The Legend of I-SPY 2 - Part A

Judith: Welcome to Berry's In the
Interim podcast, where we explore the

cutting edge of innovative clinical
trial design for the pharmaceutical and

medical industries, and so much more.

Let's dive in.

Scott Berry: All right.

Welcome everybody to, in the interim,
and Don Berry has joined me again and

we, we have a, a really cool topic
today to talk about in the interim.

And that is the I SPY two trial.

Many people may have heard of
the I Spy two trial, so we're

gonna go back and remember a
little bit of, uh, of this trial.

We'll, we'll walk through how it became,
what it was, what happened in the trial.

So Don, you ready to tell
us the story of I Spy two?

Don Berry: I sure am.

Uh, thanks Scott.

So I, SPY two has a, has a history, um,
uh, and it revolves around Laura Essman.

Of course.

Laura Esserman was the Pi I.

I was the co-PI

of the trial.

And this is back, we're talking,
uh, like, uh, 2007 and eight.

And she and I were part of a
cooperative group called the C-A-L-G-B,

the Cancer and Leukemia Group B,
which itself has a long history.

And we were in the breast committee,
uh, of that, uh, cooperative group.

And we had designed trials and
commiserated with each other,

uh, about the lack of innovation
and about what we wanted to do.

And, um, we designed a trial.

I Spy one.

You may wonder how the two
came about, but I spy one.

Um.

Where we were looking
at neoadjuvant disease.

Now this is part of the story.

Um, in breast cancer, there were two kinds
of categories of disease at the time.

One was adjuvant, which was, uh, patients
who, uh, had typically had, uh, uh,

uh, high risk disease, but were early
patients, were newly diagnosed patients,

uh, and they had, uh, bad tumors,
you know, large tumors, um, uh, uh,

tumors with, uh, positive lymph nodes
and, uh, were receiving chemotherapy.

Those with, um, estrogen receptor
status positive were, we're receiving.

Endocrine therapy, uh,
Tamoxifen at that time.

Um, and that, and, and what you
would do is you would treat the

patients, uh, after they had surgery
and after they had surgery, means

that for almost all of them, the
tumor's gone because we took it out.

Uh, and so what was the endpoint?

The endpoint was when it comes back.

Now, you know, if you're into
clinical trials, that when it

comes back is a dicey thing.

Uh, it sometimes takes a long time and
moreover, uh, which is of course good

for the patient, uh, and moreover,
it got better for the patient over

time because these therapies worked.

Scott Berry: and that

Don Berry: Um, and that was a problem.

The problem was that pharmaceutical
companies and cooperative groups

and the like, um, uh, sponsors
of, uh, these things and patients,

patient advocacy groups, it was
difficult to run these trials.

The other, just to touch the base, the
other part, uh, is metastatic disease.

Once it recurs, it becomes, um,
uh, let's say it recurs distantly

and, and, uh, uh, the lungs or the
brain or the bone or something.

Uh, then it becomes metastatic disease.

So that's a different category.

We're talking about the early,
so-called early, uh, breast cancer.

Scott Berry: So, so let me see if I,
so the, the setting is typically when

the diagnosis is made, the tumor's
removed, and that's what adjuvant means.

You go in, remove the tumor, and then
you treat them with different therapies.

Don Berry: the training
is called the Avant.

Scott Berry: adjuvant.

Is

called

Don Berry: a adjuvant to

Scott Berry: To

surgery.

To the surgery, and that trials ended
up being extremely large because

the endpoints took a long time.

The treatments were reasonably
effective, so clinical trials became

incredibly uh, onerous to run in this
patient population in breast cancer.

Don Berry: Yeah.

E expensive, uh, onerous.

Nobody wanted to do it.

Uh, what could we do
and what the CLGB said?

Uh, and I was the faculty statistician
and working with the committee and.

Uh, par parcel of these decisions.

Um, we should try the
neoadjuvant approach.

Uh, now this was novel.

It was led by, uh, other ki uh, something
called the, uh, N-S-A-B-P, another one of

the cooperative groups that had led this.

