WEBVTT

00:00:00.060 --> 00:00:01.780
The following
content is provided

00:00:01.780 --> 00:00:04.019
under a Creative
Commons license.

00:00:04.019 --> 00:00:06.870
Your support will help MIT
OpenCourseWare continue

00:00:06.870 --> 00:00:10.730
to offer high quality
educational resources for free.

00:00:10.730 --> 00:00:13.340
To make a donation or
view additional materials

00:00:13.340 --> 00:00:17.217
from hundreds of MIT courses,
visit MIT OpenCourseWare

00:00:17.217 --> 00:00:17.842
at ocw.mit.edu.

00:00:26.570 --> 00:00:30.580
DOUG LAUFFENBURGER:
So we shall start.

00:00:30.580 --> 00:00:33.092
I haven't had the pleasure
of meeting most of you.

00:00:33.092 --> 00:00:34.050
I'm Doug Lauffenburger.

00:00:34.050 --> 00:00:40.700
I'm gratefully invited for
a guest presentation here.

00:00:40.700 --> 00:00:44.229
So I'll definitely enjoy it.

00:00:44.229 --> 00:00:45.520
There should be plenty of time.

00:00:45.520 --> 00:00:48.510
I'm not racing through
a lot of material,

00:00:48.510 --> 00:00:52.080
so feel free to interrupt
me with questions.

00:00:52.080 --> 00:00:55.155
And of course I'll try
to respond as best I can.

00:00:57.621 --> 00:00:58.120
OK.

00:00:58.120 --> 00:01:02.460
Who has looked at the
background materials that

00:01:02.460 --> 00:01:06.340
were posted on the web a
long time ago, last night?

00:01:06.340 --> 00:01:08.770
Who already will admit
to having looked at it?

00:01:11.290 --> 00:01:11.790
Good.

00:01:11.790 --> 00:01:13.606
All right.

00:01:13.606 --> 00:01:15.200
I guess that means
I should do this

00:01:15.200 --> 00:01:17.320
because otherwise if
you've read it already then

00:01:17.320 --> 00:01:18.480
there'd be no point, right?

00:01:18.480 --> 00:01:20.550
OK.

00:01:20.550 --> 00:01:21.300
OK.

00:01:21.300 --> 00:01:24.210
Well, where we are
in your semester--

00:01:24.210 --> 00:01:28.290
you're learning a lot of things
across the whole spectrum

00:01:28.290 --> 00:01:30.940
of computational
systems biology.

00:01:30.940 --> 00:01:33.870
I hope I'll add
something in here.

00:01:33.870 --> 00:01:36.620
It's actually a
very specific topic.

00:01:36.620 --> 00:01:40.020
We talk about modeling of
cell signaling networks,

00:01:40.020 --> 00:01:44.490
and in particular, one approach
is worth going through today

00:01:44.490 --> 00:01:47.190
and that's the logic
modeling framework.

00:01:47.190 --> 00:01:51.710
So I'll give you a little bit
of a conceptual background

00:01:51.710 --> 00:01:54.090
for the first 10 or 15 minutes.

00:01:54.090 --> 00:02:00.030
Then we'll launch into
the particular example

00:02:00.030 --> 00:02:02.630
that was in the main paper.

00:02:02.630 --> 00:02:05.900
And a little side light
with an application of it

00:02:05.900 --> 00:02:08.830
to a particular cancer problem.

00:02:08.830 --> 00:02:11.290
And then that should take
us pretty much to the end.

00:02:11.290 --> 00:02:11.790
OK.

00:02:16.060 --> 00:02:16.600
OK.

00:02:16.600 --> 00:02:21.940
The biological topic here
is cell signaling, primarily

00:02:21.940 --> 00:02:23.930
mammalian cells.

00:02:23.930 --> 00:02:27.220
Certainly applicable
to microbial cells

00:02:27.220 --> 00:02:28.990
in a simpler sense.

00:02:28.990 --> 00:02:36.940
So just to place the context,
in mammalian cell biology,

00:02:36.940 --> 00:02:40.430
I'm a bio-engineer and a cell
biologist at the same time.

00:02:40.430 --> 00:02:43.720
We're very interested in what
controls the cell behavior,

00:02:43.720 --> 00:02:45.880
their phenotypic response.

00:02:45.880 --> 00:02:48.680
We know that it's
in fact controlled

00:02:48.680 --> 00:02:51.110
by what it sees in their
environment, growth

00:02:51.110 --> 00:02:55.890
factors, hormones, extracellular
matrix, mechanical forces,

00:02:55.890 --> 00:02:59.320
cell-cell contacts.

00:02:59.320 --> 00:03:03.730
A variety of queues
in the environment

00:03:03.730 --> 00:03:08.670
and the way these govern
phenotype or control phenotype

00:03:08.670 --> 00:03:11.750
is that they influence,
they regulate

00:03:11.750 --> 00:03:14.350
what I would call the
execution processes.

00:03:14.350 --> 00:03:17.960
The crucial execution processes
such as gene expression,

00:03:17.960 --> 00:03:22.620
transcription, and
translation are

00:03:22.620 --> 00:03:24.860
governed by
extracellular factors.

00:03:24.860 --> 00:03:28.830
Metabolism, synthesis
of new molecules,

00:03:28.830 --> 00:03:32.130
cytoskeleton, motors,
forced generation.

00:03:32.130 --> 00:03:34.200
These things all
carry out phenotype

00:03:34.200 --> 00:03:37.980
governed by the extracellular
stimuli or cues.

00:03:37.980 --> 00:03:41.710
And it happens via these
biochemical signaling

00:03:41.710 --> 00:03:45.550
pathways that are activated
primarily by cell surface

00:03:45.550 --> 00:03:49.000
receptors in the plasmid
membrane-- cascades

00:03:49.000 --> 00:03:52.930
of biochemical reactions,
mostly enzymatic.

00:03:52.930 --> 00:03:56.580
Some protein-protein
docking, mostly

00:03:56.580 --> 00:03:58.620
post-translational
modifications.

00:03:58.620 --> 00:04:02.100
Kinase phosphate reactions
adding and taking off

00:04:02.100 --> 00:04:06.020
phosphate groups that change
protein activities at locations

00:04:06.020 --> 00:04:07.119
and so forth.

00:04:07.119 --> 00:04:09.535
It could be other types of
post-translation modifications.

00:04:09.535 --> 00:04:15.800
It could be second messengers,
calcium, ATP aces and so forth.

00:04:15.800 --> 00:04:19.860
So, the extracellular--
now my battery's dead.

00:04:19.860 --> 00:04:20.970
That's not good.

00:04:20.970 --> 00:04:22.079
Oh, there we go.

00:04:22.079 --> 00:04:25.760
Extracellular stimuli,
generate the signals.

00:04:25.760 --> 00:04:29.610
They regulate gene expression,
metabolism, cytoskeleton.

00:04:29.610 --> 00:04:32.770
They carry out phenotype.

00:04:32.770 --> 00:04:33.270
OK.

00:04:36.060 --> 00:04:41.840
So we want to learn about cell
signaling network operations.

00:04:41.840 --> 00:04:43.790
There's actually multiple
pathways involved.

00:04:43.790 --> 00:04:46.180
We really need to
study many of them

00:04:46.180 --> 00:04:49.380
in concert to understand
what the cells are doing.

00:04:49.380 --> 00:04:52.680
And a big question is,
what kind of information

00:04:52.680 --> 00:04:54.630
do we need to study this?

00:04:54.630 --> 00:04:56.880
And on the end, we'd
be interested in how

00:04:56.880 --> 00:05:01.750
phenotypic behavior does arise
from variations and mutations

00:05:01.750 --> 00:05:05.270
in the genomic content of cells.

00:05:05.270 --> 00:05:10.050
But that genomic content,
of course, is not modified,

00:05:10.050 --> 00:05:12.630
but its effects are
influenced by what's

00:05:12.630 --> 00:05:15.610
the environment to these
extracellular cues,

00:05:15.610 --> 00:05:17.480
log ins and so forth.

00:05:17.480 --> 00:05:22.070
So they influence what
message is expressed.

00:05:22.070 --> 00:05:23.700
From that message
they influence what's

00:05:23.700 --> 00:05:25.630
actually translated
into protein.

00:05:25.630 --> 00:05:27.230
From those proteins
they influence

00:05:27.230 --> 00:05:28.730
the post-translational
modifications

00:05:28.730 --> 00:05:31.420
and what the proteins
are actually doing.

00:05:31.420 --> 00:05:34.280
And so, in the
end, the phenotype

00:05:34.280 --> 00:05:37.210
is carried out by these
protein operations.

00:05:37.210 --> 00:05:39.430
And the question
is, what information

00:05:39.430 --> 00:05:41.880
level that we might
want to study.

00:05:41.880 --> 00:05:44.250
And of course you would
love to have the information

00:05:44.250 --> 00:05:47.310
content at all levels--
genomic information,

00:05:47.310 --> 00:05:50.114
transcriptional information,
translational information,

00:05:50.114 --> 00:05:51.405
post-translational information.

00:05:54.180 --> 00:05:57.570
So integrating all those
different data levels

00:05:57.570 --> 00:05:59.232
can be extremely valuable.

00:05:59.232 --> 00:06:01.440
In terms of the models I'm
going to talk about today,

00:06:01.440 --> 00:06:03.610
they've essentially
been living at the level

00:06:03.610 --> 00:06:07.880
of protein activities in
these signaling pathways.

00:06:07.880 --> 00:06:08.380
OK.

00:06:08.380 --> 00:06:09.796
That will be the
kind of data sets

00:06:09.796 --> 00:06:13.310
you'll see that will be analyzed
with respect to the models.

00:06:13.310 --> 00:06:18.540
Obviously, they arise from
these underlying mechanisms

00:06:18.540 --> 00:06:22.440
that, as influenced by
the environmental context,

00:06:22.440 --> 00:06:25.070
altering the signaling
protein activities.

00:06:27.790 --> 00:06:28.290
OK.

00:06:28.290 --> 00:06:31.930
And what's very
interesting and there's

00:06:31.930 --> 00:06:36.170
going to be more and more
progress in the coming years

00:06:36.170 --> 00:06:39.520
is relating what's in the
genomic information-- mutations

00:06:39.520 --> 00:06:42.810
and variations to what's
happening at the protein level.

00:06:42.810 --> 00:06:45.487
And some of the other
instructors in this class

00:06:45.487 --> 00:06:47.070
are really some of
the world's experts

00:06:47.070 --> 00:06:48.640
in figuring out how to do this.

00:06:48.640 --> 00:06:50.750
I'd like to just
show this example

00:06:50.750 --> 00:06:54.830
as a motivation for
this kind of approach.

00:06:54.830 --> 00:06:58.890
And that is if you do gene
sequencing of many patient

00:06:58.890 --> 00:07:00.550
tumors-- in this
case, I believe this

00:07:00.550 --> 00:07:02.539
was a paper on
pancreatic tumors.

00:07:02.539 --> 00:07:05.080
This has been shown for pretty
much every other type of tumor

00:07:05.080 --> 00:07:06.190
since then.

00:07:06.190 --> 00:07:09.740
In any given patient tumor,
each one of these bars,

00:07:09.740 --> 00:07:13.340
there's dozens of
mutations in each tumor.

00:07:13.340 --> 00:07:15.850
And a variety of
types-- deletions,

00:07:15.850 --> 00:07:19.140
amplifications mutations
and by and large, they're

00:07:19.140 --> 00:07:20.710
all different.

00:07:20.710 --> 00:07:22.780
There's very few
mutations themselves

00:07:22.780 --> 00:07:25.515
that really carry over to
a substantial proportion

00:07:25.515 --> 00:07:27.770
of one patient's
tumor to another.

00:07:27.770 --> 00:07:31.160
There's some special cases
that are fairly pervasive,

00:07:31.160 --> 00:07:34.210
but the predominant of these
dozens and dozens of mutations

00:07:34.210 --> 00:07:39.190
and variations are
different from one patient

00:07:39.190 --> 00:07:43.020
to another, and even
in the same patient.

00:07:43.020 --> 00:07:46.260
So what's emerging
as a productive way

00:07:46.260 --> 00:07:51.220
to think about this-- How
do all these different types

00:07:51.220 --> 00:07:52.990
of mutations and
specific mutations

00:07:52.990 --> 00:07:56.950
actually lead to classes
of similar pathologies?

00:07:56.950 --> 00:07:59.800
And that is they
tend to reside in

00:07:59.800 --> 00:08:05.362
what can be identified
as pathways-- circuits,

00:08:05.362 --> 00:08:07.320
machines, things that
are actually carrying out

00:08:07.320 --> 00:08:09.220
function at the protein level.

00:08:09.220 --> 00:08:12.150
So for instance-- I'm
losing this again.

00:08:12.150 --> 00:08:14.510
For these pancreatic
cancers on this wheel

00:08:14.510 --> 00:08:18.430
are about a dozen different
signaling pathways

00:08:18.430 --> 00:08:20.720
and self-cycle control
pathways and apoptosis

00:08:20.720 --> 00:08:22.790
controlled pathways.

00:08:22.790 --> 00:08:26.110
And if you look at any
individual patient tumors,

00:08:26.110 --> 00:08:29.505
like this green one or this red
one-- two different patients.

00:08:29.505 --> 00:08:33.409
If you actually look at the
mutations at the genomic level,

00:08:33.409 --> 00:08:36.250
they're entirely different
in the green patient

00:08:36.250 --> 00:08:38.169
tumor versus the
red patient tumor.

00:08:38.169 --> 00:08:40.919
So if you're just trying
to match gene mutation

00:08:40.919 --> 00:08:43.171
to pancreatic cancer,
these two patients

00:08:43.171 --> 00:08:44.420
would look entirely different.

00:08:46.960 --> 00:08:50.630
But, it turns out, that you
can line up their mutations

00:08:50.630 --> 00:08:53.260
into the same
pathways and say, OK,

00:08:53.260 --> 00:08:54.660
the red tumor and
the green tumor

00:08:54.660 --> 00:08:58.515
both have mutations that
affect the TGF beta pathway.

00:08:58.515 --> 00:09:00.390
They're different
mutations, but they've just

00:09:00.390 --> 00:09:02.090
regulated that pathway.

00:09:02.090 --> 00:09:04.740
And similarly, you can
do that with pretty much

00:09:04.740 --> 00:09:06.610
all of the other mutations.

00:09:06.610 --> 00:09:09.380
That these tumors
have been dysregulated

00:09:09.380 --> 00:09:11.740
in terms of particular pathways.

00:09:11.740 --> 00:09:13.350
But patient to
patient to patient,

00:09:13.350 --> 00:09:16.960
it's happened by different
genomic gene sequence

00:09:16.960 --> 00:09:18.190
mutations.

00:09:18.190 --> 00:09:21.550
So that the ability to
look at these protein

00:09:21.550 --> 00:09:23.780
level pathways is
a way of making

00:09:23.780 --> 00:09:28.170
really good productive sense
of the gene sequencing data.

00:09:28.170 --> 00:09:31.560
So there's lots of labs trying
to go from gene sequence

00:09:31.560 --> 00:09:33.815
up to pathway modulation.

00:09:33.815 --> 00:09:37.760
In our case, we're not
going to show you that here.

00:09:37.760 --> 00:09:39.600
We're going to say,
this is a motivation

00:09:39.600 --> 00:09:42.940
for starting at
the protein level.

00:09:42.940 --> 00:09:45.450
And I'd like to show
this picture too.

00:09:45.450 --> 00:09:49.120
Number one because it's
such an anachronism.

