Nimbus Therapeutics and computational drug discovery: A promising start with much data to plumb

Nimbus Therapeutics, which as far as I know is the only drug company based on a purely computational model of drug discovery (with all experiments outsourced), just handed over one of their key programs to Gilead for a good sum of money. This program was aimed at discovering inhibitors of the protein acetyl-CoA carboxylase (ACC) which is implicated in Non-Alcoholic Steatohepatitis (NASH) and went from start to finish in about 18 months.

To discover this inhibitor Nimbus brought all of Schrodinger's resources and computing capability (they essentially have unlimited access to Schrodinger's licenses and hardware) to bear on the problem. They (presumably) used techniques like virtual screening of millions of compounds, molecular dynamics, calculating unstable water molecules and the newest Schrodinger tool - free-energy perturbation (FEP) which in its ideal incarnation can allow you to rank order compounds by their free energy of binding to a protein target.

I think this is a very promising development for the applications of computation to drug discovery, but as a scientist in the field myself I am even more excited about the volume of useful data this effort must have generated. This is simply because, based on their model, it seems that every molecule that Nimbus prioritizes or discards necessarily goes through their computational pipeline: this would be rather unique. The corpus of data this process generates is presumably locked inside Nimbus's virtual vaults, but I think they and Schrodinger should release at least a high level abstraction of it to let everyone figure out what worked and what did not. At the very least it would transform Nimbus's success from an impressive black box to a comprehensible landscape of computational workflows.

There are several questions whose answers are always very valuable when it comes to assessing the success of any computational drug discovery protocol. One of the most important insights is to get an idea of the domain of applicability of particular techniques, and try to tease apart the general features of problems where they successfully seem to work. Also since any computational protocol is a model, one also wants to know how well it compares to other models, and preferably simpler ones. If a particular technique turns out to be one that's general, accurate, robust and consistently better than others, then we're in business.

Here are a few questions which apply not only to Nimbus but which I think can be asked of any project in which computational methods are thought to play a predominant role.

1. Did they have a crystal structure or did they build a homology model? If they built a model, what was the sequence identity with the template protein?
2. Was the protein flexible or was it fairly rigid? Were there missing loops or other parts and did they computationally build these loops? If loops were actually reconstructed, what was their length?
3. What was the success rate of their virtual screening campaign? How many top ranked molecules were actually made by the chemists?
4. How much did Schrodinger's WaterMap protocol help with improving binding affinity? Were the key displaceable or replaceable water molecules buried deep in hydrophobic cavities or were there also a few non-intuitive water molecules at the protein-solvent interface?
5. Did they test any of the false negatives to know if the accuracy was what they thought it was?
6. How well did the new FEP algorithm work in terms of rank ordering binding affinity? Did they compare this method with other simpler methods like MMGBSA?
7. In what way, if at all, did molecular dynamics help in the discovery of these inhibitors? How did MD compare to simpler techniques?
8. How well did methods to predict ADME work relative to methods to predict binding affinity? (The latter are usually far more accurate).

I think it's great that a purely computation-driven company like Nimbus can discover a lead for a tough target and an important disease in a reasonable amount of time. But it would be even better to distill general lessons from their success that we could apply to the discovery of drugs for other important diseases.

Derek Lowe to world: "Beware of von Neumann's elephants"

"With four parameters you can fit an elephant to a curve,
with five you can make him wiggle his trunk" - John von Neumann

That was one of many cogent messages delivered by In the Pipeline's Derek Lowe at a meeting organized today by the Boston Area Group Modeling and Informatics (BAGIM) group. The session was well attended - by my count there were sixty to seventy people. Most were probably modelers, with a healthy number of medicinal chemists and a few biologists thrown in. Below I offer a brief report on his talk, along with some personal commentary. Memory is a malleable and selective thing, so if someone who attended the talk wants to add their own observations in the comments section they are quite welcome.

The nice thing about Derek's talk was that it was really delivered from the other side of the fence, that of an accomplished and practicing medicinal chemist. Thus he wisely did not dwell too much on all the details that can go wrong in modeling: since the audience mainly consisted of modelers presumably they knew these already (No, we still can't model water well. Stop annoying us!). Instead he offered a more impressionistic and general perspective informed by experience.

Why von Neumann's elephants? Derek was referring to a great piece by Freeman Dyson (who I have had the privilege of having lunch with a few times) published in Nature a few years back in which Dyson reminisced about a meeting with Enrico Fermi in Chicago. Dyson had taken the Greyhound bus from Cornell to tell Fermi about his latest results concerning meson-proton scattering. Fermi took one look at Dyson's graph and basically demolished the thinking that had permeated Dyson and his students' research for several years. The problem, as Fermi pointed out, was that Dyson's premise was based on fitting a curve to some data using four parameters. But, quipped Fermi, you can get a better fit to the data if you add more parameters. Fermi quoted his friend, the great mathematician John von Neumann - "With four parameters you can fit an elephant to a curve; with five you can make him wiggle his trunk". The conversation lasted about ten minutes. Dyson took the next bus back to Ithaca.

And making elephants dance is indeed what modeling in drug discovery runs the risk of doing, especially when you keep on adding parameters to improve the fit (Ah, the pleasures of the Internet - turns out you can literally fit an elephant to a curve). This applies to all models, whether they deal with docking, molecular dynamics or cheminformatics. This problem of overfitting is well-recognized, but researchers don't always run the right tests to get rid of it. As Derek pointed out however, the problem is certainly not unique to drug discovery. He started out by describing the presence of "rare" events in finance related to currency fluctuations - these rare events happen often enough and their magnitude is devastating enough to cause major damage. Yet the models never captured them, and this failure was responsible at least in part for the financial collapse of 2008 (this is well-documented in Nassim Taleb's "The Black Swan").

Here are the charges brought against modelers by medicinal chemists: They cannot always predict but often can only retrodict, they cannot account for 'Black Swans' like dissimilar binding modes resulting from small changes in structure, they often equate computing power with accuracy or confidence in their models, they promise too much and underdeliver. These charges are true to varying degrees under different circumstances, but it's also true that most modelers worth their salt are quite aware of the caveats. As Derek pointed out with some examples however, modeling has over-promised for a long time: a particularly grating example was this paper from Angewandte Chemie from 1986 in which predicting the interaction between a protein and small molecule is regarded as a "spatial problem" (energy be damned). Then there's of course the (in)famous "Designing Drugs Without Chemicals" piece from Fortune magazine which I had highlighted on Twitter a while ago. These are all good examples of how modeling has promised riches and delivered rags. 

However I don't think modeling is really any different from the dreams spun from incomplete successes by structure-based design, high-throughput screening, combinatorial chemistry or efficiency metrics. Most of these ideas consist of undue excitement engendered by a few good successes and inflated dreams of what's sometimes called "physics envy" - the idea that your discipline holds the potential to become as accurate as atomic physics if only everyone adopted your viewpoint. My feeling is that because chemistry unlike physics is primarily an experimental discipline, chemists are generally more inherently skeptical of theory, even when it's doing no worse than experiment. In some sense the charge against modeling is a bit unfair since it's also worth remembering that unlike synthetic organic chemistry or even biochemistry which have had a hundred and fifty years to mature and hone themselves into (moderately) predictive sciences, the kind of modeling that we are talking about is only about three decades old. Scientific progress does not happen overnight.

