Drug Design 2. Oliver Kohlbacher. Winter 2009/ QSAR Part 4: Selected Chapters

Transcription

1 Drug Design 2 Oliver Kohlbacher Winter 2009/ QSAR Part 4: Selected Chapters Abt. Simulation biologischer Systeme WSI/ZBIT, Eberhard-Karls-Universität Tübingen

2 Overview GRIND GRid-INDependent Descriptors Idea Construction Results How good is QSAR really?

3 Problems in 3D QSAR Core problem in 3D QSAR studies is the correct alignment of the correct conformations Which conformers should be aligned? Same problems as for structure comparison The larger the number of degrees of freedom in the molecule, the more difficult it gets to align the molecules properly Bultnick et al., p. 589

4 Problems in 3D QSAR Core problem in 3D QSAR studies is the correct alignment of the correct conformations Meaningful alignments can only be constructed if the there is structural knowledge of the binding mode It is often not clear, which property is most relevant for the superposition (hydrophobic interactions? electrostatics?) Wrong alignments lead to wrong energies in the grids and thus to incorrect QSAR models Bultnick et al., p. 589

5 GRIND/ALMOND GRid-INDependent descriptors for 3D QSAR aim at alleviating the alignment problem Idea Compute virtual description of the receptor in the form of interaction points Encode the geometry of the interaction points in an translation- and rotation-invariant representation Use this representation as descriptors for PLS GRIND is implemented in the (commercial) software package ALMOND Pastor et al., J. Med. Chem., 2000, 43, 3233

6 VRS Virtual Receptor Sites GRIND uses the GRID potentials of six different probe groups to build a virtual representation of the receptor From this grid-based representation a sparse representation based on interaction points is computed Filter algorithms thin out this point set to about 100 points per probe group Pastor et al., J. Med. Chem., 2000, 43, 3233

7 MIFs Compute MIFs Molecular Interaction Fields, i.e. GRID potentials for various probe groups around the ligand In CoMFA: grid entries are directly used as matrix entries -> alignment issues Instead: reduce grid values to interaction points (similar to what LUDI does, see lecture Drug Design 1) Filtering reduces number of interaction points further MIF of an amide probe around caffeine Pastor et al., J. Med. Chem., 2000, 43, 3233

8 VRS Virtual Receptor Sites Areas in the MIFs with significant negative energies are interesting We search for a sparse representation of these areas Points in these regions can be considered places, where the receptor might have interaction sites They are thus called virtual receptor sites (VRSs) Bultnick et al., p. 593

9 VRS Virtual Receptor Sites In contrast to the position of a VRS, the distance between pairs of VRSs is rotation- and translationinvariant We use a representation based on distances alone Compute distribution of distances between certain VRSs In analogy to a radial distribution function Disadvantages Loss of information Not invertible Bultnick et al., p. 593

10 Recap: 3D Descriptors RDF The distribution of all intramolecular distances is captured by the radial distribution function (RDF) Because the distances are typically discrete, they are turned into a density function: B is a constant describing the width of the Gaussian contribution to the density function RDF is a one-dimensional description of the threedimensional geometry of a molecule

11 Recap: 3D Descriptors RDF GE, p.416

12 Recap: 3D Descriptors RDF Isomers and even conformers (which, by definition, yield identical 2D descriptors) can be distinguished using the RDF Disadvantage: RDF is a vector, not a single number Similar to RigFit, physicochemical properties can also be integrated into RDFs: where p i, p j are the respective values for the property of atoms i and j

13 GRIND - Correlations GRIND applies the RDF idea to interaction points Instead of correlating atom coordinates, we use the interaction points Correlations can be computed just like RDFs Where RDFs multiply properties, GRIND multiplies the respective interaction energies Pastor et al., J. Med. Chem., 2000, 43, 3233

14 Auto- and Cross-Correlations ALMOND typically uses six different correlograms that are concatenated This results in a vector of scalar values used to encode the molecule Similar interaction options in similar distances will result in similar correlograms Probe 1 Probe 2 Interaction between DRY DRY Hydrophobic O O Hydrogen bond (HB) donor N1 N1 Hydrogen bond acceptor DRY O Hydrophobic and HB donor DRY N1 Hydrophobic and HB acc. O N1 HB acceptor and donor Pastor et al., J. Med. Chem., 2000, 43, 3233

15 Auto- and Cross-Correlations Pastor et al., J. Med. Chem., 2000, 43, 3233

16 Results Set of 10 superimposed correlograms (N1-N1) of glucose analogs The compounds are inhibitors of glycogen phosphorylase Points are color-coded according to biological activity (active = red, intermediate = white, inactive = green) Interaction energies correlate with activity! Pastor et al., J. Med. Chem., 2000, 43, 3233

17 Results Several cpds with serotoninergic (5- HT2A) activity GRIND model based on PLS Model: 3 LV R 2 = 0.93 Q 2 = 0.81 Average error: 0.35 Pastor et al., J. Med. Chem., 2000, 43, 3233

18 Benchmarking Descriptors It is often very hard to decide, what type of descriptors should be chosen There are also very few studies systematically comparing different descriptor sets on large independent datasets One study was published in 2006 by Gedeck et al. Datasets 944 in-house datasets (Novartis) 570,000 data points in total 143,000 different compounds Gedeck et al, J. Chem. Inf. Model., 2006, 46:

19 Benchmarking Descriptors Descriptors GRIND AlogP atom counts Avalon fingerprints (hashed, graph-based fingerprints) Pipeline Pilot fragment counts (hashed structurebased fingerprints) Hologram QSAR (graph based fragment counts) MDL public keys (binary structural fingerprint, 166 bit) Similog descriptors (pharmacophoric triplet counts) Gedeck et al, J. Chem. Inf. Model., 2006, 46:

20 Comparing Descriptors Gedeck et al, J. Chem. Inf. Model., 2006, 46:

21 Results Gedeck et al, J. Chem. Inf. Model., 2006, 46:

22 Caveats So far no consistent single set of descriptors has been found Most of these studies yield contradictory results Descriptor choice also varies with the target/scaffold: descriptors good on one target might be poor on another Also, careful data preparation plays a key role: Gedeck et al. write: Interestingly, the computationally most involved model based on the Almond descriptors, which uses the three-dimensional structure of the compound for the prediction, performed poorest. This might be due to the fact that the description of the three-dimensional structure of compounds is difficult, in general [...], and that the protocol to generate the structures in the present study was inappropriate, in particular. Gedeck et al, J. Chem. Inf. Model., 2006, 46:

23 References Books Bultnick et al. (Hrsg.), Computational Medicinal Chemistry for Drug Discovery, Marcel Dekker Inc., New York, 2004 Papers Pastor M, Cruciani G, McLay I, Pickett S, Clementi S. GRid- INdependent descriptors (GRIND): a novel class of alignmentindependent three-dimensional molecular descriptors. J Med Chem. 2000, 43(17): Fontaine F, Pastor M, Zamora I, Sanz F. Anchor-GRIND: filling the gap between standard 3D QSAR and the GRid-INdependent descriptors. J Med Chem. 2005,48(7): Gedeck P, Rohde B, Bartels C. QSAR how good is it in practice? Comparison of descriptor sets on an unbiased cross section of corporate data sets. J Chem Inf Model. 2006, 46(5):