Skip to Content

Unicorn Software for Drug Discovery

So here’s the dream. You sit down at the keyboard and load a file of the structure of your new drug target – you’ve discovered that inhibition of Whateverase II or a ligand for the Type IV Whazzat receptor would be a good candidate for modifying some disease. You type out a few commands, and your speedy, capable virtual screening program goes to work fitting useful conformations of all the molecules in your company’s collection into the active site of the protein. When it’s finished with that – it doesn’t take that long, you know – it will go on to the current commercially available set of small molecules and do the same for them. If you want more, it has a function to enumerate new structures that it has reason to believe would be potent hits. Come back in a little while and the whole list will be rank-ordered for you.

I guess I should stipulate that you’re also young, extremely well-paid, and ferociously good-looking, and that Stripebutt, your rainbow-colored pet unicorn, is looking over your shoulder and whinnying appreciatively while you get all this done. Because sometimes it looks like Stripey’s going to make an appearance before that software ever does, pesky unicorn droppings and all – we’ve been trying to realize something like this for decades, and anyone who tells you that we’re there is trying to sell you something.

Here’s more or less the state of the art, a current paper in JACS. The authors are looking at selectivities of compounds across the bromodomain enzymes, an area that’s gotten a lot of attention the last few years. It’s a good proving ground – there are a lot of proteins, they’re related in a number of different ways, and selectivity between them is bound to be important. They’re trying absolute binding free energy (ABFE) calculations, which will vacuum up all the spare processing capacity you might have, even more so (I believe) than the relative-binding free energy calculations discussed here. Protein conformations are taken from X-ray structures, removing only the crystallographic water molecules that clashed with the ligands coming in.

Running three ligands (RVX-OH, RVX-208, and bromosporine) across 22 different bromodomains gave a list of predicted affinities and selectivities. The fit is pretty good when compared to experimental calorimetry data (“All of the predicted binding free energies were within 2 kcal/mol of the ITC values, and about two-thirds were within 1 kcal/mol.”) Keep in mind, though, that 1.4 kcal/mol is a tenfold difference in binding affinity, so a bench chemist’s assessment of these predictions and a computational chemist’s might well differ. One of the key features of RVX-OH versus RVX-208 is that the latter has selectivity for the second BET bromodomain, and the calculated data do seem to point that way (although it’s not a strong signal that you might be inclined to bet on).

Bromosporine has affinities all across the bromodomain proteins, so it’s a different sort of test, and in this case the calculations were more scattered (“Roughly a third of the results were within 1 kcal/mol of the ITC values, and two-thirds within 2 kcal/mol. However, this left another third of the results being off by at least 2 kcal/mol, which corresponds to about a 30-fold error in the dissociation constant“. One big part of that might be that good X-ray structures are available for the RVX compounds, so a lot was already known about their poses, whereas no such data are around for any of the bromosporine/protein combinations.

The fit got better (although still rocky) when they went back in and reparameterized the sulfonamide group, but that’s an interesting point in itself. A sulfonamide is not an exotic functional group, but treating it computationally can be a real challenge (“We then focused on the parameters of the soft dihedrals present in the molecule, since such terms are known to have limited transferability across different molecules and might lead to inaccurate sampling of ligand conformations. In particular, chemical groups such as sulfonamides are especially challenging when considering that also quantum effects like the interaction of the nitrogen lone pair with antibonding orbitals involving the sulfur affect the torsional energy around the N−S bond“). As long as this is as hard as it is, it’s going to be similarly hard to get actionable predictions for a lot of molecules, and not just sulfonamides.

This paper is a good filled-glass test, then. Computational chemists will probably see said glass as half full or more, because this sort of thing really is better than we’ve been able to achieve with earlier approaches. Experimental medicinal chemists, though, may well be looking for Stripebutt the unicorn to gallop in, dangling the Answers To All Their Problems from his radiant mane. We really would like to be able to screen compounds computationally and not have to run all those assays, because that would mean that we don’t have to make so many compounds that simply don’t work. That’s what we all spend most of our time in the lab doing now, and the idea that there might be something better is very appealing. But we’re still listening for those unicorn hoofbeats in the distance. . .

