AFMoC Enhances Predictivity of 3D QSAR: A Case Study with DOXPreductoisomerase

1
AFMoC Enhances Predictivity of 3D QSAR A Case
Study with DOXP-reductoisomerase

• K. Silber, P. Heidler, T. Kurz, G. Klebe
• J. Med. Chem. 48(2005) 3547-3563

Journal Club, Presented by Lei Xie
2
Motivations

• No other papers read
• Interested in interface of biology and chemistry
• Inspired by a talk on nuclear receptors in the
  group meeting

3
Motivations

• No other papers read
• Interested in interface of biology and chemistry
• Inspired by a talk in the group meeting

4
Outlines

• Basic of 3D-QSAR
• Discussion of the paper
• Ideas derived from the paper

5
Concept of 3D QSAR

• Correlates spatially located features across a
  chemical series with biological activity
• Procedures
• 1. Selection of active conformation
• 2. Alignment of conformers
• 3. 3D field calculation in a box and grids
• 4. Derivation and validation of models

6
Step 1 Selection of Active Conformation
O

• Global minimum-enery conformation
• Active analogs and pharmacophores
• Protein-ligand complex
• Intuitions

N
7
Step2 Alignment of conformers

• Manuel pharmacophore or moments of inertia
• Semi-automatic overlap of steric and
  electrostatic
• Fully-automatic fragment re-construction
• No alignment

O
N
8
Step 3 3D field calculation in grids

• CoMFA
• Lennard-Jones and
• Coulomb interactions
• CoMSIA
• molecular similarity

O
N
9
Step 4 Derivation and validation of models
M1 M2 M3 . Mn

• Regression, classification or clustering
• in a mxn matrix
• Challenges in
• experimental design,
• feature selections and model validations

W1 E1 W2 E2 . . . Wm Em
Y
10
Adaptation of fields for Molecular Comparison
(AFMoC)A technique to combine protein-ligand
interaction and 3D-QSAR to improve selection of
active conformation, alignment quality, affinity
prediction
11
Concept of AFMoC Step 1

• Active conformations and molecular alignments are
  determined with protein-ligand docking

O
N
12
Concept of AFMoC Step 2

• Decompose the general scoring function at each of
  grids in the binding pocket
• Interaction fields are yielded by considering the
  contribution of the binding ligand to each of the
  grids

O
N
13
Concept of AFMoC Step 3
M1 M2 M3 . Mn

• Regression with experimental affinity data to
  obtain binding models for specified protein
  family
• Balance between the general and specific scoring
  function depending on the training data

W1 E1 W2 E2 . . . Wm Em
Y
14
DOXP-reductoisomerase(DXR)A new drug target for
multidrug-resistance malaria
15
DXR structures

• Challenges in structure-based drug design
• Contains metal-ligand coordination
• Requires cofactor binding
• - Possesses a large flexible loop

16
Training and Testing Data

• 27 for model building
• 14 for testing
• Testing data set includes functional groups that
  are not included in the training data

17
Results Regression Models

• CoMFA, CoMSIA, and AFMoC all derived
  statistically-significant models

Figure 5 (a) Experimentally determined binding
affinities versus fitted predictions using the
derived 3D QSAR models for the training set.
Results are shown applying the optimal number of
components, which is 5 for CoMFA ( ) and 4 for
CoMSIA ( ), respectively. (b) Experimentally
determined binding affinities versus fitted AFMoC
predictions for the 27 training set
DOXP-reductoisomerase inhibitors. Both
experimental and calculated values are shown
considering only the part of binding affinity
(pIC50PLS) used in PLS analysis ( ) or
considering the total binding affinity ( ). In
addition to the line of ideal correlation, dashed
lines are depicted to indicate deviations of one
logarithmic unit from ideal prediction.
18
Results Prediction Power

• CoMFA and CoMSIA fail for ligands comprising
  functional groups not present in or exhibiting
  minor structural differences from the training
• AFMoC is capable to correlate structural changes
  with affinity
• General docking methods do not perform as well as
  AFMoC

a Values are given considering only the pair
potentials for the prediction or considering also
the solvent-accessible surface term (values in
parentheses).b Squared correlation coefficient.c
In logarithmic units.d Values are given for the
optimal number of components, which are 4 for
AFMoC, 5 for CoMFA, and 4 for CoMSIA.e Values in
italics denote lacking correlation.
19
Results General vs. Adapted

• Best mixture of general and adapted model depends
  on the available training data.

Figure 8 Dependence of squared correlation
coefficient (r2) for AFMoC models on the mixing
coefficient between original DrugScore
potentials ( 0) and specifically adapted AFMoC
fields (PLS-model, 1).
20
Summary on the Paper

• Protein structure information, even not perfect,
  is valuable for 3D-QSAR studies
• Protein family adapted scoring functions is
  superior to general one
• A general framework can be extend to other studies

21
Genome-wide Prediction of Protein-Ligand
Interactions
22
Scopes

• Prediction of protein-ligand binding specificity
  given a protein sequence and a ligand

23
Procedures

• Decomposition of scoring functions with amino
  acid residues and identification of hot spot
  with known protein structures for a protein
  family
• Derivation of regression or classification models
  that correlate binding ligands and evolutionary
  profiles at the hot spot
• Exploration of genome sequences based on
  sequence/structure alignments and phylogenetic
  analysis
• Extension with functional site characterization
  and analysis

24
Applications

• Drug Discovery
• Target verifications, Lead discovery, Drug
  resistance (HIV, Gleevac), Off-target
  identifications
• Protein design
• Nature product identification

25
Data

• Protein structure (not necessary complex)
• Binding affinity
• Availability
• PDBBind - Structure abundant families (protein
  kinase, matrix metalloproteases etc.)
• Structural genomics structure
• NIH chemical genomics initiative binding
  affinity
• High throughput screening and combinatorial
  library
• Protein chips and other techonologies

26
THANKS