Prediction Enhancement of ProteinWater Binding Conservation through Evolutionary Computation

1
Prediction Enhancement of Protein-Water Binding
Conservation through Evolutionary Computation

• Michael Peterson, Travis Doom, Michael Raymer
• Abstract
• The design of drugs to combat various
  diseases is an extremely expensive and
  time-consuming process. Potentially,
  computational ligand screening will reduce the
  time and expense associated with drug lead
  discovery. Correctly predicting sites of water
  conservation on a protein surface can
  significantly increase the accuracy of ligand
  screening efforts. Traditional classification
  methods make correct predictions with
  approximately 60 accuracy. The goal of our
  research is to improve prediction accuracy by
  applying evolutionary computing (EC) to
  traditional methods of data classification. We
  present a method that improves accuracy by
  applying EC feature selection and extraction
  techniques to k-nearest neighbor and naïve Bayes
  classifiers. In order to facilitate this
  research, a versatile EC engine was developed in
  Java. Despite Javas object oriented nature, few
  general-purpose Java-based EC engines exist. Our
  engine, with several unique features, will
  therefore be useful to the EC community, and will
  be available via the World Wide Web.

2
I. Protein-Ligand Binding

• When a ligand binds to the surface of a protein,
  water molecules near that position will either be
  conserved or displaced.
• Our goal is to accurately predict sites of water
  conservation with high accuracy.

Protein surface
Water molecule
Ligand
3
II. Protein-Water Measurements

• Within a Set of 30 Proteins, There Are
• 8 Features Measured for Each Water Molecule
• Temperature factor (BVAL)
• Atomic Density (ADN)
• Atomic Hydrophilicity (AHP)
• Hydrogen bonds to protein (HBDP)
• Hydrogen bonds to water (HBDW)
• Mobility (MOB)
• ABVAL (Avg. B-val of protein atom neighbors)
• NBVAL (Net B-val of protein atom neighbors)

4
III. Feature Weighted knn-classification
a.
Applying weights to measured features can improve
the accuracy of a k-nearest neighbor classifier.
Weights can be optimized by a genetic algorithm.
P
W
W
W
Feature 1
P
W
P
P
W
W
P
P
P
Feature 2
P
b.
W
W
W
Feature 1
P
W
P
W
W
P
P
P
P
Feature 2
Scale Extended
5
IV. A Parameterized Discriminant for a Bayes
Classifier

• A confidence value for each class P is computed
  for each class.
• The class with the greatest value for P is
  selected.
• When C1 C2 .. Cd 1, the discriminant
  function is equivalent to the naïve bayes
  classifier.

The values for the coefficients, C1..Cd, are
supplied by an evolutionary computation optimizer.
6
Va. A Simple Evolutionary Algorithm
Over a number of generations, the values in a
population of individuals are optimized via
operations abstracted from natural selection.
?
Increasing
Fitness
Evaluation
Selection
Genetic Operators
7
V. GA/Classifier Hybrid Architecture
Mask Vector Masks used to hide features during
classification as a method of feature selection.
Genetic Algorithm
W1W2...W8 M1 M2...M8
KNN Classifier
W1W2...W8 M1 M2...M8
W1W2...W8 M1 M2...M8
Population of feature weight mask sets
W1W2...W8 M1 M2...M8
W2
...
...
W1
Fitness Based on the number of correct predictions
using the weight vector the number of masked
features
Weight Vector Weights to use for each
feature axis during classification.
8
VI. Original EC Engine

• In order to optimize weights masks, we
  implemented a new EC engine with several useful
  features.
• Chromosome consists of a feature vector and an
  optional mask vector. Mask bits are used to
  block features from being passed to the fitness
  function.

F1 F2 Fn M1 M2 Mn

• Groups feature vector may consist of groups of
  real, integer, or boolean values. Each group may
  have its own mutation rate and reproduction
  method.
• Mutation Each group may have its own mutation
  rate and method, and the mask vector has a
  separate mutation rate and method. Random, range
  based, and variance mutations are permitted.
• Reproduction 1-point, 2-point, or uniform
  crossover is permitted, either on a per group
  basis, or across the entire feature chromosome.
  These methods also apply to the mask chromosome.

9
VII. Original EC Engine

• Our EC engine is implemented in Java, adding
  versatility portability.
• The user provides the fitness function, which
  can be implemented in either Java or C/C.
• Our engine is ideally suited for feature
  selection extraction problems, such as this
  protein-water binding example.
• We use it to optimize feature weights and remove
  unnecessary features for both knn bayes
  classifiers.
• The engine can be used for many problems which
  genetic algorithms have been been applied in the
  past.

10
VIII. Conserved vs. Non-Conserved
0.000
0.000
0.000
0.000
0.000
0.683
59.83
60.68
60.25
3.40
2.48
69
0.000
0.000
0.000
0.336
0.000
0.000
0.000
0.664
This data is the result of previous work on
this problem by Mike Raymer, using a knn
classifier with a genetic algorithm implemented
by the University of California, San Diego.
11
VIIIb. Determinants of Solvation

• A Continuum of Favorability
• Binding
• Atomic hydrophilicity
• Atomic density
• Hydrogen bonding potential
• Conservation
• Low B-Value, High Occupancy
• Many hydrogen bonds to protein atoms

12
IX. Other EC Applications

• Structural Bioinformatics
• Prediction of metal binding sites
• Location of protein active sites
• Prediction of drug lead activity QSAR
• Gene Classification Discovery
• Prediction of gene function from microarray data
• Ligand Screening and Docking
• SLIDE, Volker Schnecke, MSU
• Crystallographic Solvent Fitting
• Xfit Duncan McRee, Scripps