The discussion of hype and failure brings us to another key part of Derek's talk, that of recognizing false patterns in the data and getting carried away by their apparent elegance, sophisticated computing power or just plain random occasional success. This part is really about human psychology rather than science so it's worth noting the names of psychologists like Daniel Kahneman and Michael Shermer who have explored human fallibility in this regard. Derek gave the example of the "Magic Tortilla" which was a half-baked tortilla in which a lady from New Mexico saw Jesus's image - the tortilla is now in a shrine in New Mexico. In this context I would strongly recommend Michael Shermer's book "The Believing Brain" in which he points out human beings' tendency to see false signals in a sea of noise - generally speaking we are much more prone to seeing false positives rather than false negatives, and this kind of Type I error makes sense when we think about how false positives would have saved us from menacing beasts on the prairie while false negatives would simply have weeded us out of the gene pool. As Derek quipped, the tendency to see signal in the data has basically been responsible for much scientific progress, but it can also contribute mightily to what Irving Langmuir called "pathological science". Modeling is no different - it's very easy to extrapolate from a limited number of results and assume that your success applies to data that is far more extensive and dissimilar to what your methods have been applied to.

There are also some important questions here about what exactly what it is that computational chemists should suggest to medicinal chemists, and much of that discussion arose in the Q&A session. One of the points that was raised was that modelers should stick their necks out and should make non-intuitive suggestions - for instance simply asking a medicinal chemist to put a fluorine or methyl group on an aromatic ring is not as useful because the medicinal chemist might have thought of that as a random perturbation anyway. However this suggestion is not as easy as it sounds since the largest changes are also often the most uncertain, so at the very least it's the modeler's responsibility to communicate this level of risk to his or her colleagues. As an aside, this reminds me of Karl Popper since one of the criteria Popper used for distinguishing a truly falsifiable and robust theory from the rest was its ability to stick its neck out and make bold predictions. As a modeler I also think it's really important to realize when to be coarse-grained and when to be fine-grained. Making very precise and limited suggestions when you don't have enough data and accuracy to be fine-grained is asking for trouble, so that's a decision you should be constantly making throughout the flow of a project. 

One questioner also asked if as a medicinal chemist Derek would make dead compounds to test a model - this is a question that is very important in my opinion since sometimes negative suggestions can provide the best interrogation of a model. I think the answer to this question is best given on a case-by-case basis: if the synthetic reaction is easy and can rapidly generate analogs then the medicinal chemist should be open to the suggestion. If the synthesis is really hard then there should be a damn good reason for doing it, often a reason that would lead to a crucial decision in the project. The same goes for positive suggestions too: a positive suggestion that may potentially lead to significant improvement should be seriously considered by the medicinal chemist, even if it may be more uncertain than one which is rock-solid but only leads to a marginal improvement.

Derek ended with some sensible exhortations for computational chemists: Don't create von Neumann's elephants by overfitting the data, don't talk about what hardware you used (effectively giving the impression that just because you used the latest GPU-based exacycle your results must be right: and really, the medicinal chemist doesn't care), don't see patterns in the data (magic tortillas) based on limited and spotty results.

Coupled with these caveats were more constructive suggestions: always communicate the domain of applicability/usability of your technique and the uncertainty inherent in the protocol, realize that most medicinal chemists are clueless about the gory details of your algorithm and therefore will take what you say at face value (and this also goes to show how it's your responsibility as modelers to properly educate medicinal chemists), admit failures when they exist and try to have the failures as visible as the successes (harder when management demands six sigma-sculpted results but still doable). And yes, it never, ever hurts for modelers to know something about organic chemistry, synthetic routes and the stability of molecules. I would also add the great importance of using statistics to analyze and judge your data.

All very good suggestions indeed. I end with only one: realize that you are all in this together and that you are all - whether modelers, medicinal chemists or biologists - imperfect minds trying to understand highly complex biological systems. In that understanding will be found the seeds of (occasional) success.

Added: It looks like there's going to be a YouTube video of the event up soon. Will link to it once I hear more.

The name's bond - reversible covalent bond.

Fifteen years ago most people would have laughed at you if you told them kinase inhibitors would become such a big deal: the received wisdom at that point was that anything that competed with ATP for one kinase would just indiscriminately hit other kinases. While that is generally true, we have found over the intervening decade that there is a wealth of detail - type II motifs, allosteric binding, relatively straightforward residue selectivity etc. - that can be tweaked to provide selectivity. In one fell swoop Gleevec upended the conventional wisdom.

A similar kind of thinking existed for covalent drugs a few years back, and I think that field too is going down the fortuitous road that kinase inhibitors took and defying the naysayers. This year especially has seen a bonanza of activity related to the discovery and fine-tuning of covalent inhibitors. The most striking and unexpected development in the field however has been the finding that one can find reversible covalent inhibitors that engage a protein target long enough to provide useful efficacy: the pessimistic thinking that had prevailed so far was mainly based on the potent off-target effects potentially arising from irreversible inhibitors. That thinking was justified...just as the thinking that kinase inhibitors would be non-selective was justified in 1998.

Reversible covalent kinase inhibitors in particular have seen a lot of interest and publications this year, and it's especially gratifying for me as a computational chemist to see that that the field is benefiting from both experimental and computational approaches. Jack Taunton at UCSF (who has started a company to exploit such kinds of inhibitors) has been a leader, but there have been others. Just recently there was a joint publication from the Taunton and Shoichet groups which used covalent docking techniques to prospectively discover JAK3 inhibitors. Taunton has also done some interesting work with Matt Jacobson from UCSF to use computational methods to tune covalent warhead reactivity - this work would be especially useful in tailoring reversible inhibitors to suit the target. A group at Pfizer did some similar work. Meanwhile the discovery of irreversible inhibitors for other important targets also isn't dead: earlier this year Nathanael Gray's group at Harvard published a report on covalent reversible inhibitors for CDK7, for instance.

It remains to be seen - it always remains to be seen - how many of these leads can survive the rigors of the clinic and how general the field becomes. But the approach in general certainly seems to be coming of age, and I will be watching its development with great interest. Quite a bit of promise there and also a real chance to defy some naysayers, which is what scientists especially relish.

Modular drug design software?

The latest issue of C&EN has an interesting article (unfortunately subscription only) about how quantum chemists are making code for standard protocols in quantum chemistry calculations available to each other as off-the-shelf modules. The movement has been driven by the realization that whenever someone develops a new quantum chemistry program he or she has to go through the tedious process of rewriting code for standardized algorithms like the Hartree-Fock method for calculation of potential energies of molecules. Why reinvent the wheel when you can simply buy it off the shelf in a centralized local tire shop?

I like this idea and I applaud the quantum chemists for having the generosity in sharing their code. But that left me wondering how soon it would be before something similar could happen in the world of computational drug design, or whether it would even be feasible.

