
both
the
irrigation
point
from
oakland
and
the
actual
like
models
themselves,
while
still
keeping
each
of
the
date.
The
data
from
each
of
the
farmer,
partners,
secure
and
private
in
their
own
accounts,
and
so
the
way
that
we
actually
do.
That
is
actually
by
running
kubernetes
in
kubernetes.
A
B
B
And
so
what
this
allows
us
to
do
is
to
create
a
single,
highly
available
kubernetes
cluster
running
in
an
account
in
aws
account
controlled
by
us,
and
what
we're
going
to
do
in
that
account
is
then
go
in
and
deploy
both
kubernetes
kubernetes
platform,
which
will
provision
all
the
clusters
and
all
the
centralized
services.
So
like
the
gitlab
instance
that
we're
using
the
aggregation
layer
to
be
able
to
manage
the
data
and
the
model
across
all
the
former
accounts.
B
And
so
what
we
can
do
with
kubernetes
is
spin
up
multiple,
like
kubernetes
clusters.
One
is
for
like
our
central
cluster
services,
that
is
running
like
the
aggregation
service
and
like
the
gitlab
instances,
and
the
other
part
is
to
actually
create
clusters
in
the
pharma
partners
so
that
they
can
actually
run
the
machine
learning
bottles.
B
But
it's
still
a
using
kubernetes
still
gives
us
the
ability
to
run
cons,
workloads
consistently
in
the
same
way
across
all
these
different
applica
different
accounts.
So,
rather
than
having
to
worry
about
inconsistencies,
we
have
a
clear
abstraction
layer
to
be
able
to
run
our
workloads
and
deliver
on
this,
and
so
we
kind
of
look
at
this
project
as
a
whole.
So
some
of
the
benefits
of
using
open
source
in
both
gramatic
kubernetes
platform,
oaken
and
kubernetes
as
a
whole,
is
that
we
can
do
that.
B
The
second
part
is
the
abstraction
of
open
source
software.
Is
that
really
anybody
can
use
it
and
run
it
wherever
they
want
to?
So
it
really
creates
a
consistent
way
for
us
to
create
and
deploy
both
infrastructure
and
the
software
across
all
of
the
companies.
It's
not
something
that's
proprietary
to
one
company,
but
it
can
be
used
in
a
consistent
way
across
all
of
our
partners.
B
So
if
we
wanted
to
in
the
future
some
of
the
models
that
we're
looking
at
is
being
able
to
run
in
private
data
centers
or
be
able
to
run
in
different
cloud
providers
and
the
great
thing
about
using
like
open
source
is
that
it
it
can
be.
We
can
use
it
wherever
we
want,
so
we
can
actually
use
the
same
model
the
same
software
again
without
having
to
run
it
on
the
exact
same
infrastructure.
B
B
B
And
so
the
cool
thing
about
this
project
is
because
it's
not
tied
to
a
specific
industry
is
that
we're
actually
looking
to
leverage
this
project
in
other
industries,
so
other
industries
that
have
the
problem
of
machine
learning
would
help
them
solve
some
of
their
problems
if
they
could
collaborate
on
their
data
to
improve
the
performance,
but
that
the
data
is
actually
still
their
competitive
advantage.
So
some
of
the
examples
of
insurance
like
can
we
create
better
risk
models,
logistics
of
how
do
we
like
efficiently
route
stuff
around
the
world
or
finance?
B
If
you
want
to
contact
the
project
directly
for
interest,
please
contact
the
project
coordinators,
either
on
the
public
side,
with
matthew
at
oaken
or
the
former
league
division
hugo
who's
at
johnson,
and
with
that
we'd
really
like
to
say
thank
you
to
all
of
our
partners
around
the
world.
I
think
for
me
how
many
people
are
actually
involved
in
the
project.
I
think
it's
over
a
hundred
at
this
point,
so
it's
not
just
one
person
who's
making
this
happen.
It's
really
like
the
power
of
the
community.