Because the neoadjuvant approach is all
you do is you exchange the surgery and the

treatment, but that's a big deal because
you're leaving the tumor in the body for

like six months, but you're pummeling
it with, um, uh, you know, lots of,

uh, toxic, uh, therapies that kill lots of
cells, including, of course, cancer cells.

Um, so it was a big deal, but.

Scott Berry: But

Don Berry: It had a benefit.

The benefit was that you got to see
whether the tumor responded to the therapy

that you gave them before you took it out.

So you left it in, you watched
it carefully, um, that you saw

whether the treatment that you
gave, the treatment you gave could

be different in a clinical trial.

Of course it should be
because you're learning.

And, uh, when you go in and do surgery
after, uh, six months of therapy, you see

whether or not the tumor is still there.

And that in the neoadjuvant approach
became the endpoint, uh, called

pathologic complete response.

You send the tissue to the pathologist and
pathologists can't find any, uh, tumor.

Okay?

So that was sort of a risk.

I described it to.

Uh, the, my colleagues at the CLGB
as betting the farm, uh, this had

this, this was novel, um, and we
did studies in different, um, uh,

categories of the disease, different
biomarker categories of the disease.

Uh, so I designed, uh, some of these
trials and meanwhile Laura and I keep

talking about what we're going to do, uh,
and, um, we both wanted to do, uh, what I

called the bandit approach, uh, where you
have multi-arm bandit, where you do lots

of therapies and you try 'em on different
patients and you see which patients work.

And this, uh, you know, I, I
had tried to do some of this

back at, uh, MD Anderson, where.

Uh, I was in am, uh, in the biostatistics
department and we designed trials, uh,

in various diseases and we, uh, Laos Push
Eye, a faculty member there now at Yale.

Uh, and I, uh, tried to do the, uh, this
kind of thing where we would look at

various therapies and go to pharmaceutical
companies and try to sell it.

And it was a hard sell, uh, to have, uh,
therapies from Eli Lilly, from Pfizer,

from Merck that you're comparing to a
control, but in the same trial as you're

comparing to each other, you know, like.

I, because I had the data, I
could look to see how the Eli

Lilly was doing versus the Merck.

Uh, and that's, uh, a, a dicey thing.

So we failed.

Lache and I failed.

Um, uh, but we, in the context of
the neoadjuvant approach, um, Laura

and I mused about maybe we can do
this in the neoadjuvant approach.

And, uh, she is an amazing person.

She, first of all, she never
takes no for an answer.

Uh, you say you can't do that, Laura.

Oh, yes, we can.

Uh, and so who's gonna fund this?

Uh, we were told by various people, uh,
Anna Barker is a good friend of ours,

and she was once the deputy director
of the NCI and she was the head of,

um, uh, committee in something called
the FNIH, the foundation for the,

uh, national Institutes of Health.

Uh, that's a, uh, that it
works with, with pharmaceutical

companies, including the government.

So it works across this.

And, uh, she said, you'll never
get this approved by the NCI.

The CLGB studies were.

CLGB was funded by the NCI
National Cancer Institute.

Um, and we would have to get the,
the, the funding from the NCI

if we were to do this at CLGB.

Um, and she told us, and others
told us, and we knew actually

this would never happen.

Uh, I had personally spent years trying
to get innovations into the NCI with

a little bit of success, uh, but this
would've been well beyond their, their,

uh, ability to imagine this could happen.

Uh, so we, we worked with the FDA, I
remember the FDA, I remember, uh, Rick

Paster, who's the, was still is the head
of, uh, the oncology, uh, telling me.

But these are

early patients.

These are.

Curable and you're experimenting
with them, how can you do that?

Scott Berry: that?

Don Berry: And so I said, uh, well,

we're

going to, uh, have a, a data monitoring
committee made up of, uh, a couple of

great statisticians and a couple of great
clinicians, and they're gonna be meeting

monthly and looking at the data, seeing
how things are going, and making sure that

nothing, and he was satisfied with that.

So they allowed us to do this study,
but then how to get the funding.

And the funding was through the foundation
for the NIH, uh, the initial funder.

Um, but, uh, actually, Laura.

Was responsible for getting
most of the funding.

You know, uh, she, she was a
surgeon who still is, um, and would

treat patients and they would,

uh, tell her that, uh, uh, she made them,
she, that they were alive because of her.