The essence of methods like Hartree-Fock is that their highly iterative and standardized nature made them instantly amenable to computation. Your code for Hartree-Fock may be faster and cleaner than the other fellow's but the basic methodology which can be captured in a well-defined flowchart is not going to change. Contrast this with 'standard' drug design software protocols like docking, similarity searching and molecular dynamics calculations. 

Even though the objective is the same in every case, every practitioner uses his or her own favorite technique for their calculations; for instance docking can be physics-based or knowledge-based or it may depend on genetic algorithms. The sampling algorithms in MD may similarly be different in every case. Docking or MD are thus not as 'standardized' as say the Hartree-Fock method so it may be difficult to offer these protocols as standardized modules.

However I cannot see why it may not be possible to offer even more specialized components that are in fact standard for the wider use of the community. For instance, certain force fields - parameters and equations for calculation of structure and energetics - are pretty standard; the MMFF force field will have a certain set of components and the MM2 will have another. Similarly in a protocol like MD, the precise methods of sampling can be much more standard compared to the overall package. So in principle these methods could be packaged as standardized modules and offered to users.

The ideal situation for computational drug design would be an age where a variety of protocols ranging from quantum chemistry, docking and MD to homology modeling, gene and protein sequence comparison tools and toxicity and PK prediction algorithms would be available for any user to patch together, rearrange and deploy in the solution of his or her particular problem. 

Going even further, we could envisage an age where the tools of systems and computational biology are thoroughly ingrained in the drug discovery process so that one can now add standard systems tools to the toolbox; for instance, in such an age, not only would I be able to snatch standard docking protocols from a website but I would also be able to combine them with some kind of a wiring diagram of the protein which I am trying to target linked to its partners, so that I know exactly which partner hubs I should additionally dock my drug against in order to maximize its efficacy and minimize its toxicity. And who knows, maybe I can even get to a stage where I can download some kind of a minimalist but accurate model of an entire cell and observe how my drug will qualitatively perturb its network of organelles and signaling pathways.

For now this sounds like a pipe dream, although I suspect that the cultural barriers to sharing algorithms with commercial potential may be much harder to overcome than the scientific hurdles to actually incorporating systems biology in drug discovery and making the whole process modular. That's where the siren song of these socialist quantum chemists would be particularly relevant.

Free Energy Perturbation (FEP) methods in drug discovery: Or, Waiting for Godot

For interested folks in the Boston area it's worth taking a look at this workshop on Free Energy Perturbation (FEP) methods in drug design at Vertex from May 19-21. The list of speakers and topics is quite impressive, and this is about as much of a state-of-the-art discussion on the topic as you can expect to find in the area.

If computational drug discovery were a series of plays, then FEP might well be the "Waiting for Godot" candidate among them. In fact I would say that FEP is a textbook case of an idea that, if it truly works, can truly transform the early stages of drug discovery. What medicinal chemist would not want to know the absolute free energy of binding of his molecules to a protein so that he can actually rank known and unknown compounds in order of priority? And what medicinal chemist would not want to know exactly what she should make next?

But that's what medicinal chemists have expected from modelers ever since modeling started to be applied realistically to drug discovery, and I think it's accurate to say that it's good they haven't held their breath. FEP methods have always looked very promising because they aim to be very rigorous, bringing the whole machinery of statistical mechanics to bear on a protein-ligand system. The basic goal is "simple": you calculate the individual  energies of the protein and the drug - in explicit water - and then you calculate the energy of the bound system. The difference is the free energy of binding. Problem solved.

Except, not really. Predicting relative free energies is still a major challenge, and predicting absolute free energies is asking for a lot. The major obstacle to the application of these methods for decades was considered to be the lack of enough computing power. But if you really thought that was the major obstacle then you were still a considerable way off. Even now there seems to be a belief that given enough computing power and simulation time we can accurately calculate the free energy of binding between a drug and a target. But that's assuming that the fundamental underlying methodology is accurate, which is a big assumption.

The "fundamental underlying methodology" in this case mainly refers to two factors: the quality of the force field which you use to calculate the energy of the various components and the sampling algorithm which you use to simulate their motions and exhaustively explore their conformations. The force fields can overemphasize electrostatic interactions and can neglect polarization, and the sampling algorithms can fail to overcome large energy barriers. Thus both these components are imperfectly known and applied in most cases, which means that no amount of simulation time or computing power is then going to be sufficient. It's a bit like the Polish army fighting the Wehrmacht in September 1939; simply having a very large number of horses or engaging them in the fight for enough time is not going to help you win against tanks and Stukas.

These problems have all been well recognized however; in fact the two most general issues in any modeling technique are sampling and energy calculation. So parts of this month's workshop are aimed exactly at dissecting the factors that can help us understand and improve sampling and scoring.

The end goal of any applied modeling technique of course is how good it is at prediction. Not surprisingly, progress on this front using FEP has been rather thin. In fact FEP is the quintessential example of a technique whose successes have been anecdotal. Even retrospective examples, while impressive, are not copious. One of the problems is that FEP works only when you are trying to predict the impact of very tiny changes in structure on ligand affinity; for instance the impact of changing a methyl group on a benzene ring to a hydroxyl group. The trouble is that the method doesn't work even for these minor changes across the board; there are projects where a CH3--->OH change will give you quantitative agreement with experiment and there are cases where it will result in error bars large enough to drive a car through them. 

But anecdotes, while not being data, are still quite valuable in telling us what may or may not work. Computing power may not solve all our problems but it has certainly given us the opportunity to examine a large number of cases and try to abstract general rules or best practices for drug discovery. We may not be able to claim consistent successes for FEP right now, but it would help quite a lot even if we know what kinds of systems it works best for. And that, to me, is as good an outcome as we could expect at this time.

"Designing drugs without chemicals"

As usual Derek beat me to highlighting this rather alarming picture from an October 5, 1981 issue of Fortune magazine that I posted on Twitter yesterday. The image is from an article about computer-aided design and it looks both like a major communications failure (chemical-free drugs, anyone?) as well as a massive dollop of hype about computer-aided drug design. In fact the article has been cited by drug designers themselves as an example of the famous hype curve, with 1981 representing the peak of inflated expectations.

It's intriguing to consider both how we are still considering pretty much the same questions about computational drug design that we were in 1981 as well as how much progress we have made on various fronts since then. I posted an extended comment about both these aspects of the issue on Derek's blog so I am just going to copy it it below. Bottom line: Many of the fundamental problems are still the same and are unsolved on the general level. However, there has been enough understanding and progress to expect solutions to a wide variety of specific problems in the near future. My own attitude is one of cautious optimism, which in drug discovery is usually the best you can have...

For anyone who wants the exact reference it's the Fortune magazine issue from Oct 5, 1981. The reference is widely considered to be both the time when CADD came to the attention of the masses, as well as a classic lesson in hype. The article itself is really odd since most of it is about computer-aided design in the industrial, construction and aeronautical fields; these are fields where the tools have actually worked exceedingly well. The part about drug design was almost a throwaway with almost no explanation in the text.