00:09:49.120 --> 00:09:52.000
This is a circuit board from
decades and decades and decades

00:09:52.000 --> 00:09:54.680
ago that none of
you would recognize.

00:09:54.680 --> 00:10:00.560
But, in the molecular biology
world, this kind of a picture,

00:10:00.560 --> 00:10:04.760
and in its modern form is viewed
as a very appealing metaphor

00:10:04.760 --> 00:10:07.390
for how to think
about these signaling

00:10:07.390 --> 00:10:10.830
pathways and signaling networks
that take the extracellular

00:10:10.830 --> 00:10:14.410
information and turn it into
governance of transcription,

00:10:14.410 --> 00:10:17.030
metabolism, cytoskeleton,
and phenotype.

00:10:17.030 --> 00:10:19.280
So, just this
metaphor of circuitry,

00:10:19.280 --> 00:10:21.360
where in white, the
extracellular ligands,

00:10:21.360 --> 00:10:24.710
growth factors are
somehow wired to the blue.

00:10:24.710 --> 00:10:28.380
The cell surface receptors,
or B for instance--

00:10:28.380 --> 00:10:30.310
they're wired too.

00:10:30.310 --> 00:10:32.920
Kinases and other
signaling proteins--

00:10:32.920 --> 00:10:37.050
they're wired to transcription
factors, self-cycle control

00:10:37.050 --> 00:10:40.990
regulators,
apoptosis regulators.

00:10:40.990 --> 00:10:44.920
So these very famous folks
in cancer biology say,

00:10:44.920 --> 00:10:47.580
what you've got to understand
is, these signalling networks

00:10:47.580 --> 00:10:49.470
as circuitry.

00:10:49.470 --> 00:10:51.740
And if the circuitry is
dysregulated somehow,

00:10:51.740 --> 00:10:54.450
the wiring is different,
then that's what's

00:10:54.450 --> 00:10:58.910
underlying malignant behavior.

00:10:58.910 --> 00:11:01.160
So, this is really
beautiful but it's

00:11:01.160 --> 00:11:02.800
pretty much useless, right.

00:11:02.800 --> 00:11:07.620
Because there's no prediction or
calculation or even hypothesis

00:11:07.620 --> 00:11:10.650
generation one can do
from a picture like this.

00:11:10.650 --> 00:11:14.310
Yes, it's circuitry, but
what do I do with it?

00:11:14.310 --> 00:11:16.110
So, what I want
to show you today

00:11:16.110 --> 00:11:18.040
are efforts to
turn them into what

00:11:18.040 --> 00:11:21.470
I would call an actionable
model, a computable model.

00:11:21.470 --> 00:11:23.789
Yes, it looks kind
of like circuitry,

00:11:23.789 --> 00:11:26.080
but in fact you would know
how to do a calculation that

00:11:26.080 --> 00:11:29.370
would fit it to data
and predict new data.

00:11:29.370 --> 00:11:32.540
And then you have, in fact, a
model rather than a metaphor.

00:11:32.540 --> 00:11:35.200
That's the idea.

00:11:35.200 --> 00:11:37.910
So, one question is, if
you want to turn that

00:11:37.910 --> 00:11:43.750
into a formal mathematical
framework for circuitry

00:11:43.750 --> 00:11:46.530
that you can calculate--
what kind of mathematics

00:11:46.530 --> 00:11:47.190
might you use?

00:11:47.190 --> 00:11:48.565
And in this class
you're learning

00:11:48.565 --> 00:11:51.730
a whole spectrum of things.

00:11:51.730 --> 00:11:54.930
And one can think
about it on one hand,

00:11:54.930 --> 00:11:56.960
if we knew all of
those components

00:11:56.960 --> 00:12:01.790
and how they interacted, and
could estimate rate constance

00:12:01.790 --> 00:12:04.340
and so forth, we could
write differential equations

00:12:04.340 --> 00:12:08.240
for maybe the dozens and dozens
of components and interactions

00:12:08.240 --> 00:12:11.690
and predict how they would
play out dynamically with time.

00:12:11.690 --> 00:12:14.440
For most systems with
the complexity that's

00:12:14.440 --> 00:12:17.280
really controlling cell
biology, at this point in time,

00:12:17.280 --> 00:12:18.990
this is almost impossible.

00:12:18.990 --> 00:12:21.530
There's only rare
cases where enough

00:12:21.530 --> 00:12:24.290
is known about
signaling biochemistry

00:12:24.290 --> 00:12:26.800
to really write
down differential

00:12:26.800 --> 00:12:29.690
equations for what's going on.

00:12:29.690 --> 00:12:31.570
At the other extreme,
of course, is the type

00:12:31.570 --> 00:12:35.830
of mathematics one gets out
of very, very large data sets,

00:12:35.830 --> 00:12:40.190
sequencing data sets,
transcriptional, and so forth.

00:12:40.190 --> 00:12:42.510
More informatics
type of analysis,

00:12:42.510 --> 00:12:44.990
where it has to do
with multivariate

00:12:44.990 --> 00:12:50.290
regression and clustering,
mutual information.

00:12:50.290 --> 00:12:55.050
And what we've been working on
is someplace up in the middle

00:12:55.050 --> 00:12:58.530
where you don't have enough
mechanistic prior knowledge

00:12:58.530 --> 00:13:02.360
to write this formal
of physics, and yet

00:13:02.360 --> 00:13:06.640
takes you someplace beyond
statistical associations.

00:13:06.640 --> 00:13:08.710
And this is one
of the areas that

00:13:08.710 --> 00:13:12.470
might be worth your
learning in this class.

00:13:12.470 --> 00:13:15.720
OK, this is really the same
set of computational methods,

00:13:15.720 --> 00:13:20.130
just like it's cast in a
little bit different form that

00:13:20.130 --> 00:13:23.380
delineates competition
modeling, really

00:13:23.380 --> 00:13:26.430
into two kinds of classes.

00:13:26.430 --> 00:13:30.070
What's traditionally appreciated
in most fields of engineering

00:13:30.070 --> 00:13:32.330
and physics are
differential equations

00:13:32.330 --> 00:13:33.910
that are very theory driven.

00:13:33.910 --> 00:13:34.810
You have a theory.

00:13:34.810 --> 00:13:37.010
You have prior knowledge
for what's happening.

00:13:37.010 --> 00:13:39.140
You're writing down the
components involved,

00:13:39.140 --> 00:13:41.570
you're writing down
how they interact.

00:13:41.570 --> 00:13:43.780
And typically,
algebraic equations

00:13:43.780 --> 00:13:47.960
for those differential
equations describe your theory,

00:13:47.960 --> 00:13:49.960
describe your prior knowledge.

00:13:49.960 --> 00:13:53.260
And now it's formalized and
you estimate rate constants

00:13:53.260 --> 00:13:55.020
and so forth.

00:13:55.020 --> 00:13:56.520
Another whole class
of information

00:13:56.520 --> 00:13:58.640
is data driven, in
which, you really

00:13:58.640 --> 00:14:01.370
don't have a good theory
about what components matter

00:14:01.370 --> 00:14:03.350
and how they interact.

00:14:03.350 --> 00:14:07.450
And so you start with
data sets and from it

00:14:07.450 --> 00:14:11.970
you do classification or
typologies or associations

00:14:11.970 --> 00:14:15.150
with different types of
mathematics that at least try

00:14:15.150 --> 00:14:18.290
to make sense and get hypotheses
out of these large data

00:14:18.290 --> 00:14:21.240
sets, where you don't
have any theory.

00:14:21.240 --> 00:14:25.440
One reason that logic
modeling appeals to me,

00:14:25.440 --> 00:14:26.960
is that it actually
can be applied

00:14:26.960 --> 00:14:31.210
in either the theory driven
or the data driven mode.

00:14:31.210 --> 00:14:34.770
You can say, I know
nothing about my system.

00:14:34.770 --> 00:14:39.370
I just generate large data sets
of signaling network activities

00:14:39.370 --> 00:14:42.000
induced by different
stimuli, but I'm

00:14:42.000 --> 00:14:44.260
going to try to fit a
logic model to it that

00:14:44.260 --> 00:14:46.620
says how the different
components influencing

00:14:46.620 --> 00:14:50.120
each other in a logic way.

00:14:50.120 --> 00:14:52.530
Or, you could say,
well, I know something.

00:14:52.530 --> 00:14:54.190
I have some prior knowledge.

00:14:54.190 --> 00:14:59.800
I may have interact ohms the say
what molecular components are

00:14:59.800 --> 00:15:02.390
present in signaling networks.

00:15:02.390 --> 00:15:05.690
And so in principle, I
kind of know who's involved

00:15:05.690 --> 00:15:07.970
and who might be
influencing whom.

00:15:07.970 --> 00:15:12.550
And I could write a logic model
based on that prior knowledge.

00:15:12.550 --> 00:15:15.170
And then run calculations
and see if it actually

00:15:15.170 --> 00:15:17.670
makes predictions about
experimental data.

00:15:17.670 --> 00:15:18.830
So that's one nice thing.

00:15:18.830 --> 00:15:21.800
It's a mathematical
formalism that

00:15:21.800 --> 00:15:25.970
can either be run in data driven
mode or in theory driven mode

00:15:25.970 --> 00:15:27.390
and go back and forth.

00:15:27.390 --> 00:15:30.665
So that's one reason--
given one lecture to offer,

00:15:30.665 --> 00:15:34.550
I've decided to offer
it on this topic.

00:15:34.550 --> 00:15:35.840
All right, with me so far?

00:15:35.840 --> 00:15:37.360
Any questions?

00:15:37.360 --> 00:15:37.860
Philosophy?

00:15:41.250 --> 00:15:41.830
OK.

00:15:41.830 --> 00:15:46.490
So, what we're
going to do today is

00:15:46.490 --> 00:15:50.510
almost take a
hybrid of these two.

00:15:50.510 --> 00:15:54.330
We're going to say, what
prior knowledge do we have,

00:15:54.330 --> 00:15:57.360
and then recognize that
it's really not enough.

00:15:57.360 --> 00:16:02.310
And so how do we now integrate
that with empirical data

00:16:02.310 --> 00:16:05.260
to now come up with
logic modeling that,

00:16:05.260 --> 00:16:08.640
in fact, is actionable
and computable?

00:16:08.640 --> 00:16:09.140
OK.

00:16:09.140 --> 00:16:12.704
So what kind of prior
knowledge do we have?

00:16:12.704 --> 00:16:14.870
Let's say we wanted to have
a logic model for what's

00:16:14.870 --> 00:16:17.400
in these signaling networks
down stream of growth factor

00:16:17.400 --> 00:16:21.620
receptors, or hormone receptors,
or things like that, that then

00:16:21.620 --> 00:16:26.030
govern gene expression,
metabolism and so forth.

00:16:26.030 --> 00:16:28.250
What prior knowledge do we have?

00:16:28.250 --> 00:16:30.129
And you folks
probably have already

00:16:30.129 --> 00:16:31.420
seen some of this in the class.

00:16:35.264 --> 00:16:36.930
There's all kinds of
databases of stuff.

00:16:36.930 --> 00:16:39.180
What's in those databases
that might be relevant here?

00:16:42.054 --> 00:16:45.547
AUDIENCE: [INAUDIBLE] that the
protein-protein interactions--

00:16:45.547 --> 00:16:48.425
if you switch proteins, they
interact with each other,

00:16:48.425 --> 00:16:50.550
but maybe not necessarily
what pathways they're in.

00:16:50.550 --> 00:16:51.758
DOUG LAUFFENBURGER: OK, good.

00:16:51.758 --> 00:16:55.506
And have you seen
databases like that?

00:16:55.506 --> 00:16:57.130
AUDIENCE: Several of
them have come up.

00:16:57.130 --> 00:16:57.630
DOUG LAUFFENBURGER: OK.

00:16:57.630 --> 00:16:59.580
Are you the only
one who's seen them?

00:16:59.580 --> 00:17:01.306
Or is there anybody
else that kind

00:17:01.306 --> 00:17:03.426
of noticed them in passing too?

00:17:03.426 --> 00:17:04.099
OK, good.

00:17:04.099 --> 00:17:06.490
Second, third,
fourth, all right.

00:17:06.490 --> 00:17:10.480
That's a critical mass
if I ever saw one.

00:17:10.480 --> 00:17:11.010
OK.

00:17:11.010 --> 00:17:16.980
So, I'm just going
to allude to those.

00:17:16.980 --> 00:17:20.520
So there are pathway databases.

00:17:20.520 --> 00:17:23.282
And this is actually an
old slide of a few years

00:17:23.282 --> 00:17:25.240
ago, so I'm sure the
numbers are all different.

00:17:25.240 --> 00:17:26.640
And, in fact, there's new ones.

00:17:26.640 --> 00:17:29.080
I just haven't taken
to updating the slide.

00:17:29.080 --> 00:17:33.540
But we'll, based on literature,
take certain numbers

00:17:33.540 --> 00:17:36.420
of gene products, a
few hundred of them,

00:17:36.420 --> 00:17:41.760
and organize them into pathways
based on biological knowledge.

00:17:41.760 --> 00:17:45.440
There's other databases that
are more interactomes, usually

00:17:45.440 --> 00:17:49.230
based on other kinds of
experimental data-- yeast II

00:17:49.230 --> 00:17:53.280
hybrid, mass spectrometry,
literature curation,

00:17:53.280 --> 00:17:58.610
that also tries to say who's
physically interacting.

00:17:58.610 --> 00:18:01.610
So these node-- these pathway
databases don't necessarily

00:18:01.610 --> 00:18:04.080
say, somebody's
physically interacting,

00:18:04.080 --> 00:18:06.890
they say somebody might
be upstream and downstream

00:18:06.890 --> 00:18:08.580
and so forth.

00:18:08.580 --> 00:18:10.060
And then they
interactome databases

00:18:10.060 --> 00:18:12.632
say, component a
and component b,

00:18:12.632 --> 00:18:15.090
there's some evidence that they
have a physical association

00:18:15.090 --> 00:18:16.230
someplace along the way.

00:18:16.230 --> 00:18:18.720
So these are two
complementary types

00:18:18.720 --> 00:18:25.280
of databases that, in
fact, can be put together.

00:18:25.280 --> 00:18:26.190
OK.

00:18:26.190 --> 00:18:29.360
So an interesting
thing about these--

00:18:29.360 --> 00:18:31.860
there's a number
of these databases.

00:18:31.860 --> 00:18:34.200
And so in principle
you could say, well

00:18:34.200 --> 00:18:37.110
if I want then to start-- if
I want to generate a logic

00:18:37.110 --> 00:18:40.219
model for signaling
networks, all

00:18:40.219 --> 00:18:42.010
I have to do is take
what's in the database

00:18:42.010 --> 00:18:43.890
and say what pathways
are there and what's

00:18:43.890 --> 00:18:46.190
known with their
interactions, and now I've

00:18:46.190 --> 00:18:47.645
got a starting point.

00:18:47.645 --> 00:18:49.660
You know, I can
actually draw a graph

00:18:49.660 --> 00:18:55.750
with lots of molecular nodes and
lots of molecular interactions.

00:18:55.750 --> 00:18:57.680
So, you can do that.

00:18:57.680 --> 00:19:01.240
And so you can choose
one of these databases

00:19:01.240 --> 00:19:03.450
and say I'm going
to draw a graph that

00:19:03.450 --> 00:19:07.890
has what's believed to be
true about nodes and pathways

00:19:07.890 --> 00:19:10.730
and interactions and
signaling networks.

00:19:10.730 --> 00:19:13.870
But then you choose a different
database and another database.

