“Biomarkers and Enrichment for Sepsis Trials” by Hector Wong for OPENPediatrics

This video is from the PICC World Congress from June 2016 where Dr. Wong talks about using biomarkers and his risk model PERSEVERE in sepsis clinical trials.

Please visit: www.openpediatrics.org

OPENPediatrics™ is an interactive digital learning platform for healthcare clinicians sponsored by Boston Children’s Hospital and in collaboration with the World Federation of Pediatric Intensive and Critical Care Societies. It is designed to promote the exchange of knowledge between healthcare providers around the world caring for critically ill children in all resource settings. The content includes internationally recognized experts teaching the full range of topics on the care of critically ill children. All content is peer-reviewed and open access-and thus at no expense to the user.

For further information on how to enroll, please email:

Please note: OPENPediatrics does not support nor control any related videos in the sidebar, these are placed by Youtube. We apologize for any inconvenience this may cause.

---

Biomarkers and Enrichment
for Sepsis Trials, By Dr. Hector Wong. Biomarkers for Enrichment,
these are my disclosures. The work that I'll show you is
funded by grants from the NIH. I have a couple of advisory
board appointments, Neither of which are at
all relevant to this talk. But what is relevant
to this talk Is that my
institution and I hold Patents for some of these
stratification biomarkers. So most of us, when we think
about biomarkers of sepsis, We think of diagnosis, right? Biomarkers to be able to
differentiate between SIRS, And sepsis, or sterile
inflammation, and sepsis, As Dr. Peters spoke
about earlier. And this is a really,
really important topic, As Dr. Peter said. But it's not what we're
going to be focusing on. It's not so much
about diagnosis, But it's more for this
concept of enrichment. OK. And so what does
enrichment mean? I think Dr. Peters
alluded to this earlier, But what enrichment
means is that one Uses patient characteristics,
any patient characteristic To select a population in which
an intervention, such as ECMO, Perhaps, is more
likely to be detected In an unselected population. All right. And then that's
the broad concept, There's some more
specific comments. There's a concept of prognostic
enrichment, in which you select A population that has a
higher likelihood of having

A disease related event, such
as mortality or organ failure. And then there's a concept
of predictive enrichment, In which you select
a population that Is more likely to respond
to your intervention based On biology, or based
on a mechanism. So those are the
concepts, and we're Going to talk about biomarkers
to address these concepts. And the way we've
approached this, Is that we've had the
good fortune of being Able to conduct
genome wide expression Studies in many, many
children, well over 200, From across the country in
a very exploratory approach. And we used bioinformatics
to try to identify Enrichment biomarkers. And I'll talk with
you about first Are prognostic
enrichment biomarkers. And basically,
without taking you Through all the details,
what this means is That we were able to find
or identify about 100 genes That seem to have some
predictive capacity For mortality. And so these genes are all
obviously at the mRNA level, So we wanted to pare this down
to something that is perhaps More translatable, or more
feasible in a clinical setting. So we said OK, among
these 100 genes, What is biologically plausible? And then perhaps
even more important, Pragmatically among those
genes, what can we actually Measure in a plasma or a serum
compartment of their protein. Because we want to try to
get to actually a feasible Clinical test. And so using those criteria, we
came up with this list of 12. It's also published,
if you're interested. And so using those
12 biomarkers,

What we've tried to
do is we've tried To develop what we call the
Pediatric Sepsis Biomarker Risk Model, which is
PERSEVERE for short. The initial iteration of
this evolved over 350 kids From about 17 institutions,
11% mortality. And basically, what we did
is that we measured these 12 Candidate stratification
biomarkers from the serum Compartment. So we're shifting
now mRNA expression To protein expression. Importantly, these biomarkers
were assayed, or measured, Or the samples came
from samples that Were drawn within the first 24
hours of coming into the ICU. And I would argue
that that's when You want to be able to make
these kinds of predictions. And in the modeling approach, we
looked at a variety of models. And in the end, we settled
on this modeling approach Called CART, which is an
acronym for classification And regression tree. And basically, what
that allows you to do Is to build a decision
tree with this process Of binary recursive partitioning
that allows you to provide A mortality probability. And so this is what
the tree looks like, It's not as complicated
as it looks. It's basically
inverted, if you will, That the root is at the top. OK? And that the root
node contains all Of the subjects in
that initial study With their respective
mortalities and survivals. And so what this
decision tree does Is that it begins to
partition patients Into different levels of risk
based on a biomarker cutoff.

