MACROSCOPE
An Engineering Approach to Translational Medicine
Physician-scientists may benefit from an approach that emphasizes solving problems over generating hypotheses
Michael Liebman
In the years since the completion of the Human Genome Project,
physician-scientists have applied new energy to translating findings
from the laboratory into better treatments for patients. Yet this
accelerated, unidirectional transfer of knowledge from the bench to
the bedside, a practice that goes by the name of translational
medicine, is hitting an obstacle: The generation of data is far
outstripping scientists' ability to convert it into usable
knowledge. I believe that, paradoxically, this problem stems from
the tightly focused approach that gives science much of its power.
Genomics, proteomics and other high-throughput technologies are
seductively powerful, but that seduction may limit our view of the
complex problems of physiology and disease.
For example, scientists can now correlate a disease with a specific
pattern of gene expression. Such experiments are straightforward and
fairly quick when the tools are available, and they provide a
massive quantity of data. However, by diverting limited resources of
time, money and personnel, mining this wealth of data may actually
lead investigators away from grasping the governing laws from which
they could build predictive models of the disease.
I am not suggesting that investigators should give up
high-throughput, brute-force methods. Today's technology is a boon
to science and a powerful component of my own research. However, as
clinical investigators, we stand to reap significant benefits on
behalf of society by expanding our focus and viewing translational
medicine not through the eyes of a scientist, but as an engineer might.
Why an engineer? Because an engineer uses the fruits of science to
feed the appetite of technology. Unlike scientists, who tend to
approach problems from a "bottom-up" perspective by
collecting data and seeking patterns, engineers take a
"top-down" approach, probing a specific system for clues,
taking it apart and considering how each component can be handled in
a tailored solution. An engineer is a problem solver rather than a
hypothesis generator.
The two perspectives are neatly symbiotic in physics and chemistry,
for which fundamental laws yield predictive models. But in the life
sciences, biologists, including physicians, must be more aware of
the gap between science and technology—we still know too
little about the complexity of living systems to make many
generalizations from first principles.
I propose that an engineering approach, what might be called
"real systems analysis," may be a better way for
scientists to identify and develop solutions for biomedical
problems. This kind of problem solving requires that
translational-medicine research place more emphasis on going from
the bedside to the bench, rather than the other way around. The
Clinical Breast Care Program (CBCP) is a collaboration between
Windber Research Institute and Walter Reed Army Medical Center, and
it is the prototype for an integrated approach to the study of
breast cancer. Here, I present some examples of how top-down problem
solving in the CBCP has provided unique insights.
Disease Is a Process, Not a State
For the purposes of diagnosis, analysis and experimentation,
academic physicians tend to focus on disease at a single point in
time. But disease needs to be treated as a process that evolves over
time through the interaction of genetic, environmental and lifestyle
factors. This view puts a premium on understanding the complex
history of a patient, and it acknowledges that most disease cannot
be tied to a single cause.
When physicians make a diagnosis, it's natural to focus on the
patient and symptoms at the time of presentation. The doctor's
knowledge of a patient's past is typically limited to major
illnesses, allergies and family history. Yet clinical assessments
could be much more meaningful if we understood the way that genes
and environment interact to produce disease. For example, we know
that certain biomarkers, such as mutations in the genes BRCA1 or
BRCA2, indicate higher risks of breast cancer. But the fact that a
woman has a mutation in BRCA1 doesn't mean that she will develop
breast cancer—it only indicates that she needs to be monitored
more closely.
Likewise, smoking, high alcohol consumption and obesity are
correlated with an increased risk of breast cancer, but we know
little about how each factor raises the risk—much less about
how two or more might work in concert to increase risk. This
situation leaves us with a circular argument: To justify the cost of
collecting a comprehensive patient history, we need proof that such
data are relevant, but we can't evaluate which data are relevant
because we don't have a database of comprehensive patient histories.
We in the CBCP think that detailed information will prove useful,
although we don't know exactly what connections will emerge from the
mass of variables. We are collecting from each patient a lengthy
history that includes her exposures to tobacco and alcohol, details
about pregnancy, childbirth and breastfeeding, and a record of
changes in her body mass. We also try to include information about
the timing of these events in a person's life. The chronology is
particularly relevant for breast cancer because the breast develops
continuously from birth through old age. This lifetime of changes
also presents an additional challenge: Not only do these factors
influence risk differently over time, but their interactions with
one another also vary with age.
An engineering perspective treats the patient as a system, or a set
of subsystems, that has been acted on, differentially, by many
elements that influence its state at critical points over time. Our
job is to identify these critical points so that they might be
controlled. Whereas many current studies identify correlations
between isolated variables, we hope that the wider scope of the
CBCP's systems-based approach will help us determine
causality, thereby improving diagnosis and treatment.
