Drug Discovery Implications of Molecular Dynamics for the Algebraically Despondent
This blog article refers to the J. Med. Chem. Perspective: Role of Molecular Dynamics and Related Methods in Drug Discovery De Vivo, Masetti, Bottegoni and Cavalli.1 Not only is it a fair review of current flavours of Molecular Dynamics (MD), within a drug discovery context, it also does the impressive task of making it reasonably accessible to those without a double first in maths and Greek. This blog article will comment on some of the basic concepts within the paper (and in the field of MD in general) but if the reader is interested, they should definitely pick up this paper – whilst there are key equations for MD contained within, they are well explained, and held within the context of what they give the modern drug discovery scientist as observables.
Classical MD was originally fathomed in the late 1950’s with landmark papers within the biochemical domain slightly later,2 but for a long time was prohibitively computationally expensive. Modern technology (namely, parallelisation via clusters and the GPU, and coding architecture utilising this such as NVidia’s CUDA)3, as well as new thinking in drug discovery has moved MD to the forefront of many drug discovery strategies. More recently, Free Energy Perturbation (FEP) with MD sampling has become a rapidly growing field with many of the big pharma picking up on this too. A colleague’s SBDD blog article4 a few weeks ago detailed an FEP keystone paper from Sherman and collaborators.5
Molecular Dynamics for the Non-Computational Mind
If we consider atoms to be balls, and the bonds to be springs, then MD is really all about how these balls and springs move around in relation to each other under Newton’s Laws of Motion over a given timespan. A force field is applied for this comprising of 5 elements (equation 1), which seeks to calculate potential energies for a predetermined time step. Knowing the starting point of each of our balls on their springs (usually from properly prepared crystallographic data), and having determined the movement energies they have, we can assess where they will be at the end of the time step using Newton’s Second Law (f(t) = ma(t)). A collection of these time steps is called a trajectory – often shown as videos or overlays such as the image atop this article.
V = ∑bond + ∑angles +∑torsions +∑van der Waals + ∑electrostatics (eqn 1)
Where V is the empirical potential energy of a system, and each of the components (e.g. bond, angles, etc) are sums of contributions across a system. This is a much simplified version of Equation 2 from De Vivo et al.’s paper.
The simulation of this movement, using various control parameters, allows us to explore the energy landscape of a protein. For example: If we know the energy of the system (and parts of the system), and we know the energy barriers for certain movements (e.g. side chain rotations), we can see which parts of the protein can move for a given energy (e.g. the energy it may have at physiological temperature). We allow the protein to move around these energetically allowed configurations. Likewise, for any given configuration, we can assess various energy terms. Figure 1 demonstrates some basic biological motions within the time domain, and what kind of things normal MD can be used to probe.
Figure 1: Timescale of various biological processes, and the techniques used to investigate them
Simulations of biological systems are particularly useful when considering the movement of a protein pockets or a protein-ligand complex – especially since understanding of drug discovery has moved rapidly (and rightly) away from Emil Fischer’s classical lock-and-key binding paradigm towards the idea of induced fit and conformational selection.6 Understanding how a pocket moves in relation to solvent and ligand, which residues are more mobile compared with others and which interactions they are open to, and how a ligand may approach and behave in the pocket becomes a very powerful collection of insights. Take a kinase for instance (Figure 2)- understanding how a DFG loop (for example) moves in response to ATP binding would be a key piece of information for drug design teams. Gosu and Choi attempted to model such movement early last year, demonstrating a change in the dynamics of the kinase in response to its ATP-bound or unbound state.7
Figure 2: An example of the scope of movement of IRAK kinase (where uIRAK4 is apo unphosphorylated, pIRAK4 is apo phosphorylated, and p-IRAK4-ATP is ATP-bound phosphorylated IRAK4.7
MD is as good as the force field it’s built upon, and the authors bring to light some limitations of this, as well as common methods to increase sampling, such as Umbrella Sampling, Replica Exchange MD (REMD), and Metadynamics approaches to getting out of landscape minima. The authors also briefly discuss FEP, which in the last few years has become a matter of significant interest. FEP is an alchemical method that can effectively allow the comparison of free binding energies between ligands in a design series. It is beyond the scope of this blog, and even this perspective article to discuss FEP to the level that it deserves, but the authors do set out a useful overview.
The authors do a pretty good job of making the rather dry subject of MD and allied techniques suitable for the med chemist or team leader who wants to understand what tools his CADD support can offer his project in terms of understanding the movement of his target, and is written in way that explains what exactly the techniques can (and cant) provide, and where the limitations are, with relevant illustrations and references to document real-world projects. As well as describing more classical usage such as conformational sampling for virtual screens and ligand design, they also detail some interesting methods such as steered MD to assess binding or unbinding events, and the use of MD in seeking out allosteric sites in proteins.8 They detail a recent study which demonstrated the ability to define binders from non-binders within a similar chemical series using a modified ligand pulling experiment (where a force is applied to pull the ligand from the binding pocket in silico. The required force is related to its binding energy, and thus peaks in force required to remove the ligand, reveal the tightness of binding.9
Fig 3: Plot of force vs time for “pulling” or “undocking” a ligand from its pocket. Two clear profiles are shown here which represent those that bind well versus those which do not. Those with the higher force profile have a proton donor which the low force do not, suggesting that the donor is important to binding.