So that initial cutoff
there is based on CCL3, And then it
continues to do this. This is the concept of
recursive binary partitioning That it continues
to generate daughter Nodes into different
risk categories Until they can no longer
divide the patients any longer. And so the way to
read these trees Is to look at the
terminal nodes. And terminal has nothing
to do with mortality. Terminal means simply
that they can no longer Divide the patient. So in this case,
the terminal note That I'm highlighting
there, if one Has those series of biomarker
based decisions or criteria, You end up in that
terminal node, Then your mortality
probability is a very low, 0.011, as opposed to ending
up in this terminal node. Again, with those sets
of biomarker criteria, Your mortality
probability's about 0.472. And so that's the way
to think about these. And so in this
particular tree, there Are three low risk
terminal nodes, Within which the mortality
probability is very low. There are these intermediate
risk terminal nodes. And if you will for
lack of a better term, Mortality probability
is about 18% to 27%. And then you have these
two high risk terminal Nodes, in which mortality
rate is about 50% or greater. Alright. And that's the way
to think about this. Again, Dr. Peters
alluded to this, We all want to see
biomarkers give us Yes, no, right, dichotomous. And life is just
not that simple.

So what I would urge you
to think about biomarkers Is Dr. Peter's alluded
to as a probability. What is the probability
of mortality? What is the
probability of sepsis? What is the probability
of a certain outcome? That's the way I
believe that we should Be thinking about these
biomarkers and these models. Nonetheless, you can
still generate a two By two contingency table,
in which you categorize All the high and
intermediate risk Patients as predicted
non-survivors, Know the low risk as
predicted survivors. That's in the initial test,
this is how this came out. Pretty good performance
characteristics there, With an AUC approaching 0.8. And these models
require testing, And we've done multiple
subsequent validation study, And we see very
similar performance Across a variety of patients. We're up to well over
700 patients now. And so how does one apply this? So what? So one way is to inform
individual patient decision Making that perhaps one
could save or identify Patients that are higher
risk and you could allocate Those patients to
higher risk therapies, Such as perhaps ECMO, or
some other intervention. This is the same concept
of situational awareness. In certain settings, I
think this kind of modeling Could assist with
allocation of ICU resources, Where patients go in a given
hospital or a certain region. Benchmarking, we talk quite a
bit about improving outcomes In sepsis, and so forth. I think models such as
this could potentially

Serve as a benchmark
or a denominator. In other words, it can estimate
the probability of mortality In your particular
population, and that Can be your denominator after
you introduce some quality Improvement efforts. But ultimately, I
think where it could Be most powerful is for
prognostic enrichment Of clinical trials, so we'll
focus a little bit on that. So this concept of
prognostic enrichment Is relatively simple, it
just has a fancy name. So when one is thinking
about a clinical trial, One of the key decisions
is your sample size. And so when you're
thinking about sample size, You're thinking–
what affects that is Your effect size of
your intervention, But also your event rate. Ok. Those are the two
main factors that Go into calculating a
sample size for your trial. And so what prognostic
enrichment does, It selects a population
with a greater event Rate, that's the concept
of prognostic enrichment. And by doing so, you
could potentially Decrease the sample
size of your trial. And so we've not been able
to actually do this yet, We hope to, but we
haven't been able to do This in a prospective manner. But we've done the
next best thing, which Is a trial simulation
in what we focused on Are kids that have septic
shock and TAMOF, this concept Of thrombocytopenia-associated
multiple organ failure. And so the purported
intervention for TAMOF Is plasma exchange,
which is not benign, It's not cannulation
through the chest,