Another way to look at the issue is to consider a presentation by Peter Goodford in 1989 (cited in a highly readable perspective by John van Drie (J Comput Aided Mol Des (2007) 21:591–601) in which he laid out the major problems in molecular modeling - things like including water, building homology models, calculating conformational changes, predicting solubility, predicting x-ray conformations etc. What's interesting is that - aside from homology modeling and x-ray conformations - we are struggling with the exact same problems today as we were in the 80s. 

That doesn't mean we haven't made any progress though. Far from it in fact. Even though many of these problems are still unsolved on a general level, the number of successful specific examples is on the rise so at some point we should be able to derive a few general principles. In addition we have made a huge amount of progress in understanding the issues, dissecting the various operational factors and in building up a solid database of results. Fields like homology modeling have actually seen very significant advances, although that's as much because of the rise of the PDB which was enabled through crystallography as accurate sequence comparison and threading algorithms. We are also now aware of the level of validation that our results need to have for everyone to take them seriously. Journals are implementing new standards for reproducibility and knowledge of the right statistical validation techniques is becoming more widespread; as Feynman warned us, hopefully this will stop us from fooling ourselves.

As you mention however, the disproportionate growth of hardware and processing power relative to our understanding of the basic physics of drug-protein interaction has led to an illusion of understanding and control. For instance it's quite true that no amount of simulation time and smart algorithms will help us if the underlying force fields are inaccurate and ill-tested. Thus you can beat every motion out of a protein until the cows come home and you still might not get accurate binding energies. That being said, we also have to realize that every method's success needs to be judged in terms of a particular context and scale. For instance an MD simulation on a GPCR might get some of the conformational details of specific residues wrong but may still help us rationalize large-scale motions that can be compared with experimental parameters. Some of the more unproductive criticism in the field has come from people who have the wrong expectations from a particular method to begin with.

Personally I am quite optimistic with the progress we have made. Computational drug design has actually followed the classic Gartner Hype curve, and it's only in the 2000s that we have reached that cherished plateau of realistic expectations. The hope is that at the very least this plateau will have a small but consistent positive slope.

Drugable.com "ranks billions of drug interactions"? Hold your horses.

Now here's a study that should make most seasoned molecular modelers cringe. Nature News reports on an effort by website Drugable.com that docked 600,000 compounds to 7,000 protein targets and predicted which ones would show activity against these targets based on docking scores:

Predicting how untested compounds will interact with proteins in the body, as Drugable attempts to do, is more challenging. In setting up the website, Cardozo’s group selected about 600,000 molecules from PubChem and the European Bioinformatics Institute’s ChEMBL, which together catalogue millions of publicly available compounds. The group evaluated how strongly these molecules would bind to 7,000 structural ‘pockets’ on human proteins also described in the databases. Computing giant Google awarded the researchers the equivalent of more than 100 million hours of processor time on its supercomputers for the mammoth effort.

But mammoth computing resources do not translate to carefully constructed protocols or correct predictions. In its current incarnation, docking is best for finding the binding pose, that is, the orientation of a drug bound into a protein pocket. Ranking compounds is far more difficult, and predicting absolute binding affinities is a very distant, currently unachievable third goal.

Anyone who has tried to run a hit to lead or lead optimization project based on docking scores would know how riddled with problems and qualifications any prediction based on these highly subjective numbers is. For starters, every modeling program gives you its own docking scores. Absolute values of these numbers (which ideally should reflect the free energy of binding but which seldom do) are almost always useless. If you are dealing with a congeneric series of molecules and are fairly confident about the binding orientation (usually confirmed by x-ray crystallography or some other technique) then maybe you could get some help from the scores in ranking the compounds, but even then mostly in terms of trends rather than quantitative differences.

Unfortunately the news piece says nothing about what method was used to generate the poses, whether there was any clustering or whether only the top pose was considered, what the false positive rate was, and most importantly, whether there was any experimental verification whatever of the ranking. The website is also not helpful in this regard. It also does not tell us if the protein structures used for docking were well-resolved or refined or whether they were homology models. In the absence of all this information the ranking of the compounds is tenuous at best and useless at worst and as it stands the study sounds little better than throwing darts in the dark and hoping some of them will stick. Ranking often fails even for similar compounds, so how well (or badly) it would work for 600,000 diverse compounds bound to 7,000 diverse protein targets is anyone's guess.

The report also compares the study to a similar activity prediction study by Brian Shoichet in which drug similarity was used to predict activity against unexpected targets. But that was a very different kettle of fish; it was a compound similarity - not docking - study so it did not have to deal with the complexities of error-ridden protein crystal structures or homology models, it verified a lot of the predictions using carefully constructed assays, and even then it gave a hit rate which did not exceed about 50%.

Either the Drugable.com study itself has failed to validate its predictions or the news report is woefully incomplete. Maybe I am wrong and in fact the study has laid the careful groundwork and validation that is necessary for trusting docking. As it stands however, the purpose of the report mainly seems to be to highlight the fact that Google generously donated 100 million hours of its computing power to the docking. This heightened, throw-technology-at-it sense of wonder and optimism is exactly what the field does not need. I would be the first one to welcome reliable predictions of drug-protein affinity based on wholesale docking of compounds to targets, but I don't think this work achieves that goal at all.

The right questions about modeling in drug discovery

There's been a fair bit of discussion recently about the role of modeling in drug design, and as a modeler I have found it encouraging to see people recognize the roles that modeling to varying extents plays in the whole "rational" drug design paradigm. But sometimes I also see a question being asked that's not quite how it should be pitched.

This question is usually along the lines of "Would you (as a medicinal chemist) or would you not make a compound that a modeler recommends?". Firstly, the way the question is set up ignores the statistical process evident in all kinds of drug design and not just modeling. You never recommend a single compound or even five of them but usually a series, often built around a common scaffold. Now suppose the chemist makes fifty compounds and finds only one of them to be active. Does this mean the modeling was useful? It depends. If the one hit that you get is a novel scaffold, or a compound with better properties, or one without that pesky unstable ester, or even a compound similar to ones that you have made that nonetheless picks up some hitherto unexplored interactions with the protein, then modeling would not have been in vain in spite of what is technically a low hit rate. The purpose of modeling is to provide interesting ideas, not rock-solid predictions. Plus you always have to ask the other question: "How likely is it that you would have come up with this structure without the aid of modeling or structural information"? As with almost everything else in drug discovery, the value of a particular approach is not absolute and you have to judge it in the context of alternatives.

The other problem I have is that it's frequently just not possible to objectively quantify the value of modeling or any specific approach for that matter in drug design. That's because anyone who has contributed to the design of a successful drug or even a potent ligand knows how haphazard and multifaceted the process is, with contributions flying in from every direction in many guises. Every project that I have contributed to which has resulted in more potent ligands has involved a patchwork process with an essential back-and-forth exchange with the chemists. Typically I will recommend a particular compound based on some modeling metric and then the synthetic chemist will modify parts of it based on synthetic accessibility, cost of building blocks, ease of purification etc. I might then remodel it and preserve or extend the synthetic modifications.