00:19:13.870 --> 00:19:16.810
And you'll actually get
different information.

00:19:16.810 --> 00:19:17.310
OK.

00:19:17.310 --> 00:19:19.143
We actually did a study
on this-- I probably

00:19:19.143 --> 00:19:21.320
should have given you the
citation of that-- that

00:19:21.320 --> 00:19:23.770
said if you look at six or
seven of these databases,

00:19:23.770 --> 00:19:25.470
they are not coincident.

00:19:25.470 --> 00:19:28.770
They have a very
small intersections.

00:19:28.770 --> 00:19:32.970
Most of their information
is non-redundant.

00:19:32.970 --> 00:19:35.189
And so you could try
to put it all together.

00:19:35.189 --> 00:19:36.730
And we did this,
again, in this paper

00:19:36.730 --> 00:19:39.760
that I'm not giving
you a citation for.

00:19:39.760 --> 00:19:42.420
And so here's a number of
nodes and signaling pathways

00:19:42.420 --> 00:19:46.680
downstream of receptors.

00:19:46.680 --> 00:19:50.940
And all the colored
nodes are those

00:19:50.940 --> 00:19:54.120
in which they appear in
only one of these-- one, two

00:19:54.120 --> 00:19:57.460
three, four, five,
six databases.

00:19:57.460 --> 00:20:00.580
So if something's colored
green, it's only in GeneGo

00:20:00.580 --> 00:20:02.175
and it's not in
any of the others.

00:20:02.175 --> 00:20:03.710
If something's
colored purple, it's

00:20:03.710 --> 00:20:06.180
in PANTHER and
none of the others.

00:20:06.180 --> 00:20:07.240
OK.

00:20:07.240 --> 00:20:09.280
If they're gray-- some
of these gray ones,

00:20:09.280 --> 00:20:11.630
they're in at least two.

00:20:11.630 --> 00:20:15.270
But out of these six, there's
an exceedingly small number

00:20:15.270 --> 00:20:18.360
of nodes interactions that
are in all six databases.

00:20:18.360 --> 00:20:22.620
Which was a real surprise
to us when we did this.

00:20:22.620 --> 00:20:24.080
So what this means
is, if you want

00:20:24.080 --> 00:20:27.400
to start with some
prior knowledge graph

00:20:27.400 --> 00:20:30.090
that you're now going to fit
a logic model to by mapping it

00:20:30.090 --> 00:20:32.540
against data, you first
even have the choice, well,

00:20:32.540 --> 00:20:33.790
what am I going to start with?

00:20:33.790 --> 00:20:35.520
What is my prior knowledge?

00:20:35.520 --> 00:20:38.540
There;s not really
consensus prior knowledge.

00:20:38.540 --> 00:20:42.810
So you can start with six
different interaction graphs.

00:20:42.810 --> 00:20:44.530
Or you could try to
put them all together

00:20:44.530 --> 00:20:46.570
and get a consensus graph.

00:20:46.570 --> 00:20:48.600
So you have all these choices.

00:20:48.600 --> 00:20:50.860
And right now, it's
not as if is there's

00:20:50.860 --> 00:20:53.640
detailed analysis of what
the best choice would

00:20:53.640 --> 00:20:57.150
be for your starting point.

00:20:57.150 --> 00:20:59.990
But I want to stress that,
with respect to our approach,

00:20:59.990 --> 00:21:04.090
this is a starting
point because one

00:21:04.090 --> 00:21:07.170
of the issues with the
database information

00:21:07.170 --> 00:21:10.990
is that it's typically
very diverse with respect

00:21:10.990 --> 00:21:12.560
to contact.

00:21:12.560 --> 00:21:15.020
What cell type did this
information come from?

00:21:15.020 --> 00:21:18.310
What treatment conditions
did it come from?

00:21:18.310 --> 00:21:21.240
If there's different cell
types, different species,

00:21:21.240 --> 00:21:22.780
different mutations.

00:21:22.780 --> 00:21:28.080
So if I see interactions or
if I don't see interactions,

00:21:28.080 --> 00:21:29.040
are they in conflict?

00:21:29.040 --> 00:21:31.650
Or they're just-- this
one was in a lymphocyte,

00:21:31.650 --> 00:21:33.330
this one was in a
hypatocye, this one

00:21:33.330 --> 00:21:38.500
was in a cardiac myocyte, and
they're actually different.

00:21:38.500 --> 00:21:42.380
OK, so if I had a cell
type specific database,

00:21:42.380 --> 00:21:44.520
or pulled that information
out, that would be good.

00:21:44.520 --> 00:21:46.202
It would be a smaller
number of things.

00:21:46.202 --> 00:21:47.910
But then under what
treatment conditions?

00:21:47.910 --> 00:21:50.795
Because remember I said starting
with the genomic content, what

00:21:50.795 --> 00:21:52.920
you actually see in terms
of molecular interactions

00:21:52.920 --> 00:21:55.580
will be very strongly
affected by what

00:21:55.580 --> 00:21:57.200
matrix were the
cells growing on?

00:21:57.200 --> 00:21:58.820
Or was this in vivo?

00:21:58.820 --> 00:22:03.750
Was this in a multicellular
culture situation?

00:22:03.750 --> 00:22:06.610
So, that's why this
is a starting point

00:22:06.610 --> 00:22:09.880
and can't really
be used to describe

00:22:09.880 --> 00:22:14.279
any particular experimental
situation with much confidence.

00:22:14.279 --> 00:22:15.695
The other thing--
and this is what

00:22:15.695 --> 00:22:17.290
I've been trying to
emphasize from the start-- is

00:22:17.290 --> 00:22:19.320
that there's no calculation
you can do on this.

00:22:22.380 --> 00:22:27.260
There's a group of
folks in this field who

00:22:27.260 --> 00:22:31.270
propose some ideas that I
think are very intriguing,

00:22:31.270 --> 00:22:33.000
but which, at least
to me personally,

00:22:33.000 --> 00:22:35.450
there's not that
much evidence for.

00:22:35.450 --> 00:22:41.040
And that is, that there's
topological characteristics

00:22:41.040 --> 00:22:45.520
of these graphs, that then
tell you what's important.

00:22:45.520 --> 00:22:47.330
So if I have a
node that's somehow

00:22:47.330 --> 00:22:49.760
connected to more
other nodes, that

00:22:49.760 --> 00:22:51.630
is going to be a
more important node,

00:22:51.630 --> 00:22:54.380
and might be associated with the
disease, versus a node that's

00:22:54.380 --> 00:22:55.930
connected to fewer.

00:22:55.930 --> 00:22:56.430
OK.

00:22:56.430 --> 00:22:59.810
Some of these are very, very
appealing ideas conceptually.

00:22:59.810 --> 00:23:03.740
If you actually look for
the experimental evidence

00:23:03.740 --> 00:23:07.450
that they're valid
notions, it's very thin.

00:23:07.450 --> 00:23:09.830
But, that's where some
folks would claim,

00:23:09.830 --> 00:23:12.950
oh, you can do predictions
on the hypotheses based

00:23:12.950 --> 00:23:15.990
on these graphs because
there are these graph theory

00:23:15.990 --> 00:23:17.630
characteristics
that somehow might

00:23:17.630 --> 00:23:20.480
be biologically meaningful.

00:23:20.480 --> 00:23:21.090
OK.

00:23:21.090 --> 00:23:24.240
But I'd say, jury's out on
whether, in fact, any of that

00:23:24.240 --> 00:23:26.130
is true.

00:23:26.130 --> 00:23:30.690
So, our view is-- OK, this
is a good starting point,

00:23:30.690 --> 00:23:33.960
but in fact, needs to be
mapped to empirical data

00:23:33.960 --> 00:23:39.540
in order to gain confidence
about calculations you can do.

00:23:39.540 --> 00:23:44.200
So that's the goal of
this kind of approach,

00:23:44.200 --> 00:23:48.870
is to say, let's
stipulate that we start

00:23:48.870 --> 00:23:51.690
with some prior
knowledge scaffold.

00:23:51.690 --> 00:23:54.710
This particular one is from
the Ingenuity database.

00:23:54.710 --> 00:23:56.850
You could get one from
any other database.

00:23:56.850 --> 00:24:00.160
You could get a consensus one
from three or four if you want.

00:24:00.160 --> 00:24:03.590
And so it has, up here,
extracellular stimuli,

00:24:03.590 --> 00:24:06.200
growth factors, cytokines.

00:24:06.200 --> 00:24:11.150
They're connected in their
interactome II receptors.

00:24:11.150 --> 00:24:14.810
They're connected to scaffolding
proteins and signaling proteins

00:24:14.810 --> 00:24:17.110
and kinases and so forth.

00:24:17.110 --> 00:24:19.520
They're connected to
transcription factors,

00:24:19.520 --> 00:24:20.970
metabolic enzymes.

00:24:20.970 --> 00:24:22.880
So you can draw this graph.

00:24:22.880 --> 00:24:25.675
Say this might be what's
going on in my cell.

00:24:25.675 --> 00:24:27.550
And then what we'd like
to do is to turn this

00:24:27.550 --> 00:24:31.970
into a formal logic
framework that's

00:24:31.970 --> 00:24:36.340
capable of then fitting
experimental data,

00:24:36.340 --> 00:24:38.380
predicting new
experimental data,

00:24:38.380 --> 00:24:41.990
and giving you a chance
at biological hypothesis

00:24:41.990 --> 00:24:43.512
and testing.

00:24:43.512 --> 00:24:46.250
All right, so
conceptually you get it?

00:24:46.250 --> 00:24:50.240
Two aspects-- some kind of
starting prior knowledge,

00:24:50.240 --> 00:24:53.230
that's kind of a scaffold,
a graph, for your network.

00:24:53.230 --> 00:24:56.450
And now you're going to turn it
into a computable logic model

00:24:56.450 --> 00:25:00.720
by mapping it against
empirical data.

00:25:00.720 --> 00:25:04.960
So, merely what it takes is
the kind of conceptual diagram

00:25:04.960 --> 00:25:08.280
you see in any cell biology
paper, any signaling paper,

00:25:08.280 --> 00:25:13.140
that says, well, a and b
both influence e positively,

00:25:13.140 --> 00:25:17.880
and b influences f negatively,
and c influences f positively.

00:25:17.880 --> 00:25:19.780
Then there's a
feedback from g to a.

00:25:19.780 --> 00:25:20.980
That's inhibitory.

00:25:20.980 --> 00:25:22.800
You can draw those.

00:25:22.800 --> 00:25:27.770
But now, how do you turn it
into a computable algorithm?

00:25:27.770 --> 00:25:30.610
So, what I'm going to spend
most of the day on is,

00:25:30.610 --> 00:25:33.680
just conversion of
this to a Boolean logic

00:25:33.680 --> 00:25:38.870
model that any one of these
interactions is and-- a and b

00:25:38.870 --> 00:25:41.870
being active makes e active.

00:25:41.870 --> 00:25:44.770
c being active, but
b not being active,

00:25:44.770 --> 00:25:47.220
allows f to be
active, and so forth.

00:25:47.220 --> 00:25:49.320
You turn these into
formal logic statements

00:25:49.320 --> 00:25:50.640
that you can compute on.

00:25:50.640 --> 00:25:52.720
At the very end,
if we have time,

00:25:52.720 --> 00:25:55.880
I'll show how to relax this from
a Boolean framework that's just

00:25:55.880 --> 00:25:59.450
on off, to something that
can be more quantitative.

00:26:02.490 --> 00:26:02.990
All right.

00:26:02.990 --> 00:26:04.050
So that's the notion.

00:26:04.050 --> 00:26:07.400
Now what I'm going to do
for the rest of the time

00:26:07.400 --> 00:26:10.820
is go through the specific
example paper that says, OK,

00:26:10.820 --> 00:26:13.665
how do we in fact do this?

00:26:13.665 --> 00:26:16.440
What is a way to
accomplish this?

00:26:16.440 --> 00:26:19.050
So now let's go back
to a biological problem

00:26:19.050 --> 00:26:21.686
where there's going to be
empirical, experimental data

00:26:21.686 --> 00:26:23.310
that we're now going
to map against one

00:26:23.310 --> 00:26:27.070
of these prior knowledge
interactome graphs.

00:26:27.070 --> 00:26:30.080
This particular study-- this was
done with Peter Sorger, who's

00:26:30.080 --> 00:26:35.380
now at Harvard Medical School--
had to do with liver cells.

00:26:35.380 --> 00:26:39.720
Liver cancer-- you'll see
some application of that

00:26:39.720 --> 00:26:44.450
at the end-- that says we
have liver cell hepatocytes.

00:26:44.450 --> 00:26:47.780
And we want to know how they
respond to different growth

00:26:47.780 --> 00:26:50.770
factors, in cytokines
in their environment.

00:26:50.770 --> 00:26:52.960
How that'll change their
proliferation or death?

00:26:52.960 --> 00:26:56.600
Or the inflammatory
cytokines that they produce.

00:26:56.600 --> 00:26:59.600
And we'd like to take-- this is
just a pictorial diagram that

00:26:59.600 --> 00:27:02.250
could be in any
cell biology paper,

00:27:02.250 --> 00:27:04.200
and make this calculable.

00:27:04.200 --> 00:27:09.040
So we could say what's
different from a primary normal

00:27:09.040 --> 00:27:11.470
hepatocyte liver cell
that's not cancerous?

00:27:11.470 --> 00:27:13.780
It might have a signaling logic.

00:27:13.780 --> 00:27:15.530
But if then we compare
the signaling logic

00:27:15.530 --> 00:27:20.030
to a liver tumor cell type, or
four different liver tumor cell

00:27:20.030 --> 00:27:21.815
types, what's different?

00:27:21.815 --> 00:27:25.270
If we can find some logic that's
different for the tumor cell

00:27:25.270 --> 00:27:29.490
lines versus the normal primary
lines-- some logic from here

00:27:29.490 --> 00:27:32.610
to there or to there-- that now
tells you biologically, where

00:27:32.610 --> 00:27:34.110
the differences
might be that have

00:27:34.110 --> 00:27:37.040
arisen from the
genetic mutations.

00:27:37.040 --> 00:27:38.870
And where good drug
targets might be,

00:27:38.870 --> 00:27:42.000
or predictions if
I intervene here,

00:27:42.000 --> 00:27:44.420
if there's no difference in
that logic, between normal

00:27:44.420 --> 00:27:46.632
and tumor, well then that
won't have any effect.

00:27:46.632 --> 00:27:48.090
I want to look for
the places where

00:27:48.090 --> 00:27:50.610
there is a difference
in the signaling logic.

00:27:50.610 --> 00:27:53.350
And that would be a
better drug target.

00:27:53.350 --> 00:27:55.810
OK, so the measurements
are made in

00:27:55.810 --> 00:28:01.860
across 17 of these different
signaling molecules

00:28:01.860 --> 00:28:05.020
here, pretty much
all by measurement

00:28:05.020 --> 00:28:06.870
of a phosphorylation state.

00:28:06.870 --> 00:28:09.280
So if you've done cell
biology or biochemistry--

00:28:09.280 --> 00:28:11.040
in these signaling
pathways, many

00:28:11.040 --> 00:28:13.890
of the activities in
these kinds of pathways

00:28:13.890 --> 00:28:16.680
that regulate this
kind of cell behavior

00:28:16.680 --> 00:28:19.680
are kinases that end up
affecting transcription factor

00:28:19.680 --> 00:28:21.250
activities and so forth.

00:28:21.250 --> 00:28:23.160
And it's the
phosphorylation state

00:28:23.160 --> 00:28:25.335
of any these proteins
that matters.