The perspective concludes that various flavours of MD can demonstrate significant insight into binding / unbinding events, long-residency waters as well as understanding target movement, but warn that more prospective studies are required to bring this into the truly into the light, as well as some well-placed concerns about force field limitations. Whether you agree with them that “FEP is ready for prime time” or not, this is a well written perspective, surely worthy of a read.
Notes from blogger (whose opinions are his own)
MD is definitely a useful tool, and with hardware advances it is becoming more common – especially as we see more and more complex drug targets going forward, where conformational changes in proteins are better understood, and we target more flexible or non-classical pockets. It is just another tool, however, and care needs to be taken to really balance its value in modern drug discovery.
Hardware changes such as GPUs (Graphical Processing Units – effectively the graphics cards in your computer), which have hundreds or thousands of processing units (albeit slower and more restricted in their calculations that your CPU), can accelerate MD calculations significantly. Many common MD packages such as NAMD, Desmond and AMBER can be parallelised on GPUs. We routinely use NVidia K40 GPUs which tend to accelerate our MD significantly (3-10 fold on Desmond/Schrodinger 2016-1, OPLS3.0). You can technically use consumer level GPU’s but beware they are not designed to fun flat out 24/7 and your normal desktop is not likely set up for the power and cooling requirements. If you are in academia, you can apply to NVIDIA for Academic Hardware Grant.10
Free Energy Perturbation has risen in popularity over the last few years, breaking out from purely academic interest to major pharmaceutical companies on-boarding this technology11 (it even made the Pipeline).12 At the time of writing, I can name at least half a dozen colleagues in different big pharma organisations investing significantly (as far as computational drug discovery goes), in setting up FEP as part of their CADD support offering. FEP is currently being invested in heavily by Schrodinger Inc., using the Desmond MD platform for it – on speaking with their development team they see it as a “game changer”. Over at CCG, when I spoke to their CEO, Paul Labute early last year, he said that they had no current plan to follow Schrodinger’s move with their platform, MOE – focussing more on the ease of use and accessibility of their offering, though this may change.
Whether you see FEP as the gateway to a new paradigm in drug discovery, or whether you’re on the fence about this “game changing” tech (I mean, we’ve all been there in presentations about “the next big thing”), it has brought significant interest into MD for Drug Discovery.
Blog written by Ben Wahab
- De Vivo et al. Role of Molecular Dynamics and Related methods in Drug Discovery. J. Med. Chem. 2016 in press. DOI: 10.1021/acs.jmedchem.5b01684
- (a) McCammon, J. A.; Gelin, B. R.; Karplus, M. Dynamics of folded proteins. Nature 1977, 267, 585−590. DOI: 10.1038/267585a0 ; (b) Levitt, M.; Warshel, A. Computer simulation of protein folding. Nature 1975, 253, 694−698. DOI: 10.1038/253694a0
- Sherman et al. Accurate and Reliable Prediction of Relative Ligand Binding Potency in Prospective Drug Discovery by Way of a Modern Free-Energy Calculation Protocol and Force Field J. Am. Chem. Soc., 2015, 137 (7), pp 2695–270. DOI: 10.1021/ja512751q
- (a) Changeux, J.-P.; Edelstein, S. Conformational selection or induced fit? 50 years of debate resolved. F1000 Biol. Rep. 2011, 3, 19. DOI: 10.3410/b3-19 ; (b) Vogt, A. D.; Di Cera, E. Conformational selection or induced fit? A critical appraisal of the kinetic mechanism. Biochemistry 2012, 51, 5894−5902. DOI: 10.1021/bi3006913 ; (c) Vogt, A. D.; Di Cera, E. Conformational Selection Is a Dominant Mechanism of Ligand Binding. Biochemistry, 2013, 52 (34), pp 5723–5729. DOI: 10.1021/bi400929b
- Choi, S and Gosu, V. Structural dynamic analysis of apo and ATP-bound IRAK4 kinase. Nature Scientific Reports, 4, No.5748, 2014 DOI: 10.1038/srep05748
- (a) Desdouits, N.; Nilges, M.; Blondel, A. Principal Component Analysis reveals correlation of cavities evolution and functional motions in proteins. J. Mol. Graphics Modell. 2015, 55, 13−24. DOI: 10.1016/j.jmgm.2014.10.011 ; (b) Kokh, D. B.; Richter, S.; Henrich, S.; Czodrowski, P.; Rippmann, F.; Wade, R. C. TRAPP: a tool for analysis of transient binding pockets in proteins. J. Chem. Inf. Model. 2013, 53, 1235−1252. DOI: 10.1021/ci4000294
- Colizzi, F.; Perozzo, R.; Scapozza, L.; Recanatini, M.; Cavalli, A. Single-molecule pulling simulations can discern active from inactive enzyme inhibitors. J. Am. Chem. Soc. 2010, 132, 7361−7371. DOI: 10.1021/ja100259r