But it can also be the other way round. The synthetic chemist will recommend a scaffold and I will suggest tweaks based on what I see on the computer, which will lead to another round of suggestions and counter-suggestions. The creative process goes both ways and the point is that it's hard to put a number on the individual contributions that go into the making of the final product. Sometimes there may be a case where one individual recommends that magic fluorine which makes all the difference, but this is usually the exception rather than the rule and more importantly, it's often not possible to trace that suggestion to a strictly rational thought process. Finally, it's key to realize that my suggestions themselves will very often be based as much on my knowledge of physical and organic chemistry as on any specific modeling technique. Thus, even internally it's not always possible for me to separate modeling-based suggestions from other kinds. On a related note, the frequent necessity of this interactive process is why I tend to find the "modeling vs medicinal chemistry" turf wars a little irksome (although they sometimes do help in fleshing out the details); the fact of the matter is that both kinds of scientists are ultimately an important part of successful drug discovery projects involving any kind of model-building.

Now this rather unmethodical approach makes it difficult to pin down the role of any specific technology in the drug discovery process, and this usually makes me skeptical every time someone writes a paper trying to assess "The contribution of HTS/virtual screening/fragment-based drug design in the discovery of major drugs during the last decade". But that's how drug discovery is, a hodgepodge of recommendations and tweaks made by a variety of rational and irrational human beings in an environment where any kind of idea from any corner is welcome.

Identifying new drug targets in zebrafish

Classically, when very little about molecular biology and protein structure was known, one of the best methods to discover drugs was to try to guess physiological effects by looking at drug similarity. The rationale was simple; drug A has this chemical structure and lowers blood pressure. Drug B which is being used for a completely different indication also has a similar chemical structure. Maybe we can use drug B for lowering blood pressure too?

In the absence of rational approaches this technique (which these days is called "repurposing") could be surprisingly fruitful. However, as molecular biology, crystallography and structure-based drug design took off in the 80s, rational drug discovery became much more focused on protein structure and drug developers began trying to guess drug function by looking at target similarity in terms of sequences and binding pockets.

But the relative lack of dividends from structure-based drug design (which nonetheless continues to be important) has led to a re-examination of the old paradigm. One of the most interesting advances to come out of this thinking was pioneered by Brian Shoichet and Bryan Roth's groups (at UCSF and UNC-Chapel Hill) a few years back. Their rationale was simple too; look at drug similarity, albeit using modern computational methods and metrics, and try to cross-correlate drug targets using these similarity metrics. Similar drugs should hit each other's targets. The method seems somewhat crude but has worked surprisingly well. It was even listed as one of Wired magazine's top ten scientific breakthroughs of 2009.

In a recent publication the authors take the method a step further and apply it to phenotypic screening which has emerged as an increasingly popular method in the pharmaceutical industry. Phenotypic screening is attractive because it bypasses having knowledge of the exact molecular target; basically you just inject different potential drugs in a test system and look at the results which are usually quantified by a phenotypic response such as a change in membrane potential, cell differentiation or even something as general as locomotion. Once a particular compound has elicited a particular response, we can start the arduous task of finding out what it's doing at a molecular level. Not surprisingly, several proteins can be responsible for a general response like locomotion and detecting all of them is non-trivial to say the least. It is for this rather involved exercise that the authors provide a remarkably simple potential (partial) solution.

The current study looks at phenotypic screening on the zebrafish, a favorite model organism of biologists. 14,000 molecule were screened to elicit a photomotor response in the zebrafish embryos. Out of these, about 700 were deemed active. To find out the targets for these molecules, the authors interrogated their chemical "similarity" against a large group of compounds for which targets are known. Importantly, the authors use a statistical technique to calculate an expectation value (E value) which indicates whether the similarity arises by chance alone. A lower E value means a higher likelihood of statistical significance. One of the most remarkable things in these studies is that the metric for similarity is a simple, computationally cheap 2D metric called a fingerprint which looks at the presence or absence of certain functional groups in the molecules. That such a metric can work at all is remarkable because we know that an accurate estimation of similarity should ideally include the 3D conformation that the drug presents to the protein target.

Nonetheless, 284 molecules predicted to be active on 404 targets were picked based on their low E values. Out of these, 20 molecules were especially interesting because they were seen to have novel drug-target associations not seen before. When the authors tested these molecules against the relevant targets, 11 among them had activities ranging from low nanomolar to 10 µM. For a computational method this hit rate is quite impressive, although a more realistic measure of the hit rate would have come from testing all 284 compounds. The activity of the molecules was validated by running comparisons with molecules that were known to elicit the same response from the predicted targets. Confirmation also came from competition experiments with antagonists. Some of the unexpected targets predicted include the beta-adrenergic receptor, dopamine receptors and potassium ion channels.

I find these studies very encouraging. For one thing, the computational method can potentially save a huge amount of time needed to experimentally uncover the target for every active compound. As mentioned above, it's also remarkable that a very simple metric of 2D similarity yields useful results. The success rate is impressive; however, an even lower rate would still have been worth the modest effort if it resulted in new drug-target relationships (which are far more useful than new chemically similar ligands for the same target). However I do think it would have been very interesting to look at the failures. An obvious source of failure comes from using the wrong measure of similarity; at least some compounds are failing presumably because their 2D similarity does not translate to the 3D similarity in conformation required for binding to the protein target. In addition there could be protein flexibility which would result in very different binding for supposedly similar compounds. Medicinal chemists are well aware of "activity cliffs" where small differences in chemical structure lead to great differences in activity. These cliffs could also lead to lack of binding to a predicted target.

Nevertheless, in an age when drug discovery is only getting harder and phenotypic screening seems to be an increasingly important technique, these computational approaches promise to be a useful complement. Gratifyingly, the authors have developed a website where the algorithm is available for free. The technique has also spawned a startup.

Image source

The future of computation in drug discovery

Computational chemistry as an independent discipline has its roots in theoretical chemistry, itself an outgrowth of the revolutions in quantum mechanics in the 1920s and 30s. Theoretical and quantum chemistry advanced rapidly in the postwar era and led to many protocols for calculating molecular and electronic properties which became amenable to algorithmic implementation once computers came on the scene. Rapid growth in software and hardware in the 80s and 90s led to the transformation of theoretical chemistry into computational chemistry and to the availability of standardized, relatively easy to use computer programs like GAUSSIAN. By the end of the first decade of the new century, the field had advanced to a stage where key properties of simple molecular systems such as energies, dipole moments and stable geometries could be calculated in many cases from first principles with an accuracy matching experiment. Developments in computational chemistry were recognized by the Nobel Prize for chemistry awarded in 1998 to John Pople and Walter Kohn.