Aging as a Background to Disease
The breast changes between a woman's time in utero and her
post-menopausal years. This maturation process is different for
women who have had children than for those who have not, and it also
varies under the influence of several variables: age of menarche,
use of hormonal birth control, number and timing of children, the
practice of breastfeeding, age of menopause and use of
hormone-replacement therapy. Thus, our definition of
"normal" varies with age and experience, and an optimal
diagnosis must use a systems-based approach to compare an individual
cancer patient's baseline (which we must guess at) to her disease
state (which we can measure during diagnosis and treatment). The
immediate aim of our project is to determine background levels of
gene and protein expression in breast tissue and to find out how
these numbers vary in a healthy population. This information will be
a significant step in the development of molecular diagnostics.
Note that a woman's life stages are not separated by fixed
boundaries. Rather, each represents a unique intersection of a
woman's age and an event. Given this complexity, it was crucial to
sieve the scientific literature for data that we could integrate
into a systems approach. This was more difficult than one might
think. Scientists have studied these stages for decades, producing a
tremendous body of work in physiology and pathology—more than
any one scientist can master. Furthermore, we recognized that even
the most encyclopedic and fair-minded review article cannot escape
the inherent bias of its author. Thus, we have harnessed some
computing power, employing text data-mining to cull the literature.
This effort has two aims: to refine the definitions of these stages
and to extract information about the underlying physiological and
developmental changes. This information becomes the foundation for
our molecular analyses and helps integrate clinical and molecular
data. We plan to augment this computational approach with a
community-based longitudinal study that includes molecular and
behavioral components.
Tumor Classification and Staging
Tumor classification is critical to the assessment and treatment of
cancer. To optimize this process of classification, the physician
must determine both the present disease state and its potential for
progression. This is a difficult task, and it will become more
difficult as more relations are established between genes,
environment and disease; an ideal representation of cancer would
reflect all of these variables. With this idealized tool, a person's
disease would become a vector in multi-dimensional space, with each
of tens or hundreds of axes representing a clinical or molecular
parameter. Perhaps we will realize this vision.
In the meantime, oncologists use three concrete variables to define
the stage of a tumor—tumor size (T), metastasis (M) and nodal
involvement (N), the finding of cancer in nearby lymph nodes. One
problem with this system is that the mapping of some TMN triples to
fixed stages is ambiguous, perhaps because the terms are imprecise
or insufficient to describe the disease. Another flaw is that these
numbers do not reflect the history of a patient's disease and
treatment. Yet the TMN system could be made into a better assessment
tool simply by setting each variable on its own axis to create a
three-dimensional TMN space. Each person's clinical trajectory can
be viewed as a unique vector in TMN space. This method highlights
the fact that although the stages of tumor progression are linear,
there are different "paths" through the disease; not all
stages may be encountered on each patient's path. Furthermore, as we
see how different vectors turn toward the origin (cancer-free) vs.
the extremity of poor outcome or reoccurrence (10,10,10 in a TMN
space where the axes run from zero to 10), we can identify paths
through TMN space that represent different responses to a given
treatment. The result is better information for clinicians to make
the best decisions for each patient.
Heterogeneity of Breast Disease
Breast tumors are usually composed of more than one type of cancer.
This is a problem when the cancers do not all respond to the same
treatment. Although scientists know about this phenomenon, it has
been difficult to quantify because pathologists use differing
diagnostic criteria. In the CBCP, we have the advantage of having a
single pathologist review all patient samples. We think it likely
that when a tumor biopsy has a specific combination of subdiagnoses,
it is more accurate to describe the tumor in terms of its
heterogeneity rather than noting only the severest cancer (the
current convention). The CBCP categorization scheme contains 135
potential subdiagnoses for tissue sections. Among 891 patient
samples, we have observed 75 of these. Although most combinations
are rare or nonexistent, others are extremely common: We found two
cancers that had a 92 percent likelihood of showing up paired rather
than alone. This finding suggests that we may need to review the
tumor-classification system to reflect this heterogeneity, thereby
refining our evaluations of tumor stage and grade and improving
treatments for patients.
An engineering perspective analyzes breast cancer by viewing the
whole patient and applying customized treatments that reflect each
person's unique confluence of biology and experience. We hope that
this practice reinvigorates the study of breast cancer and other
diseases to enhance patient care—the ultimate goal of
translational medicine. To my basic-science colleagues, I say that
our engineering counterparts have been looking at the world through
somewhat different glasses, and perhaps it is time to share the view.