51 comments on “Unicorn Software for Drug Discovery”

  1. Me says:

    Looks like they are all looking at delta H. Where are we as regards TdeltaS? Assuming that’s a little more difficult, since it requires a decent understanding of the fluid environment inside the protein, and it’s response to the presence of the ligand.

    …..I wonder if there’s any software out there that can predict what hand looks like by the shape of an empty glove…..

    1. Me2 says:

      read it again, they are doing delta G, not delta H. I do not think it is easy to separate delta H and S, even from an ITC experiment, let along calculation.

      1. David Mobley says:

        Indeed, it’s DeltaG; DeltaH and DeltaS are substantially more difficult to calculate for reasons I could go into if anyone is interested. Mike Gilson’s group has done some work on getting DeltaH and DeltaS also, but this is still only practical for rather small systems (e.g. host-guest binding).

  2. Kent G. Budge says:

    Well, shoot. Why not just dream of a program where you type in “Find a cure for Type II diabetes” and it does the additional step of identifying the ligands or kinases you need to modify to cure the disease, based on the sequenced patient genome?

    My unicorn is bigger than your unicorn.

    1. Derek Lowe says:

      You can find people out there who think that this can be done, unfortunately. . .

      1. Pedro says:

        that’s because they’ve been watching to much star trek— where you can diagnose diseases instantly with what is an oversized flip phone (tricorder)— or ask the computer a complicated question and receive a succinct answer immediately

        1. zero says:

          Once we have a post-scarcity society (Kardashev type 2 at least) with strong AI, FTL travel/communications and matter transporter/replicators then flip-phone diagnostics and instant cures would be routine. That’s a little like saying “If I had all the money that will ever exist then I could buy this cheeseburger” or “this nuclear weapon will surely kill that pesky ant”.

          Even if we had all that, there’s no guarantee that the Biology rabbit-hole doesn’t end in a door marked “Incalculable quantum-mechanical effects, Keep out.”

        2. JSC says:

          I have to disagree, Star Trek has singlehandedly inspired technologies that wouldn’t have other wise existed by convincing dreamers to dismiss the naysayers and go for their dreams. It’s entirely possible that one day this technology will see great improvements because of and not in-spite of the dreamers.

      2. Anon2 says:

        If anyone is convinced by Derek’s tirade here, take a look at the very next blog post http://blogs.sciencemag.org/pipeline/archives/2017/01/19/a-first-look-at-reproducibility-in-cancer-biology There, Derek discusses how biology replicates even less than this computational work, but he manages to do so without condescension.

        Why is computation so threatening, Derek?

        1. devicerandom says:

          Well, I worked in computational biology until last year (docking GPCR ligands and performing molecular dynamics) and I would say that what Derek says is spot on.

        2. Mol Biologist says:

          If anyone can convince Derek that biology is the subject which need some basics education first. I like to make an citation where one not biologist with high IQ who red some alternative articles and said to me that HIV virus does not exist. Well, it can be true if only unicorns does exist. When your choice in the drug discovery is targeting the receptors or particular proteins or the same protein-ligand pairs ever. Your choice to be is battling the same problems for an century. Altered metabolism does not come from singe target as way to modify the disease . It is coming from the mechanism of action of the drug and from knowledge of basic biology. There are different forms of life on Earth and they have different metabolism.
          https://www.mskcc.org/blog/license-build-new-theory-cancer-puts-metabolism-center

          1. Mol Biologist says:

            Thank you, Derek for your attention to the problem.
            I do recognize the pioneering nature of Drs. Kaelin and Ivan’s contributions to the field of HIF and PHDs Biology.
            I also have great respect for their next work. https://patentscope.wipo.int/search/en/detail.jsf?docId=US77261175&recNum=1&tab=PCTClaims&maxRec=7&office=&prevFilter=&sortOption=Pub+Date+Desc&queryString=FP%3A%28Kaelin%2C+JR.+William+G.%29
            Time is changing everything. There are piles of facts that structural/allosteric inhibitors for PHDs did not work. Your blog is excellent source to get insight on it. Merck, Pfizer, GSK etc had many attempts to develop it but still NO drugs are available neither to treat a cancer neither Cardiovascular Disease or by other words their never been created. If you really want to benefit the drug discovery you MUST uncover the mechanism. IMO it would be in a contrary position to Dogma of Oxygen Sensing by PHDs.

