Skip to Content

Modeling in Drug Discovery: Questions?

I’ll have the opportunity to sit in on a few talks during a conference on free energy calculations in drug design. Since I’m not a computational guy myself, I’ll be picking my sessions carefully, but I am interested in hearing what the state of the art is.
If we could just walk right up and calculate the free energies of binding events reliably, that would mean that the era of high throughput screening would begin to come to its end – well, in the physical world, anyway. Depending on how lengthy the computations needed to be, we could (in theory) just sit back and let the hardware hum while it ran through all the compounds we could think up – then we’d come back in on Monday and see who the winners were. Despite what some of you outside the field of medicinal chemistry might have read, we are not exactly to this point yet. That phrase “in theory” covers an awful lot of ground. But progress is apparently being made (here’s a recent paper (PDF) with background).
So here’s a question for the readership: what would you most want such calculations to be able to do for you? What would convince you that they’re actually believable? And how close to you think that we actually are to that? Your comments will go directly to the ears of a roomful of high-powered modelers, so feel free to unload.
That thought of a roomful of computational chemists, though, reminds me inexorably of a story about Robert Oppenheimer that Freeman Dyson retells here. At a theoretical physics conference in Vancouver, the attendees were on a boat ride among the islands when the weather turned impenetrably foggy. Someone asked what the consequences for physics would be if the boat sank, and Oppenheimer instantly said “It wouldn’t do any permanent good”. There, that should ensure me a warm welcome at the meeting!

62 comments on “Modeling in Drug Discovery: Questions?”

  1. milkshake says:

    what woul convinvce me about usefulness of the stuff: prediction of a induced-fit binding mode, from first principles. That means without supplying a precedent or fudging with the software to make things come out right,
    I have in mind two isoforms of a particular kinase, and a class of molecules we found, that are isoform selective. The isoforms have exactly identical ATP binding site but one of them has it more plastic (because of some minor sequence difference way back in the protein) so the arylamino on our molecule can poke out a new pocket for itself in the ATP binding site, whereas this does not happen in another kinase isoform.
    I should mention that my ex-boss is a computer modeler by trade and when he was without job he consulted for a startup that was developing the software packages. He was particularly interested how the software was coping with kinases of which he saw X-ray of a co-crystal with a ligand…. it did not do very well, to the point of the predicting power being nearly useless. It has been few years so maybe the commercial packages improved since then, but I would stil like to see some nontrivial validation of the software predictions – that means comparison with X-ray data that the software authors never saw

  2. Sili says:

    Heh. Physicists tend to peak young, so there might have been less of a loss in that case. I don’t know about chemists.
    You could always go for the classics: “What do you call a room full of comp chemists on the bottom of the ocean?”

  3. FormerMolecModeler says:

    Docking has an intrinsic accuracy of about 60%. That is, if you take a protein crystal structure with a bound ligand, take the ligand out, then re-dock it back, you will only get a good RMSD (say

  4. insilico_skeptic says:

    As someone who spent good years of my life drinking the FEP cool-aid and not getting it to work, it seems to me that FEP’s promise of estimating delta G is still a long way off.
    I think section of 7.1 of the book chapter you reference shows it nicely: the problem is the force field. In particular, I don’t think we understand water very well.
    If science can be divided in to “physics and stamp collecting” then I would say that stamp collecting often wins. That has been the case in protein folding, QSAR modeling, and steam tables.

  5. RandChemist says:

    I’ve had two different experiences with modelers.
    One was highly positive, a definite collaboration. I learned a lot working with him. We pursued the data, made hypotheses and tested them. In the end, we made tremendous strides in SAR.
    Another was so tightly bound by convention that he did not want to pursue the data in the early phases. Sometimes you just try things. It probably won’t be the final answer, but it might teach you something along the way.
    Modeling is like any tool, it is how you use it. It can be quite powerful, but it is not the final determinant.
    So my advice is on how to approach interactions (philosophical) rather than a conceptual level. Work with the chemists making the molecules! Seems obvious, but…
    My question is: how reliable is a crystal structure in determining which molecules to make? Isn’t it better to look at things in solution, as they are?

