In this week’s journal club, we reviewed a paper by Lippow et al. in Nature Biotechnology, which features a computational pipeline that is capable of maturing antibodies (Abs) by up to 140-fold. The paper itself discusses 4 test case Abs (D44.1, cetuximab, 4-4-20, bevacizumab) and uses changes in electrostatic energy to identify favourable mutations. Up to the point when this paper was published back in 2007, computational antibody design was an (almost) unexplored field of research – except for a study by Clark et al. in 2006, no one else had done anything like the work presented in this paper.

The idea behind the paper is to identify certain positions within the Ab structure for mutation and hopefully find an Ab with a higher binding affinity.

Pipeline

Briefly speaking, the group generated a mutant Ab-antigen (Ag) complex using a series of algorithms (dead-end elimination and A*), which was then scored by the group’s energy function for identifying favourable mutations. Lippow et al. used the electrostatics term of their binding affinity prediction in order to estimate the effects of mutations on an Ab’s binding affinity. In other words, instead of examining their entire scoring function, which includes terms such as van der Waal’s energy, the group only used changes in the electrostatic energy term as an indicator for proposing mutations. Overall, in 2 of the 4 mentioned test cases (D44.1 & cetuximab), the proposed mutations were experimentally tested to confirm their computational design pipeline – a brief overview of these two case studies will be described.

Results

In the case of the D44.1 anti-lysozyme Ab, the group proposed 9 single mutations by their electrostatics-based calculation method; 6/9 single mutants were confirmed to be beneficial (i.e., the mutant had an increased binding affinity). The beneficial single mutants were combined, ultimately leading to a quadruple mutant structure with a 100-fold improvement in affinity. The quadruple mutant was then subjected to a second round of computer-guided affinity maturation, leading to a new variant with six mutations (effectively a 140-fold improvement over the wild-type Ab). This case study was a solid testimony to the validity of their method; since anti-lysozyme Abs are often used as model systems, these results demonstrated that their design pipeline had taken, in principle, a suitable approach to maturing Abs in silico.

The second case study with cetuximab was arguably the more interesting result. Like the D44.1 case above, mutations were proposed to increase the Ab’s binding affinity on the basis of the changes in electrostatics. Although the newly-designed triple mutant only showed a 10-fold improvement over its wild-type counterpart, the group showed that their protocols can work for therapeutically-relevant Abs. The cetuximab example was a perfect complement to the previous case study — it demonstrated the practical implications of the method, and how this pipeline could potentially be used to mature existing Abs within the clinic today.

Effectively, the group suggested that mutations that either introduce hydrophobicity or a net charge at the binding interface tend to increase an Ab’s binding affinity. These conclusions shouldn’t come with huge surprise, but it was remarkable that the group had reached these conclusions with just one term from their energy function.

Conclusions

Effectively, the paper set off a whole new series of possibilities and helped us to widen our horizons. The paper was by no means perfect, especially with respect to predicting the precise binding affinities of mutants – much of this error could be bottled down to the modelling stage of their pipeline. However, the paper showed that computational affinity maturation is not just a dream – in fact, the paper showed that it’s perfectly doable, and immediately applicable. Interestingly, Lippow et al.’s manipulation of an Ab’s electrostatics seemed to be a valid approach, with recent publications on Ab maturation showing that introducing charged residues can enhance binding affinity (e.g. Kiyoshi et al., 2014).

More importantly, the paper was a beautiful showcase of how computational analyses could inform the decision making process in an in vitro framework, and I believe it exemplified how we should approach our problems in bioinformatics. We should not think of proteins as mere text files and numbers, but realise that they are living systems, and we’re not yet at a point where we fully understand how proteins behave. This shouldn’t discourage us from research; instead, it should give us the incentive to take things more slowly, and develop a method/product that could be used to solve greater, pragmatic problems.

This week’s topic of the Journalclub was about Molecular Mechanics Poisson−Boltzmann Surface Area (MMPBSA) binding free energy calculations between ligand and receptor using Molecular Dynamics simultions (MD). As an example I selected:

David W. Wright, Benjamin A. Hall, Owain A. Kenway, Shantenu Jha, and Peter V. Coveney. Computing Clinically Relevant Binding Free Energies of HIV-1 Protease Inhibitors. J Chem Theory Comput. Mar 11, 2014; 10(3): 1228–1241

The first question is: Why do we need such rather complex and computationally expensive approaches if other (e.g. empirical) scoring functions can do similar things? The main challenges thereby is that simple scoring functions often do not work very well for systems where they were not calibrated on (e.g. Knapp et al. 2009 (http://www.ncbi.nlm.nih.gov/pubmed/19194661)). The reasons for that are manifold. MD-based approaches can improve two major limitations of classical docking/scoring functions:

1) Proteins are not static. Ligand as well as receptor can undergo various slightly different configurations even for one binding site. Therefore the view of scoring one ligand configuration against one receptor configuration is not the whole picture. The first improvement is to consider a lot of different configurations for one position score of the ligand:

2) A more physics based scoring function can be more reliable than a simple and run-time efficient scoring function. On the basis of the MD simulations a variety of different terms can be deduced. These include:

– MM stands for Molecular Mechanics. It’s internal energy includes bond stretch, bend, and torsion. The electrostatic part is calculated using a Coulomb potential while the Van der Waals term is calculated using a Lennard-Jones potential.
– PB stands for Poisson−Boltzmann. It covers the polar solvation part i.e. the electrostatic free energy of solvation.
– SA stands for Surface Area. It covers the non-polar solvation part via a surface tension weighted solvent accessible surface area calculation.
– TS stands for the entropy loss of the system. This term is necessary because the non-polar solvation incorporates an estimate of the entropy changes implicitly but does not account for an entropy change upon receptor/ligand formation in vacuo. This term is calculated on the basis of a normal mode analysis.

If all these terms are calculated for each single frame of the MD simulations and those single values are averaged an estimate of the binding free energy of the complex can be obtained. However, this estimate might not represent the actual mean of the spatial distribution. Therefore at least 50 replica MD simulations are needed per investigated complex. In this aspect replica means an identically parameterized simultion of the same complex where only the inital forces are assinged randomly.

On the basis of the described MMPBSA-TS approach in combination with 50 replicas the authors achieve a reasonable correlation (0.63) for the 9 FDA-approved HIV-1
protease inhibitors with know experimental binding affinities. If the two largest complexes are excluded the correlation improves to an excellent value (0.93).

In a current study we are using the same methodology for peptide/MHC interactions. This system is completely different from the protease inhibitor study of Wright et al.: The ligands are peptides and the binding site is a groove consisting of two alpha-helices. The methods was applied as it is (without calibration or any kind of training). Prelimiary data still shows a high correlation with experimental values for the peptide/MHC system. This indicates that this MMPBSA approach can yield reliable predictions for very different systems without further modification.

In 1996 Gary Kasparov, the reigning world chess champion, played IBM’s Deep blue, a computer whose sole purpose was to play chess better than any human. Losing the first match, Gary sprung back swiftly defeating Deep Blue 4-2 over the remaining matches. However, his success was short lived. In a rematch with an updated Deep Blue the following year, the score was 3.5-2.5 to the computer. The media (and IBM) declared this as a pivotal moment in history, where a machine had proven itself better than humanities champion at a game deemed a highly intellectual pursuit. The outcry was that the age of machines had arrived. Was it true? Should humanity have surrendered to machine overloads at that moment? Obviously the answer is a large and resounding no. However, this competition allows for insightful comparison between the manner in which humans and computers play chess and think. By comparing the two, we learn the strengths and weaknesses of both parties from which we can make combined approaches that may exceed either.

Firstly, lets discuss the manner in which a computer “plays” chess. They simply search all possible configurations of moves that are available and pick the most optimal. However, things are not that simple. Consider only the opening sequence, there are 20 possible moves a player can make, so after only a single move by each player there is 400 possible chess positions. This count grows exponentially fast, after 5 moves by each player there is approximately 5 million combinations. For example, it was estimated that Deep Blue could analyse 2 million positions per second. However, since this is not nearly fast enough to examine all possible games from start to end in a reasonable time scale, computers cannot foresee lines of plays which are far in the distance. To overcome this, in the early game the computer will use a reference table developed by grandmasters that list both common openings and the assumed best manner to respond to them. Obviously, these are only assumed as optimal and have never been completely tested. In short, machines participate through a brute force, utilising their intricate ability to perform calculations at high speed to find the best move. However, the search is too large in the initial and end stages of a game to be completely thorough, a reference table is instead used to “inform” of the correct move at these times.

While a human can quite easily see that the following board leads to a draw, computers cannot draw the same conclusion without huge effort.

