1. The Basic Technology Research Programme Proof of Concept Studies & Consortia Building Networks

2. Background Cross research council endeavour –administered by EPSRC Funding for research to create a new technology Change the way we do science Underpin the future industrial base

3. Background 15 research projects funded up to April 2003 Total funding for this period - £41M To support large, long term, high risk, high impact research consortia Encourage investigation of speculative ideas

4. Background Two levels of funding –One year start up –Full grant up to five years Two types of start up funding –Proof of concept –Consortia building networking

5. Proof of Concept Studies One year funding up to £100K Research to investigate feasibility of developing the new technology Output – a business case for the next step of investigation to be submitted in May 2004 –Basic Technology Programme –Existing Research Council initiatives –DTI programmes

6. Consortia Building Networks Involvement of the users of the new technology at a very early stage Funding to form networks & hold workshops

7. ParaSurf – in silico Screening Technology Basic Technology Funding for October 2003 to September 2004 –Proof of concept –Consortia building networking Academic partners –University of Portsmouth –University of Erlangen –University of Southampton –University of Oxford –University of Aberdeen

8. ParaSurf – Proof of Concept Research Programme Development of techniques to describe irregular solids & surfaces Development of projection & pattern recognition techniques for non- planar colour-coded surfaces –spherical harmonics, molecular topology Conformational analysis Rigid body dynamics incorporating surface features –rigid parts of molecule treated as anisotropic solids linked by rotatable bonds Investigate how best to generate prediction models using surface properties that define a low dimensional chemical space –QSAR, pattern recognition, artificial intelligence, analysis of surfaces Bench marking using Grid computing

9. ParaSurf – Proof of Concept Research Programme

10. Potential applications of the in silico screening technology High throughput virtual docking Physical property mapping ADMET prediction Long time-period simulation techniques Crystallisation and solubility Prediction of tautomers Chemical reactivity and metabolism

12. ParaSurf Progress Report Letchworth, 16 th March 2004

13. Main Areas 1. Molecular Surfaces and Property Calculation 2. RGB Encoding & Pattern Recognition 3. Conformational Analysis 4. Rigid Body Molecular Dynamics 5. Analysis of Variables & QSAR models 6. Grid Computing 7. Consortium Building

14. Datasets Small Consensus Set of 74 Drug Molecules (diverse) QSAR set (31 CoMFA steroids) Medium WDI subset (2,400 comps) Harvard Chembank dataset (2,000 comps) Large WDI (50,000) Maybridge (50,000)

15. Example Molecule Allopurinol

16. Surface Definition & Local Property Calculation

17. Calculations 3D co-ordinates from CORINA QM calculations with VAMP Local Properties and surfaces from ParaSurf

18. ParaSurf v1.0 Surfaces Isodensity Surfaces Shrink Wrap Marching Cube Surfaces fit to Spherical Harmonics Properties MEP, LIE, LEA and LP Encoded at points on the surface Encoded as Spherical Harmonic Expansions

19. Small molecule

20. RGB Encoding & Pattern Recognition

21. RGB Encoding Each Local Property encoded as a colour LIE encoded on Red channel LEA encoded on Green Channel LP encoded on Blue Channel

22. Allopurinol RGB Surface

23. RGB Encoding Alternative Encoding LIE LEA Absolute value of MEP

24. Allopurinol RGB Surface

25. Conformational Analysis

26. Efficient All Atom MD analysis (DASH) Treated as time series (not Cluster Analysis) Scales linearly with simulation length No need for arbitrary choice of number of clusters Can be analysed using Markov Chain methodology

27. MD studies of Rosiglitazone

29. Rigid Body Molecular Dynamics

30. Rigid body molecular dynamics Well founded methodology e.g. CNS / XPLOR ( Axel T. Brunger, Stanford University ) Idea is to use rigid groups to model flexibility: In the ligand and the protein binding site. Allows time-steps of 10fs to 20fs.

31. QSAR models

32. Distribution of Properties

33. Correlation Matrix 1-0.10.470.39MEP -0.110.580.26LP 0.470.5810.44LEA 0.390.260.441LIE MEPLPLEALIE

34. Descriptors 34 descriptors based on Normal Distribution Principal Components Spherical Harmonic Co-efficients

35. Descriptors for LIE Maximum value of the local ionization energy Minimum value of the local ionization energy Mean value of the local ionization energy Range of the local ionization energy Variance in the local ionization energy

36. Other Descriptors Moments Order 1 – Mean Order 2 – Variance Order 3 – Skewness Order 4 – Kurtosis Overlapping Gaussians Derived from previous work on MD analysis

37. QSAR models Models derived from Local Properties Surface Integral Model for Solvation Energy RMS Error ~ 0.75 Kcal Drug Likeness SOMs trained on WDI (drugs) & Maybridge (general) Parameters from PC of Local Property Descriptors Medium sized datasets superimposed on SOMs

40. GRID Computing

41. ParaSurf compiled on SGI IRIX Windows Linux (SUSE) IBM AIX Future Platforms SUN Solaris GRID enabling at Portsmouth (Mark Baker), Southampton and Oxford.

42. Provisional Timings SGI R10k, 256MB VAMP ~ 30s/compound ParaSurf ~ 10s/compound Intel 1.8 Xeon/ AMD Athlon XP-2000+ ParaSurf ~ 2s/compound SGI FUEL Workstation R14K ParaSurf ~ 2s/compound

43. Conclusions

44. Properties can be calculated Properties can be RGB encoded Properties are local Properties can be used for QSAR models

46. Computer vision methods for comparing molecular surfaces Comparing and recognising 3D objects is an active research area in robotics and AI. Fast methods have been developed for database indexing. Rotationally invariant descriptors of 3D objects are possible.