  6. RB Woodweird says:

    Long ago as our organization was being trained for Class A or JIT or whatever the consultant-pushed KoolAid was that month, we had to watch a video from Oliver Wight. On this tape, a stout old guy was retelling a tale from his very first job with Bendix. He was a chemical engineer, and his mandate was to make sure that there were enough raw materials to produce what had been ordered. There were several liquids involved which were stored in tanks sunken into the plant grounds. The contents were supposed to be monitored on index cards kept in the office, but the veteran employees always went out into the yard and stuck a long stick down into the tank outlet to visually confirm how much was inside. When the young pup asked why they did this seemingly redundant activity, he was told “If it’s on the stick, it’s in the tank.”
    When you really really have to get accurate binding data, does using a computational method qualify as the index card or the stick?

  7. sgcox says:

    interesting review on the comp. med. chem:
    http://www.nature.com/nrd/journal/v9/n4/pdf/nrd3139.pdf

  8. darwin says:

    …the many failed experiments that screamed success on paper

  9. David Formerly Known as a Chemist says:

    I’ve always felt modeling was a useful, though minor, tool. Modeling helps one understand the SAR, which allows you to maintain (or enhance) potency against target. This is especially useful when trying to optimize for drug-like properties (solubility, PK, metabolism, etc. For the programs I was involved in over the years, achieving potency was never a big problem, though selectivity, favorable PK, metabolic stability, and a clean tox profile were the real hurdles. Computational modeling never helped much in clearing these hurdles.
    A conference on free energy calculations? Gah, I’d rather be waterboarded!

  10. darwin says:

    …the many failed experiments that screamed success on paper

  11. darwin says:

    …the many failed experiments that screamed success on paper

  12. DLIB says:

    So this field has been around for around 30 years now. The forces of nature that are involved have been known and many force fields have been developed to capture them. Given that, I looked into where the leading edge research is going and it seems algorithmic efficiencies is where most of the improvements are being made. I used to go to talks and at the end of them I’d see this slide of Moore’s law – see the promise of the future – that was a long time ago. The truth is the things like conformational entropy will likely be very elusive to efficient computation and the error bars for the calculated binding energies will still be large ( remember 0.5 Kcal = 2X Kd ) since it is common to have the differences between two large energies to yield a small difference.

  13. It seems to me that a major limitation is practical evaluation of results. Most modelers spend a lot of time living in a world where the end result is known, so they can say 30% of their models were close to correct. For the rest of us, how can we tell in advance which 30% to use? How do we know which results are believable? How can we come up with accuracies for binding energies by various methods? In a list of docked ligands, where should we stop believing the results?

  14. Fellow Old Timer says:

    Many years ago a published author feted in this blog had to earn his bread making analogs for evaluation as herbicides. He made one suggested by a computational chemist who was convinced it would be a winner. When it turned out a dud, the computational chemist alerted the analytical department. He wanted proof that the chemist made the wrong compound. When theory disaggreess with the facts, guess what?

  15. Cassius says:

    insilico_skeptic: “In particular, I don’t think we understand water very well.”
    I agree…. how water is handled is my main concern with modeling, and one of the first things I think about when comparing different docking programs. I’m just getting started with virtual screening, so I have too many questions to list here. My main questions would be:
    (1) What docking suites best mimic an aqueous environment, taking into account solvation and desolvation of the receptor (protein, nucleic acid) and ligands?
    (2) Is there a way to accurately balance electrostatic interactions with hydrophobic interactions? (back to water in the hydrophobic effects)
    (3) What is the best way to introduce flexibility? Are NMR structures the best way to get at the dynamics in solution?
    Thanks for the offer Derek! I hope the talks are not as brutal as they sound.

