The human problems with molecular modeling

Molecular modeling and computational chemistry are the neglected stepchildren of pharmaceutical and biotech research. In almost every company, whether large or small, these disciplines are considered "support" disciplines, peripheral to the main line of research and never at the core. At the core instead are synthetic chemistry, biology and pharmacology, with ancillary fields like formulations and process chemistry becoming increasingly important as the path to a drug progresses.

In this post I will explore two contentions:

1. Unless its technical and human problems are addressed, molecular modeling and simulation will remain peripheral instead of core fields in drug discovery.

2. The overriding problem with molecular modeling is the lack of a good fit between tools and problems. If this problem is addressed, molecular modeling stands a real chance of moving from the periphery to, if not the very core, at least close to the core of drug discovery.

There are two kinds of challenges with molecular modeling that practitioners have known for a long time - technical and human. The technical problems are well known; although great progress has been made, we still can't model the details of biochemical systems very accurately, and even key aspects of these systems like protein motion, water molecules and - in case of approaches like machine learning - lack of adequate benchmarks and datasets continue to thwart the field. 

However, in this piece I will focus on the human problems and explore potential ways of mitigating them. My main contention is that the reason modeling often works so poorly in a pharmaceutical setting is because the incentives of modelers and other scientists are fundamentally misaligned. 

In a nutshell, a modeler has two primary objectives - to make predictions about active, druglike molecules and to validate the models they are using. But that second part is actually a prerequisite for the first - without proper validation, a modeler cannot know if the exact problem space they are applying their models to is actually a valid application of their techniques. For proper validation, two things are necessary:

1. That the synthetic chemist actually makes the molecules they are suggesting.

2. That the synthetic chemistry does not make molecules which they aren't suggesting.

In reality the synthetic chemist who takes up the modelers' suggestions has little to no interest in model validation. As anyone who has done modeling knows, when a modeler suggests ten compounds to a synthetic chemist, the synthetic chemist would typically pick 2 or 5 out of those 10. In addition, the synthetic chemist might pick 5 other compounds which the modeler never recommended. The modeler typically also has no control and authority over ordering compounds themselves.

The end result of this patchwork implementation of the modeler's predictions is that they never know whether their model really worked. Negative data is especially a problem, since synthetic chemists are almost never going to make molecules that the modeler thinks will be inactive. You are therefore left with a scenario in which neither the synthetic chemist nor the modeler knows or is satisfied with the utility of the models. No wonder the modeler is relegated in the back of the room during project discussions.

There is another fundamental problem which the modeler faces, a problem which is actually more broadly applicable to drug discovery scientists. In one sense, not just modeling but all of drug discovery including devices, assays, reagents and models can be considered as a glorious application of tools. Tools only work if they are suited to the problem. If a practitioners thinks the tool will be unsuited, they need to be able to say so and decline using the tool. Unfortunately, incentive structures in organizations are rarely set up for employees to say "no". Hearing this is often regarded as an admission of defeat or an unwillingness to help out. This is a big mistake. Modelers in particular should always be rewarded if they decline to use modeling and can gives good reasons for doing so. As it stands, because they are expected to be "useful", most modelers end up indiscriminately using their tools on problems, no matter what the quality of the data or the probability of success is. This means that quite often they are simply using the wrong tool for the wrong problem. Add to this the aforementioned unwillingness of synthetic chemists to validate the models, and it's little surprise that modeling so often fails to have an impact and is relegated to the periphery.

How does one address this issue? In my opinion, the issue can be mitigated to a significant extent if modelers know something about the system they are modeling and the synthesis which will yield the molecules they are predicting. If a modeler can give sound reasons based on assays and synthesis - perhaps the protein construct they are using for docking is different from one in the assay, perhaps the benchmarks are inadequate or perhaps the compounds they are suggesting won't be amenable to easy synthesis because of a weird ring system - other scientists are more likely to both take their suggestions more seriously as well as respect their unwillingness to use modeling for a particular problem. The overriding philosophy that a modeler utilizes should be captured not in the question, "What's the best modeling tool for this problem?" but "Is modeling the right tool for this problem?". So, the first thing a modeler should know is whether modeling would even work, but if not, he or she will go a long way in gaining the respect of their organization if they can say at least a few intelligent things about alternative experimental approaches or the experimental data. There is no excuse for a computational chemist to not be a chemist in the first place.