00:28:25.335 --> 00:28:29.760
If a phosphate is on some
particular amino acid,

00:28:29.760 --> 00:28:31.340
the enzyme might be active.

00:28:31.340 --> 00:28:33.950
If it's not there it might
be inactive and so forth.

00:28:33.950 --> 00:28:36.075
So, just measurement of
phosphorylation states

00:28:36.075 --> 00:28:38.620
of 17 different proteins
in these pathways

00:28:38.620 --> 00:28:42.550
distributed across
multiple pathways.

00:28:42.550 --> 00:28:45.420
I've made these measurements on
five different cell types, four

00:28:45.420 --> 00:28:47.949
tumor cell types,
and the primaries

00:28:47.949 --> 00:28:50.240
in order to try to see what's
different between primary

00:28:50.240 --> 00:28:51.820
and tumor.

00:28:51.820 --> 00:28:55.970
And then what might be
different, patient to patient.

00:28:55.970 --> 00:28:59.500
In response to seven different
extracellular stimuli,

00:28:59.500 --> 00:29:01.340
some of them growth
factors, some of them

00:29:01.340 --> 00:29:06.850
cytokines, some of them actually
bacterial metabolic products.

00:29:06.850 --> 00:29:10.780
We all know about the effects
of microbiome these days.

00:29:10.780 --> 00:29:14.050
And, to further
populate a database that

00:29:14.050 --> 00:29:16.570
might be capable
of helping validate

00:29:16.570 --> 00:29:20.856
a model, a number of seven,
in fact-- intercellular

00:29:20.856 --> 00:29:21.355
inhibitors.

00:29:21.355 --> 00:29:23.990
A small molecule,
these things in black.

00:29:23.990 --> 00:29:25.680
One that might
inhibit this kinase.

00:29:25.680 --> 00:29:27.350
One might inhibit that kinase.

00:29:27.350 --> 00:29:29.420
One might inhibit that kinase.

00:29:29.420 --> 00:29:31.950
So now if you add all
those inhibitors too,

00:29:31.950 --> 00:29:34.310
then you start to change
the network activities

00:29:34.310 --> 00:29:38.160
and the downstream behavior.

00:29:38.160 --> 00:29:40.440
So that's how
extensive the data is.

00:29:40.440 --> 00:29:45.910
And this is actually for a
few different time points.

00:29:45.910 --> 00:29:48.190
So the data looks
something like this.

00:29:48.190 --> 00:29:50.230
Let's focus on the
one on the left.

00:29:50.230 --> 00:29:53.180
This is just the primary,
normal, human cells.

00:29:53.180 --> 00:29:56.940
It came from a liver donor.

00:29:56.940 --> 00:29:57.720
OK.

00:29:57.720 --> 00:30:01.890
Each row is one of the
17 different signals,

00:30:01.890 --> 00:30:04.650
essentially measurement
of the phosphorylation

00:30:04.650 --> 00:30:12.150
state of Akt or CREB
or P52 of staph 3.

00:30:12.150 --> 00:30:12.650
OK?

00:30:12.650 --> 00:30:14.270
So measurements of
its phosphorylation

00:30:14.270 --> 00:30:17.740
state that has something to do
with its signaling activity.

00:30:17.740 --> 00:30:20.620
Each of the big columns are the
seven different treatments--

00:30:20.620 --> 00:30:24.780
the different growth factors
and cytokines and so forth.

00:30:24.780 --> 00:30:25.730
And the control.

00:30:25.730 --> 00:30:27.580
No stimulation.

00:30:27.580 --> 00:30:31.950
And within each one
of these treatments,

00:30:31.950 --> 00:30:34.270
in each one of these
stimuli, then there's

00:30:34.270 --> 00:30:35.770
seven different
inhibitors that were

00:30:35.770 --> 00:30:38.010
used for the different pathways.

00:30:38.010 --> 00:30:43.970
So seven stimuli by seven
inhibitors plus controls.

00:30:43.970 --> 00:30:46.480
And then three
different time points.

00:30:46.480 --> 00:30:49.880
Sort of zero, 30
minutes, and three hours.

00:30:49.880 --> 00:30:52.520
So the data looks
something like this.

00:30:52.520 --> 00:30:56.230
If there's really no change,
due to the stimulation

00:30:56.230 --> 00:30:58.550
or the inhibitor, you'll
see something in gray.

00:30:58.550 --> 00:31:01.080
So in these gray bars,
there was already

00:31:01.080 --> 00:31:03.840
phosphorylation of this
transcription factor

00:31:03.840 --> 00:31:07.980
[INAUDIBLE] and it didn't really
change under most treatments.

00:31:07.980 --> 00:31:11.550
If it was yellow,
what it meant was,

00:31:11.550 --> 00:31:15.140
whatever the treatment was,
you got a quick activation

00:31:15.140 --> 00:31:19.080
of that signal and
then it went away.

00:31:19.080 --> 00:31:22.740
If it's late-- purple, then it
didn't happen in the first half

00:31:22.740 --> 00:31:25.500
hour, but it started to
show up a few hours later.

00:31:25.500 --> 00:31:28.310
And if it's green it showed
up in the first half hour

00:31:28.310 --> 00:31:29.450
and it stayed sustained.

00:31:29.450 --> 00:31:30.741
So that's what the color means.

00:31:30.741 --> 00:31:33.240
But this is the real
experimental data.

00:31:33.240 --> 00:31:36.720
And over here on the right is
one of the tumor cell lines.

00:31:36.720 --> 00:31:39.010
And you can just
see by inspection,

00:31:39.010 --> 00:31:40.300
it's different, right.

00:31:40.300 --> 00:31:42.730
The colors here are different
from the colors there.

00:31:42.730 --> 00:31:45.470
All the same treatments,
stimuli inhibitors.

00:31:45.470 --> 00:31:46.809
The colors are very different.

00:31:46.809 --> 00:31:48.850
You know, therefore that
the signaling activities

00:31:48.850 --> 00:31:50.080
are very different.

00:31:50.080 --> 00:31:50.580
OK.

00:31:50.580 --> 00:31:53.760
Just by visual inspection.

00:31:53.760 --> 00:31:54.260
OK.

00:31:54.260 --> 00:31:56.030
So what we're going to try
to do is build a logic model

00:31:56.030 --> 00:31:56.780
for this.

00:31:56.780 --> 00:31:58.580
A logic model for this.

00:31:58.580 --> 00:32:01.700
Compare them and say, oh where
are the key differences in how

00:32:01.700 --> 00:32:03.930
the signaling pathways
are getting activated?

00:32:03.930 --> 00:32:07.740
Downstream of the same stimuli.

00:32:07.740 --> 00:32:10.570
So, we start with
our prior knowledge.

00:32:10.570 --> 00:32:15.260
This is from the Ingenuity
database, which actually

00:32:15.260 --> 00:32:19.130
happened to be missing,
even basic information

00:32:19.130 --> 00:32:20.540
about insulin signaling.

00:32:20.540 --> 00:32:22.460
So we just added
our own information

00:32:22.460 --> 00:32:24.902
about what the
insulin receptor does.

00:32:24.902 --> 00:32:26.110
It's kind of hard to believe.

00:32:26.110 --> 00:32:27.901
This is a database that
cost a lot of money

00:32:27.901 --> 00:32:29.890
and they didn't have
really much information

00:32:29.890 --> 00:32:31.720
about insulin
receptor signaling.

00:32:31.720 --> 00:32:34.190
Very strange.

00:32:34.190 --> 00:32:38.330
So, downstream of
our seven stimuli,

00:32:38.330 --> 00:32:40.640
down to the transcription
factors of interest,

00:32:40.640 --> 00:32:44.390
there are about 82 molecular
nodes and a hundred some edges

00:32:44.390 --> 00:32:46.530
that you'd pull out of
the Ingenuity database.

00:32:46.530 --> 00:32:50.180
So here's our starting guess
at what this looks like.

00:32:50.180 --> 00:32:53.200
There's no logic in here, but
this is just, potentially,

00:32:53.200 --> 00:32:55.760
the things that the
logic might operate on,

00:32:55.760 --> 00:33:00.286
downstream of stimuli, and
when inhibited, and so forth.

00:33:00.286 --> 00:33:01.240
All right.

00:33:01.240 --> 00:33:02.500
So here's the process.

00:33:02.500 --> 00:33:04.900
This was the actual
algorithmic process

00:33:04.900 --> 00:33:07.730
that I'll walk you through.

00:33:07.730 --> 00:33:10.870
On the left-hand side
is the computer part.

00:33:10.870 --> 00:33:13.310
It said, OK, from the
Ingenuity database,

00:33:13.310 --> 00:33:16.620
we had this prior
knowledge about who

00:33:16.620 --> 00:33:20.110
was upstream, downstream,
who affected whom.

00:33:20.110 --> 00:33:22.810
We strip this down
some, because in terms

00:33:22.810 --> 00:33:25.610
of the measurements
on the perturbations,

00:33:25.610 --> 00:33:28.400
there are some of the nodes
that you just would not

00:33:28.400 --> 00:33:30.190
be able to see any
measurable difference.

00:33:30.190 --> 00:33:36.550
OK, there was no stimulus
upstream, or no perturbation.

00:33:36.550 --> 00:33:38.220
And it was not
measured so you really

00:33:38.220 --> 00:33:40.136
wouldn't be able to tell
if it changed or not.

00:33:40.136 --> 00:33:42.070
So you just take those out.

00:33:42.070 --> 00:33:45.070
Of everything remaining, now
you don't know the logic.

00:33:45.070 --> 00:33:46.210
You know the potential.

00:33:46.210 --> 00:33:48.730
And so you say, well, of all
the nodes and interactions

00:33:48.730 --> 00:33:52.320
remaining, I could have AND
gates, I could have OR gates,

00:33:52.320 --> 00:33:53.690
you could have NOTS.

00:33:53.690 --> 00:33:56.000
And you say, OK, in
principle, I could have,

00:33:56.000 --> 00:34:00.560
then, many, many, many,
many, many different logic

00:34:00.560 --> 00:34:06.050
models that could work.

00:34:06.050 --> 00:34:07.424
So how do I know which one?

00:34:07.424 --> 00:34:09.090
Well now you skip
over to the other side

00:34:09.090 --> 00:34:11.520
and say, well, but we have
all this experimental data.

00:34:11.520 --> 00:34:13.610
We have the data from
all the different stimuli

00:34:13.610 --> 00:34:17.500
and all the different inhibitors
for any given cell type.

00:34:17.500 --> 00:34:21.033
And so, we have that data under
all these different conditions.

00:34:21.033 --> 00:34:22.449
And what we're
going to do is just

00:34:22.449 --> 00:34:28.270
run hundreds or thousands of
these potentially appropriate

00:34:28.270 --> 00:34:29.980
models.

00:34:29.980 --> 00:34:33.320
Compare them to the data of
whether any given node is

00:34:33.320 --> 00:34:36.060
activated or not, activated
under treatment conditions,

00:34:36.060 --> 00:34:38.030
stimuli inhibitors.

00:34:38.030 --> 00:34:39.920
And we'll calculate the air.

00:34:39.920 --> 00:34:41.800
How good was any
one of those models

00:34:41.800 --> 00:34:44.460
at actually matching those data?

00:34:44.460 --> 00:34:46.396
Simple as that.

00:34:46.396 --> 00:34:47.770
And then it's a
matter of finding

00:34:47.770 --> 00:34:50.530
what are the best fit ones
from the best fit ones.

00:34:50.530 --> 00:34:52.860
Could you improve them and
make them fit even better?

00:34:52.860 --> 00:34:57.350
And in the end, how did you go
from an initial prior knowledge

00:34:57.350 --> 00:35:02.200
scaffold to something that, in
fact, fit the data really well,

00:35:02.200 --> 00:35:04.710
from which you could
make new predictions.

00:35:04.710 --> 00:35:05.210
OK.

00:35:05.210 --> 00:35:09.370
So you get the approach here?

00:35:09.370 --> 00:35:12.460
All right, good.

00:35:12.460 --> 00:35:16.320
Now, in terms of figuring
out how well any given

00:35:16.320 --> 00:35:20.630
model matches the data and how
to go through model selection,

00:35:20.630 --> 00:35:24.059
there's a myriad of
different approaches to this.

00:35:24.059 --> 00:35:25.600
And I'm not claiming
that what we did

00:35:25.600 --> 00:35:28.730
was the absolute best approach.

00:35:28.730 --> 00:35:30.850
There's alternatives to
it that one could consider

00:35:30.850 --> 00:35:33.560
and then perhaps could
work even better.

00:35:33.560 --> 00:35:35.660
If you read the paper,
you'll read the reasons

00:35:35.660 --> 00:35:37.215
for these choices.

00:35:37.215 --> 00:35:38.400
OK.

00:35:38.400 --> 00:35:40.450
So I'll let you do that.

00:35:40.450 --> 00:35:44.520
The way the model
quality was calculated

00:35:44.520 --> 00:35:47.060
was to have an
objective function that

00:35:47.060 --> 00:35:50.820
said we want to minimize
some number, theta.

00:35:50.820 --> 00:35:52.590
And how do we calculate theta?

00:35:52.590 --> 00:35:56.100
Well, first of all, for
whatever that model is,

00:35:56.100 --> 00:35:59.250
we're going to fit-- whether
the model says some nodes

00:35:59.250 --> 00:36:02.690
should be on or
off, one or zero.

00:36:02.690 --> 00:36:06.800
And we're going to compare
it to the experimental data.

00:36:06.800 --> 00:36:10.220
Now the experimental
data, I need to emphasize,

00:36:10.220 --> 00:36:12.850
isn't one or zero,
it's normalized

00:36:12.850 --> 00:36:14.160
to go between one and zero.

00:36:14.160 --> 00:36:19.050
But the actual measurement
might be 0.7 or 0.25.

00:36:19.050 --> 00:36:23.135
OK, so you're going to have
error against the Boolean model

00:36:23.135 --> 00:36:25.260
even if all the edges are
absolutely correct you'll

00:36:25.260 --> 00:36:29.500
still going to get some
quantitative error.

00:36:29.500 --> 00:36:30.920
So you calculate that.

00:36:30.920 --> 00:36:32.550
The Boolean model
says zero or one.

00:36:32.550 --> 00:36:35.710
The experimental
data says 0.250, 0.7.

00:36:35.710 --> 00:36:39.680
And you say, OK,
I'll calculate that.

00:36:39.680 --> 00:36:43.600
But then you might
think, all right, well,

00:36:43.600 --> 00:36:46.280
somehow I've got to penalize
bigger models with more

00:36:46.280 --> 00:36:49.160
nodes and more edges because
surely the more nodes and edges

00:36:49.160 --> 00:36:53.390
I put in, I could
capture more of the data.

00:36:53.390 --> 00:36:55.770
And I don't want to make
the model infinitely large

00:36:55.770 --> 00:36:57.240
just to get the best fit.

00:36:57.240 --> 00:37:00.260
So I need to penalize that.

00:37:00.260 --> 00:37:04.760
Turns out it's not true, but
nonetheless it's worth doing.

00:37:04.760 --> 00:37:08.180
So, you take a parameter
that's the size of the model.

00:37:08.180 --> 00:37:10.020
It's basically just
the number of nodes.

00:37:10.020 --> 00:37:12.350
The more nodes in
it, the more you

00:37:12.350 --> 00:37:14.540
would be suspicious of
the model for just fitting

00:37:14.540 --> 00:37:17.480
because it has too
many components.

00:37:17.480 --> 00:37:23.400
And you multiply that size by
a penalty parameter, alpha.

