A key challenge in the treatment of viral and bacterial pathogens is the emergence of drug-resistant mutations. What if we applied computational thinking to this problem? Could advanced algorithms help in the design of new drugs that not only are effective against specific disease agents, but also against any variants that are likely to arise due to drug-induced evolution? In this PROBE, Chris Langmead will explore the use of machine learning techniques to predict likely mutations and thus allow better drugs to be designed.

PROBE on Designing Resistance-Evading Drugs

Organized by Christopher James Langmead

The goal of this PROBE is to use Computational Thinking in the design of antiviral and antibacterial drugs.  One of the most difficult challenges in the battle against infectious agents is the emergence of resistance.  The essence of resistance is that mutations in a viral or bacterial genome can render an otherwise effective compound useless.  A successful drug is one that remains effective despite these mutations.  Such resistance-evading drugs are normally discovered after a time-consuming and expensive process. We believe that by selecting the right abstraction, this process can be automated to a large degree, saving time, money, and hopefully lives.

If you want to design a drug that evades resistance, you need to do four things. First, you need to understand exactly how the drug works.  This is often accomplished by performing detailed studies of the biochemistry and physics of interactions between the drug and the target molecule. Second, you need to identify mutations in the target molecule that appear to provide resistance to the drug. This is most often accomplished by waiting until the resistant strain arises in the population. Third, you need to understand why a given mutation confers resistance.  This too involves detailed studies of the interactions between the drug and the (mutated) target. Finally, one searches for a new drug that is both effective and isn’t susceptible to the mutation.

We believe these four steps can be combined by imagining that the design process is a game of strategy. On one side, we have the pharmaceutical company, and on the other side we have the disease. The pharmaceutical company makes a ‘move’ by designing a drug. The disease makes a move by evolving away from the drug. Each side must obey certain rules. The pharmaceutical company, for example, must avoid drugs that have harmful side-effects. The disease, on the other hand, must avoid mutations that reduce its own viability.  The key to winning this game is to force the opponent into a corner --- a sort of biochemical checkmate.

This PROBE focuses on one aspect of this problem. Namely, we are developing tools that, figuratively, will allow the pharmaceutical company to look ahead a few moves. We have developed a computational method that can model simultaneously: (i) the physical interactions between drugs and target, and (ii) the evolution of the target.  The abstraction we are using is known as a Markov Random Field (MRF), which is an efficient way to represent and compute over complex probability distributions. In our case, we will be modeling the joint probability distribution over molecular sequence and structure.  Then, we can automatically solve the design problem by treating it as an inference problem. Specifically, we want a drug that has favorable interactions with the target, and any viable mutant of the target.

We will demonstrate our method by considering the design problem on HIV-1 Protease (HIV PR). This protein plays an essential role in the replication of HIV. Protease inhibitors are drugs that bind to HIV PR and disrupt its normal function. A great deal is known about HIV PR, and the mutations that confer resistance to commercially available HIV drugs. Over the next year, we will validate our method by demonstrating that it can accurately predict known mutations.

This research has direct ties to ongoing research at Microsoft Research (MSR). MSR has conducted pioneering research in the application of Machine Learning and Formal Methods to key problems in Biology and Medicine. Among the many contributions are cutting-edge Machine Learning techniques for studying the evolution of the HIV genomic sequence in the context of vaccine design. The proposed research also uses techniques from Machine Learning, but studies the evolution of proteomic sequences from a biophysical perspective.

Papers

A graphical model approach for predicting free energies of association for protein-protein interactions under backbone and side-chain flexibility
Hetunandan Kamisetty, Chris Bailey-Kellogg, Christopher James Langmead
December 2008, CMU-CS-08-162

Modeling and Inference of sequence-Structure Specificity
Hetunandan Kamisetty, Bornika Ghosh, Chris Bailey-Kellogg, Christopher James Langmead

A Bayesian Approach to Protein Model Quality Assessme
Hetunandan Kamisetty, Christopher J. Langmead