In parallel with these theoretical advances, another thread started developing in the 80s which attempted something much more ambitious- to apply the principles of theoretical and computational chemistry to complex systems like proteins and other biological macromolecules and to study their interactions with drugs. The practitioners of this paradigm wisely realized that it would be futile to calculate properties of such complex systems from first principles, thus leading to the initiation of parametrized approaches in which properties would be "pre-fit" to experiment rather than calculated ab initio. Typically there would be an extensive set of experimental data (the training set) which would be used to parametrize algorithms which would then be applied to unknown systems (the test set). The adoption of this approach led to molecular mechanics and molecular dynamics - both grounded in classical physics- and to quantitative structure activity relationships (QSAR) which sought to correlate molecular descriptors of various kinds to biological activity. The first productive approach to docking a small molecule in a protein active site in its lowest energy configuration was refined by Irwin "Tack" Kuntz at UCSF. And beginning in the 70s, Corwin Hansch at Pomona College had already made remarkable forays into QSAR.

These methods gradually started to be applied to actual drug discovery in the pharmaceutical industry. Yet it was easy to see that the field was getting far ahead of itself and in fact even today it suffers from the same challenges that plagued it thirty years back. Firstly, nobody had solved the twin cardinal problems of modeling protein-ligand interactions. The first one was conformational sampling wherein you had to exhaustively search the conformation space of a ligand or protein. The second one was energetic ranking wherein you had to rank these structures, either in their isolated form or in the context of their interactions with a protein. Both of these problems remain the central problems of computation as applied to drug discovery. In the context of QSAR, spurious correlations based on complex combinations of descriptors can easily befuddle its practitioners and create an illusion of causation. Furthermore, there have been various long-standing problems such as the transferability of parameters from a known training set to an unknown test set, the calculation of solvation energies for even the simplest molecules and the estimation of entropies. And finally, it's all too easy to forget the sheer complexity of the protein systems we are trying to address which display a stunning variety of behaviors, from large conformational changes to allosteric binding to complicated changes in ionization states and interactions with water. The bottom line is that in many cases we just don't understand the system which we are trying to model well enough.

Not surprisingly, a young field still plagued with multiple problems could be relied upon as no more than a guide when it came to solving practical problems in drug design. Yet the discipline saw unfortunate failures in PR as it was periodically hyped. Even in the 80s there were murmurs about designing drugs using computers alone. Part of the hype unfortunately came from the practitioners themselves who were less than cautious about announcing the strengths and limitations of their approaches. The consequence was that although there continued to be significant advances in both computing power and algorithms, many in the drug discovery community looked at the discipline with a jaundiced eye.

Yet the significance of the problems that the field is trying to address means that it will continue to be promising. What's its future and what would be the most productive direction in which it could be steered? An interesting set of thoughts is offered in a set of articles published in the Journal of Computer-Aided molecular design. The articles are written by experienced practitioners in the field and offer a variety of opinions, critiques and analyses which should be read by all those interested in the future of modeling in the life sciences.

Jurgen Bajorath from the University of Bonn along with his fellow modelers from Novartis laments the fact that studies in the field have not aspired to a high standard of validation, presentation and reproducibility. This is an important point. No scientific field can advance if there is wide variation in the presentation of the quality of its results. When it comes to modeling in drug discovery, the proper use of statistics and well-defined metrics has been highly subjective, leading to great difficulty in separating the wheat from the chaff and honestly assessing the impact of specific techniques. Rigorous statistical validation in particular has been virtually non-existent, with the highly suspect correlation coefficients being the most refined weapon of choice for many scientists in the field. An important step in emphasizing the virtue of objective statistical methods in modeling was taken by Anthony Nicholls of OpenEye Software who in a series of important articles laid out the statistical standards and sensible metrics that any well-validated molecular modeling study should aspire to. I suspect that these articles will go down in the annals of the field as key documents.

In addition, as MIT physics professor Walter Lewin is fond of constantly emphasizing in his popular lectures, any measurement you make without knowledge of its uncertainty is meaningless. It is remarkable that in a field as fraught with complexity as modeling, there has been a rather insouciant indifference to the estimation of error and uncertainty. Modelers egregiously quote numbers involving protein-ligand energies, dipole moments and other properties to four or six figures of significance when ideally those numbers are suspect even to one decimal point. Part of the problem has simply been an insufficient grounding in statistics. Tying every number to its estimated error margin (if it can be estimated at all) will not only give experimentalists and other modelers an accurate feel for the validity of the analysis and the ensuing improvement of methods but will also keep semi-naive interpreters from being overly impressed by the numbers. Whether it's finance or pharmaceutical modeling, it's always a bad idea to get swayed by figures.

Then there's the whole issue, as the modelers from Novartis emphasize, of spreading the love. The past few years have seen the emergence of several rigorously constructed datasets carefully designed to test and benchmark different modeling algorithms. The problem is that these datasets have been most often validated in an industry that's famous for its secrecy. Until the pharmaceutical industry makes at least some efforts to divulge the results of its studies, a true assessment of the value of modeling methods will always come in fits and starts. I have been recently reading Michael Nielsen's eye-opening book on open science, and it's startling to realize the gains in advancement of knowledge that can result from sharing of problems, solutions and ideas. If modeling is to advance and practically contribute to drug discovery, it's imperative for industry - historically the most valuable generator of any kind of data in drug discovery - to open its vaults and allow scientists to use its wisdom to perfect fruitful techniques and discard unproductive ones.

Perhaps the most interesting article on the future of modeling in drug discovery comes from Arizona State University's Gerald Maggiora who writes about a topic close to my heart - the limitations of reductionist science in computational drug design. Remember the physics joke about the physicist suggesting a viable solution to the dairy farmer, but one that applies only to spherical cows in a vacuum? Maggiora describes a similar situation in computational chemistry which has focused on rather unrealistic and high simplified representations of the real world. Much of modeling of protein and ligands focuses on single proteins and single ligands in an implicitly represented solvent. Reality is of course far different, with highly crowded cells constantly buffeting thousands of small molecules, proteins, lipids, metals and water in an unending dance of chemistry and electricity. Molecules interact constantly with multiple proteins, proteins interact with others proteins, changes in local environments significantly impact structure and function, water adopts different guises depending on its location and controlled chaos generally reigns everywhere. As we move from simple molecules to societies of molecules, unexpected emergent properties may kick in which may be invisible at the single molecule level. We haven't even started to tackle this level of complexity in our models and it's a wonder they work at all, but it's clear that any attempts to fruitfully apply computational science to the behavior of molecules in the real world can only be possible when our models include these multiple levels of complexity. As Maggiora says, part of the solution can come from interfacing computational chemistry with other branches of science including biology and engineering. Maggiora's version of computation in drug discovery thus involves seamlessly integrating computational chemistry, biology, fluid dynamics and other disciplines into an integrated model-building paradigm which bravely crisscrosses problems at all levels of molecular and cellular complexity, unifying techniques from any field it chooses to address the most important problems in drug design and discovery.

Maggiora does not discuss all the details of effort that would lead to this computational utopia, but I see at least two tantalizing signs around. Until now we have focused on the thermodynamic aspects of molecular design (more specifically, being able to calculate free energies of binding) and while this will remain a challenge in the foreseeable future, it's quite clear than any merging of traditional computational chemistry with higher-order phenomena will involve accurately modeling the kinetics of interactions between proteins and small ligands and between proteins themselves. Modeling kinetics can pave the way to understanding fluxes between various components in biochemical networks, thus leading directly to the interfacing of computational chemistry with network biology. Such advances could provide essential understanding of the non-linear mechanisms that control the production, reactions, down-regulation and turnover of key proteins, a process that itself is at the heart of most higher-order physiological processes.