00:37:23.400 --> 00:37:25.450
So you have a bad
objective function

00:37:25.450 --> 00:37:27.480
if there's a lot of
error with the data,

00:37:27.480 --> 00:37:30.020
or if your model's too big.

00:37:30.020 --> 00:37:35.060
A better model would be, better
fit to the data and smaller.

00:37:35.060 --> 00:37:38.560
That's the calculation.

00:37:38.560 --> 00:37:39.090
OK.

00:37:39.090 --> 00:37:42.250
And in the end-- and I'm going
to show you how we did this.

00:37:45.190 --> 00:37:47.620
And I think the field is
now really believing this.

00:37:47.620 --> 00:37:51.170
That what you're not after
is a single best fit model.

00:37:51.170 --> 00:37:55.810
That one single model that gives
you the very smallest data.

00:37:55.810 --> 00:37:58.250
Because honestly,
within the uncertainty

00:37:58.250 --> 00:38:02.710
of the experimental
data-- OK, there's

00:38:02.710 --> 00:38:04.480
a substantial number
of models that

00:38:04.480 --> 00:38:09.080
could fit the data
within that noise.

00:38:09.080 --> 00:38:11.210
So if you demanded
the single best one,

00:38:11.210 --> 00:38:13.940
you say, well, but
these other 50 actually

00:38:13.940 --> 00:38:16.759
fit it almost as good and within
the uncertainty of the data.

00:38:16.759 --> 00:38:18.050
How can you really reject them?

00:38:18.050 --> 00:38:18.990
And you can't.

00:38:18.990 --> 00:38:23.090
So in the end, what's
being striven for in most

00:38:23.090 --> 00:38:25.522
of the field is a
family of models.

00:38:25.522 --> 00:38:26.980
And then you see
what the consensus

00:38:26.980 --> 00:38:31.670
is and the differences
within that family.

00:38:31.670 --> 00:38:35.180
The particular algorithm
for generating and running

00:38:35.180 --> 00:38:38.240
through different
potential models--

00:38:38.240 --> 00:38:39.720
because you just
can't exhaustively

00:38:39.720 --> 00:38:40.630
sample all of them.

00:38:40.630 --> 00:38:45.520
OK, these networks are so large,
that you can't exhaustively

00:38:45.520 --> 00:38:48.640
test all possibilities of
all their logic and so forth.

00:38:48.640 --> 00:38:50.720
It's really prohibitive.

00:38:50.720 --> 00:38:52.989
So there's many different
ways you can go about it.

00:38:52.989 --> 00:38:54.780
This particular method
maybe you've already

00:38:54.780 --> 00:38:57.780
learned this in class
for other applications

00:38:57.780 --> 00:38:59.560
as a genetic algorithm.

00:38:59.560 --> 00:39:01.300
So you start with
some population.

00:39:01.300 --> 00:39:03.350
You start with your
Ingenuity scaffold

00:39:03.350 --> 00:39:07.790
and then you randomly remove
or take edges and things

00:39:07.790 --> 00:39:08.300
like that.

00:39:08.300 --> 00:39:09.841
So that if you've
got a whole family,

00:39:09.841 --> 00:39:11.870
that's slightly different.

00:39:11.870 --> 00:39:14.810
For each one of them you
evaluate the objective function

00:39:14.810 --> 00:39:16.690
against the data.

00:39:16.690 --> 00:39:20.900
And you get some of those that
then are the most attractive.

00:39:20.900 --> 00:39:22.910
They seem to be the best fit.

00:39:22.910 --> 00:39:26.580
But, by no means would you
imagine they are yet optimal.

00:39:26.580 --> 00:39:30.530
So, now you create a next
generation from this population

00:39:30.530 --> 00:39:34.907
by the analog of genetics.

00:39:34.907 --> 00:39:36.740
Some of the very best--
you say, OK, they're

00:39:36.740 --> 00:39:39.800
going to survive so I'm just
going to take them as is.

00:39:39.800 --> 00:39:41.640
Some I'm going to
mutate, I'm going

00:39:41.640 --> 00:39:44.830
to have a probability of
mutating an edge here or there.

00:39:44.830 --> 00:39:46.970
You can have crossover,
actually mating

00:39:46.970 --> 00:39:49.290
between one model
and another model,

00:39:49.290 --> 00:39:50.890
so that the daughter
model gets some

00:39:50.890 --> 00:39:53.140
of the arcs from the mother
model and some of the arcs

00:39:53.140 --> 00:39:54.580
from the father model.

00:39:54.580 --> 00:39:57.920
So you just generate an
ex-population, do it again.

00:39:57.920 --> 00:40:00.610
And once you've
reached a set of models

00:40:00.610 --> 00:40:04.120
that fit your data
within the criteria

00:40:04.120 --> 00:40:07.490
that you want, then you say,
this is now my population.

00:40:07.490 --> 00:40:09.130
And these are now
my best-fit models.

00:40:09.130 --> 00:40:11.849
So it's not exhaustive.

00:40:11.849 --> 00:40:13.640
You can definitely find
local minimum here.

00:40:13.640 --> 00:40:14.931
There's no question about that.

00:40:14.931 --> 00:40:15.608
Yeah?

00:40:15.608 --> 00:40:22.082
AUDIENCE: Do you always take the
best model into the next round?

00:40:22.082 --> 00:40:23.078
Or do you--

00:40:23.078 --> 00:40:25.474
DOUG LAUFFENBURGER: Yeah,
that's the elite survival.

00:40:25.474 --> 00:40:26.890
If you don't
incorporate that, you

00:40:26.890 --> 00:40:30.075
might lose the best
ones in any given round.

00:40:30.075 --> 00:40:33.940
But this ensures you
take the best subset.

00:40:33.940 --> 00:40:34.980
Let them go for it.

00:40:34.980 --> 00:40:36.438
AUDIENCE: Is there
a worry that you

00:40:36.438 --> 00:40:38.141
might get stuck in [INAUDIBLE]?

00:40:38.141 --> 00:40:39.140
DOUG LAUFFENBURGER: Yes.

00:40:39.140 --> 00:40:41.750
Yes, absolutely.

00:40:41.750 --> 00:40:46.990
So now you run this with a
number of different starting

00:40:46.990 --> 00:40:48.120
populations.

00:40:48.120 --> 00:40:52.010
And you see if you get to
similar consensus models.

00:40:52.010 --> 00:40:54.030
Yeah, because
absolutely, this does not

00:40:54.030 --> 00:40:56.240
guarantee any kind
of a global minimum.

00:40:56.240 --> 00:40:58.260
You will always get local.

00:40:58.260 --> 00:41:01.622
So you have to condition
it on a different set

00:41:01.622 --> 00:41:02.580
of initial populations.

00:41:06.660 --> 00:41:07.160
OK.

00:41:07.160 --> 00:41:10.730
Once you do this-- I'm going
to show you some results

00:41:10.730 --> 00:41:15.810
first and then dig into some
other ways to think about it.

00:41:15.810 --> 00:41:17.210
So it's plotted here.

00:41:17.210 --> 00:41:20.860
This is one of the
tumor cell lines.

00:41:20.860 --> 00:41:24.160
What's plotted here, is again,
all the rows or all the signals

00:41:24.160 --> 00:41:25.310
that were measured.

00:41:25.310 --> 00:41:28.550
All the big columns or
all the different stimuli,

00:41:28.550 --> 00:41:32.910
and all the little columns
are the different inhibitors.

00:41:32.910 --> 00:41:36.920
And I should point out, this
is only for the 30 minute data.

00:41:36.920 --> 00:41:37.420
OK.

00:41:37.420 --> 00:41:39.840
This isn't for the
three hour or both,

00:41:39.840 --> 00:41:41.780
this is just the 30 minute data.

00:41:41.780 --> 00:41:45.450
And basically where
there's green,

00:41:45.450 --> 00:41:53.530
the model and data
fit was considered OK.

00:41:53.530 --> 00:41:55.820
Where it's red, it's not OK.

00:41:55.820 --> 00:41:59.240
Where it's pink it's less bad.

00:41:59.240 --> 00:42:01.010
So by the shaded.

00:42:01.010 --> 00:42:03.950
And the yellow
actually, the model

00:42:03.950 --> 00:42:05.897
really couldn't
make a prediction.

00:42:05.897 --> 00:42:07.980
Now, why that's the case
is what's showing up here

00:42:07.980 --> 00:42:10.300
is just the initial
Ingenuity scaffold.

00:42:10.300 --> 00:42:14.200
The very best one that
didn't add or remove any

00:42:14.200 --> 00:42:17.165
arcs or nodes from the
Ingenuity prior knowledge.

00:42:17.165 --> 00:42:18.790
It's that all we're
going to do is just

00:42:18.790 --> 00:42:23.130
run the best fit Boolean
logic model we can on that.

00:42:23.130 --> 00:42:24.130
And it wasn't very good.

00:42:24.130 --> 00:42:27.110
It was about 45% error.

00:42:27.110 --> 00:42:30.020
Almost half of the
nodes it got wrong.

00:42:30.020 --> 00:42:32.500
So what that tells you if you
just take a scaffold from one

00:42:32.500 --> 00:42:37.890
is interactive databases
and without adulterating it,

00:42:37.890 --> 00:42:41.190
just fit the best logic
model to some data--

00:42:41.190 --> 00:42:43.850
OK, at least in this case, and
we've done a number of others,

00:42:43.850 --> 00:42:46.770
it actually doesn't
fit very well.

00:42:46.770 --> 00:42:49.040
And the reasons being,
you're trying to fit this now

00:42:49.040 --> 00:42:52.670
to a very specific
biological context.

00:42:52.670 --> 00:42:54.870
Hepatocyte tumor
cells under these

00:42:54.870 --> 00:42:57.090
grow factor and
cytokine treatments.

00:42:57.090 --> 00:42:58.990
That network is
likely very different

00:42:58.990 --> 00:43:01.850
from whatever aggregate you
got from literature curation

00:43:01.850 --> 00:43:03.510
and so forth in a database.

00:43:03.510 --> 00:43:05.980
There's going to be a lot of
stuff in the database that's

00:43:05.980 --> 00:43:08.580
not applicable, because it
came from a different cell

00:43:08.580 --> 00:43:10.950
type, a different
condition, or there just

00:43:10.950 --> 00:43:15.210
wasn't enough experiments in
the literature for hypatocytes.

00:43:15.210 --> 00:43:17.160
Maybe it was never
measured under treatment

00:43:17.160 --> 00:43:18.840
with interferon gamma.

00:43:18.840 --> 00:43:21.750
So there's data here
that the database never

00:43:21.750 --> 00:43:23.960
had access to literature
that it had explored.

00:43:23.960 --> 00:43:26.950
So lots of reasons.

00:43:26.950 --> 00:43:29.460
Now when you go through the
processes we just talked about,

00:43:29.460 --> 00:43:33.020
and in the end,
the best fit models

00:43:33.020 --> 00:43:35.370
give you something like
less than 10% error.

00:43:35.370 --> 00:43:39.079
So less than 10% of these
squares are red or pink.

00:43:39.079 --> 00:43:40.620
OK, so that's the
kind of improvement

00:43:40.620 --> 00:43:44.420
that you can take by
generating an improved model.

00:43:44.420 --> 00:43:48.050
By adding and subtracting
arcs and nodes.

00:43:51.340 --> 00:43:54.820
So this is what the model looks
like in the end for this tumor

00:43:54.820 --> 00:43:55.460
cell line.

00:43:55.460 --> 00:44:02.240
And this is a consensus model
from the 20 or so best fit.

00:44:02.240 --> 00:44:05.470
And so the thickness
of a line is

00:44:05.470 --> 00:44:08.050
how strong the consensus was.

00:44:08.050 --> 00:44:11.200
The strongest would
be all 20 had it.

00:44:11.200 --> 00:44:16.470
And the point here being,
you see some purple.

00:44:16.470 --> 00:44:18.810
And I wish my pen wasn't
fading in and out.

00:44:18.810 --> 00:44:23.090
If anybody has a pointer
I'll be happy to have it.

00:44:23.090 --> 00:44:26.450
Where you see purple,
those were arcs that

00:44:26.450 --> 00:44:28.240
weren't in the
Ingenuity database

00:44:28.240 --> 00:44:34.166
and had to be put in to get
the data to fit this well.

00:44:34.166 --> 00:44:35.540
And it turns out,
if you actually

00:44:35.540 --> 00:44:40.110
go back to the literature, you
find that those purple arcs

00:44:40.110 --> 00:44:41.940
were already described
in the literature.

00:44:41.940 --> 00:44:44.190
It's just that they weren't
captured in that database.

00:44:48.850 --> 00:44:51.470
Well that's green and purple.

00:44:51.470 --> 00:44:56.820
Then you see some
blue and they were

00:44:56.820 --> 00:44:58.880
in some of the other
tumor cell types

00:44:58.880 --> 00:45:00.740
but now in this
particular hep G2

00:45:00.740 --> 00:45:04.530
But you can generate a
model that works very well.

00:45:04.530 --> 00:45:08.990
And see that it's consistent
with much of literature.

00:45:08.990 --> 00:45:11.880
It's a more stripped down
than what's in the databases.

00:45:11.880 --> 00:45:13.920
And there's some new
things in it, that in fact,

00:45:13.920 --> 00:45:15.836
if you go back to the
literature you can find,

00:45:15.836 --> 00:45:18.521
because they just were
captured in the database.

00:45:18.521 --> 00:45:19.060
All right.

00:45:19.060 --> 00:45:22.570
A few insights
about the analysis.

00:45:22.570 --> 00:45:26.150
So I want to show you, here
is the objective function.

00:45:26.150 --> 00:45:30.100
How well the model
fit and that's in red.

00:45:30.100 --> 00:45:30.810
OK.

00:45:30.810 --> 00:45:35.690
And in blue is the actual
fit to the experimental data.

00:45:35.690 --> 00:45:37.770
And again, the hirer
it is the worse it is.

00:45:37.770 --> 00:45:40.320
And the green gives you
essentially the size.

00:45:40.320 --> 00:45:43.730
And this is plotted
against the size penalty.

00:45:43.730 --> 00:45:46.600
And what's very interesting,
is even for very small size

00:45:46.600 --> 00:45:54.090
penalties, almost negligible,
that the size of the model that

00:45:54.090 --> 00:45:58.270
turns out to be best fit
is substantially smaller

00:45:58.270 --> 00:46:00.186
than what was in the database.

00:46:00.186 --> 00:46:03.575
OK, so you actually generate
a small model immediately.

00:46:03.575 --> 00:46:07.770
A smaller model immediately,
even without any size penalty.

00:46:07.770 --> 00:46:10.270
So your intuition
that a bigger model

00:46:10.270 --> 00:46:13.820
was going to be better actually
turns out to be incorrect.

00:46:13.820 --> 00:46:17.290
That even without a size
penalty, the model strips down.

00:46:17.290 --> 00:46:20.400
And why is that?

00:46:20.400 --> 00:46:20.980
Why is that?

00:46:20.980 --> 00:46:23.120
Let me make that
question number two.

00:46:23.120 --> 00:46:26.050
Just to see who's still awake.

00:46:26.050 --> 00:46:30.910
Why, in fitting this
hepatocyte data,

00:46:30.910 --> 00:46:34.100
would a model that leaves out
a lot of stuff in the Ingenuity

00:46:34.100 --> 00:46:38.780
database that's presumably
going on actually

00:46:38.780 --> 00:46:40.582
fit the data better?

00:46:40.582 --> 00:46:42.636
A smaller model fits better?

00:46:42.636 --> 00:46:46.110
Why is that?