  16. LeeH says:

    Historically, compchem has been an oversold science/art, but you shouldn’t throw the baby out with the bathwater.
    CompChem is like blackjack. If you use the strategy tables in blackjack, you reliably minimize your losses over time (and if you play enough and don’t count cards, you ARE losing). Using the tables does not guarantee that you will win any given hand, but the strategies maximize the outcomes over a long time. If you’re not keeping track carefully, you may not see those benefits over time, but they are there without a doubt.
    CompChem is the same. Applied correctly and over many compounds, it can give you a statistical advantage. The art for the modeler is understanding which methods are really appropriate for which problem, and when the investment is worth the payback.

  17. Maks says:

    It’s a nice tool which has to be taken with caution, i.e. finding an high affinity inhibitor of a kinase doesn’t exclude the possibility that is also binds with the same affinity to any other protein in your assay. Just like Lipinsky’s rules it can give you an idea which compounds are more likely to work, computational work is quite cheap compared to wet work, so increasing your hit rate in an initial screen by 20% is worth the virtual screening.
    Another approach which I really appreciate is “scaffold hoping” by virtual screening, i.e. taking an initial hit, looking for compounds with different chemical properties but similiar 3D shape, for a review look here: http://www.ncbi.nlm.nih.gov/pubmed/19442066

  18. JC says:

    Pretty pictures that make for nice presentations, but entirely useless if not misleading otherwise except to explain how something is good after the fact.

  19. Anonymous says:

    While a modeller can contribute a lot to the drug discovery process, doing free energy calculations isn’t one that makes any significant contributions.
    The calculations isn’t predictive in the general case.

  20. chris says:

    There is a recent paper looking at virtual screening 2D descriptors http://pubs.acs.org/doi/abs/10.1021/ci100062n, the results are pretty competitive with docking and take a fraction of the time.
    Solvation is a major problem for comp chemists, but also for Medicinal Chemists, try explaining why a hydrogen bond between a ligand and an exposed residue may not actually increase affinity.
    That said I think these tools can be used to generate ideas, often ideas that chemists thinking more about the synthesis might not come up with.

  21. deekai says:

    Derek, ask them whether they can generate a good result independent of the amount of CPU time required – ie if the the physics is there. If this is the case, code and hardware can be optimized. If the physics isn’t there, it has to be developed first.

  22. Computationally entertained says:

    agree with #19 …
    one can take a look at the performance of free energy calculation methods in predicting the solvation free energies of organic molecules. the average accuracy is in the range of 1~2 kcal/mol at best (last time I checked). with proteins included, the sampling convergence and the error in the force fields are likely to become worse in general.
    reliably predicting absolute/relative binding affinities is like the holy grail of CADD, and it’s the ideal to work towards. but we’re not there yet. however, the contribution of computational chemistry and modeling doesn’t need to be limited by that.
    speaking of models, what is not in pharmaceutical R&D?

  23. Anon says:

    I’d be curious to hear more examples of FEP work going on in Pharma & Biotech companies — or, if nothing is happening, it would be good to know that too. My sense is that few companies are working in this space. Is that what most of us are seeing?

  24. Chrispy says:

    Above all, I wish the field had more honesty. You see all the time stories about how in-silico docking and modeling led to some winning series of compounds, yet what really happened is that the chemists banged on a screening lead. (Are you listening, Merck HIV protease team?) Overselling this has led to many of us not believing the good stuff, either.

  25. Mutatis Mutandis says:

    Screening teams at conferences are usually of the opinion that targeted libraries selected by computational chemists don’t have a higher average hit rate than random libraries. I wonder whether a statistically valid study of this has been made, comparing different techniques? That could be a more reliable indicator. Although no doubt some targets can be modeled better than others.
    The skeptic’s view is that so far all forms of modeling become most useful when you already have a crystal structure. A bit of predictive SAR is possible, or at least can help to avoid dead ends.
    Apart from the obvious challenges of calculating binding energies for all possible off-target effects as well, and taking into account physico-chemical properties, the case has been made that many successful drugs don’t reach binding equilibrium anyway (at least not in a patient), so it’s not necessarily Kd that matters most, but perhaps kon or koff, depending on the kinetics of the target.

