VIEWS: 4 PAGES: 1 POSTED ON: 2/20/2010
Molecular Similarity as a guide to Safety Pharmacology Fiona T. Mc Cahey1*, Philip M. Marsden1, Andreas Bender1, Jos W. M. Tissen2, Werner Klaffke2, Robert C. Glen1 1 Unilever Centre for Molecular Informatics, Chemistry Department, University of Cambridge, Cambridge CB2 1EW, United Kingdom 2 Unilever Research Vlaardingen Laboratory, Olivier van Noortlaan 120, NL-3133AT Vlaardingen, The Netherlands * Corresponding author email address firstname.lastname@example.org http://www-ucc.ch.cam.ac.uk/ Introduction Results The prediction of both the main therapeutic activity and side-effects (Safety Pharmacology) of both existing and new drugs is of major importance to the Using 5-fold cross-validation, the average correct hit rate (number of compounds correctly predicted to be active and inactive) pharmaceutical industry. A recent estimate suggests that it takes 10-15 years to using this method is 90.37% over 154 activity classes (Table 1). Ranking of classes shows that 76% of compounds were correctly produce a new drug at an average cost of U.S.$800 million (1). Much of this cost predicted to be in the top ranked class and 90% were in the top 3 (Table 2). Therefore it appears that SARs created by MOLPRINT could be reduced if there was a reliable and inexpensive way in which to determine 2D can differentiate between drugs that act at different target receptors. Centroid-linkage hierarchical clustering indicates that data the in vivo effects of a compound. We present our work to create an in silico screen, fall into clusters corresponding to receptors of similar type (e.g. aminergic GPCR receptors) (Figure 3) . ALL of the main 1. 5HT reuptake inhibitor which is analogous to an in vitro receptor screening assay, using fragment-based therapeutic effects of the top 10 selling drugs were predicted correctly using the model. Some of the main side-effects can also be 2. 5HT1A agonist molecular similarity methods. We have used circular fingerprints to analyse explained using the model (Table 3). 3. 5HT1A antagonist approximately 50,000 biologically active compounds. The top 10 selling drugs in the 4. 5HT2A antagonist U.S. in 2004 (2) have been tested in the model, and their predicted therapeutic and 5. Adrenergic (alpha1) blocker Rank position % correct side-effect profiles have been compared to their known clinical effects. 6. Dopamine D4 antagonist 1 75.7 7. Dopamine D2 antagonist % predicted correct incorrect Top 2 86.1 Figure 3: Centroid-linkage hierarchical clustering model, created in R (6). Small section enlarged to show Method active 84.36 15.64 Top 3 89.7 seven individual classes in a cluster. Database Used Approximately 50,000 biologically active compounds in 154 activity classes from the MDL Drug Data Report inactive 96.38 3.62 Top 5 92.9 database (MDDR) (3) were chosen for characterisation. Discussion MOLPRINT 2D Average 90.37 9.63 Top 10 96.4 Molecular similarity does appear to be useful both in identifying the principal mode of The MOLPRINT 2D method uses atom environment descriptors (Figure 1), information-based feature selection and a naïve Bayesian classifier to classify compounds in each activity class (4). action of a drug molecule and in identifying some side effects. The correct hit rate of Descriptor Generation: Atom Environments, Table 1: Results of 5-fold cross-validation Table 2: Results of ranking all activity classes for each molecule 90.37% over 154 classes (as shown by cross-validation) using MOLPRINT2D can be (Circular Fingerprints) for every atom in the molecule are compared to that of the PASS (Prediction of Activity Spectra for Substances) program, calculated using bond distances from 0 up to 2 bonds. SYBYL® ( atom types are used to derive the environments as which has a reported hit rate of 85% over 1000 types of biological activity (8). To improve they partially capture physicochemical properties this hit rate further, the occurrence of false positives (compounds that have been (hybridisation, geometry and electronic