00:46:46.110 --> 00:46:48.430
Yeah.

00:46:48.430 --> 00:46:50.020
AUDIENCE: This is a [INAUDIBLE].

00:46:50.020 --> 00:46:53.114
Maybe the strength
of the attractions

00:46:53.114 --> 00:46:54.960
aren't really taken
into account here?

00:46:54.960 --> 00:46:59.830
And so the moving things
out, essentially means

00:46:59.830 --> 00:47:03.180
that you're not sealing
everything in the [INAUDIBLE].

00:47:03.180 --> 00:47:05.129
You're just taking one.

00:47:05.129 --> 00:47:06.170
DOUG LAUFFENBURGER: Yeah.

00:47:06.170 --> 00:47:07.260
That's essentially it.

00:47:07.260 --> 00:47:11.550
I think you've casted it an
almost quantitative term,

00:47:11.550 --> 00:47:15.190
but I think it's true
even in qualitative terms.

00:47:15.190 --> 00:47:18.490
And one way to think
about it is-- let's

00:47:18.490 --> 00:47:21.090
say I have an extra
arc or extra node.

00:47:21.090 --> 00:47:22.180
OK.

00:47:22.180 --> 00:47:24.480
I might capture some
more true positives.

00:47:24.480 --> 00:47:26.910
I might actually
capture more of my data,

00:47:26.910 --> 00:47:32.590
but I could actually now, gain
more complex with my data.

00:47:32.590 --> 00:47:35.840
Because now I've put
in logic that, yes,

00:47:35.840 --> 00:47:38.370
it captures this
measurement, but now maybe it

00:47:38.370 --> 00:47:41.180
messes up these other two
or three measurements.

00:47:41.180 --> 00:47:43.020
So you actually can
make your model worse

00:47:43.020 --> 00:47:47.250
trying to capture some small
piece, that in fact, adversely

00:47:47.250 --> 00:47:51.020
influences the effects on
the other measurements.

00:47:51.020 --> 00:47:54.150
So you get you get false
positives, false negatives,

00:47:54.150 --> 00:47:56.110
along with anything
and that's true.

00:47:56.110 --> 00:48:00.090
And it just so happens that
in these kind of situations

00:48:00.090 --> 00:48:01.910
those can outweigh.

00:48:01.910 --> 00:48:04.810
Then of course, as you
increase the size penalty

00:48:04.810 --> 00:48:08.770
you can drive your model to
be even smaller, fewer arcs,

00:48:08.770 --> 00:48:11.630
and now that of course
does come at the expense

00:48:11.630 --> 00:48:13.340
of not fitting the data better.

00:48:13.340 --> 00:48:14.670
OK.

00:48:14.670 --> 00:48:18.130
So where we decided that
the size penalty best lived

00:48:18.130 --> 00:48:22.930
was where it was large enough
to ensure stripping down

00:48:22.930 --> 00:48:26.040
of nonessential nodes and
arcs, but not large enough

00:48:26.040 --> 00:48:29.100
to start compromising the
actual experimental fit.

00:48:29.100 --> 00:48:29.600
OK.

00:48:29.600 --> 00:48:31.308
And so that lived
someplace around there.

00:48:34.890 --> 00:48:35.390
OK.

00:48:35.390 --> 00:48:37.014
An important thing--
and this goes back

00:48:37.014 --> 00:48:38.600
to the consensus model.

00:48:38.600 --> 00:48:40.700
If you think about, quote,
model identification,

00:48:40.700 --> 00:48:45.560
can you uniquely specify
one model a best fit model?

00:48:45.560 --> 00:48:46.450
You really can't.

00:48:46.450 --> 00:48:48.590
What's plotted here is
for any of the arcs that

00:48:48.590 --> 00:48:50.892
would end up in a model.

00:48:50.892 --> 00:48:53.100
Let's say we let's say we
numbered them from one to I

00:48:53.100 --> 00:48:55.980
think it was 113
in the first place.

00:48:55.980 --> 00:48:58.040
One arc, another arc,
another arc, another arc.

00:48:58.040 --> 00:48:59.970
And you say, how
frequently did they

00:48:59.970 --> 00:49:03.880
end up in the best fit models?

00:49:03.880 --> 00:49:07.190
Basically, only a small
proportion of them

00:49:07.190 --> 00:49:08.870
were in all the best fit models.

00:49:08.870 --> 00:49:13.610
Some of them were in
some models and some not.

00:49:13.610 --> 00:49:17.650
Of course the higher
the tolerance,

00:49:17.650 --> 00:49:20.310
the more air you
allowed and now you

00:49:20.310 --> 00:49:22.130
started to get
models that all fit

00:49:22.130 --> 00:49:27.330
to within whatever that criteria
was in which most of their arcs

00:49:27.330 --> 00:49:28.160
weren't the same.

00:49:28.160 --> 00:49:30.326
You could have a lot of
different network structures

00:49:30.326 --> 00:49:32.370
that give you that same fit.

00:49:32.370 --> 00:49:37.080
If you require a very, very
tiny fit, compared to air,

00:49:37.080 --> 00:49:40.830
something like this, then
more of the arcs in the models

00:49:40.830 --> 00:49:42.600
have to be in common.

00:49:42.600 --> 00:49:43.100
OK.

00:49:43.100 --> 00:49:44.670
So that makes some sense.

00:49:44.670 --> 00:49:48.220
But you can't really completely
identify a unique model.

00:49:48.220 --> 00:49:50.800
That goes to what I said before.

00:49:50.800 --> 00:49:51.300
OK.

00:49:51.300 --> 00:49:53.217
I was talking before
about trade-offs

00:49:53.217 --> 00:49:55.050
between false positives
and false negatives.

00:49:55.050 --> 00:49:59.180
You must know, I'm sure from
previous things in this class,

00:49:59.180 --> 00:50:01.870
the receiver operating
characteristic curves,

00:50:01.870 --> 00:50:05.530
where for every of your
model parameter choices,

00:50:05.530 --> 00:50:07.200
you say, what are
my results in terms

00:50:07.200 --> 00:50:10.390
of false positives
versus true positives?

00:50:10.390 --> 00:50:14.620
And you're trying to
find the optimal location

00:50:14.620 --> 00:50:19.750
along this type of path.

00:50:19.750 --> 00:50:27.400
And so, what's shown here is
that the best predictive model,

00:50:27.400 --> 00:50:32.160
in fact, is the one where
we have the size penalty

00:50:32.160 --> 00:50:34.760
to be right on the edge of not
making the experimental data

00:50:34.760 --> 00:50:38.060
fit worse, but still
strips out the most arcs.

00:50:38.060 --> 00:50:41.670
So again, that demonstrates
that the smaller model actually

00:50:41.670 --> 00:50:47.540
is in fact better, in terms
of finding this type of--

00:50:47.540 --> 00:50:50.650
And this shows if we actually
put in some more arcs that

00:50:50.650 --> 00:50:53.140
tried to capture some
more data, yes we decrease

00:50:53.140 --> 00:50:55.680
the false negatives,
but, in fact, we

00:50:55.680 --> 00:50:56.900
increase the false positives.

00:50:56.900 --> 00:50:59.780
We actually shift
ourselves on this curve.

00:50:59.780 --> 00:51:03.310
And so you decide whether
that's desirable or not.

00:51:03.310 --> 00:51:04.620
Where you'd like to live.

00:51:04.620 --> 00:51:08.940
So you can analyze what you
like about your best fit class

00:51:08.940 --> 00:51:11.270
of models in this kind of way.

00:51:14.550 --> 00:51:17.920
OK, so now we have some
confidence in this.

00:51:17.920 --> 00:51:19.360
What are you going
to do with it?

00:51:19.360 --> 00:51:22.950
And one thing I'd like to do is
just make a priori predictions.

00:51:22.950 --> 00:51:26.040
Say I now believe that on
these hepatocytes or tumor

00:51:26.040 --> 00:51:29.810
cells stimulated with
these kind of things,

00:51:29.810 --> 00:51:32.870
I can calculate what the
experimental signaling

00:51:32.870 --> 00:51:35.121
activities should be.

00:51:35.121 --> 00:51:35.620
All right.

00:51:35.620 --> 00:51:37.020
Let's see if we
do that a priori.

00:51:37.020 --> 00:51:43.440
So let's now use new inhibitors
that hadn't been used before.

00:51:43.440 --> 00:51:45.650
Combination of inhibitors,
especially in cancer.

00:51:45.650 --> 00:51:47.960
People are always interested
in combinatorial drugs.

00:51:47.960 --> 00:51:50.460
Experimentally it's
prohibitive to run

00:51:50.460 --> 00:51:52.190
through all possible
combinations.

00:51:52.190 --> 00:51:54.190
So this is one thing in
the pharmaceutical field

00:51:54.190 --> 00:51:56.900
people believe these kind of
models are really useful for.

00:51:56.900 --> 00:51:58.780
Let's try all possible
drug combinations

00:51:58.780 --> 00:52:01.270
and see which ones
are most promising.

00:52:01.270 --> 00:52:04.810
And instead of just one
ligand growth factor cytokine

00:52:04.810 --> 00:52:07.660
at a time, do
different combinations.

00:52:07.660 --> 00:52:10.270
So this is all an
entirely new data set.

00:52:10.270 --> 00:52:12.830
So different treatments that
are different combinations,

00:52:12.830 --> 00:52:16.110
different inhibitors, different
combinations of inhibitors.

00:52:16.110 --> 00:52:18.760
And now you just run the model--
it's not trained on this.

00:52:18.760 --> 00:52:20.520
It was trained on
the previous data.

00:52:20.520 --> 00:52:23.300
And now a priori
predicts this data set.

00:52:23.300 --> 00:52:26.760
And now, again, you look for
the model fit in the bottom.

00:52:26.760 --> 00:52:30.660
And again, you want the smallest
number of red and pink boxes.

00:52:30.660 --> 00:52:33.430
In effect it predicted to
within about 11% error.

00:52:33.430 --> 00:52:38.360
About 11% of the boxes didn't
fit well, but 89% percent did.

00:52:38.360 --> 00:52:41.050
And that's, in fact
pretty close to the 9%

00:52:41.050 --> 00:52:43.430
that was on the
original training model.

00:52:43.430 --> 00:52:47.140
So in terms of this, in
this realm of studies,

00:52:47.140 --> 00:52:49.970
these a priori treatment
conditions-- drug combinations,

00:52:49.970 --> 00:52:52.480
growth factors,
cytokine combinations--

00:52:52.480 --> 00:52:56.120
this is a pretty good validation
that this model wasn't

00:52:56.120 --> 00:52:58.330
just kind of trained and fit.

00:52:58.330 --> 00:53:00.130
That it, in fact,
could predict then

00:53:00.130 --> 00:53:02.316
what was happening
in these pathways.

00:53:02.316 --> 00:53:04.190
And then of course, what
it allows you to do,

00:53:04.190 --> 00:53:05.898
where all the red
boxes are-- it say, OK,

00:53:05.898 --> 00:53:08.042
that's where we need
more intensive study.

00:53:08.042 --> 00:53:10.000
Now maybe we go back to
the literature and say,

00:53:10.000 --> 00:53:12.050
is there more known
about those nodes that

00:53:12.050 --> 00:53:15.780
was captured in whatever
our interactive database

00:53:15.780 --> 00:53:17.390
that we started with?

00:53:17.390 --> 00:53:21.880
Maybe we need to supplement the
scaffold with more information.

00:53:21.880 --> 00:53:24.730
That's out in the literature
where more and more dedicated

00:53:24.730 --> 00:53:25.690
experiments are done.

00:53:25.690 --> 00:53:29.190
So it narrows down where the
next set of investigations

00:53:29.190 --> 00:53:32.030
need to be, whether from the
literature or from yourself.

00:53:35.286 --> 00:53:35.785
OK.

00:53:40.900 --> 00:53:43.240
So this is just then
some biological results.

00:53:43.240 --> 00:53:48.820
If you do this for the four
different hepatocellular lines.

00:53:48.820 --> 00:53:50.810
Some of the signaling
activities are the same,

00:53:50.810 --> 00:53:51.768
and some are different.

00:53:54.700 --> 00:53:58.400
I think I'll skip that.

00:53:58.400 --> 00:53:59.750
All right, let me show this.

00:53:59.750 --> 00:54:03.850
So this says, where are the
similarities and differences

00:54:03.850 --> 00:54:07.220
between the normal hepatocytes
versus the tumor lines.

00:54:07.220 --> 00:54:09.160
Because this is
where you would want

00:54:09.160 --> 00:54:11.540
to get the ideas for where
the right drugs would be.

00:54:11.540 --> 00:54:15.340
Where is the logic different,
between a normal liver

00:54:15.340 --> 00:54:19.420
cell and one of these
transformed types.

00:54:19.420 --> 00:54:22.220
So, this is the same
kind of scaffold.

00:54:22.220 --> 00:54:24.210
It'll get us the
consensus models

00:54:24.210 --> 00:54:26.820
and the thickness of
the line is how strong--

00:54:26.820 --> 00:54:30.330
what proportion of the models
did that arc show up in?

00:54:30.330 --> 00:54:31.840
Along the best.

00:54:31.840 --> 00:54:37.760
If it's black, the arc was
in the primary hepatocytes

00:54:37.760 --> 00:54:39.690
and all the cell lines.

00:54:39.690 --> 00:54:44.380
So black is just sort
of consensus core.

00:54:44.380 --> 00:54:45.970
This is just invariably there.

00:54:49.020 --> 00:54:53.480
The blue was in the models
for the primary hepatocytes,

00:54:53.480 --> 00:54:57.030
but for some reason didn't
exist in the tumor cell lines.

00:54:57.030 --> 00:55:00.300
So we're signaling logic
that normal hepatocytes use,

00:55:00.300 --> 00:55:03.870
that the tumor cell
lines have somehow lost.

00:55:03.870 --> 00:55:07.977
Red, are arcs that weren't
in the primary cells,

00:55:07.977 --> 00:55:09.560
but showed up in the
tumor cell lines.

00:55:09.560 --> 00:55:12.160
So was logic that the normal
liver cells apparently

00:55:12.160 --> 00:55:16.912
didn't use, but now showed
up in the tumor cell lines.

00:55:16.912 --> 00:55:18.620
And why would there
be these differences?

00:55:18.620 --> 00:55:21.850
Well this is where it goes back
to then the genetic mutations

00:55:21.850 --> 00:55:24.170
and variations.

00:55:24.170 --> 00:55:27.880
Because going from a primary
to some tumor cell line,

00:55:27.880 --> 00:55:32.540
there's enough of the genetic
mutations, that in this case

00:55:32.540 --> 00:55:36.460
said, OK, I've got some genetic
mutation that interrupts

00:55:36.460 --> 00:55:39.680
the link between map
three kinase and Ikk.

00:55:39.680 --> 00:55:41.700
There was some docking
protein or something

00:55:41.700 --> 00:55:45.250
that's now missing, not
expressed as highly.

00:55:45.250 --> 00:55:46.730
It's got a mutation
of amino acids

00:55:46.730 --> 00:55:48.300
and no longer docks right.

00:55:48.300 --> 00:55:50.506
It has a lower
enzymatic activity.

00:55:50.506 --> 00:55:51.880
So now you can go
back and trace.

00:55:51.880 --> 00:55:54.410
Can I find some
genetic mutation that

00:55:54.410 --> 00:55:57.290
has to do with the
loss of that arc?

00:55:57.290 --> 00:55:59.015
Or if I've got a red
arc that shows up--

00:55:59.015 --> 00:56:00.640
like I said because
there was something

00:56:00.640 --> 00:56:06.230
in my genetic mutations that
now adds an activity here

00:56:06.230 --> 00:56:07.270
that wasn't there.