  26. Malcolm says:

    If we’re talking the more expensive MD or MC free-energy calculation methods, then these are done with explicit modelling of the solvent i.e. a water box, perhaps with some ions floating around.
    The water models are OK. In fact, we have some really good water models (I’m thinking the more recently tuned TIP4P or TIP5P variants, or perhaps even something exotic with polarisability). Unfortunately you trade off computation time with quality of the model, but hey SPC water is probably good enough for many purposes.
    My observation is that the force fields may be flawed, but the bigger problem is to do with sampling and time-scales, i.e. ergodicity. This is where Michael Shirts et al have been doing some great work. Maybe not much practical help to modellers just yet, but fundamentally a step in the right direction.

  27. Brian fred says:

    Once again I tend to agree with JC. Malcolm M? Polarisability….smile on my face.

  28. Anonymous says:

    # 24. Yes I am now.

  29. Matt says:

    I spent a couple of years working on modeling damaged nucleic acids with MD in explicit water. I even got personal advice from David Mobley in trying to apply “alchemical” transformations to correlate models with experimental strand melting temperatures. No dice.
    At one point I went back to a very simple case from older literature, modeling the solvation energy of bromide ion. I parameterized the system just like the authors from the early 1980s and got the same answer to within three decimal places in Gromacs. This also was also a near-perfect match for available experimental data.
    Then I tried several standard water models instead of the special one the old paper used. The energy of the transformation bounced around by nearly 10 kcal/mol depending on the model. At least I got my wrong answers orders of magnitude faster than people working in the 80s.
    I think computational chemistry still has a lot to offer with high-accuracy quantum methods, but I now view force fields with the greatest skepticism. Unfortunately, the accurate quantum methods are nowhere near fast or scalable enough to handle biological macromolecules.
    I’d need some serious convincing to trust force fields as predictors of experimental results rather than crib notes for them, like a counterintuitive prediction from FF models that is investigated and confirmed experimentally only *after* the computational results are published.

  30. Malcolm says:

    @Matt did your bromide ion solvation simulation involve a net change in charge?
    My memory is that changes in ionisation state are a sore point for FF-based models.
    DNA ranks as a higher degree of difficulty than proteins; the FFs haven’t had nearly as much TLC and the need for counter-ions complicates things.
    I do speak however from a fading memory of the modelling world circa 2002.

  31. cliffintokyo says:

    Maybe I am a bit behind the times, but IMHO:
    We need computer-assisted 3D molecular modeling to help turn a theory into an hypothesis.
    What I mean is, we don’t usually need absolute validation by comparison of molecular modeling of ligand-receptor interaction results with the experimental results of structure analysis by X-ray, etc.
    What we need is pragmatic relative validation, whereby a SAR theory founded on 3D intermolecular interactions with a *receptor* for a particular series of compounds shows good correlation with biological activity parameters for those compounds, and enables reasonably useful predictions for the biological activity of compounds in the series not yet synthesised and screened.
    Applying all the *useful* molecular interactions parameters and their associated calculations simultaneously to a ligand-receptor interaction is a vast improvement on correlating biological activity with only 1 or 2 parameters, such as the nmr chemical shift(s) of (a) key H-atom(s).
    Oh, and molecular modeling also reminds medicinal chemists to think about their compounds in 3D.
    We should make 3D models of every compound, period.

  32. michael says:

    any suggestions on a good molecular modelling software? im a medicinal chemist looking to try my hand a comp chem. much appreciated

  33. Wavefunction says:

    Schrodinger is one of the most accessible, comprehensive, user-friendly and well-validated suites, especially their docking, homology modeling and MD programs. The MD program is from DE Shaw.