And therefore, since I'm the CEO of
this big corporation will fund you.

Uh, she got lots of funding that way.

It was originally funded by, um,
uh, the donations and philanthropy.

Uh, as time went along and as we
got more and more, uh, companies

involved with their drugs, uh, we
passed the funding off onto them.

So they had to, uh, pay to play.

Scott Berry: So, so let's back
up a little bit, but Sure.

So you, I is, is I spy
one funded or I Spy one?

You, you described this as a bit of a
pilot and from there you went out and

were able to get this additional funding
to bring in the first investigation arm.

Don Berry: Uh,

yes.

Um, uh, exactly.

I Spy one was funded by the government.

It was funded by the NCI.

Uh, and it, it was not randomized

or adaptively randomized.

It was looking at patients, uh, in
the neoadjuvant setting because, you

know, the surgeons had to learn how
to do, uh, the, this, how this thing,

you know, not do surgery right away,
but then come in and do surgery later.

It was a new thing to surgeons.

I.

And, um, so it, I spy
one had two endpoints.

Uh, I mean we, we were still looking
at path pathologic, complete response,

PCR, uh, but we were interested
in could we predict PCR from an

MRI that we give, uh, intermediate

in time between the initial
presentation and the initial therapies.

Uh, you know, after, uh,
three weeks, for example, on a

therapy, maybe there's an effect.

And maybe that effect could predict not
only path cr, but also, uh, survival.

So that's what I SPY one was about.

It was, uh, uh, uh, kind of a registry
if you like, or, uh, it, it, it

wasn't, uh, a randomized, we weren't
trying to learn about therapies.

There was a specific therapy that patients
got who, who met the eligibility criteria.

Scott Berry: So, so interesting
people listening to this.

You've described the neoadjuvant as a
huge advantage to clinical trials and

the ability to learn about the treatment.

What do we know now about it?

For the patient benefit?

Um, there's no sense that this is
better or worse for the patient,

that that goes through neoadjuvant
as opposed to adjuvant care.

Don Berry: Uh, boy, I'm, I'm gonna answer
the question, but the only way I can

answer the question is, uh, with, um,

with, with hindsight.

Uh, so at the time we thought that the
benefit for the patient was this business

about learning what benefits the tumor.

When we would give therapies different
experimental therapies, and we do

the MRI, sometimes we looked after
three weeks and the tumor was gone.

Scott Berry: Yeah.

Don Berry: Um, and that, I mean, is
an obvious advantage to the patient

that, uh, you can do something.

Uh.

Uh, else if it, if it hasn't
gone, I mean, you could stop.

If it's not defected at all, you
might want to do something else.

So it was getting the information that
would help the individual patient, but

it was also getting the information
that would help in understanding which

therapies are benning, benefiting which
patients, and then we could, you know,

emphasize those, those pairings, the
right patient for the, or the right

therapy for, for the individual patient.

Um, and so it was this, uh, PCR that was
the attraction for the clinical trialists

that perhaps you could get approval.

Um, and in fact, this has
happened since perhaps you could

get approval for a therapy.

Uh, that had a, uh, a, a
great benefit on the path.

CR rate, you know, improves a 30% rate
to 50%, uh, because patients who got

complete responses did extremely well.

And it didn't matter what the
therapy was, it didn't matter what

the biomarker characteristics were.

If you got a, if you were disease free
as far as they could tell at the time

of surgery, uh, and then follow them,
they did extremely well for overall

survival and event-free survival.

Now, I mentioned fast forward,
uh, we're gonna have to go

back to tell you about ipy two.

You know, what it did.

But, but, but fast forward
with the neo adjunct approach.

What happened?

Was there are patients who come out
of the surgery, and as I said, when

they're Pat cr they do extremely well.

But if they don't have a pat
cr, if they still have residual

disease, they did poorly.

The pharmaceutical companies latched
onto this possibility and said,

maybe we should, uh, treat those
patients who have residual disease.

And to your point earlier, Scott, about
the size of the trials, when we have, uh,

residual disease, um, the, uh,
event rate is a lot greater

because they, you know, they recur.

Um.

