Watching drugs move: Markov models of protein interactions

For decades drug designers learned about their targets from a frozen picture. X-ray crystallography shows where the atoms of a protein sit. Pharmacophores, simple maps of the features a drug needs to bind, are built from that picture. Docking, the most common computer method, also works on a static structure.

Real proteins move. They wiggle, open and close. Molecular dynamics (MD) simulations show this motion by calculating how every atom moves over time. As computers got faster, MD became a standard part of drug design.

The thesis argues that this creates a conflict. Models built for still pictures do not fit what moving simulations show.

It builds on two earlier ideas. Ensemble QSAR methods learn from many protein shapes instead of one. Dynophores, dynamic pharmacophores, track how the contacts between a drug and a protein change during a simulation. Dynophores had recently replaced classic pharmacophores in a virtual screening study on the enzyme DNA topoisomerase IIa.

The thesis goes one step further and drops the still picture entirely. It derives models directly from MD simulations by combining Dynophores with Markov modelling.

A Markov chain describes a system that jumps between states, where the next step depends only on the current state. Here a state can be a pattern of contacts between drug and protein. The chain then tells you how often each pattern shows up and how likely it is to switch to another. That turns a long, messy simulation into a map you can read.

The thesis also used lean project management methods to run the work, which is unusual for a chemistry thesis.

The method was tested on three targets.

The first is the Zika virus protease, an enzyme the virus needs to copy itself. It served as a benchmark against expert knowledge used to design new inhibitors.

The second is human cyclin-dependent kinase 2 (CDK2), studied in complex with three different ligands. The thesis notes that how CDK2 works and how it is blocked is still not fully understood, so it served as a test for new insights from experimental data.

The third is the hepatitis C virus protease NS3/4A. Four complexes were studied: the normal (wild type) enzyme and three mutants, each bound to the drug vaniprevir. The question was whether the method can explain why mutations make the virus resistant.

According to the abstract, the approach produced new insights for all three test systems. The detailed results are in the 60-page thesis on Refubium, the open repository of FU Berlin.

Molecular dynamics: a computer simulation that moves every atom of a system step by step according to physical forces.

Pharmacophore: a map of the features a molecule needs to bind a target, such as a spot that donates a hydrogen bond.

Ligand: a molecule that binds to a protein, such as a drug.

Protease: an enzyme that cuts other proteins. Many viruses need a protease to mature, which makes it a drug target.

Kinase: an enzyme that attaches phosphate groups to other molecules and so switches cell processes on or off.

Resistance mutation: a change in a virus protein that makes a drug bind less well.

This is a master's thesis, not a peer-reviewed journal paper. The systems are three well-known targets, and simulations always depend on the force fields and settings chosen. Insights from simulations need lab work to confirm them.

Later work by Julian M. Kleber moved from simulations to learned molecule representations. See CARATE and the CACHE #1 challenge.