structure). predicted as active at a receptor, but are not reported as such in the MDDR database) Figure 1. Atom Environment Descriptor generation could be due to the presence of many highly correlated receptor classes for which there Calculations Performed Rank Structure Proprietary Name Clinical effect Main side-effects Predicted Target from Clinical effects Explained? may only be one activity reported in the database (of course, this dataset is not (generic name) and Target Molprint analysis comprehensive and contains activities, not inactivities). All of the therapeutic effects of 1.MOLPRINT2D was used to create qualitative Structure Activity relationships (SARs) for each activity class. 1 LIPITOR Treats hypercholesterolemia by Reversible myositis, myalgia, HMG CoA reductase (beta) Therapeutic effect – yes. Side effects the top 10 selling drugs were predicted correctly using the models. Some side-effects competitively inhibiting HMG CoA myopathy inhibitor are not well understood (Atorvastatin) can also be explained, in particular those side-effects that can be characterised by 2.Euclidean distances and Pearson’s correlation coefficients between all (154) activity classes were reductase calculated. predicted activity due to specific receptors (e.g. blood pressure lowering from agonists at 2 ZOCOR Treats hypercholesterolemia by Reversible myositis, myalgia, HMG CoA reductase (beta) Therapeutic effect – yes. Side effects the 5HT2A receptor). Side-effects such as “headache” which have many physiological competitively inhibiting HMG CoA myopathy inhibitor are not well understood 3.Centroid-linkage hierarchical clustering was carried out on the data using R (6) (Figure 3). (Simvastatin) reductase and psychological reasons for occurrence may not be specific to individual receptors and are not particularly easy to explain by this approach. However, further literature searches 4.Pearson’s correlation coefficients between ALL (452) activity classes in the MDDR were calculated to create an in vitro-in vivo (target-effect) matrix. This can be used to find if the data suggests that certain targets (e.g. 3 PREVACID Ulcer healing by inhibiting “proton GI disturbances (nausea, vomiting, H+/K+ -ATPase Inhibitor Therapeutic effect- yes. Side-effects could indeed identify measured effects similar to those predicted. Using hierarchical pump” abdominal pain, flatulence) caused by many factors HMG-CoA reductase inhibitor) can be linked to certain effects (e.g. Hypolipidemic). (Lansoprazole) clustering (based on the occurrence of the fragments present in the compounds making up each of the classes) many clusters appear to contain very similar receptor types. This 5.A comparison of all activity classes was performed by combining the SARs from each individual activity 4 NEXIUM Ulcer healing by inhibiting “proton GI disturbances (nausea, vomiting, H+/K+ -ATPase Inhibitor Therapeutic effect- yes. Side-effects is consistent with similar binding and activation characteristics being present in these class as determined by MOLPRINT 2D. A 5-fold cross-validation and ranking of all classes against individual (Esomoprazole) pump” abdominal pain, flatulence). caused by many factors compounds was carried out. The rationale behind assigning a molecule as belonging to an active or inactive Dermatitis molecules (pharmacophores). set in the cross-validation can be seen in Figure 2. 5 PROTEIN PROCRIT CANNOT be predicted by Future Work 6. A list of the top 10 selling drugs in the U.S. in 2004 was created. Structures were collected from various (Epoetin alfa) model sources and information on therapeutic effects and side-effects was collected from the British National A more specific and comprehensive database containing activities at each receptor Formulary (BNF) (7). 