WhyIsMathematical
BiologySoHard?
Michael C. Reed
338 NOTICES OF THE AMS VOLUME 51, NUMBER 3
Although there is a long history of the applications
of mathematics to biology, only recently has
mathematical biology become an accepted branch
of applied mathematics. Undergraduates are doing
research projects and graduate students are writing
Ph.D. dissertations in mathematical biology,
and departments are trying to hire them. But what
should the Ph.D. training consist of? How should
departments judge work in mathematical biology?
Such policy questions are always important and
controversial, but they are particularly difficult
here because mathematical biology is very different
from the traditional applications of mathematics
in physics. I’ll begin by discussing the
nature of the field itself and then return to the
policy questions.
Where’s Newton’s Law? The phenomena that
mathematical biology seeks to understand and predict
are very rich and diverse and not derived from
a few simple principles. Consider, in comparison,
classical mechanics and continuum mechanics.
Newton’s Law of Motion is not just a central explanatory
principle; it also gives an immediate way
to write down equations governing the important
variables in a real or hypothetical physical situation.
Since the Navier-Stokes equations express
Newton’s Law for fluids, they are fundamental and
have embedded in them both the fundamental
principle and the complexity of the fluid phenomena
that we see. Thus a pure mathematician
who proves a theorem about the Navier-Stokes
equations and an applied mathematician who
develops new numerical tools knows that he or she
has really contributed something. Alas, there are
no such central fundamental principles in biology.
There are principles of course—some would say
dogmas—such as “evolution by natural selection”,
“no inheritance of acquired characteristics”, or
“DNA → RNA → proteins”. But these are not translatable
into mathematical equations or other structures
without hosts of additional facts and assumptions
that are context-dependent. This means
that mathematical biology is very unsatisfying for
pure mathematicians, who usually are interested
in discovering fundamental and universal structural
relationships. It also means that there is no “mathematics
of biology” in the same way that ordinary
differential equations is the mathematics of classical
mechanics and partial differential equations
is the mathematics of continuum mechanics.
Diverse, yet special. Because of evolution, biological
systems are exceptionally diverse, complex,
and special at the same time, and this presents several
difficulties to a mathematician. The first is
choosing what to work on. There’s too much biology!
How do changes in the physics or chemistry
of a particular environment affect the species that
live in the environment (ecology)? How do diseases
spread within a population (epidemiology)? How
do the organ systems of the human body work
(physiology)? How do the neurons in our brain
work together to allow us to think and feel and
calculate and read (neurobiology)? How does our
immune system protect us, and what are the
dynamical changes that occur when we are under
attack by pathogens (immunology)? How do cells
use physics and chemistry to accomplish fundamental
tasks (cell biology, biochemistry)? How does
the genetic code, inscribed in a cell’s DNA, give
rise to a cell’s biochemical functioning (molecular
biology and biochemistry)? How do DNA sequences
evolve due to environmental pressures and random
events (genomics and genetics)?
The second difficulty is that a priori reasoning
is frequently misleading. By “a priori reasoning” I
Michael C. Reed is professor of mathematics at Duke
University. His email address is reed@math.duke.edu.
MARCH 2004 NOTICES OF THE AMS 339
mean thinking how we would design a mechanism
to accomplish a particular task. As a simple
example, both birds and planes burn something
to create energy that can be turned into potential
energy, and both have to use the properties of
fluids that are implicit in the Navier-Stokes equations.
But that doesn’t mean that one understands
birds if one understands planes. To understand how
a bird flies, one has to study the bird. Modelers are
sometimes satisfied that they have created a
mathematical model that “captures” the biological
behavior. But that is not enough. Our purpose is
to understand how the biological mechanisms give
rise to the biological behavior. Since these biological
mechanisms have been “designed” by evolution,
they are often complicated, subtle, and very special
or unusual. To understand them, one must
immerse oneself in the messy, complex details of
the biology; that is, you must work directly with
biologists.
Thirdly, different species (or different tissues or
different cells) may accomplish the same task by
different mechanisms. An astounding array of special
mechanisms allows animals to exploit special
niches in their environments. For example, a diverse
set of locomotory mechanisms are used at different
size scales. Thus, when you have understood
bird flight completely, you have not even started
on the butterfly or the fruit fly. So, even when one
is successful, one may have provided understanding
only in particular cases.
We can already draw some conclusions. Don’t do
mathematical biology to satisfy a desire to find universal
structural relationships; you’ll be disappointed.
Don’t waste time developing “methods of
mathematical biology”; the problems are too
diverse for central methods. What’s left is the biology.