47. Pattern matching on molecular surfaces Can we recognise similar surfaces? Can we recognise similar surface fragments? Can we identify the most similar surface to our target? How do we compare field descriptors on the molecular surface?

48. Rotationally invariant 3D object descriptors Internal coordinates e.g. a distance matrix. Energy distributions based on the spherical harmonics. The spherical harmonic coefficients. Radial integration, radial scanning, and invariant moments.

49. Surface comparison Two different approaches: 1.Using spherical harmonic molecular surfaces [ J. Comp. Chem. 20(4) 383-395; Ritchie and Kemp 2000; University of Aberdeen ]. 2.Partial molecular alignment via local structure analysis [ J. Chem. Inf. Comput. Sci. 40(2) 503-512 ; Robinson, Lyne and Richards 1999; University of Oxford ].

50. An example grid of surface points A grid is placed on a ParaSurf surface in order to reduce the number of surface points from 4038 to 55.

51. Partial molecular alignment We do not know which points on the two surfaces need to be aligned with each other. The essential approach is: all surface points on one surface are compared with all points on the other. For two surfaces, with M and N points, MN possible alignments are possible: – we want to reduce this large search space!

52. Voting pairs are possible alignments The voting pairs can have a critical effect on the quality of the surface alignment.

53. The voting table A voting table may list all matching pairs of surface points (i.e. all possible alignments). A smart editing of votes within the voting table can enable speed and accuracy. –We want to only consider alignments between similar local features on the surfaces. –The more false votes we have in the voting table the harder it is to find the optimum alignment.

54. A distance matrix can be used to describe local surface features P1 P2 The internal distance matrix can be used to distinguish between surface points. By comparing rows and columns from distance matrices of different surfaces we can detect similar surface features. P3

55. Selecting the voting pairs Similar local features, or interest points, on the molecular surface can be identified using a distance matrix. For a point on each surface: 1.Arrays of internal surface point distances are calculated for both points i.e. dist1[], dist2[]. 2.After a crude alignment, the absolute difference of dist1[] and dist2[] indicates the similarity of this pair of points.

56. Scoring the possible alignments The optimum alignment is composed of a rotation R and a translation T. Apply the current rotation r: 1.Score the translation vectors t = p – q of all voting pairs (p,q) using a gravitational potential: 2.High potentials identify clusters of similar translation vectors. 3.The vector with the highest potential is the optimum translation T. Scoring all r gives R and T.

57. Scoring with a gravitational potential Translation vectors (x,y coordinates plotted) Some voting pairs for example rotations

58. Can we use the potential to compare aligned structures?

59. Can we get better alignments with more voting pairs?

60. Example alignments 1 3 4 2

61. Example 1: RMSD = 0.75 A B

62. Example 2: RMSD = 1.05 A B

63. Example 3: RMSD = 1.20 A B

64. Example 4: RMSD = 1.89 A B

65. Matching with the surface field descriptors: example 1 Surfaces are aligned (using a quick search method; e.g. 45º rotations). Best N alignments are selected. Each alignment is gently perturbed and optimised using the field descriptors.

66. Matching with the surface field descriptors: example 2 Align using the field descriptors’ values to identify suitable voting pairs: –only match on similar field descriptors. Filtering can be achieved by aligning the fields separately. More accurate alignments can be generated by combining field values.

67. Parameterisation Voting pairs: –The distance between points in surface grid. –The number of voting pairs. –Identifying and selecting local features. –How to represent the fields at interest points. Scoring: –Scoring function to identify the correct rotation and translation (e.g. gravitational potential). –Target function to compare different surface alignments (e.g. RMSD). Optimising the alignments.

69. Molecular Surface Property Graphs Characterize the behaviour of a property f : S   on a molecular surface S, in terms of a directed graph G on S derived from the gradient vector field x = grad f(x) Vertices (G) = fixed points of grad f (= critical points of f ). Edges (G) = stable and unstable manifolds of the saddle points.

70. Gradient Flow minima saddles maxima

71. Molecular Surface Property Graph

72. Applications Similarity – Pattern recognition methods – Maximal common subgraphs Complementarity – Compare ligand graph with graph induced on ligand by receptor QSAR – Topological indices

73. Example S = Connolly Surface f(x) = Electrostatic Potential = ∑ q(i) / d(x,i) Method Locate critical points of f (Newton-Raphson). Linearize at saddles, find eigenvectors of Hessian( f ). Integrate gradient vector field forward in time from 2 points on unstable eigenvector, backward in time from 2 points on stable eigenvector (Runge-Kutta). Integrate to boundary of Connolly surface patch, then continue on adjacent patch until reaching another critical point.

74. Allopurinol 8 maxima 7 minima 13 saddles #maxima – #saddles + #minima =  (S) = 2

75. Work in Progress … Implementation for S = spherical harmonic surface f = MEP, LIE, LEA and LP –Use images of triangulation points as starting points for Newton-Raphson search for critical points. –Automatic differentiation.

77. Summary Molecular surfaces QM properties presented on surface Compound screening Pattern matching on surfaces Martin Swain Critical features Dave Whitley Data reduction and QSAR Brian Hudson Spherical harmonic representation Dave Ritchie

78. Future directions High-throughput ligand docking –Superimposition of ligand and a “negative” of the receptor Use of the fields to drive simulation –Use of the fields to derive intermolecular forces –Rigid-body motions – long time-step MD –Free energy calculations

79. A hierarchy of methods Rapid screening using computationally fast approaches –3D fields – Andy Vinter On reduced set: –Semi-empirical property calculations and alignments On most interesting molecules: –Density-functional or ab-initio calculations and alignment More accurate molecular representations are used as appropriate, as resources allow