00:56:07.270 --> 00:56:09.920
Maybe something is now
constituently active.

00:56:09.920 --> 00:56:12.360
Maybe something is just
expressed at a higher level.

00:56:12.360 --> 00:56:15.179
And all of a sudden that
pathway comes into play.

00:56:15.179 --> 00:56:16.220
So that's the cool thing.

00:56:16.220 --> 00:56:19.270
You can trace what's actually
in the genetic mutations

00:56:19.270 --> 00:56:21.400
if you have some
methodology for that,

00:56:21.400 --> 00:56:24.510
to what's actually been
altered in the network logic.

00:56:24.510 --> 00:56:26.576
Yeah?

00:56:26.576 --> 00:56:28.950
AUDIENCE: Are the primary
lines considered healthy lines?

00:56:28.950 --> 00:56:29.530
Or are they--

00:56:29.530 --> 00:56:29.970
DOUG LAUFFENBURGER: Yes.

00:56:29.970 --> 00:56:30.887
AUDIENCE: OK, so the--

00:56:30.887 --> 00:56:31.928
DOUG LAUFFENBURGER: Yeah.

00:56:31.928 --> 00:56:33.370
So they're from
donors but they're

00:56:33.370 --> 00:56:40.120
mainly like motorcycle accident
donors that don't either

00:56:40.120 --> 00:56:43.370
liver anymore but
the liver was fine.

00:56:43.370 --> 00:56:45.130
So, yeah, they're
from healthy donors.

00:56:45.130 --> 00:56:45.810
AUDIENCE: [INAUDIBLE].

00:56:45.810 --> 00:56:46.360
DOUG LAUFFENBURGER: Yeah.

00:56:46.360 --> 00:56:46.860
Yeah.

00:56:46.860 --> 00:56:49.530
It was the lines at some
point came from a tumor

00:56:49.530 --> 00:56:51.591
and have been propagated
in a culture, yeah.

00:56:56.310 --> 00:56:56.950
OK.

00:56:56.950 --> 00:57:01.040
What do I want to-- got
a little bit more time.

00:57:01.040 --> 00:57:01.740
Let me do this.

00:57:01.740 --> 00:57:02.250
OK.

00:57:02.250 --> 00:57:05.100
So here's another interesting
thing that can happen.

00:57:05.100 --> 00:57:06.730
If you take these
models seriously,

00:57:06.730 --> 00:57:10.050
it can tell you something
about the biochemistry, perhaps

00:57:10.050 --> 00:57:10.910
of what's going on.

00:57:16.440 --> 00:57:21.120
So see there's this dashed line
here that I want to emphasize

00:57:21.120 --> 00:57:23.510
and we'll emphasize it
again on another slide.

00:57:23.510 --> 00:57:24.950
That was one that
had to be added.

00:57:24.950 --> 00:57:30.970
It just wasn't in the
Ingenuity pathway, scaffold.

00:57:30.970 --> 00:57:33.440
Actually couldn't find it
in any literature anywhere.

00:57:33.440 --> 00:57:37.100
But nonetheless you needed
it to fit some data.

00:57:37.100 --> 00:57:38.760
So we kind of kept
our eye on that one.

00:57:38.760 --> 00:57:41.090
What the heck is going on here?

00:57:41.090 --> 00:57:45.450
This dashed line from I kappa
kinase up to step three.

00:57:45.450 --> 00:57:49.620
No evidence for that signaling
linkage in the literature

00:57:49.620 --> 00:57:51.200
anywhere.

00:57:51.200 --> 00:57:52.272
What could that tell you?

00:57:56.150 --> 00:57:56.650
All right.

00:57:56.650 --> 00:57:58.730
Well, you go back to the
data now and you say,

00:57:58.730 --> 00:58:01.590
well what of the data set, of
the experimental measurements

00:58:01.590 --> 00:58:04.400
that we made, caused
that arc to have

00:58:04.400 --> 00:58:06.071
to be there to
fit the data well?

00:58:06.071 --> 00:58:06.570
OK.

00:58:06.570 --> 00:58:08.590
You can now ask that
kind of question.

00:58:08.590 --> 00:58:11.437
Well remember I said in the
data set were inhibitors.

00:58:11.437 --> 00:58:13.520
Some small molecule
inhibitors against this kinase

00:58:13.520 --> 00:58:16.660
or that kinase or that kinase
that would perturb the network

00:58:16.660 --> 00:58:19.005
and then give us relationships
at the logic model

00:58:19.005 --> 00:58:21.820
and had to account for.

00:58:21.820 --> 00:58:24.810
Well, this one had
to be there, mainly

00:58:24.810 --> 00:58:29.270
to account for data that came
from an inhibitor of Ikk.

00:58:29.270 --> 00:58:32.250
That one of the kinases
that we had a small molecule

00:58:32.250 --> 00:58:36.000
inhibitor for,
inhabited this kinase.

00:58:36.000 --> 00:58:37.690
And somehow there
turned out to be

00:58:37.690 --> 00:58:40.740
an effect on staph
3 phosphorylation.

00:58:40.740 --> 00:58:44.810
And so you needed
that arc to be there.

00:58:44.810 --> 00:58:46.810
So either the explanation
that either there's,

00:58:46.810 --> 00:58:49.800
in fact, some real
mechanism going on here.

00:58:49.800 --> 00:58:51.360
It might have been
transcriptional

00:58:51.360 --> 00:58:53.880
that somehow the
activity of this kinase

00:58:53.880 --> 00:58:57.760
affects the levels of expression
and the responsiveness

00:58:57.760 --> 00:58:59.260
of staph 3.

00:58:59.260 --> 00:59:02.370
Or you say, ah, maybe it's
a problem with the drug?

00:59:02.370 --> 00:59:04.074
It's a problem
with the inhibitor.

00:59:04.074 --> 00:59:06.490
That, in fact, what you thought
was an inhibitor that just

00:59:06.490 --> 00:59:09.320
affected this kinase, has
an off-target target effect

00:59:09.320 --> 00:59:10.980
on that kind of that kinase.

00:59:10.980 --> 00:59:12.880
And it's just an artifact.

00:59:12.880 --> 00:59:14.820
That's an alternative
explanation.

00:59:14.820 --> 00:59:16.820
Right, so that's the sort
of thing you can test.

00:59:16.820 --> 00:59:19.420
And we did test it.

00:59:19.420 --> 00:59:21.110
And here's the data here.

00:59:21.110 --> 00:59:26.130
At the bottom is the kinase that
you wanted the inhibition 2.

00:59:26.130 --> 00:59:30.230
And in the blue was the
inhibitor that was actually

00:59:30.230 --> 00:59:33.990
used in the study, both
in vivo and en vitro

00:59:33.990 --> 00:59:37.080
and it inhibited that kinase.

00:59:37.080 --> 00:59:39.930
But then we looked at the
potential off target effect

00:59:39.930 --> 00:59:43.550
on that other-- the
JAK2 [? stat ?] 3

00:59:43.550 --> 00:59:46.560
and it also did have
activity on that.

00:59:46.560 --> 00:59:50.400
So it meant that
that inhibitor had

00:59:50.400 --> 00:59:55.410
an effect, not just on the Ikk,
but also on the JAK [? stat ?]

00:59:55.410 --> 00:59:57.030
3.

00:59:57.030 --> 01:00:01.380
And so that's why that
arc had to be there,

01:00:01.380 --> 01:00:03.130
is because, in fact,
that inhibitor,

01:00:03.130 --> 01:00:05.240
inhibited this kinase as well.

01:00:05.240 --> 01:00:07.830
So if we took that into account
in terms of the algorithm,

01:00:07.830 --> 01:00:09.980
then we wouldn't have to
have that arc because it

01:00:09.980 --> 01:00:11.900
was spurious and came
from the arc, in fact,

01:00:11.900 --> 01:00:13.090
of that inhibitor.

01:00:13.090 --> 01:00:15.670
But the interesting thing
is that, by taking the model

01:00:15.670 --> 01:00:18.500
seriously, we can
actually find that.

01:00:18.500 --> 01:00:22.400
Because it was not previously
known that this inhibitor had

01:00:22.400 --> 01:00:25.560
an off-target effect
on that kinase.

01:00:25.560 --> 01:00:28.620
In effect, the interesting
thing, pharmacologically,

01:00:28.620 --> 01:00:34.840
was that this
small molecule that

01:00:34.840 --> 01:00:37.550
was aimed to be an inhibitor
against this kinase

01:00:37.550 --> 01:00:42.180
was the best by far in treating
lung airway inflammation,

01:00:42.180 --> 01:00:43.960
compared against
a whole other set

01:00:43.960 --> 01:00:46.690
of other types of inhibitors
for the same kinase.

01:00:46.690 --> 01:00:48.520
So now the reason
might be is, it's

01:00:48.520 --> 01:00:51.010
better because it's also
hitting this other kinase.

01:00:51.010 --> 01:00:52.610
That this off-target
effect actually

01:00:52.610 --> 01:00:55.920
is therapeutically
efficacious and in fact

01:00:55.920 --> 01:00:58.830
a combination of drugs
against this kinase

01:00:58.830 --> 01:01:01.270
and the other kinase
is what's required

01:01:01.270 --> 01:01:03.600
for the therapeutic benefit.

01:01:03.600 --> 01:01:05.750
So that's something
that could be explored.

01:01:05.750 --> 01:01:07.920
And that's the sort of
thing this model leads to.

01:01:10.770 --> 01:01:11.270
OK.

01:01:14.096 --> 01:01:20.220
Let me end by digging into
this difference a little bit.

01:01:20.220 --> 01:01:24.700
Because I said, you
see these differences

01:01:24.700 --> 01:01:31.080
between primary hepatocytes
and the tumor cell lines.

01:01:31.080 --> 01:01:34.290
And the model said, just
from examining the data sets,

01:01:34.290 --> 01:01:36.175
that the logic is different.

01:01:36.175 --> 01:01:36.820
OK.

01:01:36.820 --> 01:01:40.280
Is there any
validation for that?

01:01:40.280 --> 01:01:43.315
Well, so let's go back and
look at those differences

01:01:43.315 --> 01:01:44.440
with respect to literature.

01:01:44.440 --> 01:01:47.920
So if you just blow up
that part of the model,

01:01:47.920 --> 01:01:50.830
there's eight edges
that are strongly

01:01:50.830 --> 01:01:53.940
disparate between the primary,
normal cell types and the tumor

01:01:53.940 --> 01:01:55.940
cells and they're
all enumerated here.

01:01:55.940 --> 01:01:59.240
One, two, three, four,
five, six, seven, eight.

01:01:59.240 --> 01:02:02.582
And they're essentially in
three different pathways.

01:02:02.582 --> 01:02:04.040
So what the model
is telling you is

01:02:04.040 --> 01:02:08.670
that there's three different
pathways that are substantially

01:02:08.670 --> 01:02:13.468
different between a normal liver
cell and a liver tumor cell.

01:02:13.468 --> 01:02:14.460
OK.

01:02:14.460 --> 01:02:19.470
So is there any evidence
that this is really true?

01:02:19.470 --> 01:02:20.750
So let's look at one.

01:02:20.750 --> 01:02:23.165
On to this pathway that
I've got differences.

01:02:23.165 --> 01:02:25.090
And you see blue
here and red here.

01:02:29.020 --> 01:02:32.385
It says that this particular
signaling node in normal cells

01:02:32.385 --> 01:02:34.650
is activated by this pathway.

01:02:34.650 --> 01:02:36.430
In the tumors, that
regulation is lost

01:02:36.430 --> 01:02:40.160
and that actually comes
through another pathway.

01:02:40.160 --> 01:02:45.550
And it turns this is
consistent with literature

01:02:45.550 --> 01:02:47.910
that, in fact, in
the tumor cells,

01:02:47.910 --> 01:02:52.350
you get a higher activity
of this downstream node.

01:02:52.350 --> 01:02:53.940
And now I've lost
my light again.

01:02:53.940 --> 01:02:56.470
This HSP27.

01:02:56.470 --> 01:02:59.560
Even though it's
over expressed, you

01:02:59.560 --> 01:03:03.930
get less activation because
this pathway is less strongly

01:03:03.930 --> 01:03:07.490
activated in red than
the blue pathway is.

01:03:07.490 --> 01:03:08.930
So if you went by
gene expression,

01:03:08.930 --> 01:03:10.500
you'd think in the
tumor cells, this

01:03:10.500 --> 01:03:12.706
is a higher activated pathway.

01:03:12.706 --> 01:03:14.080
Turns out the
logic is different,

01:03:14.080 --> 01:03:15.870
and you actually get
less activation of it

01:03:15.870 --> 01:03:17.745
because it's coming from
a different pathway.

01:03:17.745 --> 01:03:21.740
So that turns out to be true
in the liver tumor literature.

01:03:21.740 --> 01:03:24.530
Another one-- I find this
one really interesting.

01:03:24.530 --> 01:03:29.110
That in normal liver cells,
to activate this Ikk pathway--

01:03:29.110 --> 01:03:30.910
that's a very important
kinase pathway,

01:03:30.910 --> 01:03:33.660
governing the transcription
factor of NF Kappa b.

01:03:33.660 --> 01:03:37.210
In a primary cell, I
need this combined logic

01:03:37.210 --> 01:03:39.455
between a pathway downstream
of insulin receptor

01:03:39.455 --> 01:03:42.060
and a pathway downstream
of a cytokine.

01:03:42.060 --> 01:03:46.520
Only if both of those pathways
are on, do I now turn this on.

01:03:46.520 --> 01:03:48.890
In the tumor cells,
that check is lost.

01:03:48.890 --> 01:03:51.690
Only one pathway is required.

01:03:51.690 --> 01:03:52.190
OK.

01:03:52.190 --> 01:03:53.700
If this one is
activated, I'm going

01:03:53.700 --> 01:03:56.140
to get this transcription
factor activated.

01:03:56.140 --> 01:03:58.820
I don't have to wait for
simultaneous activation

01:03:58.820 --> 01:03:59.800
of this pathway.

01:03:59.800 --> 01:04:02.650
Where as a normal
says I have to.

01:04:02.650 --> 01:04:03.470
OK.

01:04:03.470 --> 01:04:06.220
That turns out to be true
that in the liver cells,

01:04:06.220 --> 01:04:09.040
the progression is
associated with a looser

01:04:09.040 --> 01:04:13.810
regulation of this
transcription factor.

01:04:13.810 --> 01:04:15.130
And one more.

01:04:15.130 --> 01:04:18.870
I won't go into too much
detail, but again, you

01:04:18.870 --> 01:04:21.310
see reds and blues here.

01:04:21.310 --> 01:04:23.280
In the tumor cell
lines, you've now

01:04:23.280 --> 01:04:25.810
got activities
downstream of insulin.

01:04:25.810 --> 01:04:27.770
That's normally just
a survival factor,

01:04:27.770 --> 01:04:30.490
that's just not found
in the primary cells.

01:04:30.490 --> 01:04:33.870
And that, in fact, is shown
in the literature too,

01:04:33.870 --> 01:04:37.500
that insulin signaling
shifts from metabolism

01:04:37.500 --> 01:04:38.785
to proliferation.

01:04:38.785 --> 01:04:40.384
It's mainly metabolic, stimulus.

01:04:40.384 --> 01:04:42.800
In the normal cells it turns
into a proliferative stimulus

01:04:42.800 --> 01:04:44.500
in the tumor cells.

01:04:44.500 --> 01:04:45.000
OK.

