Physicist
and code specialist Dr. Sandipan Mohanty has been working on molecular
biology simulations for the world's fastest supercomputers for 20 years.
Such simulations help to unravel the building blocks of life and
provide new insights into cellular machinery.
Together
with researchers at Sweden's Lund University, he has now gone one step
further and taken the problem of protein folding to a quantum computer.
D-Wave's quantum annealer JUPSI of the quantum computer user facility
JUNIQ at Forschungszentrum Jülich has more than 5,000 qubits and is the
first device of this size outside North America. In an interview,
Sandipan Mohanty gives insight into the pioneering work.
I
would say what we really achieved is to show the viability of quantum
computers for non-trivial research questions in our field. Quantum
computers are a quite new technology and it is not yet clear how to
program them when you try to solve real scientific tasks with these new
machines. For example, it is rather different from solving the problem
with classical high performance computing.
More
specifically, we successfully studied protein folding using a very
simple model. Proteins are important building blocks of life. They
fulfill a wide variety of tasks. These include, for instance, the
transport of substances and cell structure. And they can only fulfill
all these functions if they have a very specific form, which they
achieve through a process called protein folding.
One
of the many reasons why there is a lot of interest in this process is
the connection between neuro-degenerative diseases like Alzheimer's or
Parkinson's disease with protein mis-folding. Our hope is that quantum
computers will have important advantages which will further advance our
understanding of such phenomena.
Proteins are long flexible chains of amino acids.
One fascinating property of these molecules is that a large fraction of
all proteins spontaneously balls up into very specific
three-dimensional shapes when you put them in a solution, think of
water. So, in principle, all you need to know is the sequence of amino acids that make up a protein chain. The chain then automatically knows which shape it has to fold into.
When
modeling this folding process on a computer, there's a lot to try out.
You can think of it like trying to calculate all the different ways you
can arrange a necklace, to search for the "best" arrangement. Moreover,
examining each arrangement is also computationally very expensive
because of the large number of particles involved. Usually this means
millions of interactions to calculate for every arrangement examined.
The
task we solved is years away in complexity from the problems we
normally solve with classical supercomputers, where large scale
atomically detailed simulations are common. On the D-Wave machine, we
used a very reduced HP model. This simplifies the problem massively,
retaining just the minimum essential physical characteristics of the
folding process. We ignore the surrounding medium, divide amino acids
into only two types, approximate each amino acid by a single ball which
can only occupy positions on a lattice.
I
should point out that as of today, even with such simplified models,
being able to study chains of 64 amino acids using a quantum computer is
a hard problem, which makes our results very satisfying!
Corresponding
simulations can also be carried out classically. A notebook is
sufficient for this. The time for the calculation does not differ much,
in both cases it takes one to two minutes. However, this value is
actually meaningless. Much more important is the quality of the results.
And here, the quantum annealer clearly performs better.
It
was rather easy to achieve 100 percent success rate in finding the
lowest energy structures on JUPSI. With classical computers, on the
other hand, comparable simulations only achieve 80 precent for a chain
of 30 amino acids. For the more complex proteins consisting of 48 or 64
amino acid blocks, they do much worse, whereas the quantum annealer
always produces the correct result here as well.
Because
it benefits from specific aspects of the research problem. The
computational effort required with classical computers, to take all
relevant protein conformations into account is astronomically high. It
grows exponentially with the length of the protein chain. With a chain
of two particles, there are maybe ten possibilities. With three
particles, there are already a hundred. But with 100 particles—which are
still rather few for a protein—one would have to calculate billions of times more variations than there are atoms in the universe.
To
make any meaningful calculation at all, lots of tricks are used. Our
group at the JSC and my collaborators in Sweden both specialize in so
called Monte Carlo simulations. It is a procedure based on statistical
physics and stochastic sampling. Although infinitely long simulations
are guaranteed to produce correct results, short runs can have large
errors. In practice, one tries to perform "long-enough" simulations, so
that the estimated errors are acceptably small. Here lies the advantage
of the quantum annealer.
This
machine can, if it is programmed correctly, perform this approximation
in a very direct way via its quantum mechanical couplings. Basically, it
is a kind of an intricate physics experiment that automatically solves
the equation. In our problem, it seems to have the effect that
comparatively smaller run times are required to obtain very good
answers. The fact that it works so well in practice did, however,
surprise us a little.
Our
work marks only a first step. Most of today's quantum computers have
only a few qubits. The D-Wave system has 5000, which is a lot. But for
most research problems, fruitful applications of quantum computers would
need even more qubits. We are still a long way from simulations such as
those used in drug research done on supercomputers. I expect that we
will have to wait for two or three more generations of devices to come
before we will be able to solve such problems on a quantum computer.
But
I am hopeful. In contrast to the existing research we learned from, our
formulation retains its simplicity with increasing system size. This
opens up a possible smoother path to the study of considerably more
complex problems on quantum computers.