            I am proposing a way of humility and start looking for evolutionary process. Promiscuous activity originated from evolution pressure and the development of secondary metabolism synthesis allowed for the extensive diversification of the enzymes. The emergence of bacterial secondary metabolism systems received a tremendous impetus from the primary oxygenation event in Earth’s history caused by cyanobacterial biosynthesis. Under new selective pressures these activities may confer a fitness benefit therefore prompting the evolution of the formerly promiscuous activity to become the new main activity. https://en.wikipedia.org/wiki/Enzyme_promiscuity
            I wrote it last year as well.
            http://blogs.sciencemag.org/pipeline/archives/2016/03/21/screening-the-mechanism-not-the-products

    2. Kelvin Stott says:

      Well, double shoot. Why not just dream of a machine where you add sick or diseased cells, press a button and your safe and effective new drug physically pops out?

      My unicorn is *even* bigger than your unicorn.

      And unlike the computer/software-based approach, it is very achievable as it does not depend on feeding in potentially false extrinsic information, because all the right information about the disease is already there within the cells. After all, that’s exactly how evolution by natural selection works. We just need to increase molecular diversity and accelerate the screening process.

      1. Kent G. Budge says:

        I kind of already have such a machine. I call it a “germinal center.” It keeps me quite healthy, most of the time. The fact that it sometimes fails is the whole reason for a drug industry — and a warning about the limitations of such machines.

        1. Kelvin says:

          Certainly evolution can’t solve all problems, especially not after reproduction when selected genes no longer propagate. But it works in 99% of cases, so just imagine if we could harness the same principles to solve 99% of the 1% cases when natural evolution has gone wrong…

  3. Peter Kenny says:

    The term ‘absolute’ when applied to free energy calculations may be a misnomer because we still need to define a standard state (even if the units are molecule/cubic Angstrom).

    One factor that may limit the accuracy of free energy calculations is the use of electrostatic models that are based on atom-centered charges. QM calculations suggest that hydrogen bond (HB) acceptors are associated with one or molecular electrostatic potential (MEP) minima which cannot be reproduced by atom centered charges (or multipoles). The value of MEP at the minimum is usefully predictive of HB basicity and I’ve linked a recent article on the topic as the URL for this comment.

  4. Chrispy says:

    Progress in the docking field has been severely hampered by people overselling their docking programs for many years. Oh, and the inability to properly model solvent effects.

    And you medicinal chemists who have sold HTS leads as coming from docking — when in fact the structures came AFTER the screen — are not doing the field any favors. There are some good (bad) examples with HIV protease inhibitors…

  5. ITGuy says:

    People used to think computer could never (or at least in their lifetime) beat human GO master.

    1. futurist says:

      I think that most skepticism here is just the manifestation of severely-tempered optimism: technology can develop in unexpected leaps and bounds but, thus far, these summed efforts still haven’t crossed the finish line for computational med chem…here’s hoping for the next breakthrough!

    2. Matt says:

      I have a feeling that, a lot like Go, progress in this field is going to be very very slow, until it goes very very fast. And the people that finally do make that leap probably won’t have to go on talking about it, the results should stand for themselves.

  6. b says:

    It looks like they used ABFE in Gromacs. Last I had heard, that wasn’t particularly accurate. How does that compare to other free energy binding calculations, like FEP+ in the Schrodinger suite?

    https://www.schrodinger.com/fep

    I would be curious to see a head-to-head comparison on this dataset!

    1. David Mobley says:

      These days, we can reproduce energies for the same system in GROMACS, AMBER, CHARMM, Desmond (used by Schrodinger’s FEP), etc. So, there’s not a problem with GROMACS.

      It’s worth noting that much of what this paper does would be rather difficult to do with relative free energy calculations like Schrodinger’s FEP+, since this is looking at binding of the same ligands to different proteins, some of which are substantially different. That part could really only be done (at present) with absolute calculations like those used here.