More significantly, my opinion is that this mismatch will not be addressed until modelers themselves are in the driver's seat, until they can ensure that their predictions are tested in their entirety. Unfortunately there's little control modelers have over testing their models; much of it simply depends on how much the synthetic chemists trust the modelers, a relationship driven as much by personality and experience as modeling success. Even today, modelers can't usually simply order their compounds for synthesis from internal or external teams.

Fortunately there are two very significant recent developments that promise modelers a degree of control and validation that is unprecedented. One is the availability of cheap CROs like WuXi and Enamine which can make many of the compounds that are predicted by modeling. These CROs have driven the cost down so significantly that even negative predictions can now importantly be tested. In general, the big advantage of external CROs relative to internal chemists is that you can dictate what the external CROs should and shouldn't make - they won't make compounds which you don't recommend and they will make every compound that you do; the whims of personal relationships won't make a difference in a fee-for-service structure.

More tantalizingly, there have been a few success stories now of fully computationally-driven pipelines, most notably Nimbus and Morphic Therapeutic and, more recently, Silicon Therapeutics. When I say "fully computationally driven" I don't mean that synthetic chemists don't have any input - the inaccuracy of computational techniques precludes fully automated molecule selection from a model - what I mean is that every compound is a modeled compound. In these organizations the relationship between modeling and other disciplines is reversed, computation is front and center - at the core - and it's synthetic chemistry and biology in the form of CROs that are at the periphery. These organizations can ensure that every single prediction made by modelers is tested and made, or conversely, that no molecule that is made and tested fails to go through the computational pipeline. At the very least, you can then keep a detailed bookkeeping record of how designed molecules perform and therefore validate the models; at best, as some of these organizations showed, you can discover viable druglike leads and development candidates.

Computational chemistry and modeling have come a long way, but they have a long way to go both in terms of technical and organizational challenges. Even if the technical challenges are solved, the human challenges are significant and will hobble the influence computation has on drug discovery. Unless incentive structures are aligned the fields will continue to have poor impact and be at the periphery. The only way for them to progress is for computation to be in the driver's seat and for computational chemists to be as informed as possible. Fortunately with the advent of the commodification of synthesis and the increased funding and interest in computationally driven drug pipelines, it seems there may be a chance for us to find out how well these techniques work after all.

Image source

Will CADD ever become relevant to drug discovery?

Let’s face some hard facts first. The good news is that in the last twenty years or so, computer-aided drug design (CADD or computational chemistry) has become a relatively standard part of drug discovery, and even in organizations not formally employing CADD scientists, some form of computation – sometimes as simple as property calculation or visualization – is used.


The bad news is that, as it is largely practiced now, CADD is not considered a core part of the drug discovery process. Period. Rather, it’s considered as a supporting part. This situation has only marginally improved in twenty years. Biology and synthetic chemistry are still the core driving disciplines of drug discovery and will remain so in the foreseeable future. If there is a marketed drug that benefits extravagantly from CADD, it comes along once every decade or so at best (good luck finding another HIV protease, for instance). As integral as CADD scientists consider themselves and as impressive as they think the rotating pictures on their screen are, the fact that their peers consider what they do as being marginally relevant to the big picture is a bitter pill that needs to be swallowed.


Why is CADD not considered a core part of drug discovery, even though it is now part of every organization’s drug discovery portfolio in one form or another? There are two reasons, one related to fact and the other related to perception. The fact is that drug discovery and biology are still very much experimental disciplines, and you simply can’t compute your way to a drug; in this sense, CADD simply is not good enough yet. The perception of CADD is that unlike synthetic chemistry and biology which generate products (compounds in bottles, assays, animal models etc.), CADD generates ideas. In that sense, experimental scientists often have the same view of CADD scientists that Ernest Rutherford had of theorists: "They play games with their symbols while we discover the mysteries of the universe."

Now one of the advantages of CADD is that the ideas can be useful because of their generality (eg. “Put a hydrophobic group here.”), but it’s this very generality that often counts against CADD. Synthetic chemists in particular want to see specific suggestions, and while plenty of them appreciate general guidance, it doesn’t quite have the same impact. The fact that CADD scientists also often make suggestions that are considered obvious – filling hydrophobic pockets is a standard suggestion – does even less to enamor the discipline to other scientists.


