Method for determining discrete quantitative structure activity relationships

Abstract

Method for developing a quantitative structure activity relationship that includes obtaining a training set of chemical compounds with molecular descriptors consisting of a number of multidimensional vectors with an activity class for each of the vectors; partitioning the multidimensional vectors into groups having interdependence; transforming the descriptors such that the interdependence of the groups is lessened; estimating a probability distribution of the descriptors by assuming that a probability distribution of a product of each of the groups is approximately equal to the probability distribution of the molecular descriptors; performing the partitioning, transforming and estimating steps for each of the activity classes; and, developing a probability distribution for the activity classes.

What is claimed is:  
     
         1 . A computer-based method of generating a quantitative structure activity relationship comprising: 
 a) calculating a numerical representation of molecules consisting of n numbers per molecule; and,    b) estimating a probability distribution that a said molecules is active.    
     
     
         2 . A method as recited in  claim 1 , wherein: 
 a) said estimating step is calculated with Bayes Theorem.    
     
     
         3 . A method as recited in  claim 1 , wherein: 
 a) said probability distribution of said estimating step comprises n one-dimensional distributions.    
     
     
         4 . A method as recited in  claim 1 , wherein: 
 a) said estimating step is performed by using a means to remove linear correlations between said n numbers per molecule.    
     
     
         5 . A method as recited in  claim 4 , wherein: 
 a) said means to remove linear correlations between said n numbers per molecule is a principal components analysis.    
     
     
         6 . A method as recited in  claim 4 , wherein: 
 a) said means to remove linear correlations between said n numbers per molecule is a matrix diagonalization.    
     
     
         7 . A method as recited in  claim 1 , wherein: 
 a) said estimating step is performed by using a means to remove dependencies between said n numbers per molecule.    
     
     
         8 . A method as recited in  claim 7 , wherein: 
 a) said means to remove dependencies between said n numbers per molecule is a principal components analysis.    
     
     
         9 . A method as recited in  claim 7 , wherein: 
 a) said means to remove dependencies between said n numbers per molecule is a matrix diagonalization.    
     
     
         10 . A method as recited in  claim 1 , wherein: 
 a) said estimating step is performed by estimating a distribution over a single number.    
     
     
         11 . A method as recited in  claim 1 , wherein: 
 a) said estimating step is performed by replacing a single observation with a Gaussian distribution.