  7. Anon says:

    Doesn’t matter how smart your algorithms/AI deep learning systems are, the ultimate problem is the reliability of the data you feed into them.

    Information quality is limited by signal-to-noise, so if you want more signal, you’re gonna get more noise. Or put simply: Garbage in, garbage out.

    1. David Mobley says:

      For what it’s worth, there’s actually no machine learning involved here, or training to the data. This is based on force fields, which are developed using much simpler data (quantum mechanical data, simple physical properties for much simpler molecules (heats of vaporization, density, etc.), and various others — properties which can be measured quite precisely. So there’s less of a concern with data quality.

      1. Anon says:

        Sure, the fundamental parameters we do have are very accurate. The problem then is too few of them (constraints) vs too many degrees of freedom. Macromolecular vibrational harmonics anyone?

        We still haven’t cracked the protein folding problem for exactly that reason, and probably never will.

  8. Barry says:

    we’ve been reading this “in-silico assay” fantasy for thirty years now. It was supposed to replace med. chem. sometime last century. It may yet, in the long run. But “in the long run, we are all dead”. How (if) this virtual screening changes what we do in 2017 is less clear.

  9. Dizzle says:

    The title of the article made me think of AKTA software Unicorn

  10. Anon says:

    It seems to me that this type of idea is more suited as a companion to medicinal chemistry than a replacement.

    You can theoretically model hundreds or thousands of compounds for every one that you make, at least to try and optimize selectivity or potency.

    It’s a good idea, just that each step needs to be verified by real physcial experimentation.

  11. Biff says:

    Ok, I confess. I did laugh out loud at the “Stripebutt” line. Sue me!

  12. LeeH says:

    There’s the old adage that “the perfect is the enemy of the good”. This is especially true when you consider docking software.

    It’s unreasonable to expect that virtual screening spits out drugs. It IS reasonable to hope that virtual screening can rank order compounds in such a way that you end up with an acceptable hit/lead at a rate that is greater than random. And as long as the cost of that savings is better than the cost of doing the calculation, you’re ahead. The problem is that it’s really hard to demonstrate that savings, since you usually don’t do both the targeted selection and the random selection. You often have to rely on retrospective analyses, which make everyone uncomfortable.

    My personal feeling is that the greatest utility of docking is to visualize the docking mode (since docking software generally does that really well), and use that to suggest the most likely sites on the molecule to make changes. However, given advances in FEP and GPU parallelized code, among others, it shouldn’t be too long before we have better faith in binding affinity estimations.

    1. Mark Mackey says:

      Chris Bayly, while at Merck Frosst, presented some work a long time ago on a ligand-based virtual screen where he’d actually persuaded someone to run the control experiment and work out the background hit rate. As I recall, the “good” virtual screen with lots of hits actually turned out to be not that impressive as the background hit rate was high, while the “poor” virtual screen with relatively few hits was hugely enriched as the background hit rate was close to zero.

      You’re right about retrospective VS experiments: they are very very very hard to do right which is presumably why virtually nobody bothers.

      1. LeeH says:

        I think the real issue is – if the benefit of virtual screening is modest (say, a 20% improvement of hit rate), is it worth doing? My feeling is yes, especially if the computational method is cheap.

        The contrary argument I’ve seen most often invoked is “well, what if you miss something because of the virtual screening”. This, to me, is a specious argument. You always miss something. In fact, you miss most things all the time. If you worry about what you miss, you’re going to be highly neurotic and haunted person. The issue is whether or not you end up with useful chemical matter. And I’ll take 20% more (or a 20% savings in resources) every time.

  13. An objective goalpost for success is when computational methods are as accurate as physical methods. Therefore, we need to have a baseline: what is the agreement when I make a measurement with two physical methods?

    Fortunately, for delta-H (rather than delta-G), this study has already been done for you (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2279960/). Table 4 reports ranges of delta-H of -10.4+/-2.5 kcal/mol by ITC and -10.6+/-1.4 kcal/mol by SPR. See Figure 5B and Table 3 for the break out.