This rather undervalued reputation of CADD impacts both the discipline and the scientists. The market for CADD scientists is usually a niche market, with low demand and low supply, and typically a small organization will have only one or two CADD scientists. But the supportive role of CADD means that when there are layoffs, a CADD scientist is far more likely to go than a cell biologist or assay developer. Because CADD scientists are not abundant, they also may not have direct reports. All this means that CADD scientists may not have the ear of senior management. They thus rarely occupy senior managerial positions like CSO, VP of drug discovery or CTO; if you take a random sample of top management at both small and large drug discovery organizations, you will very rarely find scientists with a background in CADD in these positions.


The supporting role of CADD combined with the lack of core influence and understanding CADD scientists have means that they often have to fight uphill battles to make their voice heard. While this is a good character-building experience, it doesn’t necessarily contribute to career progression or a growing influence on the part of the field. So how can CADD make a bigger impact both on the facts of drug discovery and the perception?


At least part of the answer – as unseemly as it seems to a lot of drug hunters today – does involve large datasets and machine learning. I am not saying that ML or some form of AI will have the kind of immediate, hype-heavy, transformational impact that often seems all-too-apparent through Silicon Valley sunglasses. But I am saying that the impact of ML and AI on drug discovery will inexorably increase as time passes, and anyone who simply chooses the option of dismissing these technologies will be left behind.


There are some important reasons why I believe this. The most important reason perhaps is that traditional CADD, as it pertains to protein structure and physics-based algorithms, has fallen far short of its promise. Part of the problem was the hype in the late 80s and early 90s, but a more realistic problem is that the data is often not good enough and even when it’s good it can impact a very limited part of early drug discovery. The biggest problem is that, even thirty years after a leading CADD scientist stated the challenges inherent in using CADD to predict binding affinity (protein flexibility, water behavior, crowding effects) we are still struggling with the same issues. Ignorance of the basic physics of protein-ligand interactions, especially as it pertains to making reliable, general predictions, is still rife.


I am not saying that physics-based approaches haven’t worked or improved at all over the last two decades – some applications like shape-based similarity searches and pose prediction – have certainly showed impressive progress, but the fact remains that we are still swimming in a sea of ignorance, either ignorance borne out of basic scientific understanding, or ignorance borne out of engineering our current understanding into viable approaches. Equally important, even when we understand these factors, they will be applicable to a very narrow part of drug development, namely improving binding affinity between a single small molecule and (usually) a single protein. At best CADD as we know it will design good ligands, not drugs.


Approaches based on large datasets in contrast are agnostic about the physics of protein-ligand binding. In principle, they can take a bunch of well-curated data points about interaction energies and predict what a new interaction will be like without ever explicitly modeling a single hydrogen bond. They don’t understand causation but can provide actionable correlations. But the second and truly novel advantage of machine learning approaches in my opinion is that, unlike traditional CADD which applies to a very narrow part of drug discovery even on its best day, ML can in theory encompass multiple aspects of drug discovery by simply adding more columns to the Excel file. Thus, if one has multiple kinds of data – in vitro and in vivo binding affinity, mutations, metabolic data, solubility, clearance etc. – traditional CADD can pretty much take advantage of only one or two kinds of data, but ML can consider all the datasets at once. Thus, machine learning provides pretty much the only way to solve the multiparameter optimization problem that is at the heart of drug discovery.


Now, as is well known, the real problem with data is not the physics but the quantity, quality and curation of the datasets (including negative data). But these are without a doubt getting better and more integrated across various phases of drug discovery every day. Whether it’s structures in the PDB, synthetic reactions for making drugs or patient data in the TCGA (The Cancer Genome Atlas), the improvement of data quantity and quality is inexorable. And as this quantity and quality improves, so would the utility of this data for machine learning. This is a certainty. The naysayers therefore are fighting against a current which is only going to get stronger, even if it doesn’t quite turn into the tsunami that its practitioners say will engulf us very soon.


Are we there yet? The simple answer is no. And we probably won’t be in terms of measurable impact for some time, especially since we have to hack our way through so much hype to get to the nugget of causal or even correlative truth. The domain of applicability of even impressive techniques like deep learning is still limited (for instance it works best on images), and the data problem is still a big one. But what I am saying is that unless CADD embraces machine learning and AI – irrespective of the hype – CADD scientists will always occupy a marginal perch in the grand stadium of drug development and CADD scientists will increasingly be cast on the sidelines, both in terms of career and discipline progress. This will especially be true as both protein and drug classes expand out into space (unstructured proteins, macrocycles etc.) that is even more challenging for traditional CADD than usual. 