Uh, much sooner than, and more
likely than patients who get, uh,

PCR patients, who gets get a PCR
don't need additional therapy.

Uh, and moreover, in a clinical trial,
it would be difficult to see that

there's a benefit for the therapy.

'cause everybody lives,

Scott Berry: Hmm.

Don Berry: uh, not everybody,
but, uh, most, most women.

And, um, so they did this and they
found, uh, depending on the, uh,

biomarker characteristics that
some therapies, immunotherapy for

example, worked extremely well for

patients who were
so-called her two negative.

Um, and the, um, the, the company
would then go to the FDA and say,

we've got this great therapy.

Moreover, that's changing the course.

Scott Berry: yeah.

Don Berry: When, and, and, and, and
it means that patient, you could, if

you've got a pharmaceutical, if you're a
pharmaceutical company and you take using

the neoadjuvant approach and you randomize
patients to get your therapy versus the

standard therapy, and you saw that there's
a PCR benefit, you're still not going

to get approved by the FDA because what
happens is the patients who don't have

a benefit, who have residual disease are
getting other therapy that's effective.

Scott Berry: Yeah.

Don Berry: So it's confounding.

And so the neoadjuvant

approach has

completely changed, um, uh, breast
cancer treatment and, uh, and approach.

And it, it means that the original
hope, it's sort of ironic, the

original hope was to get a PCR.

And now the, um, the, the benefit
for the neoadjuvant approach

is it really doesn't matter.

You're just getting information
about what therapy works and then

you give some other therapy and they
call it adjuvant after the surgery.

Scott Berry: Mm.

Don Berry: So that's a whole different
story that's not really part of

IFI two, but it's part of, uh, the
breast cancer, uh, uh, story RA.

Scott Berry: Okay.

So, so I, in the timeline of this,
I Spy one is proof of concept that

you can do neoadjuvant, there's
no investigational therapies.

You can get MRIs, you're
getting data on PCR.

So meanwhile, if, if a company wants
to do a phase two trial before I

SPY two in the adjuvant setting.

It would be incredibly hard
because disease-free survival

after surgery is so very long.

These trials would be huge and long.

It's hard to do drug development,
so you turn it upside down

and you're doing neoadjuvant.

So you're ready for I SPY two,
which is intervention in the

neoadjuvant space and the funding.

Are we ready to describe
the I SPY two trial?

Yes.

Don Berry: Yes.

Scott Berry: Okay.

Okay.

Okay.

So, um, do you want to describe
what this trial looks like?

You, you want to do, you described
you want to do the bandit approach,

you want multiple therapies
simultaneously, but this is also

about precision medicine in a sense.

Don Berry: Exactly.

So the precision medicine aspect is,
uh, breast cancer is a, probably the

poster child of precision medicine.

Um, it has biomarkers that
are extremely important.

Estrogen receptor status,
progesterone receptor status, her two.

Uh, status, um, of, uh, something called
MammaPrint, or, which is very similar

to Oncotype dx, which is a 21 gene.

MammaPrint is a 70 gene, uh, uh, uh,
biomarker multi, you know, poly marker.

Um, uh,

Scott Berry: So, so I-SPY 2 classifies.

You described these, there's
HER2 status, hormone receptor

status and MammaPrint status.

Each one is dichotomous, uh, for this.

So women who come into I-SPY 2 fit into
a subgroup, and they're, or a subtype,

sorry, I SPY two created all kinds of
new terminology, and there are eight

subtypes by the classification of
that when a woman comes in that their,

their, their, uh, tumors classified
in eight, eight different subtypes.

Don Berry: Uh, right.

And then we want to see which
subtypes benefit from which therapies.

But now if you, uh, we will, what we
want to do is, is, uh, tell the company.

Your therapy is beneficial in this,
the subtype, but not in that subtype.

Now, how can you do that?

There are, if there are eight subtypes,
there are 255 different combinations

of those subtypes and we can't do 255
different partitions of of cancer.

So we looked at what we call
signatures, which are subsets

of subtypes, and we looked at 10
that were the primary analysis.

So every month we analyzed how well
the various therapies in the trial were

doing in these 10 different signatures.

