300 likes | 838 Views
Computer-aided drug design – the next twenty years. A talk in commemoration of Yvonne C. Martin, given at the ACS session Mar 2007 in honor of her ‘retirement’. John H. Van Drie Novartis Institutes for BioMedical Research Cambridge, MA. Hugo, YCM, and Han.
E N D
Computer-aided drug design – the next twenty years A talk in commemoration of Yvonne C. Martin, given at the ACS session Mar 2007 in honor of her ‘retirement’. John H. Van Drie Novartis Institutes for BioMedical Research Cambridge, MA
Why 20 years? Predicting the future?? This commemorates work that Yvonne and I and many others at Abbott did twenty years ago, the first successful virtual screen (a pharmacophore search of a 3D database, which yielded a novel D1 agonist)1,2,3. “Predictions are hard… esp. about the future” – Yogi Berra Nonetheless, I think it’s safe to predict that At the 2027 CADD Gordon Conference, we’ll hear a talk on “Progress in scoring functions” And at the 2028 Comp. Chem. Gordon Conference, we’ll hear “Progress in polarizable force-fields” 1 JH Van Drie, D Weininger, YC Martin, JCAMD, 1989; 2 JW Kebabian et al, Am J Hypertens. 1990 3:40S2 YC Martin, JMC, 1992. The events described therein took place in the summer of 1987.
Why 20 years? Predicting the future?? Yvonne’s compatriot Peter Goodford concluded his conference on computational drug design in Erice, Sicily in 1989 with a list of things we had to work on. This list looks very modern, e.g. “we must improve homology modelling”, “we must predict solubility”. Today, I’m really not trying to predict the future. My aspiration is to provoke some thinking in all of you about where our field is heading (as Yvonne has so often done herself).
All new technologies tend to follow a similar path Peak of Hype EXPECTATION Asymptote of Reality Naive Euphoria True User Benefits TIME Overreaction to Immature Technology J. Bezdek, IEEE Trans. Fuzzy Sys., 1, 1-5 (1993) His figure put into PPT by J. D. Baker Depth of Cynicism
Designing drugs by computer at Merck In CADD, one can put dates on each of these turns Peak of Hype – 1989-1991 “we can design drugs atom-by-atom” EXPECTATION True User Benefits – 2000 and beyond Oct 5, 1981 Overreaction to Immature Technology TIME Depth of Cynicism – 1994-6, the era of “make ‘em all, let the assay sort ‘em out” Also, P Gund et al, Science, 1980
Yvonne chairs 2nd QSAR Gordon Conference But Yvonne was at work on QSAR far before 1980… Yvonne begins working w/ Corwin Peak of Hype – 1989-1991 “we can design drugs atom-by-atom” 2001 - QSAR GRC becomes CADD GRC EXPECTATION Yvonne publishes Quantitative Drug Design TIME 1960 1980 Corwin Hansch devises QSAR Depth of Cynicism – 1994-6, the era of “make ‘em all, let the assay sort ‘em out”
This forms the basis for my main projection for the future of CADD – this has been only a warmup Dramatically higher expectations EXPECTATION TIME 2027 1987 2007 I can’t say when the new wave will begin, nor can I imagine what will be the stimulus to kick it off
The key drivers of the evolution of CADD CADD is emerging as a sub-discipline of computational chemistry, distinct in its own right. Computational chemistry itself is an off-shoot of physical chemistry, sharing its paradigm of aiming to achieve atomic-level understanding of experimental phenomena. Like comp. chem., CADD aims to present explanations of experimental phenomena, but in addition aims to provide answers to the fundamental question of medicinal chemistry: What molecule(s) should be made next? This leads to things like virtual screening, virtual library design, de novo design, etc. – heresy to many academic comp. chemists.
The key drivers of the evolution of CADD • CADD focuses on the design and discovery of ligands and drugs. • To design a potent ligand, “all” it takes is: • To understand molecular recognition, and • To exploit that understanding in proposing new molecules to make. • Our understanding, #1, is astonishingly primitive, and #2 works best today in lead discovery (where lots of options are available, and lots of predictions are tested), less well in lead optimization. However, recall too “It’s relatively easy to discover a potent ligand, it’s damned tough to discover a drug” – E. H. Cordes
Outlook #1: To gain a more accurate understanding of molecular recognition… We’ve relied too long on molecular dynamics (MD) to handle thermodynamics of ligand-protein interactions, e.g. free-energy perturbation. The results have fallen short of our high hopes. At a fundamental level, ligand-receptor interactions often display non-additivity. Yet, almost all of our energy functions1 and scoring functions2 are linear, i.e. implicitly assume additivity: However, many structure-activity relationships display non-additivity, like this Raf kinase SAR, that led to sorafenib3: 1 CHARMm force field; 2 HJ Böhm , JCAMD, 1994 3 RA Smith et al, BMCL, 2001
Outlook #1: …we’ll finally need to learn thermodynamics A proper thermodynamic treatment naturally leads to a description of non-additivity. 1,2 One area in a hot ‘naïve euphoria’ phase are methods for treating thermodynamics of ligand-protein interactions better: - Gibbs’ ensemble methods used by LOCUS (F. Guarnieri originally), and related things at other companies (Bioleap, Vitae, SolMap). Stems from work of M Mezei at Mt Sinai and others. - Internal coordinate methods (R. Abagyan, M. Jacobson) allow greatly increased sampling vis-à-vis MD. - Ken Dill, Rob Phillips et al. published in 2006 new equations for statistical dynamics of non-equilibrium systems (“principle of maximum caliper”, Am J Phys, 74:123, 2006) – a bolt of lightning with as yet no thunder. 1 K. Dill, JBC, 1997; 2 JH Van Drie, manuscript sitting on my desk for years