As Dr. McClaren
just spoke about. But nonetheless, it's exposure
to extracorporeal circuit, And so forth. And so we used the data
from this quasi trial that's Listed in
clinicaltrials.gov, and so we Looked at our
database, and we found 108 patients who met
the same criteria that Are listed for that trial. And so they would be
eligible theoretically, For plasmapheresis. And this group, just based
on meeting those criteria, Had a mortality rate of 38%. And so to each of
those patients we Assigned a PERSEVERE
mortality risk, And we found 39 true
positives, 35 false positives, And you can read the rest. So theoretically, in
this trial simulation, Again, this is all theory,
this is all computers, We would exclude these
patients that are low risk, Because they have a very low
probability of mortality. And you would
include the patients That are higher risk,
based on that model. And so what this does is
it selects a population With an overall
mortality rate of 53%, Remembering that the initial
mortality rate was 38%. So that's the concept of
increasing your event rate. And so then the trial
simulation really, We just assumed a range of
effects of plasmapheresis, Mortality reductions
ranging from 10% to 50%, Standard power and alpha. And those are the
number of patients That you would require if the
patients were not stratified. However, if you selected the
patients based on the model, You can see that the number of
patients for each size effect

Gets reduced by
about 40% or 45%. Not a very difficult
concept, but this Is the concept of prognostic
enrichment, in which you select A population with a
greater event rate, And then you can reduce
your effect size– I'm sorry, your sample size. So now, let's turn to
predictive enrichment. And again, what we've done here
through this bioinformatics And transcriptomics
is that we've Tried to identify what we
call septic shock endotypes. So what is an endotype? So an endotype is a subclass
of a disease or a syndrome As defined by biology or
some other characteristic. And so if you think about
septic shock as a syndrome– That hasn't really come
up yet in this session. Right. We're talking a
little bit about it As sort of a singular disease,
and it's anything but. It's really a syndrome, it's
a constellation of problems, And the patients
are very different. So if you think
about septic shock In that way as a syndrome,
it implies the existence Of endotypes. And so we hypothesize
in that these endotypes May have distinct
gene expression Patterns and biological
processes that may then in turn Need to distinct
clinical phenotypes. And then we asked the
simple question is, "Can by doing transcriptomics,
whole genome expression Profiling, can we identify
endotypes of septic shock?" Alright. And so this isn't
looking at gene variants, This is looking at gene
expression, very different. Alright.

And so I'm not going to drag
you through all the details, But basically, the
initial iteration Of this, we were able to
identify what we thought Were endotypes of septic
shock based on the expression Patterns of over 8,000 genes. And basically, what we
did there is we just Simply using clustering
algorithms that Were agnostic completely
to patient outcomes Or patient characteristics,
it basically Just clustered
patients together based On similarities or dissimilarity
of gene expression. Post-hoc we said, OK, so here's
a few groups of patients that Behave similarly from a
gene expression standpoint, But that's all well
and good, but are they Different clinically? Right, cause in the end,
that's what really matters. And so when we went
back and looked, It turned out that
what we call endotype A patients have much
higher illness severity. Their organ failure
burden is higher, And the mortality rate
is about threefold That of the other groups,
so 36% versus 11%. So that's all well and good. 8,000 genes, it's a lot of fun. I can do a lot of cool things
on my computer and things Like that, but doesn't help
us very much at the bedside. And so we're
interested in trying To get this to be a little
bit more clinically feasible, Trying to get this
to the bedside. And so we've approached
this now with this goal Of developing a clinical
test that can meet The time sensitive demands. Because we need to
get these kind of data

Back very, very, very quickly,
within a couple of hours. So step one is we went from
8,000 genes to 100 genes. 100 genes, measuring 100 genes
starts to get a little bit more Clinically feasible for the ICU. And then we're starting to
express these genes using This thing called GEDI, the Gene
Expression Dynamics Inspector. Which basically, what it does
is it gives complex array data, Or gene expression data a face. Alright. So there's an example
from their website. On your left is normal,
on the right is cancer, And without knowing the genes,
you can tell they're different. The other step that
we've taken is now We're measuring gene expression
using a multiplex mRNA Quantification platform that
gives you a digital readout. And this is standard
hybridization Of nucleic acid using both
reporter and capture probes. And basically, what
it allows you to do Is generate a count
of mRNA expression. The nice thing about
this is that you Can do it in solution phase. So it takes away another step,
there's no amplification, Such as PCR. So this is what the
pediatric endotypes Look like based on 100 genes
and that digital platform. Ok. So again, without knowing
anything about bioinformatics, Gene expression studies, and
so forth, you can look there. Patients on your
right, endotype A, Look very different than
the endotype B patients. And so the endotype A patients
tend to be younger, sicker, And have worse outcomes. But this isn't just a fancy
CBC, because those kids actually Have higher lymphocyte counts
than the endotype B patients.