You should do mathematical biology only if
you are deeply interested in the science itself. If you
are, there’s lots of good news. We mathematicians
are experts at thinking through complex relationships
and formulating scientific questions as mathematical
questions. Some of these mathematical
questions are deep and interesting problems in
pure mathematics. And most biologists know that
the scientific questions are difficult and complicated,
so they want our help. There’s some bad news
too; there are three more reasons why the field is
so hard.
The problem of levels. In many biological problems
one is trying to understand how the behavior
of the system at one level arises from structures and
mechanisms at lower levels. How does the coordinated
firing of neurons give rise to the graceful
motion of an arm? How does the genetic code in
DNA create, maintain, and adjust a cell’s biochemistry?
How does the biochemistry of a cell
allow it to receive signals, process them, and send
signals to other cells? How does the behavior of
groups of cells in the immune system give rise to
the overall immune response? How do the properties
of individual bees give rise to the behavior of
the hive? How do the cells in a leaf “cooperate” to
turn the leaf towards the sun? How does the
varied behavior of individuals contribute to the
spread of epidemics?
We are familiar with these types of questions
from physics. What are the right variables to
describe the behavior of a gas, and how do the
values of these variables arise from the classical
mechanics of the molecules making up the gas?
The behavior at the higher level is relatively simple,
and Newton’s law suggests the few important
variables at the lower level; even so, the proofs are
not easy. In the case of biological systems these
questions are even more difficult, because the
objects at the lower level have been designed by
evolution (or trained by feedback control; see below)
to have just the right special properties to give
rise to the (often complicated) behavior at the
higher level. And it is usually not easy to decide
what the important variables are at the lower level.
If your model has too few, you will not be studying
the “real” biological mechanism. If your model
has too many, it may be so complicated that a
lifetime of computer simulations will not give new
biological understanding. You need ideas, guesswork,
experience, and luck. You need to be able to
deduce the consequences from the assumptions.
That is what mathematicians are good at.
The difficulty of experimentation. We mathematicians
often have an overly simple view of experiments
and the role they play, probably because
we don’t conduct them ourselves. A theory is tested
by deciding on a few crucial variables and designing
the right experimental setup. For example, one
measures how fast metal beads and feathers fall
in a vacuum or the angle subtended by two stars.
However, the complicated histories of interaction
between theory and experiment in quantum mechanics,
nuclear physics, and elementary particle
physics in the twentieth century show that this
simple view is naive. And for several reasons the
experimental situation is even more difficult in
biology.
First, one is often interested in how the behavior
at one level arises from lower levels. Typically,
this “emergent” behavior cannot be seen in any of
the parts at the lower level but arises because of
complex interactions among the parts. Unfortunately,
it can be misleading to study the parts in
isolation. For example, I try to understand how
certain biochemical networks in mammalian cells
function. The networks give rise to systems of
ODEs in which the nonlinear terms depend on the
enzyme kinetics for each separate enzyme. The
enzymes can be isolated and their reaction kinetics
studied “in vitro” in experiments that combine
340 NOTICES OF THE AMS VOLUME 51, NUMBER 3
pure enzyme with pure substrate. But in the soup,
in the real cell, the enzymes and substrates are binding
to or being affected by other chemicals too, so
one is unsure whether the “in vitro” experiments
reflect the true “in vivo” kinetics. That is, each of
the parts at the lower level behaves differently in
isolation than when it is connected to the other
parts.
Second, chance plays a role, not only in experimentation,
but perhaps in explanation. Why are
two neighboring fields dominated by black-eyed
Susans and poppies respectively? Are the soils
different? Are the local plant and animal species
different? Or perhaps the “explanation” is a chance
event in the past (or all of the above).
Third, individuals (whether cells or flowers or
people) are both similar and different. How does
one know whether data collected is “special” or
“typical”? How does one assure oneself that rat data
tells us something about humans or, indeed, something
about other rats?
Finally, it is characteristic of living systems that
the parts themselves are not fixed but ever changing,
sometimes even affected by the behavior of the
whole (see feedback control, below). A simple true
story illustrates this point. An experimenter (#1)
who used rats in his experiments was getting very
unusual results, and the results were not repeatable
week to week. After six months of this he had
to stop his experiments and investigate the rats.
His rats were housed in his university’s vivarium.
It turned out that another experimenter (#2) had
a mean technician, and when the other experimenter’s
technician came to get #2’s rats, #2’s rats
would cry out, upsetting the rats belonging to #1.
The behavior of #1’s rats in experiments depended
on whether #2 was doing an experiment the same
day! Whew, I’m glad I became a mathematician.