So one signature, uh, would be HER
two positive, that as you described

it, Scott has four subtypes.

Uh, but we're looking not at those
individual subtypes, but at the,

at the, uh, all of them together.

Uh, we also look at the HER two negative.

We look at the estrogen
receptor positive, the negative.

We look at the, uh, estrogen receptor
negative, HER two negative, which

is called triple negative disease.

And that has two subtypes, namely
MammaPrint positive, MammaPrint

negative, um, that we don't,

we're not advertising those subtypes
as being an indication for your drug.

It's the signatures which are
indications for your drug.

So that was, uh, uh, a, a big deal to
go to the FDA and say, we have these,

uh, 10 different possible indications.

Uh, and that in itself is a huge
innovation, which by the way, uh, I SPY

two is a phase two trial, and we're,
we're trying to ready it for phase three.

Uh, we've since then designed similar
trials in, uh, GBM glioblastoma, uh,

pancreatic cancer, uh, that have a
smaller number of subtypes because they

have, you know, different biomarker,
uh, uh, science associated with them.

Um, and, but that the FDA has agreed
with looking at these multiple

signatures in a registration trial.

Uh, so it's, it's, it's pie in
the sky, but we're eating the pie.

Scott Berry: mm.

So, uh, a woman comes in and she belongs
to one of these eight subtypes and is

randomized among control and, uh, call
it 20% chance she goes to control.

And the other 80% she could go
to the various investigational

therapies that are there at a time.

So this is your randomized bandits.

In a way that if there are three
therapies in the trial right now, she

gets randomized among the three therapies.

And so, uh, you probably need to
describe the response, adaptive

randomization aspect and which
is so critical to icey two.

Don Berry: Yeah.

So,

um, the bandit problem is that you have
these, uh, multi arms and, and, uh, it's

usually posed in the context of a single.

Uh, subtype, um, where you want to
treat patients effectively, uh, in the

trial, which is, uh, you know, trying
to blow, I've been trying to blow up

the, uh, notion that you can't learn.

You, you, you, you're not treating
patients in a clinical trial.

You're learning about patients.

And I say, why can't we do both?

Um, and we had done that.

Um, uh, I've been writing
about it for many years.

Uh, and, and when I went to MD
Anderson, we actually started to do it.

So we had trials where we,
uh, adaptively, randomized.

Uh, you get a, a randomization, but if
a, if that therapy is doing well for your

subtype, if you're the patient, you get
that therapy with a higher probability.

And what it means is that,
uh, patients in the trial,

Scott Berry: the trial,

Don Berry: um,

Scott Berry: um,

Don Berry: on average have a higher,

uh, PA PCR rate

than

uh, a, a standard trial

because they're more likely to
get a therapy that's doing well.

And moreover, you're not exposing
those patients to a therapy that's

not doing well for your subtype.

Um, and it it means that patients
in, in the trial do better.

Um, it also, uh, means that
the, uh, overall patient.

Uh, population does better.

Um,

and you get more information about
the therapies that are doing well

in the subtype signature because
they're getting it more often.

So you get a bigger sample size.

Um, because in, in and faster you
learn faster, you learn better about

where you really wanted to learn.

You don't care how well it does in
a therapy, in, in a subtype that

doesn't do very well with it.

Uh, you don't use that
for, for those patients.

You use it only for those
patients who it is doing well.

Scott Berry: okay.

So, uh, when women come in, they're
classified in their eight groups and

the randomization, uh, probabilities
for one subtype are different

than another because the drugs
are modeled as potentially having

differential effect for those women.

Don Berry: Exactly.

Scott Berry: And it might

even be that they have no chance
to get a particular therapy

'cause it's not doing well.

And a very high chance to get another
one that's doing well for women like them

from, from, the drug side.

The drug can focus on the women
where it's having an effect and not

randomize to those, that they're not.

What's amazing about this
is you can't really do this.

You can't do it very well if it's a single
drug trial, because now you can't change

the prevalence of various subsets without
stopping and rolling them all together.

It's almost undoable to do this precision
unless you've got multiple things to give.

Now, now you're doing these, I
I, I know you wanna jump at this,

but you're doing these adaptations
and you're analyzing the data.

Uh, the algorithms are
being run and resetting.