    If it’s true that “All of the predicted binding free energies were within 2 kcal/mol of the ITC values, and about two-thirds were within 1 kcal/mol.”, and if you’re willing to trust experimental results from another lab, why wouldn’t you trust these results?

    1. someone else says:

      Note that in that ITC/SPR study the error propagation from the van’t Hoff enthalpy calculations magnifies the original measurement error of the dissociation constant. The error in binding affinity is plenty small enough to allow meaningful SAR comparisons at much finer detail than a 1-2kcal/mol error.

  14. Garrett Wollman says:

    You should well know, Derek, that unicorn hooves don’t make beats, they *chime*.

    1. Kent G. Budge says:

      *Tinkle.*

  15. AK says:

    Take two different labs to measure the delta G’s of the same protein-ligand pairs and you’ll get greater fluctuations than 2 kcal/mol. Take two different assay techniques and it might differ even more. Why are people expecting more accurate results from compchem than from their own pharmacology department? Just take ChEMBL and look at the distribution of bioactivities for the same protein-ligand pairs.

    Yes, granted, there are still (very) big strides to be made in compchem, but also be aware of the experimental error of the values that are used to compare these techniques to.

    1. DLIB says:

      Here’s a link to a study of that point…the Std dev of 12 labs for ITC was about 1.6Kcal/mol. They used different instruments with different run parameters ( the labs were just given the proteins. Not bad for such varied setups

      https://abrf.org/sites/default/files/temp/RGs/MIRG/mirg2002_summary.pdf

  16. entropyGain says:

    “Keep in mind, though, that 1.4 kcal/mol is a tenfold difference in binding affinity, ” – Therein lies the rub!

    As a recovering computational chemist who spent some years in the 1980s and 1990s in quest of StripeButt I’ve come to the conclusion that “accurate” docking predictions are a giant waste of time. For one, the underlying biological data people often try to fit is a hodgepodge from different labs that is frankly non-comparable. But the true futility is in applying molecular mechanics force fields which usually neglect fundamental forces that can easily contribute more than 1.4 kcal (polarizability anyone? long range electrostatics?).

    Sadly, with the exception of ubiquitous faster computers and color graphics, little has fundamentally changed in computational chemistry since 1980. As a field, we should stop wasting our time asking this futile question and actually invest a little effort making the user experience/interface of these software packages usable by real medicinal chemists who might appreciate some help developing hypotheses based on structure and perhaps tickling their imagination to help come up with fresh ideas.

  17. stripebutt believer says:

    Even though the answers can swing as much as 2kcal/mol, I’m….failing to see the issue? The criticisms I’m reading here seem to indicate that the prevailing sentiment is that if a calculation isn’t 100% as accurate as the wet experiment, then why bother–which seems pretty flawed to me. The point of computational chemistry has always been to get rid of ‘bad ideas’, and promote ‘probably good ideas’ and/or preventing medchemists/chemists/whomever from wasting time on making ‘obviously bad’ compounds.

    In a theoretical situation where a virtual screen using absolute free energy calculations suggests a compounds has a Ki/IC50 of ~0.5nM, but in actuality was actually ~50nM (difference of ~2.7kcal/mol), was that a bad result? Of course not. Of course the waters get a bit more murky when you want to differentiate or trust results with only moderate potency (or a small difference in potency between two compounds), but in that case the reasonable thing to do is to make & test the compound if you already have a hypothesis you’re testing. As is often the case in science, so long as you’re asking an appropriate question (or have reasonable expectations of your method), you’ll be fine. Now, the real question, it seems to me, is ‘how do we know which compounds to run this method on?’

    1. entropyGain says:

      StripeButt believer–The level of accuracy your talking about is fine for idea generation, and can be useful as I noted above. Though the usability of all the software packages I’ve checked is so abysmal that you need to be a full time user with programming skills to be productive.

      However, I strongly disagree about the role of computational chemistry as saving chemists from themselves. I lost a bet with one of my chemists recently when he proposed a really “dumb” idea (in my opinion) that in all likelyhood may now be our next development candidate. Good thing he didn’t listen to me!