Sadly, as with many things these days, the mention of AI or ML in CADD continues to evoke two extreme reactions: the first one which is more common is to say that it’s going to solve everything and is the Next Big Thing, but the second one is sadly also common, the one which says that it’s all hype and good old thinking will always be better. But good old thinking will always be around. The data is getting better and the machine learning is getting better; these are pretty much certainties which we ignore at our own peril. The bottom line is that unless CADD starts including machine learning and the related paraphernalia of techniques that either accurately or misleadingly fall under the rubric of “AI” in its very definition, its role in drug discovery and development threatens to dwindle to a point of vanishing relevance. This will do an enormous disservice both to CADD and drug discovery in general. 

Every year when the Nobel Prize in chemistry is announced, there's some bellyaching when the work leading to the prize turns out to be as much related to biology or physics as to chemistry. "But is it chemistry?", is the collective outcry. However, a more refined understanding of the matter over the years leads us to realize that chemistry has as much absorbed biology or physics as they have absorbed chemistry. Chemistry's definition has been changing ever since it emerged from the shroud of alchemy. Chemical biology, chemical physics and chemical engineering are as much chemistry as they are biology, physics and engineering. The same thing needs to happen to CADD in my opinion. By embracing computer science or automation or AI, by hiring ML engineers or data scientists in their midst, CADD is not losing its identity. It's showing that its identity is much bigger than anyone thought. That it is vast and contains multitudes.

The same and not the same: Carboxylic acids and tetrazoles

Here's a useful comparative study from Donna Huryn, Carlo Ballatore and others at Penn which looks at a problem encountered by almost every medicinal chemist in their careers: finding the right isosteres for certain functional groups. I certainly have faced this challenge in my work, and this work illustrates that the solutions for addressing the challenge can be more subtle than we think.

For non-medicinal chemists, an isostere is a set of atoms that can be swapped with another set of atoms in a molecule because it possesses physicochemical properties that are similar to the parent. This swap can be done for a variety of reasons: most commonly to improve physical properties like cell permeability. Carboxylic acids pose special problems for permeability due to their charge (since the cell membrane is hydrophobic and the charge renders them hydrophilic), so finding uncharged or other suitable charged replacements for them has been a routine consideration in drug design.

The current study looks at 35 different isosteric derivatives of phenylpropionic acids. It measures common properties like permeability (in PAMPA), logD, pKa and plasma protein binding (PPB) which can influence permeability. It also calculates the logD and pKa for comparison with the experimental values.

What the authors find is that by and large there are several acid substitutes like sulfonamides, hydroxamic acids and oxadiazoles which are superior to the parent acid in terms of permeability, and by and large this trend seems to correlate well with their logD and pKa values.

I say "by and large" above because there is one glaring exception which sticks out and which is perhaps not obvious: tetrazoles. Tetrazoles are very common replacements for acids; intuitively one would think that locking up the atoms in a ring and adding in a carbon would improve their ability to get past a membrane. But as the study finds, tetrazoles have permeabilities that are generally worse than those of the acids, sometimes by as much as 1.5 or 2 log units. And this is perhaps not too surprising if you think a bit more about their structure: While you may be locking up atoms in a ring, you are also adding two more polar atoms. In addition the tetrazole NH group is as ionizable as the COOH group of an acid. The paper says that these effects lead to a bigger desolvation penalty for the tetrazole, and that seems right; the tetrazole just seems to love water more, and a quantum mechanical study that studies the details of its solvation pattern relative to the acid might be interesting. Tetrazoles also seem to present more plasma protein binding.

There are some other interesting subtleties noted in the paper that are worth remembering. For instance the calculated logD and pKa values generally correlated well with the experimental values, except for two systems (oxadiazole-thiones and 1,3 dicarbonyls). They also noted that bigger desolvation penalties aren't always correlated with lower logD values. It's another slightly counterintuitive observation which I have seen myself in a kinase inhibitor study before.

Overall these are nice lessons to remember. Medicinal chemistry is full of useful general rules, but exceptions such as these show that the field is far from explored and will always stay interesting; there will always be enough subtleties to keep us occupied. More generally this comparison illustrates one of my favorite principles in chemistry: similarity can be both useful and deceptive.

Nobelist John Pople on using theories and models the right way

John Pople was a towering figure in theoretical and computational chemistry. He contributed to several aspects of the field, meticulously consolidated those aspects into rigorous computer algorithms for the masses and received a Nobel for his efforts. The Gaussian set of programs that he pioneered is now a mainstay of almost every lab in the world which does any kind of molecular calculation at the quantum level.