6 ZOLOFT Antidepressant -SSRI Nausea, vomiting, GI bleeding Glutamate receptor Therapeutic effect-yes. Target has high (positive and negative) would be beneficial in model building and validation. Access to disorders antagonist correlation to Antidepressant and (sertraline) Antipsychotic effects this type of data will be pursued. Also, investigation of additional/alternative descriptors 7. The top 10 selling drugs were tested in the model to find if their therapeutic and side-effects were correctly predicted. (such as clogP or MOLPRINT3D) will be tested in order to improve the current models. 7 PLAVIX Prevention of athersclerosclerotic Dyspepsia, abdominal pain, Somatostatin Antagonist, Therapeutic effect-yes. Targets have (clopidogrel) events diarrhoea, GI bleeding, intercranial Bradykinin BK2 antagonist high correlation to Anticoagulant and Additional comparisons with other methods and data will be informative in determining bleeding antiaggregatory effects. This may also S = SCORE be the cause of bleeding side effects the usefulness of this approach. Testing of compounds such as the top 200 selling S drugs, new-to-the-market drugs, compounds that failed in clinical trials and traditional NA = total number in active set Chinese medicines will be pursued to ascertain if the model is reliable and makes NI = total number in inactive set useful testable predictions. 8 ADVAIR DISKUS Nasal allergy Dryness, irritation and ulceration of Corticosteroid Therapeutic effect-yes. Target has high NI NAO = number of actives below (fluticasone nose and throat. Raised Intra- correlation with Anti-inflammatory effect NA S proprionate) ocular pressure and glaucoma and Ophthalmic effect. The latter may be the reason for ocular side-effects Development of a comprehensive in-vitro and in-vivo (target-effect) ontology for both NAO NIO NIO = number of inactives above therapeutic effects and side-effects will allow better relationships to be developed INACTIVES ACTIVES S between receptor activities and medicinal effects. 8 ADVAIR DISKUS Bronchodilator- selective beta 2 Fine tremor (particularly in hands), Adrenergic (beta) agonist Therapeutic effect-yes. Side-effects- MOLPRINT Score (S) PA = NAO/NA (salmeterol) agonist headache, nervous tension Target has high correlation with spasmolytic effect (tremor) and References PI = NIO/NI anxiolytic (nervous tension) (1) IFPMA, 15 Ch. Louis-Dunant, PO Box 195, 1211 Geneva 20, Switzerland When PA / (PA + PI) > C, molecule is active C = cutoff, a user defined (2) NDCHealth, NDC Plaza, Atlanta, GA 30329-2010, USA PA / (PA + PI) < C, molecule is inactive 9 ZYPREXA Atypical antipsychotic many 5HT3 agonist, D1 Therapeutic effect-yes. Atypical number which biases model (olanzapine) antagonist, 5HT2A antipsychotic drugs act at many (3) MDL Information Systems, Inc., 14600 Catalina St., San Leandro, CA 94577-6608 antagonist, D2 antagonist receptors in the CNS towards assigning a compound (4) Bender, A.; Mussa, H. Y.; Glen, R. C.; Reiling, S.; J. Chem. Inf. Comput. Sci. 2004; 44(1); 170-178 as active or inactive. C was (5) SYBYL 7.0, Tripos Inc., 1699 South Hanley Rd., St. Louis, Missouri, 63144, USA optimised to reduce numbers of false positives and false 10 CELEBREX NSAID to treat rheumatic disease and GI discomfort, nausea, bleeding, Neurotensin receptor Therapeutic effect-yes. Target has high (6) Development Core Team (2004). R: A language and environment for statistical computing. R Foundation for (celecoxib) gout. Also treats familial adenomatous ulceration antagonist correlation with anti-arthritic effect, anti- Statistical Computing, Vienna, Austria negatives polyposis (FAP) inflammatory effect. Side-effect: target has high correlation with agent for (7) Joint Formulary Committee. British National Formulary. 49 ed. London: British Medical Association and Royal inflammatory bowel disease effect Pharmaceutical Society of Great Britain; 2005 Figure 2: Assigning membership of a molecule, based on its MOLPRINT score to the active or inactive set (8) Poroikov, VV; Filimonov, D.A; Borodina, Y.V.; Lagunin, A.A.; Kos, A. J. Chem. Inf. Comput. Sci.2000, 40,1349 Table3: Results of testing the top 10 selling drugs in the U.S. in 2004 in the model. Acknowledgments We would like to thank Unilever for funding
"Slide 1 - Informatik-Rechnerbetri"