This document discusses the automation of computational chemistry calculations and protocols to reliably generate molecular property data. It addresses validating computational methods, analyzing results for errors and outliers, and comparing output to experimental data. The goal is to provide high-quality "experimental" data through automated high-throughput computation while ensuring valid results and identifying unusual computations. Workflows, parsing tools, and dissemination methods are presented for managing large numbers of jobs and analyzing results.

1. Computational Chemistry Robots ACS Sep 2005 Computational Chemistry Robots J. A. Townsend, P. Murray-Rust, S. M. Tyrrell, Y. Zhang [email_address]

2. Can high-throughput computation provide a reliable “experimental” resource for molecular properties? Can protocols be automated? Can we believe the results?

3. Aspects of complete automation Humans must validate protocols rather than individual data Low rates of error must be addressed Users should know the rates of error and degree of conformance

4. Approaches to conformance Explore limits of job behaviour (times, convergence, etc.) Analyse reproducibility Vary and analyse effects of parameters and algorithms Compare output with other “measurements” of same quantity

5. The overall view molecules computation dissemination

6. The overall view molecules computation dissemination Check results

7. Components of System Workflow for management of jobs (Taverna) Natural Language Processing based parsing of outputs (JUMBOMarker) Pairwise comparison of data sets (R) Analysis of mean and variance Detection and analysis of outliers

8. Computing the NCI database MOPAC PM5 a a MOPAC PM5 – collaboration with J.J.P. Stewart

9. Protocol Log Files Parse System Crashes Science Errors Analysis Pathological Behaviour Statistics Other Science Disseminate Results Unsuitable Data Program Crashes Inform Developer

10. Taverna Workflow programs allow a series of small tasks to be linked together to develop more complex tasks Open Source myGRID, eScience European Bioinformatics Institute University of Manchester

11. An Example Taverna Workflow

12. Parsing Log Files to CML Coordinates Molecular Formula Calculation Type Point Group Dipole Total Energy Computational Chemistry Log Files

13. CompChem Output Coordinates Energy Levels Vibrations Coordinates Energy Level Vibration CML File CMLCore CMLCore CMLComp CMLSpect Input/jobControl General Parsers

14. Dissemination of results LOG FILE CML FILE HUMAN DISPLAY WWMM* Server and DSpace Outside world JUMBOMarker NLP-based log file parser * World Wide Molecular Matrix

15. InChI: IUPAC International Chemical Identifier A non-proprietary unique identifier for the representation of chemical structures. A normal, canonicalised and serialised form of a chemical connection table. InChI FAQ: http://wwmm.ch.cam.ac.uk/inchifaq/

16. Proteus molecules * Calculation JUNK Cured by MOPAC * Proteus was a shape changing ocean deity

17. Proteus molecules Calculation Input JUNK

18. How do we know our results are valid? Computational Method 1 Computational Method 2 Experiment

19. J.J.P. Stewart’s example Calculated  H f – Expt  H f

20. GAMESS MOPAC results GAMESS a 631G* B3LYP Log Files a Project with Kim Baldridge and Wibke Sudholt

21. Protocol Log Files Parse System Crashes Science Errors Analysis Pathological Behaviour Statistics Other Science Disseminate Results Unsuitable Data Program Crashes Inform Developer

22. Repeat runs, different methods Multiple runs give same final structure from same input Changing memory allocation doesn’t make a difference

23. Pathological behaviour - Early detection 100 min 631G*, B3LYP 200 min 15 min 631G*, B3LYP 10080 min divinyl ether trans-Crotonaldehyde Z matrix

24. Times to run jobs

25. Analysis of different computational methods Mean - Overall difference Normality - Distribution of values Outliers - Unusual molecules? Variance - Spread of the data, depends on both distributions. (standard deviation)

26. Probability Plot (Normal QQ plot)

27. Mean of distribution (Approx - 0.03 Å ) Range over which sample distribution is approximately normal Outliers Probability Plot (Normal QQ plot) S.D. 0.020 Å

28. All bonds*  r (MOPAC – GAMESS) / Å * Excludes bonds to Hydrogenc

29. All bonds*  r (MOPAC – GAMESS) / Å Good agreement Nearly normal Outliers S.D. 0.005 Å * Excludes bonds to Hydrogenc

30. 2- Bad molecules and data usually cause outliers Na P O O H H

31. Mean  r (M - G) / Å Standard Error of the Mean / Å All values given to 3 significant figures C N O F S Cl C -0.006 0.020 -0.010 -0.014 -0.040 -0.037 0.000 0.000 0.000 0.001 0.001 0.001 N 0.006 -0.037 -0.055 0.001 0.001 0.009 O -0.087 -0.070 0.004 0.014

32.  r CC bonds (M - G) / Å

33.  r CC bonds (M - G) / Å Good agreement Nearly normal Outliers S.D. 0.013 Å JUNK

34. Selection of molecules with C C  r (M - G) > 0.05 Angstroms

35. Y = 0.0277 X – 0.0061 Non aromatic C C bonds adjacent to CF n

36.  r NN bonds (M - G) / Å

37. Good agreement Nearly normal Kink S.D. 0.022 Å  r NN bonds (M - G) / Å

38. Density plot of  r NN bonds (M - G) / Å

39. LEFT RIGHT Density plot of  r NN bonds (M - G) / Å

40. Most common fragments found in Left set but not Right set C(sp 3 ) C(sp 3 ) (sp 3 ) S(sp 2 ) N(ar) N (ar) C(sp 2 ) S(sp 2 ) N(ar) N (ar) C(sp 2 ) Or

41. GAMESS Log Files Comparison of theory and experiment CIF* CIF* CIF* CIF* CIF* CIF 2 CML * CIF: Crystallographic Information File

42. Reading Acta Crystallographica Section E

43. All bonds*  r (Cryst. – GAMESS) / Å Single molecules, no disorder * Excludes bonds to Hydrogenc

44. All bonds*  r (Cryst. – GAMESS) / Å Single molecules, no disorder Mean  r - 0.011 Å Nearly normal Outliers S.D. 0.014 Å * Excludes bonds to Hydrogenc

45.  r CC bonds (C – G) / Å

46. Mean  r - 0.01 Å Nearly normal S.D. 0.009 Å  r CC bonds (C – G) / Å

47.  r CO bonds (C – G) / Å

48. Good agreement Nearly normal Outliers ? S.D. 0.011 Å  r CO bonds (C – G) / Å

49.  r = +0.08 Å Chemistry can cause outliers H movement

50. Conclusions Protocols can be automated Machines can highlight unusual behaviour, geometries and distribution of results for humans to consider Computational programs can provide high quality “experimental” molecular properties

51. Thanks J.J.P. Stewart Kim Baldridge Wibke Sudholt Simon Tyrrell Yong Zhang Peter Murray-Rust Unilever

52. Questions Homepage: http://wwmm.ch.cam.ac.uk InChI FAQ: http://wwmm.ch.cam.ac.uk/inchifaq R: http:// www.r-project.org Taverna: http://taverna.sourceforge.net/ MOPAC 2002: http://www.cachesoftware.com/mopac/ GAMESS: http:// www.msg.ameslab.gov/GAMESS/GAMESS.html