On the website of Gaussian is a tribute to Pople from his former student (and president of Gaussian) Michael Frisch. Of particular interest are Pople’s views on the role of theories and models in chemistry which make for interesting contemplation, not just for chemists but really for anyone who uses theory and modeling.

• Theorists should compute what is measured, not just what is easy to calculate
• Theorists should study systems people care about, not just what is easy or inexpensive to study.

Both these points are well taken as long as one understands that it’s often important to perform calculations on ‘easy’ model systems to benchmark techniques and software (think of spherical cows…). However it’s also a key aspect of modeling that’s often lost on people who simulate more complex systems. For instance the thrust of a lot of protein-small molecule modeling is in determining or rationalizing the binding affinity between the protein and the small molecule. This is an important goal, but it’s often quite insufficient for understanding the ‘true’ binding affinity between the two components in the highly complex milieu of a cell, where other proteins, ions, cofactors and water jostle for attention with the small molecule. Thus, while modelers should indeed try to optimize the affinity of their small molecules for proteins, they should also try to understand how these calculations might translate to a more biological context.

• Models should be calibrated carefully and the results presented with scrupulous honesty about their weaknesses as well as their strengths.

This is another aspect of theory and modeling that’s often lost in the din of communicating results that seem to make sense. The importance of training sets in validating models on known systems is well-understood, although even in this case the right kind of statistics isn’t always applied to get a real sense of the model’s behavior. But one of the simpler problems with training sets is that they are often incomplete and miss essential features that are rampant among the real world’s test sets (more pithily, all real cows as far as we know are non-spherical). This is where Pople’s point about presenting the strengths and weaknesses of models applies: if you are unsure how similar the test case is to the training set, let the experimentalists know about this limitation. Pople’s admonition also speaks to the more general one about always communicating the degree of confidence in a model to the experimentalists. Often even a crude assessment of this degree can help prioritize which experiments should be done and which ones should be best assessed against cost and implementation.

• One should recognize the strengths as well as the weaknesses of other people's models and learn from them.

Here we are talking about plagiarism in the best sense of the tradition. It is key to be able to compare different methods and borrow from their strengths. But comparing methods is also important for another, more elemental reason: without proper comparison you might often be misled into thinking that your method actually works, and more importantly that it works because of a chain of causality embedded in the technique. But if a simpler method works as well as your technique, then perhaps your technique worked not because of but in spite of the causal chain that appears so logical to you. A case in point is the whole field of molecular dynamics: Ant Nicholls from OpenEye has made the argument that you can’t really trust MD as a robust and ‘real’ technique if simpler methods are giving you the answer (often faster).

• If a model is worth implementing in software, it should be implemented in a way which is both efficient and easy to use. There is no point in creating models which are not useful to other chemists.

This should be an obvious point but it isn’t always one. One of the resounding truths in the entire field of modeling and simulation is that the best techniques are the ones which are readily accessible to other scientists and provided to them cheaply or free of cost. Gaussian itself is a good example – even today a single user license is offered for about $1500. Providing user-friendly graphical interfaces seems like a trivial corollary of this principle, but it can make a world of difference for non-specialists. The Schrodinger suite is an especially good example of user-friendly GUIs. Conversely, software for which a premium was charged died a slow death, simply because very few people could use, validate and improve it.

There do seem to be some exceptions to this rule. For instance the protein-modeling program Rosetta is still rather expensive for general industrial use. More importantly, Rosetta seems to be particularly user-unfriendly and is of greatest use to descendants of David Baker’s laboratory at the University of Washington. However the program has still seen some very notable successes, largely because the Baker tribe counts hundreds of accomplished people who are still very actively developing and using the program.

Notwithstanding such exceptions though, it seems almost inevitable in the age of open source software that only the easiest to use and cheapest programs will be widespread and successful, with lesser competitors culled in a ruthlessly Darwinian process.

Image link

If computational recipes were like food recipes...

...the menu at Boston's elegant L'Espalier might look something like this.

Maine beef tenderloin: Carrot tagliatelle al ragù, gorgonzola mornay, sunny side up quail egg, crispy shallots

10 ns molecular dynamics run: Replica exchange, orthorhombic box, embedded sodium ions, 1 fs time step, Nose-Hoover thermostat. Happy water molecules. Smooth trajectory.

