Reported by NCI Press Office Staff
February 7, 2002
A fast-tracked paper on a possible new ovarian cancer screening
tool by Lance Liotta , M.D., Ph.D., of the National Cancer Institute
(NCI), and Emanuel Petricoin, M.D., of the Food and Drug Administration
(FDA), appeared electronically in The Lancet on February
7, 2002. BenchMarks interviewed the scientists about this new tool,
which employs a technique called proteomics. The researchers talk
about the tool, how it was developed, and its potential applicability
to screening and treatment for ovarian and other cancers.
Dr. Liotta and Dr. Petricoin run the FDA/NCI Clinical Proteomics
Program.
Are there cancers other than ovarian cancer that would be
obvious targets for this new technology? If so, why would those
cancers be better to study than others?
Dr. Liotta: As we speak we are conducting additional research
on other cancers, and we're planning to do more in the future. An
example would be prostate cancer. We are looking for cancers that
have an indolent phase -- an asymptomatic, early silent stage of
progression. If we can pick up the cancer at this early stage we
may be able to treat the tumor before it has metastasized. We chose
ovarian cancer and prostate cancer because of the urgent clinical
need. We were very fortunate to be able to collaborate with David
Fishman at Northwestern University, who has a high-risk clinic and
a very good population of samples with five-year follow-up. This
is a valuable resource.
Are there other cancers you will eventually focus on?
Dr. Petricoin: Breast, pancreatic, colon and lung -- some
of the most common solid human tumors. We're now focusing our attention
to validating ovarian cancer in larger study sets. However, we're
not just focusing on early cancer detection. We have a clinical
trial that Dr. Elise Kohn of NCI is running (see accompanying story)
where we're going to assess the ability of this new method to detect
relapse in women with ovarian cancer.
We are also using this technique to look at early drug toxicity.
In the future, if we can predict by protein patterns which patients
are likely to have an early toxic response, we could treat them
at lower doses or switch to a new therapy choice.
What are the advantages of looking at protein patterns rather
than looking at a single biomarker to detect disease?
Dr. Petricoin: In the past, scientists have taken a reductionist
view of biomarker and drug target discovery and the subsequent use
of these biomarkers, one at a time, to detect disease. The new concept
we are proposing is that patterns of serum proteins can be used
as diagnostic endpoints -- the pattern itself is the diagnostic.
Instead of looking at one biomarker at a time, you can take an
information archive, such as the serum proteome, and look at tens
of thousands of proteins and peptides at once -- a proteomic portrait,
a barcode, if you will -- that we generate very rapidly from the
mass spectrometer.
By eye you would not see any underlying hidden patterns -- it would
look like a chaotic picture. The human brain doesn't do well with
that kind of pattern recognition. We've taken a machine-based approach
-- an artificial intelligence-based algorithm to look at hundreds
of millions of combinations of patterns of proteins very quickly
and find the most optimal pattern that segregates cancerous samples
from noncancerous samples in the training sets.
In this case, the sets are sera from normal healthy women and from
women with active ovarian cancer at the time the serum is taken.
The algorithm is trained on those samples first. The outcome is
already known in the samples in the training sets.
This algorithm was trained on this complex pattern of tens of thousands
of proteins and peptides in the serum and it found a combination
of five proteins that segregated those sets 100 percent of the time.
Then we challenged that algorithm with blinded, unknown sera specimens,
and it was able to take that combination of five proteins, and say,
'Does this pattern look like the pattern from normal subjects or
does the pattern look like patients with cancer?' In this blinded
set, we found these great results! And now, we're going to attempt
to validate these findings in much larger sets and find out how
well we really can do.
Dr. Liotta: Let's say you look at a population of patients
with cancer. In 20 percent, maybe, one marker works, but it's not
good in the other 80 percent. In another 20 percent a different
marker is best.
The best way to predict cancer in an overall population might be
a combination of markers that transcends that heterogeneity. The
marker is a pattern formed by a group of proteins, rather than one
protein. Perhaps one marker is an old idea. The new idea is to look
for patterns or groups of markers, rather than one marker at a time.
We have shown support for that hypothesis in ovarian cancer.