And so if we do logistic
regressions that Accounts for age, illness
severity, comorbidity, And so forth, something
about being allocated To endotype A means you're
going to have a worse outcome. So those kids have almost
a threefold worse outcome. So I haven't told you
what the genes are. So what those genes represent
are adaptive immunity And the glucocorticoid
receptor signaling pathway. Adaptive immunity and the
glucocorticoid receptor Signaling pathway,
and those genes Are repressed in the
endotype A patients. That's the blue coloring that
you see relative to the red. So if this is true,
imagine that this is true, This may have implications
or might have implications For precision medicine. The new math now is
that we're saying well, All the old stuff of inhibiting
inflammation and so forth Hasn't worked. Now, we actually have to augment
the adaptive immune system. Biologically plausible,
but if this is true, One could imagine that these
two classes of patients Are going to have a
very different response To any kind of adaptive
immune enhancing strategy. Glucocorticoids. Ok. I think it's embarrassing
that it's 2016 And we still don't have an
answer of whether we should Be using hydrocortisone or not. So in any case, again,
if this is true, One could imagine that
these children, A versus B, Would have a different
response to hydrocortisone Or glucocorticoids because
the genes that correspond To the receptor in the
whole signaling pathway Are differentially expressed.

And we were actually able to
look this in a post-hoc manner. And surprisingly,
what it turns out Is that the kids that are
in endotype A, if they get Corticosteroids, they have
four times the risk of dying. And that's independent
of the illness severity, Independent of
comorbidities and so forth By logistic regression. It's all post hoc, though. So we think that because
these endotype-defining genes Correspond to adaptive immunity
and a glucocorticoid receptor Signaling pathway that
they may be able to serve, Or might be able to serve
as predictive enrichment Biomarkers. And so recently, we asked
a question is, "Can we Identify kids who may actually
benefit from corticosteroids If we combine both prognostic
and predictive enrichment?" And so we did that for
prognostic enrichment, We classified the kids
based on PERSEVERE, In which we assign them a
baseline mortality probability. For predictive
enrichment, we allocated Them to endotype A and
endotype B. Again, remember That the endotype
B patients actually Have expression
of the genes that Correspond to this pathway. And I'm not going to take
you through all the math, But among endotype
B patients who Have an intermediate to high
mortality risk at baseline, Adjunctive
corticosteroids actually Decrease mortality tenfold. Alright, so here,
that's the concept. Endotype B, predictive
enrichment, right. We're selecting the kids that
have increased expression, Actually are
expressing these genes, And we're also
selecting the kids that

Have a higher likelihood of
mortality increasing the event Rate that's
prognostic enrichment. My caution now, this
is all post hoc. All post hoc analyses have
lots and lots of problems, But it's the best that
we can do right now. But the size effect is hard
to argue with, tenfold. So we're trying to look
at this now prospectively, And hopefully, I can have
some more data for you At some other point. So one of the next
steps moving forward Is developing assays now that
can generate data quickly. We often forget about that
when we talk about biomarkers, We need to be able
to generate data. And I would argue that the
technology to measure genes And to measure proteins
quickly is here, It's just a matter of
investing the resources, And funds, and the time,
and energy to do it. And so for us,
that's the next step, Is generating rapid
assay platforms That can generate these data
within one or two hours, And then be able to test
some of these concepts Prospectively through
clinical trials. Thank you for listening. These are the
contributing centers That are currently enrolling
patients in our database. Some of you are in
the audience here, And we couldn't thank you
enough for that collaboration. Otherwise, this
wouldn't be possible. Thank you very
much for listening. Please help us improve the
content by providing us With some feedback.