  34. Matt says:

    @Malcolm: I don’t remember absolutely if there were charge changes involved for bromide, but there probably were. This was about 5 years ago.
    I remember that there were certainly charge changes for the actual DNA cases, but I can’t remember exactly how that was handled with the counterions. At the end I was running simulations as suggested in the PDF posted at the start of this article, with separate transformations for charges and VdW radii and using “soft cores.” I still couldn’t get any sort of reliable correlation with experimental strand melting temperatures.
    Without throwing yet more computer power at it — and I was already using a lot for the time — I couldn’t see if the results improved with further sampling or if it was an inherent limitation of the model. I was trying to stay as close to standard AMBER 95 as possible for the actual force field parameters, but assigned new RESP partial charges to atoms near the lesion.

  35. Chris says:

    @32, I use MOE (http://www.chemcomp.com/) and would recommend it highly. You can use “skins” to change the interface to suit the task in hand, there is a Medchem skin you could look at.

  36. anon the II says:

    @#32 Michael
    If you don’t have any money, you might try Avogadro.

  37. FreeEnergyGuy says:

    The easiest and slickest full service modeling package is MOE. The package with the strongest science is probably Schrodinger. Neither is free, although MOE is inexpensive (relatively speaking) for personal use.
    If you have no cash, you can try PyMol for visualization, Amber for minimization/MD/free energy/etc., DOCK (UCSF) for docking, etc. All are free, though, except for PyMol, all have a steeper learning curve than the commercial packages.

  38. Evorich says:

    All attempts to make free energy calculations faster (i.e. MMPBSA) have largely failed. Proper FEP calculations are still not really possible in a reasonable timescale for med.chem.
    Faster QM methods are now coming to the fore and, as long as you keep a handle on when and where the changes you’re making actually are relative to solvent, can be phenomenally predictive.
    Of course this is just target binding! The real power of comp. chem. would be if the ADMET predictor models were better. But given that this is never going to be more than empiracle, it’s unlikely the models will improve too much. The key is just having a model based as close to each chemical series as possible. Of course simple but well fed models like clogP have helped the field a lot.
    In general, good comp. chem. is about good communication with the chemists and good understanding, on both sides, of the power and the limitations of the methods.

  39. FreeEnergyGuy says:

    Productive modeling is a two way street and requires both a chemist willing to listen and a modeler capable of distilling minutia with error bars into broader suggestions.
    Anyone can carry out modeling calculations, and unfortunately, many times, anyone does.
    If you’re a chemist and the modeler you’re working with says “yes” every time you ask whether (s)he can answer a question, find another modeler. If your modeler tells you the difficulties and then suggests how to better construct a question that can be at least qualitatively addressed through modeling, you’ve got a keeper.

  40. Dr Van Nostrand says:

    @32
    Schrödinger is state-of-the-art nowadays, outperforming competitors in most if not all applications. Expensive, but tt’s like with red wine, you get what you pay for.

  41. Wavefunction says:

    In my experience, some of the blame also goes to experimentalists, of which I have mostly seen two extreme kinds; those who think modeling is sacrosanct and can solve any problem, and those who think it can’t do diddly squat. There has to be a reasonable, balanced outlook towards the role of modeling.

  42. I absolutely agree with 41. Wavefunction. If all positives in the experiment are false positives, then the experiments must be wrong. Brian Shoichet’s lecture is very interesting in this regard. Again, it’s more important to have tools that would organize and annotate these results, rather than to have better tools for free energy calculations.

