Analyzing Success Rates of Supposedly Easy Reactions

Transcription

1 Analyzing Success Rates of Supposedly Easy Reactions Roger Sayle and Daniel Lowe NextMove Software, Cambridge, UK

2 Symposium overview This symposium is entitled Retrosynthesis, synthesis planning, reaction prediction: When will computers meet the needs of the synthetic chemist?. I believe that different chemists (academic chemists, industrial chemists, process development chemists, medicinal chemists) each have different needs and perhaps different expectations for synthesis tools. Traditionally, retrosynthetic analysis/planning software has aimed to satisfy the top 1% of experts. Instead, I wonder about a different use-case.

3 Analysis of pharmaceutical elns NextMove Software s HazELNut software is used to export and analyze ELN content at 6 of the top 10 large pharmaceutical companies. In-house analysis of this data, across the industry, reveals a surprisingly high rate of synthesis failure, not indicated in the published literature (journal articles, patent applications or reaction databases). Understanding the causes of these failures is perhaps more significant than attempting to access new chemistries.

4 Eli lilly s automated synthesis lab Alexander G. Godfrey, Thierry Masquelin and Horst Hemmerle, A Remote-Control Adaptive MedChem Lab: An Innovative Approach to Enable Drug Discovery in the 21 st Century, Drug Discovery Today, Vol. 18, Nos , September 2013.

5 Synthesis failures at lilly At the 2013 Sheffield Cheminformatics conference, Christos Nicolaou highlighted the technical challenge with predicting compounds potentially accessible by the Lilly s Advanced Synthesis Lab (ASL). Predicted > 150 billion compounds with MW < 500. In a proof-of-concept pilot project, only 25 of 90 compounds suggested by Lilly s Annotated Reaction Repository (LARR) rule-set could be successfully synthesized in practice.

6 Synthesis failures at gsk Alas, poor success rates are not unique to Lilly. For example, Pickett et al describe the parallel synthesis of a 50x50 library of MMP-12 inhibitors by an iodo-suzuki coupling reaction. 50 R1 x 50 R2 = 2500 cmpds Stephen D. Pickett, Darren V.S. Green, David L. Hunt, David A. Pardoe and Ian Hughes, Automated Lead Optimization of MMP-12 Inhibitors using a Genetic Algorithm, ACS Medicinal Chemistry Letters, Vol. 2, pp , 2011

7 Libraries in practice Objective was to be diverse complete, so variations of standard route and conditions were required to avoid compromising diversity. Despite best efforts, 566 not made due to poor reactivity, unanticipated side reactions or product stability compared to 26 not assayed, 28 assay failed and 176 inactive, giving 1704 assay results. Interestingly, R1=21&R2=7 and R1=31&R2=25 were the two most active compounds (pic50=8), R1=21&R2=25 had a pic50=7.8 and R1=31&R2=7 was never made!

8 Is there anything that can be done? Perhaps the only significant source of information on failed and low yield reactions are industry electronic lab notebooks. The published literature, and databases derived from it, almost entirely describe only synthesis successes. Patents disclose best practice, and are therefore self selecting for highest yield examples. Academic research is often in paper notebooks. This lack of negative data introduces a belief bias.

9 Context: the hazelnut pipeline Accelrys Pipeline Pilot (AstraZeneca, AbbVie & Hoffmann-La Roche) HazELNut Filbert NameRXN Cobnut ChemAxon JChem Cartridge (GlaxoSmithKline & Novartis) Perkin Elmer Informatics (formerly CambridgeSoft) enotebook v9, v11 or v13 or Symyx ELN v5.x or v6.x Oracle Server version 10, 11 or 12 Microsoft Windows, Linux or Mac OS X Elsevier Reaxys (Hoffmann-La Roche, AstraZeneca, Merck) Chicken and Egg: The same data analysis that identified the high rates of reaction failures, can also be used to diagnose and understand the causes of these failures.

10 Transforms vs. reactions Importance of Reaction Mechanism Example: Ullman-type Coupling Reactions SMIRKS: [H][N:1].[Cl][c:2]>>[N:1][c:2] The SMIRKS transform alone is insufficient to predict the products and by-products in this example. A measure of nucleophility is desirable for each atom in a molecule. Without this software may be misled into believing that protecting groups are required.

11 Learning from failure Nadine et al [1] hypothesize that low LogP is a major cause of synthesis failure in parallel synthesis of combinatorial libraries. Analysis confirms that this is indeed a significant factor for the GSK MMP-12 library compounds measured, mean logp = 3.56 (1.44) 566 compounds not made, mean logp = 2.83 (1.52) Student s t-test for different distributions, p<2x Nadine, Hattotuwagama and Churcher, Lead-Oriented Synthesis: A New Opportunity for Synthetic Chemistry, Angew. Chem. Int. Ed, 51:

12 Nadine-churcher hypothesis Sucessful Reactions Failed Reactions 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% < >6.0 Reaction Product Predicted clogp 63 The clear trend between Suzuki coupling success rate and predicted octanol-water partition co-efficient.