This paper focuses on protein patterns, but could this technique
help elucidate pathways or signals between proteins?
Dr. Liotta: Another application of this technology is to
look at patterns of protein changes, such as phosphorylation or
cleavage, or patterns of interactions of proteins. If we sort through
large number of interactions and then correlate the numbers with
the state of a signal or look at changes before and after drug treatments,
then we can use this method to build cell circuit diagrams, because
we can figure out who is talking to whom and who is connected to
whom.
Dr. Petricoin: We're developing platforms that can look
at early-stage serum-based portraits and we're inventing platforms
that can look at hundreds of different phosphorylation endpoints
simultaneously. And given that constellation of changes in patterns
of protein phosphorylation, we can begin visualizing the critical
'gates' in the cellular circuitry or in the wiring diagram of the
cell.
We can look at the changing patterns of phosphorylation, relate
that back to what we see in the phosphorylation patterns when a
patient is treated with a therapeutic, such as Herceptin. We can
look before and after the treatment and determine whether that patient
responded to treatment and correlate that back with the signaling
patterns.
Patient-tailored therapy can then be achieved as each patient is
profiled beforehand for the state of their cellular circuitry in
say, a biopsy, and the correct choice of therapy is matched to the
pattern seen.
The success of the test in correctly diagnosing cancer is
remarkable in this small sample. How important is the five percent
false positive rate?
Dr. Petricoin: I think it's especially important to be cautious
about the utility of this system for general population screening.
Our false positive rate is only five or six percent right now. But
when a large population of women is screened, this will produce
many women who test positive but are actually without ovarian cancer
who will be worked-up unnecessarily.
The first application would be for women in high-risk clinics.
It's very important in ovarian cancer, because it's a rare type
of cancer, to have almost 100 percent specificity (identifying every
woman in the group who has cancer), even if that means sacrificing
some sensitivity (not identifying correctly every woman who doesn't
have cancer) .
This is why you see this first application being applied to high-risk
clinics. Even with a false-positive rate of five or six percent,
you can have a positive impact among women who have, say, a mother
or sister with ovarian cancer or are known to have a BRCA1 mutation.
These women are already thinking about prophylactic oopharectomies.
They're enrolling or referred to these high-risk clinics because
their anxiety level is very high. If we were able to give the physician
something else to help him go in and find out for sure that patient
has something wrong or could assure the patient that they are okay,
that would be a dramatic help in and of itself. Also, in that setting
you could accept a higher false-positive rate because you would
be looking at an enriched population with the potential to harbor
the occult (asymptomatic) disease.
Dr. Liotta: A diagnostic test that is negative can provide
reassurance to a patient who is worried about cancer risk.
This technique is minimally invasive, requiring a finger
prick's worth of blood. Do you see screening for cancer evolving
to methodologies such as these and away from radiology and more
complex and costly techniques?
Dr. Liotta: We view this as supplementary and complementary
to other tests. The most accurate means of diagnosis will be a combination
of all of these methods. You can screen patients with this kind
of method and then more efficiently choose who should get a thorough
radiologic work-up or physical examination. This kind of methodology
combined with other marker tests or genetic screening might end
up being a good combination, much like a combination of different
protein testing methods might be better than one marker.
Dr. Petricoin: We don't see this as being a stand-alone
endpoint. Rather, we see it as something that hopefully will give
physicians something to refer to in addition to all the other tests
they have at their disposal. The physician is the ultimate intelligence
tool to decide what's best for the patient.
Is there evidence indicating a possible function of any of
the key proteins in the pattern used to diagnose ovarian cancer?
Dr. Liotta: We don't know what the proteins are. We're investigating
their identity. Nevertheless, even though we don't know what they
are, the pattern is still discriminatory. When we find out what
these proteins are, they may or may not provide a clear understanding
of why they're altered. They could be several orders removed from
the pathological source.
Dr. Petricoin: PSA is a very widely used biomarker and the
fact that it is a cystein protease does not add additional value
to its utility as a biomarker for prostate cancer or prostate volume.