  43. Higgsboson says:

    Hmmmm, it can be done but it depends on how robust your binding site docking is, how many nodes your cluster has and how many compounds you want to screen. My experience with these kinds of programs, like liaison, is that changes to the ligand structure has to be small in order to have real predictive value between analogs. And we are talking between analogs in a series here. Between series? Ehhhhh. If you really want to narrow it down, do a first pass virtual screen with a docking program, then a more rigorous free energy screen with your clustered hits.
    Can probably be done in a couple of weeks provided you start with a cross validated docking site.
    One thing you can’t account for is the presence of waters in your binding site. That adds another layer of complexity that makes virtual screening more of an art than a science,

  44. I am glad to be a visitor of this thoroughgoing blog! , regards for this rare information! .

  45. Thank you John, I said I would not check this forumanymore due to all the negative talk. However my friend kept reading and bought your thoughts to my attention. I back you and you defiantly make more sense than anyone. The one factor that you did not go into that upsets me so much is the reasoning some people have used that ect is given to patients to make money! That is crazy. Thank you for showing more positive aspects of ect. I can honestly say without the ect I would have had a successful suicide attempt, instead now I smile again, laugh, have goals. And the severe memory loss ro me was a gait trade for my life back for me and my family.

  46. I love looking through a post that can make people think. Also, many thanks for allowing me to comment!

  47. I could not refrain from commenting. Perfectly written!

  48. Hello there! I could have sworn I’ve visited this website before but after going through some of the posts I realized it’s new to me. Nonetheless, I’m definitely delighted I discovered it and I’ll be book-marking it and checking back regularly!

  49. Everything is very open with a precise description of the issues. It was really informative. Your website is very useful. Thanks for sharing!

  50. inforegistry says:

    You are so interesting! I do not suppose I have read through a single thing like this before. So good to discover somebody with some original thoughts on this topic. Seriously.. many thanks for starting this up. This site is one thing that is needed on the internet, someone with a bit of originality!

  51. Everything is very open with a precise explanation of the challenges. It was really informative. Your site is extremely helpful. Thanks for sharing!

  52. eliran oved says:

    Excellent article! We are linking to this great post on our site. Keep up the great writing.

  53. love spells says:

    Hello there! This post couldn’t be written much better! Going through this post reminds me of my previous roommate! He always kept preaching about this. I’ll send this information to him. Pretty sure he’s going to have a good read. I appreciate you for sharing!

  54. fdpod.com says:

    Hi there! This blog post could not be written any better! Reading through this article reminds me of my previous roommate! He constantly kept talking about this. I most certainly will forward this article to him. Fairly certain he’ll have a very good read. Thank you for sharing!

  55. cigar says:

    Whether your unit is being foreclosed or you are searching for a storage foreclosure unit to purchase

  56. I truly love your blog.. Great colors & theme. Did you create this amazing site yourself? Please reply back as I’m attempting to create my own site and would like to know where you got this from or just what the theme is named. Kudos!

  57. Oh my goodness! Incredible article dude! Thank you so much, However I am going through issues with your RSS. I don’t know why I cannot join it. Is there anyone else having the same RSS problems? Anybody who knows the solution can you kindly respond? Thanks!!

  58. pvc-banners says:

    Saved as a favorite, I really like your website!

  59. I blog frequently and I truly appreciate your information. Your article has really peaked my interest. I am going to bookmark your website and keep checking for new details about once a week. I opted in for your Feed too.

  60. pvc-banners says:

    An impressive share! I have just forwarded this onto a co-worker who was conducting a little homework on this. And he actually ordered me breakfast simply because I stumbled upon it for him… lol. So allow me to reword this…. Thanks for the meal!! But yeah, thanx for spending the time to discuss this subject here on your blog.

  61. silk road 2 says:

    Can I simply say what a comfort to discover someone who actually understands what they’re discussing on the internet. You certainly understand how to bring an issue to light and make it important. More people really need to check this out and understand this side of the story. It’s surprising you are not more popular given that you definitely possess the gift.

  62. lorenzo says:

    It’s hard to come by knowledgeable people on this subject, but you sound like you know what you’re talking about! Thanks

Comments are closed.