How can such an understanding come about? Earlier this year, scientists at D. E. Shaw Research performed a study in which they ran a completely blind molecular dynamics simulation of a drug molecule placed outside a kinase protein. By running the simulation for a sufficiently long time, the drug could efficiently sample all configurations and finally lock into its binding site where it stayed put. The model system was simple and we don't know yet how generalizable such a simulation is, but one tantalizing possibility such simulations offer is to be able to correlate calculated protein-drug kinetic binding data to experimental on and off binding rates. The D. E. Shaw method certainly won't be the only computational method to estimate such data, but it offers a glimpse of a future in which accurate computed kinetic data will be available to estimate higher-order fluxes between biomolecules. Such approaches promise revealing insights

The other auspicious development concerns the modeling of crowded environments. As mentioned above, no simulation of molecular structure and function can be considered truly realistic until it takes into account all the myriad molecular partners that populate a cell's interior. Until now our methods have been too primitive to take these crowded conditions into account. But recent proof-of-principle studies have provided the first glimpse into how protein folding under crowded conditions can be predicted. Hopefully at some point, improved computing power will afford us the capability to simulate multiple small and large molecules in proximity to each other. It's only when we can do this can we claim to have taken the first steps in moving from one level of complexity to another.

The future abounds with these possibilities. Whatever the hype, promises and current status of modeling in drug discovery, one thing is clear: the problems are tough, many of them have been well-defined for a long time and all of them without exception are important. In the sense of the significance of the problems themselves, computational chemistry remains a field pregnant with possibilities and enormous score. Calculating the free energy of binding of an arbitrary small molecule to an arbitrary protein still remains a foundational goal. Honing in on the correct, minimal set of molecular descriptors that would allow is to predict the biological activity of any newly synthesized molecule is another. So is being able to simulate protein folding, or calculate the interaction of a drug with multiple protein targets, or model what happens to the concentration of a protein in a cell when a drug is introduced. All these problems really present different sides of the same general challenge of modeling biomolecular systems over an expansive set of levels of complexity. And all of them are tough, involved, multifaceted and essentially interdisciplinary problems that will keep the midnight oil burning in the labs of experimental and computational scientists alike.

But as Jack Kennedy would say, that's precisely why we need to tackle them, not because they are easy but because they are difficult. That's what we have done. That's what we will continue to do.

References:

Bajorath, J. (2011). Computational chemistry in pharmaceutical research: at the crossroads Journal of Computer-Aided Molecular Design DOI: 10.1007/s10822-011-9488-z

Martin, E., Ertl, P., Hunt, P., Duca, J., & Lewis, R. (2011). Gazing into the crystal ball; the future of computer-aided drug design Journal of Computer-Aided Molecular Design DOI: 10.1007/s10822-011-9487-0

Maggiora, G. (2011). Is there a future for computational chemistry in drug research? Journal of Computer-Aided Molecular Design DOI: 10.1007/s10822-011-9493-2


Image source

Scientific challenges in drug discovery (Part 1): We are still infants

In the face of a flagging pharmaceutical industry and depleted pipeline of new drugs, you will inevitably find someone who tries to put what he or she thinks is a positive spin on the whole depressing situation. That person will say something to the effect that it's probably ok if we don't have more drugs since we already have very good drugs for many disorders and reasonably good ones for most others, so we shouldn't worry too much if we don't continue to come up with miracle breakthroughs.

The recent deaths of Steve Jobs (56) and Ralph Steinman (63, who tragically died only two days before he was awarded the Nobel Prize) should put any such pseudocheerful beliefs to rest. These were people who had access to the best treatment money can buy. Jobs could have afforded any treatment anywhere in the world and Steinman was a professor at one of the world's leading medical research institutes. Yet both were mowed down in the prime of their lives and careers by pancreatic cancer, a disease for which our current pharmaceutical tools are as blunt and primitive as stones and twigs were for fighting wars. Cancer is the great equalizer, unsparing and all-encompassing enough to compete with death itself as one of the great facts of life.

Pancreatic cancer however is just one ailment which we are light years from conquering. Include many other cancers, Alzheimer's disease and ALS, antibiotic-resistant infections and diseases like malaria that are still rampant in developing countries, and it's quite clear that nobody should have to make a case for a thriving pharmaceutical and biotech industry that needs every inch of financial and societal support that we can muster. That we still have to do so is of course a tragic reflection on the state of our myopic short-term vision and misguided financial goals.

But even if we disregard this massive mass of humanity that still very much needs the fruits of pharmaceutical research, there are still enough unsolved problems and challenges on a purely scientific level that should keep researchers - both in academia and industry- hungry for more ideas and solutions. These challenges are longstanding and we can be assured that, whatever happens to drug research in the new century, they will always stick around waiting for us to tackle them. The scope and diversity of these problems widely vary, but you would be hard-pressed to find a researcher who is satisfied with their current status. In this post and the next, I hope to briefly dwell on some problems that we will surely need to solve if we want to radically improve the chances of discovering a potent organic molecule and taking it to the market as a drug.

1. We are still far from being able to predict toxic side effects for any drug

As those in the know are well aware, most drugs fail in the clinic because of unacceptable adverse effects and it's obvious that being able to predict toxic side effects will work wonders for medical science. The truth however is that it's rather shameful that, even after achieving so much sophistication in our knowledge of biology and chemistry, exuberant advertisements for drugs have to be tempered with a laundry list of side-effects (proclaimed sotto voce of course) that outnumber the benefits by at least five to one. It's rather shameful that all these years of unprecedented funding for cancer have resulted in hundreds of clinical compounds, even the best of which evidence rather horrible side effects like nausea, immune system debilitation and loss of fertility. No wonder the bar for the approval of cancer therapeutics is so low; any higher and almost nothing from our current list would make the cut. If this is really the best we can do to fight cancer, then the war on cancer has not even started. A Martian who has conquered most diseases in his own land would call us primitive savages if he saw our present arsenal of anticancer drugs.

As far as prediction goes, we are barely able to predict the effects of unknown drugs with a modest degree of success, and that too using statistical models based on empirical data. Calculating side-effects from first principles is pretty much impossible right now. We will have to develop exceedingly clever and complicated model systems approximating whole-organism physiology to do any kind of respectable toxicity prediction. Of course we don't have to predict side effects in order to minimize them, but any such kind of optimization is currently largely a matter of trial and error. Black box approaches can only get us so far and at some point we will need a healthy understanding of the interplay of various proteins and systems in our body that contributes to toxicity. Systems biology could help us pin down these interactions, but ultimately there would be no substitute for a detailed understanding of biology at a molecular level. For now doing all this is a dream, and toxicity prediction remains one of the great challenges of drug discovery that should keep researchers busy for decades.