01:04:45.000 --> 01:04:50.090
So, what this says is,
just by mapping this logic

01:04:50.090 --> 01:04:53.670
scaffold, the scaffold
against empirical data,

01:04:53.670 --> 01:04:55.390
developing a logic
model, you in fact

01:04:55.390 --> 01:04:59.650
can find loci of differences
between the normal cell

01:04:59.650 --> 01:05:02.060
signaling logic and tumor
cells signalling logic

01:05:02.060 --> 01:05:05.200
for which there's evidence in
the literature, none of which

01:05:05.200 --> 01:05:07.085
was in the original databases.

01:05:09.726 --> 01:05:11.100
Finally, I'm going
I'm just going

01:05:11.100 --> 01:05:15.560
to say that it turns out in
another study, what you could

01:05:15.560 --> 01:05:19.259
show is those three pathways
that the model predicts

01:05:19.259 --> 01:05:21.050
are the differences
between the liver tumor

01:05:21.050 --> 01:05:22.800
cells and the normal cells.

01:05:22.800 --> 01:05:27.410
That in order to kill
these liver tumor cells,

01:05:27.410 --> 01:05:30.460
you need inhibitors against all
three pathways simultaneously.

01:05:30.460 --> 01:05:32.220
You actually need
combination drugs

01:05:32.220 --> 01:05:35.760
of three different pathway
inhibitors to kill these cells.

01:05:35.760 --> 01:05:37.730
And it's exactly
the three pathways

01:05:37.730 --> 01:05:40.610
that the model predicted of the
differences between the normals

01:05:40.610 --> 01:05:42.030
and the tumor cells.

01:05:45.790 --> 01:05:46.600
OK.

01:05:46.600 --> 01:05:48.700
All right, so I will
end here and then see

01:05:48.700 --> 01:05:50.510
if there's any more questions.

01:05:50.510 --> 01:05:53.380
Something that
comes up a lot is--

01:05:53.380 --> 01:05:56.520
there's discomfort with Boolean
logic because of zero, one.

01:05:56.520 --> 01:05:58.880
It's off, on, and of
course we know biology,

01:05:58.880 --> 01:06:01.260
biochemistry doesn't
work that way.

01:06:01.260 --> 01:06:03.430
And so there can be
so many artifacts,

01:06:03.430 --> 01:06:05.390
so many places that you
can get things wrong,

01:06:05.390 --> 01:06:09.210
because you're trying to fit a
model where the measurement is

01:06:09.210 --> 01:06:11.147
supposed to be
either zero or one,

01:06:11.147 --> 01:06:13.230
and you're comparing it
against a measurement that

01:06:13.230 --> 01:06:15.840
might be 0.6.

01:06:15.840 --> 01:06:19.900
Well, 0.6, is that closer
to 1, is it closer to 0?

01:06:19.900 --> 01:06:22.624
Is there some normalization
that would shift it

01:06:22.624 --> 01:06:23.540
from one to the other.

01:06:23.540 --> 01:06:26.510
And instead of being a correct
fit, it's now an incorrect fit.

01:06:26.510 --> 01:06:28.360
So you can see the
room for artifacts

01:06:28.360 --> 01:06:34.470
by mapping quantitative data
against a qualitative model.

01:06:34.470 --> 01:06:37.490
So, one thing done more
recently is to admit that

01:06:37.490 --> 01:06:40.820
and say, well, let's say
just relax this a bit.

01:06:40.820 --> 01:06:45.470
And instead of having step
functions from off to on,

01:06:45.470 --> 01:06:46.820
that they're more graded.

01:06:46.820 --> 01:06:50.600
It's like an analog
transfer function.

01:06:50.600 --> 01:06:53.940
So what you've essentially
done is add one more parameter

01:06:53.940 --> 01:06:57.010
to every node, to every gate.

01:06:57.010 --> 01:06:58.920
Because of Boolean
logic, there's

01:06:58.920 --> 01:07:00.670
essentially one
hidden parameter.

01:07:00.670 --> 01:07:02.910
That's where you shift
from off to on, right?

01:07:02.910 --> 01:07:07.280
There's some location of
the level of the signal

01:07:07.280 --> 01:07:09.554
that you've decided is 0 or 1.

01:07:09.554 --> 01:07:11.470
So there's some parameter
that you shift from,

01:07:11.470 --> 01:07:13.710
saying it's off to on.

01:07:13.710 --> 01:07:16.430
Well here now in this
formalism there's that,

01:07:16.430 --> 01:07:19.570
but there's also then the slope
of shifting from off to on.

01:07:19.570 --> 01:07:21.580
Is it still fairly steep?

01:07:21.580 --> 01:07:23.040
Is it really mild?

01:07:23.040 --> 01:07:25.611
Is it someplace in between?

01:07:25.611 --> 01:07:26.110
OK?

01:07:26.110 --> 01:07:29.310
And this can go with
AND and OR gates too.

01:07:29.310 --> 01:07:30.830
Now, instead of
just one dimension,

01:07:30.830 --> 01:07:34.020
one component being
off to on or on to off,

01:07:34.020 --> 01:07:38.460
now you got AND and OR gates
that have these slopes as well.

01:07:38.460 --> 01:07:42.960
So what this means is
you require more data

01:07:42.960 --> 01:07:46.090
to fit this-- we call it a
constrained fuzzy logic model

01:07:46.090 --> 01:07:49.120
because you've got-- if I've
got 50 nodes in my system,

01:07:49.120 --> 01:07:51.985
I've got 50 more
parameters I've got to fit.

01:07:51.985 --> 01:07:54.450
OK, so that requires more data.

01:07:54.450 --> 01:07:59.930
What's the benefit of it, is
that your predictions now,

01:07:59.930 --> 01:08:01.441
in fact, can be quantitative.

01:08:01.441 --> 01:08:02.940
So you can go into
the model and say

01:08:02.940 --> 01:08:05.340
here's a transcription
factor CREB.

01:08:05.340 --> 01:08:08.320
I'm going to predict its
phosphorylation state

01:08:08.320 --> 01:08:10.600
and its transcriptional
activity, perhaps,

01:08:10.600 --> 01:08:13.694
based on the activities
of two upstream kinases.

01:08:13.694 --> 01:08:16.069
And so if I had had an inhibitor
for one of these kinases

01:08:16.069 --> 01:08:19.279
or another, how much would
I shift the phosphorylation

01:08:19.279 --> 01:08:21.270
of this transcription factor?

01:08:21.270 --> 01:08:25.390
And what you actually see is
these gradual curves, that if I

01:08:25.390 --> 01:08:29.109
start to inhibit
[INAUDIBLE], OK, it gradually

01:08:29.109 --> 01:08:31.399
changes the
phosphorylation of CREB.

01:08:31.399 --> 01:08:34.789
Or if I inhibit the
activity of P38,

01:08:34.789 --> 01:08:39.250
it even more gradually
effects the activity of CREB.

01:08:39.250 --> 01:08:42.295
So you can turn these into
quantitative predictions

01:08:42.295 --> 01:08:45.000
of strong effects, weak effects.

01:08:45.000 --> 01:08:47.149
And again, look at
drug combinations.

01:08:47.149 --> 01:08:51.649
So that's the advantage of going
to this more analog transfer

01:08:51.649 --> 01:08:53.950
function logic model.

01:08:53.950 --> 01:08:57.430
You can deal with
quantification much better,

01:08:57.430 --> 01:08:59.938
but at the cost of
requiring more data.

01:08:59.938 --> 01:09:01.229
OK, I think I'll leave it here.

01:09:01.229 --> 01:09:03.140
It's about 3:15
and so if there's

01:09:03.140 --> 01:09:08.090
more questions we can take
them about any aspect of this.

01:09:08.090 --> 01:09:11.933
Most of you have stayed awake,
I think that's a good thing.

01:09:11.933 --> 01:09:12.710
OK.

01:09:12.710 --> 01:09:13.366
More questions?

01:09:20.104 --> 01:09:21.020
AUDIENCE: [INAUDIBLE].

01:09:21.020 --> 01:09:25.970
When you have the model for
the template of the Ikk story.

01:09:25.970 --> 01:09:28.069
And then it seems
like it may not

01:09:28.069 --> 01:09:31.822
be as easy to back out
the original data that

01:09:31.822 --> 01:09:33.485
led to that specific mode.

01:09:33.485 --> 01:09:35.100
For example, you
showed that one arc

01:09:35.100 --> 01:09:38.035
was from this one treatment.

01:09:38.035 --> 01:09:41.182
But because if you trained
the model the same as that

01:09:41.182 --> 01:09:45.000
and it's not deterministic,
then what-- could you just add--

01:09:45.000 --> 01:09:48.720
DOUG LAUFFENBURGER: I think
that's a great question.

01:09:48.720 --> 01:09:52.827
So let's say there's
new arcs that you add,

01:09:52.827 --> 01:09:54.410
that weren't in the
original scaffold.

01:09:54.410 --> 01:09:56.660
I mean that's what you got
the biggest questions from.

01:09:56.660 --> 01:09:58.620
If you delete one,
you say, ah, it's

01:09:58.620 --> 01:10:01.420
easy to believe why
you would delete one.

01:10:01.420 --> 01:10:03.480
Any arc that you add
to get a best fit,

01:10:03.480 --> 01:10:05.920
I think you've got to
ask questions about.

01:10:05.920 --> 01:10:07.450
So in all those
cases where there

01:10:07.450 --> 01:10:10.320
are arcs that were added that
led to a better fit model,

01:10:10.320 --> 01:10:12.610
the first thing we did
was go to the literature.

01:10:12.610 --> 01:10:17.480
Say, OK, is there literature
on some affect of this node

01:10:17.480 --> 01:10:19.160
to that node?

01:10:19.160 --> 01:10:20.950
And it's just that
that literature

01:10:20.950 --> 01:10:22.560
wasn't curated
into that database

01:10:22.560 --> 01:10:24.830
or something like that.

01:10:24.830 --> 01:10:27.730
And most of the time
we could find it there.

01:10:27.730 --> 01:10:28.490
OK.

01:10:28.490 --> 01:10:31.720
So then there were the cases,
and this was the most prominent

01:10:31.720 --> 01:10:35.270
one, where from some
added arc, we just

01:10:35.270 --> 01:10:36.770
couldn't find it
in the literature.

01:10:36.770 --> 01:10:38.280
In this particular
case, it was very

01:10:38.280 --> 01:10:40.830
easy to trace it to
this particular effect

01:10:40.830 --> 01:10:43.412
of this inhibitor.

01:10:43.412 --> 01:10:45.870
I would say there's no reason
to believe that that's always

01:10:45.870 --> 01:10:46.960
going to be the case.

01:10:46.960 --> 01:10:49.790
I don't have another example to
show you where it was harder.

01:10:49.790 --> 01:10:52.804
Everything else we actually
found in the literature.

01:10:52.804 --> 01:10:55.220
But you could imagine, having
some new arc that you really

01:10:55.220 --> 01:10:57.250
couldn't find in the
literature and there's

01:10:57.250 --> 01:11:00.330
no artifactual
explanation for it.

01:11:00.330 --> 01:11:04.250
And now, how you trace it
back to what the data was

01:11:04.250 --> 01:11:07.274
that might give you
a more nuanced hint.

01:11:07.274 --> 01:11:08.190
It's a great question.

01:11:08.190 --> 01:11:11.760
I don't really know
how we'll do that.

01:11:11.760 --> 01:11:14.450
I think, we and other
practitioners who use this,

01:11:14.450 --> 01:11:17.310
I'm sure we'll run
into it at some point.

01:11:17.310 --> 01:11:22.108
That's a great challenge
to be thinking about.

01:11:22.108 --> 01:11:22.608
Yeah?

01:11:22.608 --> 01:11:24.596
AUDIENCE: I might have
missed this earlier,

01:11:24.596 --> 01:11:30.680
but I was wondering,
is this model actually

01:11:30.680 --> 01:11:34.600
able to incorporate the
heterogeneity of a tumor,

01:11:34.600 --> 01:11:35.580
for example?

01:11:35.580 --> 01:11:40.010
Or the population heterogeneity?

01:11:40.010 --> 01:11:44.068
DOUG LAUFFENBURGER: That's also
a really interesting question.

01:11:44.068 --> 01:11:47.470
Let me try to show
something here.

01:11:47.470 --> 01:11:50.380
Yeah.

01:11:50.380 --> 01:11:52.320
So two things.

01:11:52.320 --> 01:11:58.210
One is, what's shown here--
this is the four different tumor

01:11:58.210 --> 01:12:00.270
cells that we did.

01:12:00.270 --> 01:12:05.480
And what's shown in color is
the arcs for each one of them.

01:12:05.480 --> 01:12:09.510
So yellow, orange, brown, red.

01:12:09.510 --> 01:12:12.840
So some places you see all
four of those colors there.

01:12:12.840 --> 01:12:15.150
In some places only two or one.

01:12:15.150 --> 01:12:17.650
It says if I had four
different tumor types,

01:12:17.650 --> 01:12:21.450
there's some slight differences
in logic among them.

01:12:21.450 --> 01:12:23.340
You could translate
to that is, well

01:12:23.340 --> 01:12:25.230
I could imagine
then having a tumor

01:12:25.230 --> 01:12:31.360
that's a mixture of sub types
and how would I discern that?

01:12:31.360 --> 01:12:35.120
One possible idea that's
attractive to me--

01:12:35.120 --> 01:12:37.450
although we didn't really
explore this in any form of

01:12:37.450 --> 01:12:37.640
[INAUDIBLE].

01:12:37.640 --> 01:12:39.350
We didn't really have
the means to make

01:12:39.350 --> 01:12:40.933
experimental
measurements on the tumor

01:12:40.933 --> 01:12:42.790
heterogeneity at the time.

01:12:42.790 --> 01:12:45.570
It's when you get to a
set of consensus models.

01:12:45.570 --> 01:12:46.070
Right.

01:12:46.070 --> 01:12:49.450
So let's say you get the 50
best fit models and you say,

01:12:49.450 --> 01:12:55.530
some arc is in 80% of them,
but it's not in 20% of them,

01:12:55.530 --> 01:12:57.980
is it possible that
that represents

01:12:57.980 --> 01:13:02.250
some of heterogeneity
because you're

01:13:02.250 --> 01:13:05.860
getting an average of
different subtypes?

01:13:05.860 --> 01:13:06.730
I don't know.

01:13:06.730 --> 01:13:09.650
Sometimes that appeals to
me as potentially valid.

01:13:09.650 --> 01:13:13.820
Sometimes I think there's
a flaw in that reasoning.

01:13:13.820 --> 01:13:17.695
Just because you get an average
that's not then as strong.

01:13:17.695 --> 01:13:21.580
Does that necessarily
reflect a sub-population?

01:13:21.580 --> 01:13:24.230
I don't know.

01:13:24.230 --> 01:13:26.700
So what we do know is,
we can see differences

01:13:26.700 --> 01:13:28.770
when there are differences.

01:13:28.770 --> 01:13:31.380
How you actually see
them then, if all you

01:13:31.380 --> 01:13:34.670
have is averaged
data, maybe it's

01:13:34.670 --> 01:13:38.110
reflected in the heterogeneity
of the consensus models.

01:13:38.110 --> 01:13:38.860
Maybe not.

01:13:41.710 --> 01:13:43.500
It would be an
interesting to explore.

01:13:48.349 --> 01:13:48.890
Anybody else?

01:13:54.940 --> 01:13:55.440
All right.

01:13:55.440 --> 01:13:55.940
All set.

01:13:55.940 --> 01:13:57.280
Thanks.