Predicting drug selectivity using computers
One of the most important requirements for the development of a safe drug is selectivity - in other words, the ability to interact preferentially with the desired biological target. If a drug interacts significantly with other targets in the body, this can lead to undesirable side effects. A major factor determining selectivity is how tightly a drug binds to its target molecule, as compared to other targets. The strength of the interaction is described by the drug's affinity, which can be measured with, for instance, isothermal titration calorimetry (ITC). It would be incredibly useful and cost-efficient if we could predict affinity in silico. However, to do this in a way that has general applicability has proven very difficult.
Figure 1. A schematic summary of the work. The affinity of a single ligand is computed against a series of different proteins from the same family, in this case human bromodomains. The predictions show good agreement with experimentally obtained affinity data (Figure taken from ). (Click to Enlarge)
One method that can potentially offer a general solution to the problem makes use of molecular dynamics simulations in order to calculate the absolute binding free energy (ABFE) through non-physical (alchemical) thermodynamic cycles. This approach, whereby contributions to the binding free energy from individual steps of the cycle are computed is thermodynamically rigorous, but until recently, suffered from the fact that vast amounts of computational power were required to ensure that the system had adequately sampled the conformational space. Even if the system can fully explore all of the conformational space, there are other issues that need to be considered. For example, it is not really known whether the accuracy of the physical models describing the protein-ligand systems (the so-called force-fields) is sufficient to be used with confidence in prospective drug-design applications.
In recent years, many groups have shown that free energy calculations based on molecular dynamics can be performed with an accuracy that would be useful in a drug-design research program. Most have focussed on assessing the relative affinities of a series of different drugs for a particular protein target. The prediction of selectivity is slightly different (and is considered more challenging) as it involves predicting the affinity of a single drug for different protein targets.
In JACS this week, Aldeghi and colleagues evaluated the performance of free energy calculations based on molecular dynamics for the prediction of selectivity by estimating the affinity profile of three bromodomain inhibitors across multiple bromodomain families. The predictions were compared to ITC. Bromodomains are a class of protein domains and epigenetic mark readers; specifically, they "read" whether lysines on histone tails have been acetylated.
The authors performed two case studies. In the first one, the affinities of two similar ligands for seven bromodomain proteins were calculated and compared to ITC data. The results were impressive, with a mean unsigned error of 0.81 kcal/mol and a strong positive correlation of 0.75 with respect to ITC measurements (Figure 2).
Figure 2. Affinity predictions for two closely related compounds (RVX-208 and RVX-OH) capture the selectivity profiles extremely well. The shaded gray areas in the plot on the left indicate where the 1 and 2 kcal/mol error boundaries lie. The RVX-208 compound shows tighter binding to members of the BD2 family of bromodomains compared to the BD1 family, whereas the RVX-OH compound is predicted to exhibit no selectivity. The pattern of affinity is summarised on the right, where Gaussian curves have been fitted to the data. The results were confirmed by ITC data. (Figure adapted from ). (Click to Enlarge)
"This was particularly satisfying as the orientation of this compound varies depending on the protein target" said Matteo Aldeghi, the DPhil student who performed the work. "The approach we have used here is the only one that can realistically be used for that kind of scenario where the pose (orientation) of the drug can change".
In the second case study, the affinities of a broad-spectrum inhibitor called bromosporine were calculated for 22 different bromodomains. The accuracy obtained was slightly worse in this instance (a mean unsigned error of 1.76 kcal/mol) but the authors show that some improvement can be obtained by refining the underlying physical model describing the small molecule. This suggests that force fields are still not quite optimal, but also that there is the possibility of improving them. Despite this, the results are still impressive as they manage to reproduce experimental affinities in absolute terms, rather than only relative ones.
"This work is an important line in the sand in terms of what is currently achievable with freely available tools" said Professor Philip Biggin, senior author of the work.
The work suggests that this approach is likely to be exploited more and more as computational power and algorithms continue to improve.