Protein Design by Provable Algorithms with Bruce Donald and Mark Hallen


Imagine the potential to cure infectious diseases
because of an algorithm. For billions of years, the process of evolution has optimized the
sequence of amino acids to suit the needs of the organisms that make them. So we ask,
“Can we use computation to design non-naturally occurring proteins that suit our needs?” Join
us, as we sit down with Mark Hallen and Bruce Donald, where we share an inside look into
protein design by provable algorithms. I had this idea that many other people did
too, that mathematics and computer science could really make a difference in how we design
molecules. I realize that it was an area that mixed portions of discrete and continuous
optimization and reasoning in a way that I was familiar with, having worked in robotics
and also nanotechnology. When I realized that we could actually, potentially
use this technique to make new reagents and new therapeutics even that could actually
help humans, that just continued the excitement for me.
The pairwise discrete model captures the most essential aspects of computational protein
design, but it falls short for many practical applications. Real proteins have significant,
continuous flexibility. What we’re trying to do with the continuous
flexibility, is more realistically model how proteins move and if there’s a small molecule,
how that moves, too and interacts with the protein. In the discrete pairwise model, that
we talk about a lot in the paper, because it’s received so much research attention,
you assume that the protein has this sort of discrete list of states, geometric states
that it can go to. But the problem of course is that these are 3-D objects, they can move
continuously. And what we’ve found is that those small, continuous motions that are kind of smaller than the resolution ot the discrete models can really be critical in predicting and correctly
modeling the energies and the functional properties of proteins and molecules that
they interact with. Research in provable algorithms not only paves
the way for advances, it also gives the opportunity for a broad range of users to participate.
With the discrete pairwise model that we talk about how methods from the way to constrain satisfaction
in the algorithm world have helped to bring a lot more efficiency for the same model,
which of course comes with intrinsically the same physical accuracy, because they are provable
algorithms, like you get in the model there is a certain answer. These algorithms are
going to find it. What we talk about a lot in the second half
of our review is increasing the accuracy while trying not to make things too much slower.
This tells us really where algorithms research can help. The user of these techniques, even
a very sophisticated user or programmer basically is choosing where they want to be on that
curve. What algorithms research lets you do is move that curve outwards. It sounds like
something for nothing, but it comes from the power of mathematics. You move the curve outwards,
and then all of a sudden you can simultaneously do larger problems with better models.
Progress is inevitable. Some believe that collaboration among sectors could be the key
to further computational advancement. We made so much progress in the past decade
or so, and being able to handle the models we have very efficiently and in the very core
of computational chemistry world bringing in some machine learning techniques, bringing
in GPUs, and I think there could be a possibility for a lot of this to come together and have
really powerful models that have predictive ability that is well beyond what we have today.
What the future holds for provable algorithms is unknown. But participation growth within
the field becomes more of a reality each day. We hope it’s a little bit of a call to action:
like if people are inspired by either the applications, by medical problems, or by the
cool algorithms maybe they’d like to work on these problems. And we saw that happen
so many times in robotics and AI, we think that it might be time in molecular design
where there is a unity between sort of machine learning, computational chemistry, and computer
science to move to the next level. Find out more in “Protein Design by Provable
Algorithms,” a review article in the October 2019 Communications of the ACM.

Leave a Reply

Your email address will not be published. Required fields are marked *