A
Hello,
so
the
first
question
was
what
is
trustless
possibility,
and
so
I'm
gonna
go
back
to
the
slide.
So
trustless
possibility
is
a
traceability
without
the
need
to
rely
on
on
one
one
organization,
only
meaning
that
the
so
I
think
it
was
in
this
side.
It's
mentioned
here
so
and,
as
you
can
see
on
this
slide,
there
is
no
organization.
I
mean
no
account
on
the
platform
which
is
responsible
for
administrating
all
the
nodes
and
for
storing
a
database
with
all
the
operations
that
are
gonna
occur
on
the
platform.
B
So
each
farmer
partner
is
actually
running
their
own
machines,
they're
completing
completely
separated
aws
accounts,
so
the
only
sharing
is
of
the
model
between
each
of
our
partners,
so
the
actual
compute
power
is
individual
to
each
farmer,
because,
obviously
you
want
to
keep
the
machines
where
the
code
is
actually
running
and
where
the
data
is
actually
stored,
separate
from
each
other.
So
each
one
is
individual.
B
A
A
A
B
A
B
And
then
the
next
one
can
you
provide
examples
of?
I
think
it's
sensitive
and
unsensitive
data,
which
things
like
pharma
industries
sharon,
which
one
they
rather
would
not.
So
in
this
specific
context
of
this
project,
what
they're,
looking
at
and
trying
to
keep
private
is.
They
have
kind
of
like
whole,
like
libraries
of
like
the
activity
of
certain
molecules,
so
they're
going
to
run
whole
because
they.
B
And
they
want
to
know
how
they
react
to
different
antigens
within
your
body,
and
so
they
can
look
at
the
specific
molecules
and
kind
of
like
their
activity
profile,
and
so
that's
kind
of
like
the
sensitive
data
that
they're
trying
to
keep
separate
is
like
how
important
it
or
how
good
is
that
molecule
for
attaching
to
certain
targets
is
right.
If
it's
really
good
at
attaching
through
a
certain
target,
then
it's
a
good
candidate
for
a
drug.
If
it's
a
good
candidate
for
a
drug,
then
you
want
to
patent
it.
B
B
A
A
A
But
but
it's
a
good,
it's
a
difficult.
It's
a
key
expert
and
a
difficult
point.
Then
there
was
a
question:
what
are
the
algorithm
the
same
across
pharma
organizations?
Yes,
they
are
the
same.
So
it's
the
same
algorithms,
but
but
models
which
are
created
then
are
not
the
same
because
they
are
created
not
exactly
with
the
same
data.
B
Kubernetes
and
kubernetes
crds
leverage
any
of
the
existing
open
source,
libraries
and
frameworks.
So
I
think
this
is
relating
to
kubernetes
kubernetes
platform
and
how
the
actual
platform
works,
and
so
what
it
is
using
is
kubernetes
operators
and
actually
all
the
code
is
open
source
on
github.
So
if
you
look
up
chromatic
on
github,
you
can
find
actually
all
the
code
there
for
both
cube
one
and
for
chromatic.
B
A
Okay
and
next
question
is:
do
different
companies
use
different
model
or
same
model
with
different
with
different
parameters,
so
we
call
a
machine
learning
model,
the
output
of
the
training
of
an
algorithm
on
data,
and
and
so
we
use
the
same
hyper,
the
same
hyper
parameters
for
for
the
different
models.
But
we
combine
a
model
that
is
shared
between
the
different
partners
and
also
having
the
same
parameters
that
are
trained
during
the
the
federated
learning
optimization.
B
Companies
will
the
slides
be
available
if
so,
where
I
can
find
them
in
the
session
agenda.
So
I
think
I
did
upload
them
to
the
session
before
the
talk,
but
if
they're
not
there
I'll
make
sure
that
they
are
after
this
talk
or
if
there's
problems
getting
them,
please
feel
free
to
message
me
too,
but
they
should
already
actually
be
on
the
kubecon
platform.