13 CHEMICAL reactions for free

14 Example reaction mining INPUT Methyl 4-[(pentafluorophenoxy)sulfonyl]benzoate To a solution of methyl 4-(chlorosulfonyl)benzoate (606 mg, 2.1 mmol, 1 eq) in DCM (35 ml) was added pentafluorophenol (412 mg, 2.2 mmol, 1.1 eq) and Et 3 N (540 mg, 5.4 mmol, 2.5 eq) and the reaction mixture stirred at room temperature until all of the starting material was consumed. The solvent was evaporated in vacuo and the residue redissolved in ethyl acetate (10 ml), washed with water (10 ml), saturated sodium hydrogen carbonate (10 ml), dried over sodium sulphate, filtered and evaporated to yield the title compound as a white solid (690 mg, 1.8 mmol, 85%).

15 Example reaction mining Output

16 reaction ontology Reactions are classified into a common subset of the Carey et al. classes and the RSC s RXNO ontology. There are 12 super-classes e.g. 3 C-C bond formation (RXNO: ). These contain 84 class/categories. e.g. 3.5 Pd-catalyzed C-C bond formation (RXNO: ) These contain ~300 named reactions/types. e.g Negishi coupling (RXNO: ) These require >675 SMIRKS-like transformations.

17 categorization of ELN reactions 2% 5% Heteroatom alkylation and arylation 1% 10% 6% 15% 3% 2% 5% 17% 34% Acylation and related processes C-C bond formations Heterocycle formation Protections Deprotections Reductions Oxidations Functional group conversion Functional group addition Resolution 1. J. Carey, D. Laffan, C. Thomson, M. Williams, Org. Biomol. Chem. 2337, S. Roughley and A. Jordan, J. Med. Chem. 54: , 2011.

18 10 most popular reactions ID Name Count Carboxylic acid + amine 26, Buchwald-Hartwig amination 22, Suzuki coupling 16, Williamson ether synthesis 15, Amide Schotten-Baumann 11, Nitro to amino 10, N-Boc deprotection 9, CO2H-Me deprotection 9, CO2H-Et deprotection 6, Sulfonamide Schotten-Baumann 6,223

19 Most/least successful reactions ID Name Mean Yield Count Diazomethane esterification 91% Carboxylic acid to acid chloride 88% Bromo to azido 85% Methyl esterification 84% Bromo to iodo Finkelstein reaction 82% N-Cbz deprotection 81% Larock indole synthesis 47% Ullmann-type biaryl coupling 44% Chan-Lam ether coupling 44% Pinner pyrimidine synthesis 39% 47

20 Suzuki couplings as a percentage of reactions in a year Trends in Reaction Types 8.0% 7.0% 6.0% 5.0% 4.0% 3.0% 2.0% 1.0% 0.0%

21 Suzuki coupling leaving groups Leaving Group Mean Yield N Observations Bromo 58.80% Chloro 57.96% 2752 Iodo 57.21% 2049 Triflyloxy 65.48% 717

22 Nadine-churcher hypothesis On 16,335 Suzuki coupling reactions extracted from US patent applications between 2001 and LogP Mean Yield N Obs < % % % % % % 2126 > % 2720

23 Occurrences Multi-step Synthetic Routes Intermediates Terminal Products Intermediates Terminal Products Number of steps

24 Rxn searching: make or Break In order to simplify the learning curve for exploiting reaction databases, an easy to use query mechanism has been developed for frequent use-cases. Query: Retrieve reactions that make indoles. Traditional interfaces provide either substructure search or require reaction centers/atom mapping. A more convenient (robust) mechanism is to check whether (or count) that a drawn substructure appears in the products but not in either the reactants or agents.

25 Avoiding uninteresting results

26 bond changes in indole synthesis Synthesis A B C D Baeyer-Emmerling M Bartoli M M Bischler-Möhlau M C M Fischer M C M Fukuyama M Hemetsberger M Larock M C M Mandelung M Nenitzescu M M Reissert M M

27 Application to planning 1 Cinnamic Acid (PhCHCHCO2) 1. Bromo Heck reaction (272) 2. Horner-Wadsworth-Emmons reaction (268) 3. Wittig olefination (129) 4. Bromo Heck-type reaction (62) 5. Iodo Heck reaction (49) 6. Triflyloxy Heck[-type] reaction (43) 7. Ester Schotten-Baumann (10) 8. Bromo Suzuki coupling (5) 9. Stille reaction (2) 10. Olefin metathesis (1)

28 Application to planning 2 p-nitrobenzoic acid p-nitrotoluene 1. Nitrile to carboxy (12) 2. CO2H-Me deprot (8) 3. CO2H-Et deprot (5) 4. Ester hydrolysis (1) 5. Nitration (1) 1. Nitration (96) 2. Bromo Suzuki-type (1) 3. Chloro Suzuki (1)

29 Experimental validation Synthesis of a novel aromatic heterocycle previously unreported in the literature. William Pitt et al., Heteroaromatic Rings of the Future, Journal of Medicinal Chemistry, 52(9): , 2009.

30 Cautionary note: scale vs yield US applications, Suzuki couplings

31 Acknowledgements NextMove Software Daniel Lowe Noel O Boyle John May Thank you for you time. Questions? AbbVie AstraZeneca Bristol-Myers Squibb Eli Lilly GlaxoSmithKline Hoffmann-La Roche Merck Novartis Royal Society of Chemistry Vernalis Vertex Pharmaceuticals