So knowing what a biomarker is and how it relates to the biologic
underpinnings of the disease is a separate issue altogether from
how well it behaves as a specific and sensitive marker for the presence
of a disease.
What trials are currently under way to further evaluate this
technique for the detection of ovarian cancer? When are these trials
expected to be completed?
Dr. Liotta: We have an ongoing trial looking at patients
who are being treated for ovarian cancer to see if this technology
can pick up early recurrence. That trial is open to accrual and
the P.I. (principal investigator) is Dr. Elise Kohn, NCI. In addition,
we are developing a whole series of future confirmatory trials in
multiple institutions and with cooperative groups to test this as
an early diagnostic tool over a larger population. It should take
about six months to initiate all these trials.
Dr. Petricoin: For a researcher who has an interesting bioinformatics
platform but doesn't have a clinician and can't get access to patients
or doesn't have a mass spectrometer, they can download the raw spectra
from our study set and evaluate their own tool with their own program.
At the same time, we'll be expanding our mass spectra proteomic
profiles with our ever-expanding studies and, like the government
did with the human genome sequencing effort, making it all open
access as we generate patient spectra whose data are posted anonymously,
under IRB (institutional review board) rules, on the Web. Hopefully
we'll have thousands and thousands of these spectra on the Web that
people can easily access, and we think it will be a great resource
for the scientific community.
Of the thousands of proteins in the blood, what criteria
did the artificial intelligence program use to identify those that
would be most useful in discriminating between cancer and non-cancer?
How many proteins or peptides were needed to create a reliable pattern?
Dr. Liotta: There was no preconceived information. All it
did was look at the patterns and correlate with the training set
and correlate with disease or not disease. It was a survival of
the fittest pattern.
Dr. Petricoin: Here are 50 women who don't have cancer,
and here are 50 women who have cancer and here are the spectra of
each of these women. Each of the spectra is comprised of 15,200
data points -- one for every protein or peptide. At a precise molecular
weight, a defined intensity value reflective of the relative abundance
of that protein is given. We can look at all the intensities of
each oneof those precise molecular weight points as individual data
points and find the optimal pattern to segregate the two training
sets 100 percent of the time using the combinations of intensities
in N-dimensional space.
Once we found that pattern we then tested it on unknown samples.
We also made sure that we tested the pattern on women with common
benign gynecologic conditions such as ovarian cysts, fibroids, endometriosis,
and general inflammatory diseases. It is much more common to have
women with these benign conditions. Markers like CA-125 can be elevated
in women with these benign conditions.
Ovarian cancer is so rare, that for every one woman with ovarian
cancer, many, many more are going to have ovarian cysts, have uterine
fibroids. So biomarkers that detect these conditions indiscriminant
of ovarian cancer are not really useful. We trained the artificial
intelligence program to see sera patterns from women with ovarian
cysts by including these samples in with the healthy cohort, therefore
the patterns that discriminate ovarian cancer from healthy also
discriminate it from benign conditions.
The exciting thing about artificial intelligence programs is that
they can learn and get better over time as more data are brought
in. As we validate this in larger trials, we envision an ever-growing
and expanding training set of spectral data. We can then build models
from this for whatever we want -- we might later want to look at
different groups of women with ovarian cancer. The models get better
and better as they see more data.
Where does this advancement in the field of proteomics leave
the study of genomics? Will it no longer be as important to know
what genes are defective or unexpressed but rather know what proteins
are non-functional or which pathways have been interrupted?
Dr. Liotta: Information we get from proteomics we cannot
get from genomics and vice versa. Genetic mutations and alterations
in DNA and chromosomes may not be known at the protein level and
the same is true the other way. Whether a gene is transcribed to
a high or low level may have nothing to do with what happens to
the protein product - whether it is phosphorylated or whether the
protein product of the gene binds to another protein partner. Genomics
and proteomics are in some ways exclusive and also complementary.
Dr. Petricoin: Putting myself in the shoes of a patient,
I would want to have the full benefit of what science has to offer.
Dr. Liotta and I believe that genomic and proteomic approaches should
be, and will be, used as complementary technologies to benefit and
positively affect patient outcome -- that's really what our proteomics
program is all about.
### |