Outlook #2: We’ll get much better at understanding what it takes to turn a potent ligand into a drug The attrition rates of drug candidates in clinical trials are staggering – we’re throwing lots of money down the drain, and, more importantly, the fruits of peoples’ creativity. The more that we understand why molecules fail, the better we’ll be able to design molecules that don’t. This is the grand challenge of drug design in the next 20 years. See, for example, S. Biller et al, “The Challenge of Quality in Candidate Optimization,in Borchardt RT, eds. Pharmaceutical Profiling in Drug Discovery for Lead Selection, 2004. Figure from Kola & Landis, Nat Rev Drug Disc, 2004
The best example of our recently-increased understanding of a liability: hERG and long QT Outlook #2: We’ll get much better at understanding what it takes to turn a potent ligand into a drug We now have atomic-level understanding of binding to the hERG channel, mediator of the clinical LQT syndrome Outlook #2.1: We’ll see a lot more of this type of stuff. R. A. Pearlstein et al, BMCL, 2003
Outlook #2: We’ll get much better at understanding what it takes to turn a potent ligand into a drug However, we don’t need to model all liabilities at the molecular levelWe have tons of data, and are getting more. We tend not to use it outside the chemical series for which it was developed. Outlook #2.2: our methods for computationally learning from data will get much better (they stink now). People thought SVM’s would be our salvation – hasn’t happened. Outlook #2.3: we’ll get much better at building empirical 3D models. Something will come along to replace CoMFA/CoMSiA, and better alignments will arise via improved pharmacophore methods. Outlook #2.4: Use of pharmacophores will grow. The science is there to create simple pharmacophore models of each receptor-mediated liability for which in vitro data is available (e.g. off-target GPCR’s). We need to just do it. Outlook #2.5 We’ll be able to find the data we need. Figure from JH Van Drie, IEJMD, 2007
Outlook #3: We’ll be led into new classes of drug targets – ones that challenge our competencies Protein-protein interactions (PPI’s) are thought to be a nearly-impossible challenge as drug targets. Yet, we’re starting to crack them. If one figures that there’s ~30,000 genes in the genome, that gives us ~30K protein targets, but 30K x 30K = ~ 1 billion PPI pairs as targets. Lots of opportunity, once we figure out how to wrestle these to the ground. This shows Novartis’ success in designing inhibitors to IAP, mimicking part of the SMAC interaction partner. (C. Straub, Keystone Symposium April 2006).
Outlook #4: wild idea: self-assembling drugs Exjade is a new Novartis drug for iron chelation therapy. Two divalent molecules together form a tetravalent complex of iron. S. Stupp et al at Northwestern are investigating something even more bizarre: molecules that self-assemble around a blood vessel to promote neovascularization: Note how this allows us to design small molecules to slip across the gut wall, but to reassemble to bigger things at the site of action. To design these, we must understand thermodynamics (G. Whitesides). K. Rajangam et al & S. Stupp, Nano Lett, 2006; GW comment made at MIT-Novartis Nanotechnology Symposium, Nov, 2006
Outlook #5: pathways and “systems biology” – it’s not enough to think about inhibiting one target A breakthrough in our understanding of how HIV causes AIDS came from mathemetical modelling of the entire system A Perelson, et al, “HIV-1 dynamics in vivo: virion clearance rate, infected cell life-span, and viral generation time.”, Science, 1996. Our target selection must take into account the entire signalling network or pathway. For example, the cellular phenotypes of inhibiting each of these kinases in the same pathway is totally different MAP MEK ERK Most of this modelling up to now has been done by analogy to electrical circuits, i.e. numerically solving coupled ordinary differential equations. But does the proper approach reside here?
Outlook #6: Things that require an expert today will be on chemists’ desktops tomorrow It’s mainly an issue of building intuitive user interfaces. We tried this with Catalyst (1990-1994) but failed. At Novartis, we’re putting sophisticated methods on chemists’ desks, called FOCUS. Also, “the slow one now will later be fast…” – Bob Dylan, 1964
Outlook #7: Virtual screening will become routine I anticipate that virtual screening will become as routine as HTS is now. The driver of that will be the growing appreciation of the importance of speed. VS can provide a chemical starting point relatively quickly. HTS is more comprehensive, but when all the assay-reformatting, etc. is accounted for, it takes much longer. It’s quite a surprise that it’s still relatively rare in Big Pharma, despite it having been introduced 20 years ago. For a recent overview of methods and applications, see Shoichet & Alvarez, Virtual Screening, 2005
In summary, these are my outlooks for the next 20 years in CADD • “Computational thermodynamics” will flower • Increased ability to turn a potent molecule into a drug • Use molecular understanding for receptor-mediated off-target liabilities, e.g. hERG • Our computer learning methods will greatly improve, to allow us to build good empirical models • We’ll get much better at building empirical 3D models • We’ll have at least pharmacophore models of each receptor-mediated off-target liability • We’ll be able to find the data we need.
In summary, these are my outlooks for the next 20 years in CADD (cont’d) • We’ll conquer challenging new classes of drug targets, e.g. PPI’s • We’ll learn to design self-assembling drugs • We’ll use our knowledge of pathways to predict which targets provide the best intervention point • Sophisticated CADD methods will be on the desktops of medicinal chemists. What is fancy today will be routine tomorrow. • Virtual screening will become routine.
And, finally… Thanks, Yvonne, for introducing me to such an endlessly fascinating line of work.