West Coast King salmon, lightly smoked, with vanilla and cardamom: faux hazelnut gnocchi and pumpkin juice

New England Induced-fit Docking protocol: Braised loop resampling, 5A residue movement, multicore run with threading. Optional polarized force field (add 10 tokens). Error bars.

Honey glazed duck for two with roasted pumpkin, chestnut, foie jus, smoked chocolate puree and toasted nuts and seeds (add 20)

Multi-ligand conformer search: Seared 10 kcal/mol energy window. Glazed harmonic potential. Monte-Carlo dihedral drive (add 20 tokens). Experimental infusion.

Licorice glazed Hudson Valley foie gras with black trumpet mushrooms, hay ash roasted banana, black sesame and almond-rice milk.

Similarity search: Shape-based scaffold hopper, optionally paired with electrostatics. Statistically validated (add 50 tokens). 

There's a reason they're called computational "recipes".

The golden age of computational materials science gives me a disturbing feeling of déjà vu

Graphene, a wonder material which was made by scientists using a version of Scotch tape (Image: Wikipedia)

I was a mere toddler in the early 1980s when they announced the “golden age of computational drug design”. Now I may have been a toddler, but I often hear stories about the impending golden age from misty-eyed veterans in the field. A cover story in Fortune magazine (which I can never seem to find online) announced that pharmaceutical scientists were now designing drugs on computers. The idea was that once you feed in the parameters for how a drug behaves in the human body, the computer would simply spit out the answer. The only roadblock was computing power limited by hardware and software advances. Give it enough time, the article seemed to indicate, and the white-coat clad laboratory scientist might be a historical curiosity. The future looked rosy and full of promise.

Fast forward to the twilight days of 2013. We are still awaiting the golden age of computational drug design. The preponderance of drug discovery and design is still enabled by white coat-clad laboratory scientists. Now let’s be clear about one thing: the computational side of the field has seen enormous advances since the 1980s and it continues to thrive. There will almost undoubtedly be a time when its contributions to drug design would be seen as substantial. Most drugs perform their magic in living systems by binding to specific proteins, and computational drug design is now competent enough so that it can predict with a fair degree of accuracy, the structure and orientation of a drug molecule bound in a protein’s deep binding pocket. Computational scientists can now suggest useful modifications to a drug’s structure which laboratory chemists can make to improve multiple properties including solubility, diffusivity across cell membranes, activity inside cells and ability to avoid getting chewed up by enzymes in the body. You would be hard pressed to find a drug design project where computational modeling does not play at least a modest role. The awarding of this year’s Nobel Prize in chemistry to computational chemists is only one indication of how far the field has advanced.

And yet it seems that computational drug designers are facing exactly the same basic challenges they faced in the 80s. They have certainly made progress in understanding these challenges, but robust prediction is still a thing of the future. The most significant questions they are dealing with are the same ones they dealt with in the 80s: How do you account for water in a protein-drug system? How do you calculate entropies? How do you predict the folded structure of a protein? How do you calculate the different structures a drug molecule adopts in the aqueous milieu of the body? How do you modify a drug compound so that cells – which have evolved to resist the intrusion of foreign molecules - don’t toss it right out? How do you predict the absolute value of the binding energy between drug and protein? And scientists are grappling with these questions in spite of tremendous, orders-of -magnitude improvements in software and hardware.

I say all this because a very similar cover story about computational materials design in this month's Scientific American evokes disturbing feelings of déjà vu in me. The article is written by a pair of scientists who enthusiastically talk about a project whose goal is to tabulate calculated properties of materials for every conceivable application: from lightweight alloys in cars to new materials for solar cells to thermoelectric materials that would convert dissipated heat into electricity. The authors are confident that we are now approaching a golden age of computational materials design where high-throughput prediction of materials properties will allow us to at least speed up the making of novel materials.

We can now use a century of progress in physics and computing to move beyond the Edisonian process (of trial and error). The exponential growth of computer-processing power, combined with work done in the 1960s and 1970s by Walter Kohn and the late John Pople, who developed simplified but accurate solutions to the equations of quantum mechanics, has made it possible to design new materials from scratch using supercomputers and first-principle physics. The technique is called high-throughput computational materials design, and the idea is simple: use supercomputers to virtually study hundreds or thousands of chemical compounds at a time, quickly and efficiently looking for the best building blocks for a new material, be it a battery electrode, a metal alloy or a new type of semiconductor.