      And remember that screening is a very minor part of the drug discovery. 99% of the effort and cost is in lead optimization and potency is usually taken for granted because you know the SAR so well. It’s pretty easy to make subnanomolar compounds all day long once you get into it, but F or CYPs or selectivity or one of the other N variables your trying to optimize will always be biting you in the butt, even if it is striped and on a unicorn.

  18. AndyM says:

    Well stated entropyGain, AK and others, including “stripebutt believer” – I think you’re more or less on the same page. A good tool (unicorn), if applied intelligently, can you get you to the goal line sooner, as long as one is mindful of both the strengths and limitations of the unicorn.

    If one simply considers the propagation of errors associated with modeling all the free-energy terms associated with computing protein-ligand-(de)solvation interactions over many degrees of freedom, it can be fortuitous to get rank-ordering correct, let alone breaking the 1.4 kcal/mol barrier. The accuracy and/or rank-ordering will always be context dependent. For example, if you’re making subtle changes (decorations) to a chemical scaffold, FEP can probably provide a reasonably quantitative (physical) result. But as soon as you move outside of a congeneric series to a new chemotype, you’d have to reestablish a baseline (a “standard state”) for a new scaffold. In some ways, the newer StripeButt methods are just fancy versions of Free-Wilson analysis. As with most retrospective analyses, it’s hard to determine how well StripeButts perform prospectively – it’s foolhardy to apply as a black-box method, without grounding in experimental reality. In practice, I’ve had to clean up after unicorns.

    A reminder about molecular mechanics forcefields – these are relatively crude parametric models for reproducing low-energy molecular confirmations and dynamics. They are fit-for-purpose for SIMULATING molecular confirmations/dynamics, for limited molecular systems, with limited accuracy. Not much has changed since the 80s other than the tweaking of fixed parameters (spring constants, atomic radii, etc).

    Regarding quantum methods, it’s also worth reading the recent Science article on the state of first principle approaches, namely DFT. Unfortunately, one finds that in the quest to achieve ever more accurate energies for more complex molecular systems, the underlying electron densities have become poorer – i.e. some DFT practitioners have become more enamored with fitting parameters, while losing site of the underlying physical basis for DFT, the electron density (via the Hohenberg-Kohn theorems) . “We found that densities became closer to the exact ones, reflecting theoretical advances, until the early 2000s, when this trend was reversed by unconstrained functionals sacrificing physical rigor for the flexibility of empirical fitting.”

    “Whither the density in DFT calculations?”
    http://science.sciencemag.org/content/355/6320/49

    In short, with all the parameters in computational chemistry and potentially myopic application to complex systems, it’s not too hard to get the “right” answers for the “wrong” reasons – eventually physics and/or Mother Nature balances that equation.

  19. Peter says:

    From a CS / mathematics point of view, this kind of thing looks, in some sense, ridiculous. Even for incredibly simplified models of chemistry, we know that the kind of problems you have to solve in order to have the Unicorn Program are at least NP-hard, if you allow a little realism you can push up the complexity classes.

    The layman’s version of that is: don’t expect the Unicorn Program to take less than exponential time in the number of atoms involved. If that’s more than about 40, forget it unless you have a supercomputer and a whole lot of time. No such program exists (if you believe reasonable conjectures).

    1. Complexity classes are probably an uninformative way of looking at this problem, since you can achieve utility by solving specific problems in a class without guaranteeing a solution to every problem in the class. For example, 3-SAT is NP-complete but heuristic SAT solvers are used in chip design and theorem-proving with tens of thousands of variables. Even more dramatically, Chess and Go are EXPTIME-complete, yet computers are much better than people at these games. If your job depended on winning at Go, you might want to use AlphaGo as a tool irrespective of the fact that Go is not in P.

      You may believe that programs can’t provide value for chemists, but your argument shouldn’t rely on the complexity class of the task.

  20. skeptism says:

    I don’t think that any of the above discussions ever touched the real issue regarding ABFE calculations in this paper. REPRODUCIBILITY !

    Regardless of how meaningful the reported accuracy is, even on the reported accuracy level, any experienced FE practitioner knows that it is unlikely to be reproducible. Due to a large computational and setup cost to generate this paper, every few would like to repeat the whole computing effort. It is just like that accuracy levels published in many JACS FEP papers in 90’s cannot be reached even now.