In contrast, human players use far more visual and spatial recognition alongside both memory and calculation to pick their moves. Like a computer, a player will analyse a portion of the moves available at any given moment. Though since a human cannot compare on computation speed to that of a computer, they cannot analyse nearly the same magnitude of moves. Hence, this subset of moves chosen for analysis must contain the most optimal move(s) to compete against the computer’s raw power. This is where the visual and spatial recognition abilities of humans come to bare. Firstly, a human can easily dissect the board into pieces worth considering and those to be ignored. For example, consider a possible move that would result in your queen being exposed and then taken. A human would conclude this as bad (normally) and discard further moves leading from such a play. A computer, however, would explore the resultant board state. One can see how this immediately and drastically reduces the required search. Another human ability is that a player will often be able to able to see sub-structures within a full set-up that are common in the game and hence can be processed in a known manner. In other words, the game is broken down into fragments which can be processed far easier and with less computation. Obviously, both of the above techniques rely on prior knowledge of chess to be useful, but they based upon our human ability to perceive both the substructure of the game and the overall picture with relative ease.

So how does all this chess talk relate to protein folding? In 2010, the Baker group and creators of the ROSETTA protein fold prediction program produced the protein folding game “Foldit”. In Foldit the general public could attempt to fold proteins for themselves and try to get closer to the native structure than the computer algorithms. Obviously, simplified in presentation to that of academic structural biology, it was hoped that the visual and spatial reasoning abilities of humans, the same ones that differentiated them from machines at chess, would prove useful in protein structure prediction. A key issue within ROSETTA drove this train of thought, the fact that is is relatively bad at exploring fully the confirmation space. Often, it will get stuck in the one general configuration and not explore the fold space fully. Furthermore, due to the size of configuration space, this is not easily overcome with simulated annealing due to the sheer scale of the problem. The ability of humans to view the overall picture meant that it should be easier for them to see other possible configurations. As end goals for Foldit, it was hoped that structures that proved unsolvable by current algorithms would be solved by humans and also that new techniques would emerge as “moves” employed by players to achieve high scores could be studied.

To make a comparison of the structures produced by Foldit players and ROSETTA viable, the underlying energy “scores” that judge a structure is the same between the programs. It is assumed, though is not always true, that the better the score the closer you are to the native fold. In addition,  Foldit players were also able to use a set of optimisation tools that were deterministic and would alter the backbone and side chains to the most optimal local configuration to the arrangement the player would make. This meant that players could focus predominantly on altering the overall structure of the protein rather than the fine detail, such as the position of sidec hains. To make the game as approachable as possible, technical terms were replaced by common analogues and visual cues where displayed to highlight poor scoring areas of the protein. For example, clashes between atoms are shown via large spiked red orbs, while the backbone is coloured from green to red depending on how well buried the hydrophobic residues on that segment are. To drive players, gamification elements were also included such as leader boards and rewarding “fireworks” as graphical effects.

To objectively compare the ability of the player base to that of the ROSETTA algorithm, they performed blind predictions on a set of 10 proteins whose structure were not in the public domain. This was run in a similar manner to CASP for those familiar with that set-up. The results exemplified the innate human ability of visual and spatial recognition. In 5 of the cases the playerbase performed significantly better than the ROSETTA program. In 3 of the cases they performed similar. And in the remaining 2 cases the ROSETTA algorithm performed better, though in both of these the model produced was still extremely far from the native structure. Looking through the cases individually, it was identified that the most crucial element used by players was that they were able to deal with large rearrangements that ROSETTA struggled to deal with, including register shifts and strand swapping. This highlights the ability of humans to view the overall picture and to persevere through “bad scoring patches” to reach a more optimal configuration.

Comparison of foldit player’s solutions (green) to ROSETTA’s solutions (red) and the native 2KPO protein structure (blue). The players correctly identified a strand swap needed to reach the native form, while this large reconfiguration was not seen by ROSETTA.

Since the release of the game and the accompanying paper in 2010, Foldit has received much praise in conveying the field of protein folding in an approachable manner to so many people. In addition, the player base has contributed to science as whole. In 2011 the player base successfully solved the structure of a M-PMV protein, a retrovirus whose structure was unobtainable via normal means. Then in 2012, by analysing the common set of moves employed by the player base, they collectively produced an algorithm that outperforms previously published fold prediction methods. Personally, I think of Foldit as a fun and relative intuitive game that introduces the core elements of the protein folding problem. As to its scientific merit, I’m unsure as to how much impact it will continue to have. As Saulo discussed last week, if infinite monkeys have infinite time then Shakespeare will be reproduced. Likewise, if enough people manipulate a protein structure, eventually the best structure will be found. Though who am I to judge, if people find the game fun, then there are far worse past-times one can have than trying to solve structures. As a finishing note I would be extremely interested in using Foldit to teach structural biology in the future, though feel it is overall too simple for a university setting.