The randomization weekly
during the, i I think it varied

during the time of Ipy two.

I think initially it was
actually daily, then weekly.

Now your endpoint is six months.

And you're trying to learn about what
works and you don't wanna wait six months.

And so you're trying to accelerate
the learning to allow all of these

things you just described to do better.

So how do you do that?

Don Berry: Uh, thanks for the setup.

Uh, uh, I mentioned earlier the MRI,
we did MRIs in I spy one to learn

about that and we, we learned that it
is predictive what we di like to do.

Is to use all of the patients.

Uh, we don't wanna wait six months.

Um, in the context of waiting for
disease-free survival or overall

survival, six months is a short time.

But in the context of, uh, uh, I
spy two, six months is a long time.

Um, we get MRIs, we got MRIs, uh, MRIs
at three weeks and at, uh, 12 weeks.

Uh, and we use that information
to predict is it gonna be a PCR.

Now here is where, uh, I reported to
the, uh, data Safety Monitoring Board.

Uh, every month.

Um, and, uh, uh, I explained
to them the trial and they

kept learning about the trial.

And after some period of time,
uh, they said to me, said, Don,

you told us about MRI and how you
use MRI, uh, at least eight times.

Could you tell us again?

So it's not an easy concept for
those people that are used to,

you know, looking at an endpoint
and focusing on the endpoint.

If you say

we're gonna, we focus on the
endpoint, being MRI, uh, first of

all, nobody would, uh, accept that.

It's not a surrogate for, uh, anything.

But we looked at MRI,

not with the notion that
what we're gonna do is.

Uh, uh, focus on MRI as the endpoint, but
as an auxiliary endpoint, as a marker of

how likely is it that it's gonna be a PCR?

So it,

if

if, there's no tumor, uh, that you can
see on the MRI, that means there's a

high probability is gonna be a PCR.

So what we do is we do
multiple imputation.

We have this probability, and when we are
doing the multiple imputation for all of

the women in the trial who don't have.

Six months, you know, who don't
have the result of surgery.

We're predicting the result of
surgery, but it's probabilistic.

So, um, uh, there was a drug
pembrolizumab, which you've seen

advertisements for, for Keytruda, uh,
from Merck, uh, which, uh, graduated.

We call it graduate when it, we've
learned what it's, what it's,

uh, uh, signature is, uh, when no
patient had results at six months.

Um, and the, the way we could do that,

we will talk about the time
machine, uh, in a minute, I'm sure.

Uh, the, the, the way we could do that is.

Uh, we had the algorithm that was
making the patient assignment and

making decisions about graduation.

The drug would graduate, uh, no longer
get, uh, patients, uh, but we would

continue follow up, uh, when only
one patient had the result of surgery

in triple negative breast cancer.

And, uh, how could we do that?

It was because the algorithm
was seeing that the therapy

was melting the tumor away.

And I went in and I

looked at the data, I
said, how can it do this?

And the answer was, you know, in the 12
patients who had, uh, results at 12 weeks.

Um, 11 of them eventually had
a PCR, uh, and 11 of them, uh,

were, uh, and it was clear that
they were gonna do extremely well.

Um, so

Scott Berry: so, so, and at the time
I think one patient had been through

six months and the model predicted
something like a 63% PCR rate for

the arm and, uh, with one patient.

And by the time all the patients
got through, it stopped for

graduation, uh, at some point.

It almost nailed exactly the
PCR rate based on the MRIs.

Don Berry: it was exactly
the sa the, the same as the,

as the MRI, but now.

Um, there's another aspect to this.

How could it do this 60% and get
it right with only one patient?

Uh, the answer is the time
machine and the controls.

Now, you,

Scott Berry: so hang on, hang on.

Let's,

let's, we're gonna be adaptive here, so
for the first time, we're gonna call this

part A of the podcast and we're gonna make
people have to tune in for part B of the

podcast and learn about the time machine.

And thank you all for joining
and join our next episode and

learn about the time machine

Don Berry: And, and other things
as well, predictive probabilities,

for example, which are usually
important in, uh, uh, being

adaptive.

Scott Berry: Alright, so thank you.

Don Berry: Thank you.