    I hope that the authors can release the docked structures for ABFE calculations and the community distributes the work load and tried to seriously check whether the accuracy level can be produced.

    1. AndyM says:

      Excellent point!

      When complex computational methods are applied to complex protein-ligand interactions, reproducibility is often a casualty. The devil is in the details of the methods (as well as the systems sampled) and in many publications the details and means to repeat and to compare calculations are insufficient. This isn’t limited to computational chemistry, but publications involving complex methods and systems in general. Your point ties nicely to Derek’s follow-on post: “A First Look at Reproducibility in Cancer Biology” http://blogs.sciencemag.org/pipeline/archives/2017/01/19/a-first-look-at-reproducibility-in-cancer-biology

      1. skeptism says:

        A system should be designed that allows such paper to be under serious scrutinization like in the experimental science world. If anyone wants to challenge such work, the authors should release necessary details for the challenger to repeat, provided that the challenger likes to contribute his/her own time and computing resource. It is feasible as long as the challenge test is encouraged to be published.

    2. Matteo Aldeghi says:

      Reproducibility is indeed a good point – as it was mentioned the devil is in the details, and oftentimes even the authors might not be aware of the impact of all their choices on the final results.

      As far as this paper is concerned, the input files used for the ABFE calculations are available as part of the Supplementary Information as a compressed archive. These include the starting coordinates of the systems, the force field parameters, and all the simulation setup options.

      It would be interesting to see if, given a likely decrease in the cost of the calculations in the future, longer timescales and better sampling still return similar or better/worse results. We did not see much change when running longer simulations, or re-running the simulations a few time; however, this was done on a single test case rather than systematically (some discussion in the SI). Reproducibility issues related to the accuracy of older publications may, in fact, be due to a combination of little sampling, small datasets, and luck. That said, anecdotally, I found the results of some ABFE publications from the late 2000s to be fairly reproducible, given the same starting structures and force field parameters.

      Relative calculations are getting more and more feasible nowadays. Absolute ones are not at the same stage, and I doubt they will be widely adopted in the very near future, but they might end up having some value in certain applications where there are little alternatives. As time goes on, hopefully they will become easier to setup and cheaper to run, so that they will also be tested more extensively. If it is eventually found that they are no better than alternative cheaper approaches, it will be fine with me: I do not have any particular reason to push this specific approach, and I am happy with whatever is shown to work better.

      PS Thanks to all for the interesting comments and feedback!

  21. Mol Biologist says:

    Thanks Derek, for your attention to the problem. I do recognize the pioneering nature of Drs. Kaelin and Ivan’s contributions to the field of HIF and PHDs Biology.
    I also have great respect for their next attempt. https://patentscope.wipo.int/search/en/detail.jsf?docId=US77261175&recNum=1&tab=PCTClaims&maxRec=7&office=&prevFilter=&sortOption=Pub+Date+Desc&queryString=FP%3A%28Kaelin%2C+JR.+William+G.%29
    Time is changing everything. There are piles of facts in the field of developing specific/allosteric inhibitors of PHDs (Merck, GSK, Pfizer etc). Unfortunately, there are still no outcome for drugs neither for cancer neither for cardiovascular disease. Your blog is great source to give insights on this problem. Why not?
    IMO If you really want to benefit the humanity you must uncover the mechanism. But you need to be in a contrary position to Dogma of Oxygen Sensing by PHDs.
    Promiscuous activity originated from evolution pressure and the development of secondary metabolism synthesis allowed for the extensive diversification of the enzymes. The emergence of bacterial secondary metabolism systems received a tremendous impetus from the primary oxygenation event in Earth’s history caused by cyanobacterial biosynthesis. Under new selective pressures these activities may confer a fitness benefit therefore prompting the evolution of the formerly promiscuous activity to become the new main activity. https://en.wikipedia.org/wiki/Enzyme_promiscuity
    And I wrote about it last year already.
    http://blogs.sciencemag.org/pipeline/archives/2016/03/21/screening-the-mechanism-not-the-products

  22. medicinal chemist says:

    waste of time.

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.