It certainly sounds optimistic. However the article seems big on possibilities and short on substance and shortcomings. This is probably because it occupies only three pages in the magazine; I think it deserved far more space, especially for a cover article. As it stands the piece appears more pollyannaish than grounded in cautious optimism.

I applaud the efforts to build a database of computed materials properties but I am far more pessimistic about how well this knowledge can be used in designing new materials in the near future. I am not a materials scientist, but I think some of the problems the computational end of the discipline faces are similar to those faced by any computational chemist. As the article notes, the principal tools used for materials design are quantum mechanics-based chemistry methods developed mainly by John Pople and Walter Kohn in the 1970s, a discovery that got the duo the 1998 chemistry Nobel Prize. Since then these methods have been coded into dozens of efficient, user-friendly computer programs. Yet these methods – based as they are on first principles – are notoriously slow. Even with heavy computing power it can take several days to do a detailed quantum mechanical calculation on an atomic lattice. With materials involving hundreds of atoms and extended frameworks it would take much longer.

I am especially not convinced that the methods would allow the kind of fast, high-throughput calculations that would substitute for experimental trial and error. One reason why I feel pessimistic is because of the great difficulty of predicting crystal structures. Here’s the problem: the properties of a material depend on the geometric arrangement of its atoms in a well-defined crystal lattice. Depending on the conditions (temperature, pressure, solvent) a material can crystallize in dozens of possible structures, which makes the exercise of assuming “a” crystal structure futile. What’s worse for computer calculations is that the energy differences between these structures may be tiny, within the error limits of many theoretical techniques.

On the other hand, the wrong crystal structure could give us the wrong properties. The challenge for any kind of computational prediction method is therefore two-fold: firstly, it has to predict the various possible crystal forms that a given material can adopt (and sometimes this number can run into the hundreds). Secondly, even if it can achieve this listing, it now has to rank these crystal forms in order of energy and predict which would be the most stable one. Since the energy differences between the various forms are tiny, this would be a steep challenge even for detailed calculation on a single material. Factoring conditions of temperature, pressure and solvent into the calculation would make it even more computationally expensive. To me, it seems like doing all this in a high-throughput manner for dozens or hundreds of materials would be an endeavor fraught with delays and errors. It would certainly make for an extremely valuable intellectual contribution that advances the field, but I cannot see how we can be on the verge of practically and cheaply using such calculations to design complex new materials at a pace which at least equals experiment.

The second problem I foresee is a common one, what almost any scientist or engineer calls the multi-parameter optimization problem. We in the drug design field face it all the time; not only do we need to optimize the activity of a drug inside cells, but we also need to simultaneously optimize other key properties like stability, toxicity and solubility and – at a higher level – even non-scientific properties like price and availability of starting materials for making the drug. Optimizing each one of these properties would be an uphill battle, but optimizing them all at once (or at least two or more at a given time) strains the intellect and resources of the best scientists and computers. I assume that new materials also have to satisfy similar multiple properties; for instance a new alloy for cars would have to be lightweight, environmentally benign, stable to heat and light and inexpensive. One of the principal reasons drug discovery is still so hard is this multi-parameter optimization problem, and I cannot see how the situation would be different for materials science on a computational level, especially if the majority of techniques involve expensive quantum mechanical calculations.

One way in which calculations can be sped up – and this is something which I would have loved to read about in the article – is by using cheap, classical mechanics-based parameterized methods. In these methods you simplify the problem by using parameters from experiment in a classical model that implicitly includes quantum effects by way of the experimentally determined variables. While these calculations are cheap they can also result in larger error, although they work almost as well as detailed quantum calculations for simpler systems. It seems to me that this database of properties they are building could be shored up with experimental values and used to build parameterized, cheaper models that can actually be employed in a high-throughput capacity.

Does all this make me pessimistic about the future of computational materials design? Not at all; we are just getting started and we need an influx of talented scientists in this area. Computational drug design followed the classic technology curve, with inflated expectations followed by a valley of disappointment culminating in a plateau of realistic assessment. Perhaps something similar will happen for computational materials design. But I think it’s a mistake to say that we are entering the golden age. We are probably testing the waters right and getting ready for a satisfying dip. And that is what it should be called, a dip, not a successful swim across the materials channel. I wish those who take the plunge all good luck.

First published on the Scientific American Blog Network.

Note: Profs. Chris Cramer and Alan Aspuru-Guzik have pointed me to some successful examples of the paradigm. It's clear that certain kinds of problems (especially involving MOFs) are more accessible to the approach than others.