For all these reasons, biological data must be
approached cautiously and critically. Since biological
systems are so diverse and everything
seems to interact with everything else, there are
many possible measurements, and enormous
amounts of data can be produced. But data itself
is not understanding. Understanding requires a
conceptual framework (that is, a theory) that
identifies the fundamental variables and their
causative influences on each other. In messy biological
problems, without simple fundamental
principles like Newton’s Law, useful conceptual
frameworks are not easy to propose or validate. One
must have ideas about how the structure of (or
behavior of) the whole is related to the assumptions
about the parts. Thinking through such ideas
and proving the consequences of the assumptions
are important ways that we mathematicians can
make contributions.
The problem of feedback control. It is common
to think of biological systems as fragile.
However, most are very stable, and it is almost a
tautology to say so, because they must all operate
in the face of changing and fluctuating environmental
parameters; so if they weren’t stable, they
wouldn’t be here. We are familiar from engineering
with the concept of feedback control, whereby
variables are sensed and parameters are then reset
to change the behavior of the system. The nephrons
in the kidney sense NaCl concentration in the blood
and adjust filtration rate to regulate salt and water
balance in the body. The baroreceptor loop regulates
blood pressure, heart rate, and peripheral
resistance to adjust the circulation to different
challenges. Numerous such control systems are
known and studied in animal and plant physiology.
There is another kind of feedback that operates
between levels that poses special problems. Here
are two examples. In the auditory system sensory
information is transformed in the cochlea to electrical
information that proceeds up the VIIIth nerve
to the cochlear nucleus and from there to various
other nuclei in the brain stem (a nucleus is a large
anatomically distinct group of cells) and on to the
midbrain and the cortex. Surprisingly, there are also
neural projections from the cortex that influence
the sensitivity of the cochlea. Second, the dogma
DNA → RNA → proteins → function has turned
out to be a naive fiction. Genes (pieces of DNA) don’t
turn themselves on or off but are activated or inhibited
by proteins. That is, proteins affect the
genes, adding a reverse loop to the simple picture,
implying of course that the genes affect each other
through the proteins. The “fundamental” objects,
that is, the objects most closely related to “function”,
may not be genes or proteins but small
networks involving both genes and proteins that
respond in certain ways to changes in the cell’s
environment. These kinds of examples show that
the nineteenth-century picture of a machine with
parts is a very inappropriate metaphor for (at least
some) biological systems. When there is feedback
between levels, it is hard to say which are the parts!
In fact, it may be hard to say which are the levels,
and therefore our traditional scientific research
paradigm of breaking things into smaller parts
(lower levels) may not be successful.
This is not just a philosophical point but a fundamental
research issue that deepens the impact
of the previous four difficulties. Take, for example,
the question of dendritic geometry. It’s been one
hundred years since Ramon y Cajal made beautiful
drawings of complicated dendritic arbors on
nerve cells. Is the geometry important? Surely it
must be, we feel, since cells in the same brain nucleus
in different individuals seem to have roughly
similar dendritic arborization. And, indeed, there
are examples where it is understood how specific
dendritic geometry creates specific neuron-firing
properties and presumably specific cell function,
MARCH 2004 NOTICES OF THE AMS 341
though it is not always clear what “function” means
for a single cell embedded in layers and layers of
a large neural network. On the other hand, suppose
a cell is part of a large neural network whose
job it is to transform its pattern of inputs into a
corresponding pattern of outputs. This neural
network may have been trained to do this job by
feedback control from a higher level, in which case
the details of the dendritic geometry (and even the
details of the neural connections) may not be important
at all. The details arose from the training,
and they are whatever they need to be to give the
behavior at the higher level. Furthermore, for large
networks there may be many choices of details
that give the same network behavior, in which case
it will be hard to infer the behavior of the whole
by studying the properties of the parts.
I now want to turn to the policy questions that
I mentioned at the beginning. I have been using
the term “mathematical biology” to refer in the
broadest way to quantitative methods in the biological
and medical sciences. Physicists, chemists,
computer scientists, and biological and medical
researchers with some mathematical training can
and do contribute to the field I have been referring
to as “mathematical biology”. But let us now narrow
the focus to mathematics education, both undergraduate
and graduate, and the mathematics job
market.
Undergraduate education. Mathematical biology
is an extremely appealing subject to undergraduate
students with good training in freshman and
sophomore mathematics. Many are naturally interested
in biology, and all know that we are in the
midst of a revolution in the biological sciences.
They are usually amazed and delighted that the
mathematical techniques that they have learned can
be used to help understand how biological systems
work. Further, mathematical biology is a perfect
subject for undergraduate research projects.
