_PROBLEM CoEPrA-2006_Classification_001

_GROUP_NAME Artem Cherkasov 

_GROUP_MEMBERS
Emre Karakoc
Cenk Sahinalp
Artem Cherkasov

_ADDRESS
University of British Columbia, Medicine Simon Fraser University, 
Computer
Science Vancouver, BC, Canada

_MODELING_PROCEDURE
Descriptors:
Based on our previous experience with QSAR modeling of peptides and 
general
QSAR clustering and classification of bioactivity properties, we 
decided to
use the following strategy. 
First, we have optimized the geometry of the studied peptides using 
MMFF94
force field; carboxylic groups have been deprotonated, amino groups -
protonated, partial charges computed according to [1].
Then we have computed QSAR descriptors that describe an entire peptide
molecule (global parameters) as well as descriptors corresponding to
constituent aminoacids considered in the context of their peptide
environment (we did not use 'isolated aminoacids' approximation).
Thus, for all peptides in the testing and training sets of
Classification_001 problem, we initially calculated > 200 various 3D 
and 2D
QSAR parameters that included:

.	50 global 'inductive' QSAR descriptors as described in [2]. 
.	10 local 'inductive' QSAR descriptors (computed toward CA atom) have
been calculated for each aminoacid of a given nano-mer; therefore, 90
additional 'inductive' QSAR descriptors have been produced. 
.	We have also computed ~90 conventional 3D and 2D global QSAR
parameters which are implemented within the MOE modeling package [3].

All 'inductive' QSAR descriptors that are described above, have been
calculated by our own SVL scripts for the MOE, that can be freely 
downloaded
through the SVL exchange. 

Descriptors Selection:
As a first step, the kNN-based linear optimization method based on a
distance measure [4] has been used for the initial selection of most
relevant QSAR parameters. As the result, 90 local 'inductive' 
descriptors
have been selected for the 'kernel' QSAR models for Clafficifcation_001
problem. 
As the second round of the training model optimization, we have 
expanded the
set of initial 90 local 'inductive' parameters by sampling global
'inductive' and conventional QSAR descriptors using greedy approach. In
particular, we have used the Partial Least Squares (PLS) method 
implemented
in MOE to identify those global parameters that can improve the QSAR 
fitting
of dependent variables for the training set's peptides. 
Thus, the final set of 101 QSAR descriptors has been identified and can 
be
found in the attachment file. 

Modeling Procedure:
In order to build a predictive QSAR model on the set of the training
peptides we employed the method of Artificial Neural Networks (ANN) as 
it is
implemented within the SNNS package [5]. 
The values of all descriptors selected in the previous section have 
been
normalized into the range [0.0, 1.0] based on the union of the external 
and
calibration data-sets.
To identify the optimal ANN setting we have conducted the Leave-One-Out
(LOO) validation of the results, changing the learning rate, number of
training cycles, ANN jogging setting, while judging the ANN performance 
by
ROC values produced by the LOO predictions. 

As the result, we have identified that the best LOO performance could 
be
achieved with the standard back-propagation training on 400 cycles and 
0.7
learning rate, input shuffling, weight decay and random assignment of
initial weights in a range of [-1.0, 1.0]. The number of ANN hidden 
nodes
has been set to 100. 
All 89 ANN models produced during the LOO validation have then been 
applied
to the external set of peptides and the outputs have been averaged and
interpreted with 0.2 threshold. 
The interpreted averaged outputs are presented in the next section. 
Since we
have used 89 ANN models (each trained on 88 peptides) - we can provide 
their
weights upon request. The attachment contains a single ANN trained on 
all 89
peptides with known activities. 

[1] Cherkasov, A. Inductive Electronegativity Scale. Iterative 
Calculation
of Inductive Partial Charges. Journal of Chemical Information and 
Computer
Sciences, 2003, 43, 2039-2047. Cherkasov, A., Z. Shi, Y. Li, S.M. 
Jones, M.
Fallahi, G.L. Hammond. 'Inductive' Charges on Atoms in Proteins: 
Comparative
Docking with the Extended Steroid Benchmark Set and Discovery of a 
Novel
SHBG Ligand. Journal of Chemical Information and Modelling, 2005, 45,
1842-1853.
[2] Cherkasov, A. 'Inductive' Descriptors. 10 Successful Years in QSAR.
Current Computer-Aided Drug Design, 2005, 1, 21-42.
[3] Molecular Operational Environment, 2005, by Chemical Computing 
Group
Inc., Montreal, Canada.
[4] Karakoc E., Cherkasov A., Sahinalp S. C.  Distance Based Algorithms 
for Small Biomolecule Classification and Structural Similarity Search. 
ISMB'06, 14th Annual International conference on Intelligent Systems for 
Molecular Biology, Fortaleza, Brazil 2006.
[5] SNNS: Stuttgart Neural Network Simulator; Version 4.0, University 
of Stuttgart, 1995.

_PREDICTION
Obj_00001 +1
Obj_00002 +1
Obj_00003 -1
Obj_00004 -1
Obj_00005 +1
Obj_00006 +1
Obj_00007 -1
Obj_00008 -1
Obj_00009 +1
Obj_00010 -1
Obj_00011 +1
Obj_00012 -1
Obj_00013 +1
Obj_00014 -1
Obj_00015 +1
Obj_00016 +1
Obj_00017 -1
Obj_00018 +1
Obj_00019 -1
Obj_00020 +1
Obj_00021 +1
Obj_00022 +1
Obj_00023 -1
Obj_00024 -1
Obj_00025 -1
Obj_00026 +1
Obj_00027 +1
Obj_00028 +1
Obj_00029 +1
Obj_00030 -1
Obj_00031 +1
Obj_00032 +1
Obj_00033 -1
Obj_00034 +1
Obj_00035 +1
Obj_00036 +1
Obj_00037 -1
Obj_00038 -1
Obj_00039 -1
Obj_00040 -1
Obj_00041 +1
Obj_00042 -1
Obj_00043 +1
Obj_00044 -1
Obj_00045 +1
Obj_00046 +1
Obj_00047 -1
Obj_00048 +1
Obj_00049 -1
Obj_00050 -1
Obj_00051 +1
Obj_00052 +1
Obj_00053 -1
Obj_00054 +1
Obj_00055 -1
Obj_00056 -1
Obj_00057 -1
Obj_00058 -1
Obj_00059 +1
Obj_00060 -1
Obj_00061 -1
Obj_00062 -1
Obj_00063 +1
Obj_00064 -1
Obj_00065 -1
Obj_00066 +1
Obj_00067 +1
Obj_00068 +1
Obj_00069 +1
Obj_00070 -1
Obj_00071 +1
Obj_00072 -1
Obj_00073 +1
Obj_00074 -1
Obj_00075 +1
Obj_00076 -1
Obj_00077 +1
Obj_00078 +1
Obj_00079 -1
Obj_00080 +1
Obj_00081 +1
Obj_00082 -1
Obj_00083 -1
Obj_00084 +1
Obj_00085 +1
Obj_00086 -1
Obj_00087 -1
Obj_00088 -1