Zn(III)? Not so fast

Chemists love rogues, oddballs which seem to defy the rules and bond, react, and exist on their own terms. The rogues are valuable because they push the boundaries, teach us about new principles of structure and reactivity and challenge coveted preconceptions. One of the most striking rogues in the history of chemistry was the compound xenon hexafluoroplatinate which shattered belief in the non-reactivity of noble gases. Another immensely productive rogue was the first stable carbocation, a species that was considered too unstable to isolate until George Olah surmounted the energy barrier. Some rogues are literal rogues in the sense that they need to be incarcerated in order to prevent their unruly bonds from going haywire; witness the classic taming of cyclobutadiene by Donald Cram. There's no doubt about it; rogues' galleries are shining gems in the chemical establishment.

A couple of months ago it seemed that another minor rogue had made its appearance. Students around the world know that the most stable oxidation state of zinc is +2. The rationale is simple; unlike transition metals like iron and nickel, zinc contains a fully filled d orbital with an electronic configuration of d¹⁰s². It is therefore quite happy to lose its outermost 2s electrons and remain stable. Thus it came as a surprise when a paper detailing theoretical calculations on zinc compounds predicted the existence of a complex with Zn as Zn(III). That fact would have raised the eyebrows of a million freshmen memorizing transition metals trends for their final exams. The existence of the unusual zinc was based on quantum chemical calculations done by Puru Jena's group at VCU and their basic rationale was that a ligand that was oxidizing and electronegative enough would essentially force the metal to donate extra electrons. They seemed to find such a complex in Zn(AuF₆)₃. On the face of it that would make sense, but it's worth keeping in mind that nature loves filled orbitals; for instance, even a ligand as oxidizing and electronegative as CF3 does not force copper to adopt the +2 state (copper's most stable oxidation state being +1, resulting from its d¹⁰s¹ configuration).

Now another paper suggests, with ample evidence it seems, that the suspected Zn(III) rogue might be an upstanding chemical citizen after all. Sebastian Riedel's group in Berlin has examined the purported Zn(III) complex and found that not only does zinc adopt its good old +2 oxidation state in the compound, but that the compound would not be thermodynamically stable. They use a quantum chemical method more sophisticated than the previous one and include more relativistic effects. Where does relativity enter the picture, you ask. It turns out that for heavy metals, the speed (in a crude sense) of the d s-electrons can be fast enough to warrant relativistic calculations; indeed, relativistic quantum chemistry has shed light on many commonplace and yet unusual phenomena, like the color of gold and the liquid state of mercury at room temperature. Relativity is not the exclusive domain of physicists.

I am not enough of an expert in quantum chemistry, but it clearly seems from the latest paper that the more accurate calculations which account for finer details of electron correlation and relativistic effects indicate something striking; the purported complex with Zn in the +3 oxidation state would undergo an exothermic reaction. Put simply, it would not be stable and would decompose to a compound with good old Zn(II). For me, a more troubling fact was the structure of the compound published in the original paper. The structure was supposed to be a low-energy minimum but it has two fluorines practically colliding with each other in an unholy, 197 pm, sub-Van der Waals radius squeeze (illustrated above). The present calculations find another low-energy structure existing only 10 kJ/mol above the previous structure that is consistent with a Zn(II) state. Other calculations indicate that the fluorines in the previously proposed structure are best described as radical anions bridging two AuF₅ units.

There's other theoretical evidence in there too, along with citations of experimental facts that point to the recalcitrance of Zn to adopt a +3 state. The study seems to be as solid a piece of evidence against Zn(III) as can be obtained using current levels of theory. But the real reason I want to point out this paper is because it illustrates one of the key roles of theory and computation in chemistry; as an invitation to experiment. Just because the complex is thermodynamically unstable does not mean it cannot be isolated under any conditions. One of the great lessons of chemical science in the last forty years has been the fact that given the right conditions almost anything can be synthesized, stored and isolated (Cram's cyclobutadiene again being a case in point) and that the words stable and unstable are highly relative constructions created by an impoverished chemical lexicon.

This tussle between two predictions illustrates the function of theory in chemistry as the ultimate teaser; if I were an experimentalist I would be rubbing my hands in glee, sending my armies of postdocs and graduate students into the lab to try to synthesize Zn(AuF₆)₃. As Excimer commented on another site, "Enough theory! Someone make the damn thing already".
Image source: ACS publications.