2. We still cannot accurately predict the free energy of binding of an arbitrary small molecule to an arbitrary protein from first principles

From clinical-level challenges to the most basic of problems, this one essentially being one in physical chemistry. It is a commentary on both the limitations and the opportunities inherent in computational modeling of biochemical systems that we have barely started to scratch the surface of being able to understand, let alone predict, the myriad factors that go into dictating the free energy of binding of a small molecule to a protein. Part of the problem is just the sheer number of variables involved, from the conformational complexity of ligand and protein to the behavior of solvent and solvation of the various parts (vida infra) to the plethora of energetic interactions that any two molecules have with each other.

The other problem involves a battle against nature herself. Because of the exponential dependence of the binding constant on the free energy of binding, errors as small as 1 kcal/mol can significantly under or overpredict small molecule binding. This problem itself subsumes a difficulty that is inherent to any kind of system susceptible to perturbation by minute changes- being able to predict small differences between large numbers; whether the numbers are economic statistics, variables influencing the weather or in this case, free energies of binding.

How far have we come in being able to predicting these energies? The good news is that we now understand a lot more of the physical chemistry of protein ligand binding than before. We have a rather good understanding of the statistical mechanics framework involved in calculating free energies. Using model systems, we have been able to do a reasonably good job of classifying the various factors - hydrogen bonding, the hydrophobic effect, entropy, electrostatic interactions- that encapsulate these energies. The other good news is that phenomenal improvements in hardware and software continue to allow us to push the boundaries of accuracy.

The bad news is that listing the contributing factors is like listing the genes in a human genome without understanding their function. For instance, listing "entropy" as a factor and being able to calculate the entropy of the protein are two quite different things, with the latter still being largely beyond our capabilities. As with other chemical systems, being able to predict protein ligand systems means predicting the precise contribution (even as a rough percentage) of each one these factors to a single resultant number. And even if we were able to calculate these variables in theory, implementing this theoretical framework inevitably involves patching all the factors together in a model. Using the model inevitably involves parameterizing it. And parametrization is a well-known fickle beast, subject to statistical and systematic errors. The worse news is that when it comes to complex systems subject to so many causes, model building is always going to be our best bet. So we can only hope for the time when we have essentially unlimited computing power and have been able to either parametrize our model to kingdom come (without overparameterizing it) or have been able to implement every minute part of the physics of protein-ligand binding in our model. While I would love to have the latter, I would be more than happy to settle for the former if it ever happens.

3. We understand very little about the behavior of solvents, especially water

This is a significant factor in its own right to separate it from the second point. My optimism about the prospects of computational modeling of proteins and molecules in general took an exponential leap downwards when I schooled myself about solvent behavior and realized that we lack the resources to calculate accurate solvation energies even for simple organic molecules, let alone proteins. It's one of life's enduring mysteries; that which is most familiar succumbs the least to our efforts in understanding it- in this case that elusive entity would be water.

It's become something of a cliche to say that water is an anomalous solvent essential to life and yet we understand so little of its depths, but the gloomy implication of this for drug discovery is that we have always been struggling to incorporate this essential factor in our understanding of biomolecular systems. The problem is again essentially one in physical chemistry and has withstood decades of computational and experimental assault. The earliest attempts to incorporate water into molecular simulation simply involved...ignoring it. Not exactly ignoring it, but replacing its discrete tapestry with a continuous electric field that duplicated its dielectric constant. This effort gave short shrift to both the role of discrete water molecules in mediating molecular interactions as well as the shimmering dynamic hydrogen bond network that is the soul of water's behavior. I am hoping that three decades from now we will look back and laugh at the barbaric simplicity of this approximation, but we have to be excused for attempting this feat in the absence of massive computing power (which could handle thousands of discrete water molecules). And we must confess we did it for a very simple reason - it worked. Even today, these so-called "implicit solvation" models can give us surprisingly satisfactory results for many systems (partly due to cancellation of errors, but let's not go there).

But the implicit solvent consigned water to become the proverbial elephant in the room and researchers, especially those developing molecular dynamics (MD) techniques, strove to replace this continuum with "real" water molecules. But in the world of modeling, even "real" water molecules correspond to models with calculated point charges, dipole moments mimicking the polarizability of the water molecule and so on and these models largely ignore the special behavior that particular water molecules in a system may evidence. Nonetheless, these modeled waters are widely used today in molecular dynamics simulations and massive computing power can now allow us to routinely handle thousands of such entities.

But the elephant has not left the room. Revealing experimental approaches (mainly spectroscopic) in the last few years have painted a very different picture for water in the bulk compared to water surrounding a biomolecule like a protein. Not surprisingly, water surrounding a protein is less mobile and more viscous than that in the bulk. This allows the protein to project something like a "ghost field", a watery extension of its shape and form into the surrounding solvent. This proxy effect can cause other molecules in the vicinity to respond, although its precise effects are still not known. This also brings us to another related big elephant which we will discuss in the next post- the understanding of molecular behavior in the cell as opposed to in a test tube or on the computer. For now it suffices to say that water in cells behaves very differently from water in dilute solution. Its electrostatics, hydrophobicity, composition and hydrogen bonding network are very different. And we have barely started scratching the surface of this "crowding" that is the hallmark of cellular systems.

A related unsolved problem is the behavior of discrete water molecules in protein active sites. We already know from our knowledge of enzymatic catalysis that water can play an essential role in enzyme reactions. In addition, new evidence indicates that hydrophobic interactions can lead to "dewetting" or the sudden expulsion of water between surfaces. Crystal structures usually don't have enough of resolution to clearly pinpoint the locations of bound water molecules. But the real problem is that we still don't know how to accurately predict the thermodynamic and other features of such trapped water. There have been some notable recent advances which attempt to calculate the thermodynamics of enclosed discrete water but these are just getting started. Ultimately the problem boils down to taking theories describing average bulk behavior and using them to calculate specific instances, a problem well-known to statisticians.

It should go without saying that we cannot aim for a realistic understanding of the physics and chemistry of small molecule-protein interactions before we can understand the behavior of water and predict solvation energies of simple organic molecules. But the truth of the matter is that modeling of biomolecular systems has proceeded largely in the absence of such detailed understanding but in the presence of ever so many clever tricks, ad hoc parameterization efforts, models shored up by experimental data and a "Shut up and just use it" attitude. And all this may be justified if our end goal is to find new drugs. But we haven't discovered very many, and while this cannot mainly be blamed on not being able to model water, there is little doubt that a better understanding of solvation will help not just computational modeling but also other drug discovery related activities like formulation, dosing and storage of drug molecules.

In the next part we will look at some other fundamental scientific challenges in drug discovery. Perhaps the people who are disillusioned by the current pharmaceutical bloodbath can take perverse pleasure in the fact that, even when the last scientist has been laid off by the last pharmaceutical organization, these problems will still be standing and needing our attention. When it comes to satisfactorily solving these problems, we are still infants.

Next post: Predicting crystal structures, understanding intracellular interactions, deconstructing the drug-protein-gene connection and more. Stay tuned.