Biology is so diverse and so little quantitative modeling
has been done that it is relatively easy to find
projects that use undergraduate mathematics in
new biological applications. The students find such
projects to be very rewarding. They know that the
undergraduate major consists mostly of nineteenthcentury
mathematics; of course they are excited by
twenty-first-century applications. Here at Duke we
have found that the availability of projects in
mathematical biology has attracted many students
to the mathematics major. Of course, it helps to
have a mathematical biologist on the faculty, but
it is not necessary. Any mathematician can create
and supervise such projects by working cooperatively
with local biologists. It requires only the
effort to make the connections and tolerance for
appearing “nonexpert” to the students (something
we are not used to!).
Graduate education. There is quite a bit of disagreement
about the proper mathematics graduate
training of a student who wants to be a mathematical
biologist. I’ll simplify the discussion into two
(extreme) positions. The first position emphasizes
maximal contact with biology and biologists as
part of graduate training. Graduate students should
take biology courses (including labs) and should participate
in or even initiate collaborative modeling
projects. This way they learn a lot of biology, and,
even more importantly, they learn modeling and
how to communicate with biologists. By doing this
they study less mathematics of course, but they can
learn what they need later when they need it. The
second position emphasizes training as a mathematician
first. Graduate students should receive the
traditional training in analysis or applied mathematics
(or other subjects), and (ideally) the thesis
should contain some applications to biological
problems. But the graduate student should not
spend too much time slogging around in the biological
details or working on collaborative projects.
It is the job of the thesis advisor to be the interface
between the graduate student and the biology.
Later, after the Ph.D., when the mathematician is established,
he or she can choose to become involved
in collaborations and learn more biology.
I guess that most mathematical biologists support
the first position. I support the second,
perhaps because that is the route that I followed
myself. Mathematical biology is really a very hard
subject (I hope I have convinced you of that), and
a great many ideas and techniques from different
branches of mathematics have proven useful. So
mathematical biologists need broad training in
mathematics. Secondly, I believe that only deep
and rigorous graduate training creates mathematicians
who can not only learn new mathematics
when they need it, but who can also recognize
what they need to learn.
Hiring issues. Hiring a mathematical biologist
posesspecialchallengesfordepartments.Mostmathematicians
have no idea how large the field “biology”
is or how large the research communities are. Two
examples illustrate this. Here at Duke, Arts and Sciences
has 469 tenure-track faculty members, and
the Medical Center has 767 (not counting clinical
faculty). My colleague Harold Layton is a mathematician
who works on the kidney, so of course he
goes to the annual meeting of the American Society
for Nephrology, where the typical registration number,
12,000, completely dwarfs the registration at
the Joint Mathematics Meetings. And those constitute
just a subset of the researchers who work on the
kidney! So, the first issue is thinking about what kind
of mathematical biologist you want. It is a good idea
to involve local biologists and medical researchers
in preliminary discussions, both to educate department
faculty and to understand the local context.
342 NOTICES OF THE AMS VOLUME 51, NUMBER 3
The question of how best to judge job candidates
is particularly difficult in mathematical biology.
First, most mathematicians know less biology than
their tenth-grade daughters. That’s just the way it
is; biology was not a mathematically related discipline
when we were growing up. But, more importantly,
mathematical biology really isn’t a “field” of
mathematics with a coherent community that can
testify meaningfully about young people. Mathematical
biology is fragmented because biology
itself is so diverse. A mathematician working on the
lung might want to talk to pulmonary physiologists
or geometric analysts who are experts on fractals,
but why would he or she want to talk to mathematical
biologists working on the kidney, the neurobiology
of hearing, or the epidemiology of AIDS?
In each area of specialization there is a tremendous
amount of biology to learn, and if one doesn’t have
the background, it’s hard to judge the strength of
an individual’s contributions. This is true even for
us, the mathematical biologists, the “experts” you
expect to consult. So hiring in mathematical biology
necessarily involves intuition and high risks as
well as high potential payoffs for the department
and the college.
The first step. The best way for departments to
overcome these difficulties is to encourage senior
faculty to become involved in bringing biological
applications and student projects into the undergraduate
curriculum. This should be done by
working cooperatively with local biologists to
create examples and projects related to their
own specialties. All mathematicians can do this,
and they do not have to give up their own research
agendas or become mathematical biologists; it
only requires effort.
This strategy for engagement with biology has
great benefits both for departments and individuals.
The faculty as a whole will become educated in
biology and thus better able to judge job candidates
in mathematical biology, and the undergraduate
curriculum will be more attractive. More importantly,
departments and individuals will be participating
intellectually in the biological revolution, the greatest
scientific revolution of our times, perhaps of all
times. The task is to understand how life, in all its
diversity and detail, works. This includes how we
act, think, and feel, and how we influence and are
influenced by other forms of life. We mathematicians
have the technical and intellectual tools to
make enormous contributions. So, surely, this is our
responsibility and our opportunity.