Data processing: artificial intelligence – Knowledge processing system
Reexamination Certificate
2000-01-19
2002-08-13
Black, Thomas (Department: 2121)
Data processing: artificial intelligence
Knowledge processing system
C514S012200, C435S069100
Reexamination Certificate
active
06434542
ABSTRACT:
BACKGROUND OF THE INVENTION
A portion of the disclosure of this patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever.
This invention relates generally to computer assisted methods of analyzing chemical or biological activity and specifically to computer assisted methods of determining chemical structure-activity relationships, and determining which species in a mixture from a chemical or biological population can be predicted to have a given biological activity or biological phenotype. This method is particularly useful in the fields of chemistry and genetics.
Combinatorial chemistry and high-throughput screening (HTS) are having a major impact on the way pharmaceutical companies identify new therapeutic lead chemical compounds. Voluminous quantities of data are now being produced routinely from the synthesis and testing of thousands of compounds in a high-throughput biochemical assay. The construction of chemical libraries has, in effect, replaced the painstaking individual synthesis of compounds for biological testing with a strategy for the multiple synthesis of many compounds about a common structural core scaffold. Since there is such a low probability of identifying new lead compounds from screening programs, it is expected that the sheer number of compounds made via a combinatorial approach will provide many more opportunities to find novel leads. However, making and testing thousands of compounds instead of fifty to one hundred per chemist per year has placed a tremendous strain on the logistical and computational infrastructure usually relied upon to store and analyze these datasets. Methods, developed in the last decade, for the statistical analysis of a relatively small number of compounds (less than 100) are not suitable for use on much larger data sets. Consequently, new technologies must be investigated.
Various methods for the storage and retrieval of chemical structure/biological activity data have been devised. Software products are now available from major vendors that address most of the logistical needs of combinatorial chemistry. Little thought, however, has been given to how the data might best be used to guide future synthetic efforts once the biological activity of chemical compounds has been learned. One possible result from the synthesis and testing of large numbers of compounds is a short list of promising new lead compounds for further consideration. Many research programs stop here and immediately revert to traditional synthesis in order to optimize the new leads. On the other hand, others are seeking to continue along a combinatorial path have employed an evolutionary approach to make best use of all the data.
Genetic algorithms have also been used to select new chemical libraries to be made. However, due to the complex and specialized nature of the software used to identify 3D pharmacophores, it is unlikely that these methods will be able to routinely handle the volume of data and/or possible multiple binding modes or sites.
For a number of years, there has been an interest in using artificial intelligence methods to deconvolute, uncover hidden rules from, or otherwise classify chemical datasets. Most have focused on reaction prediction. Others have used neural networks, fuzzy adaptive least squares and the like to analyze structure-activity datasets or predict chemical properties. Most of these methods are generally much too complex for routine structure-activity-relationship (SAR) analysis of large heterogenous data sets.
Recursive partitioning (RP) is a simple, yet powerful, statistical method that seeks to uncover relationships in large data sets. These relationships may involve thresholds, interactions and nonlinearities. Any or all of these factors impede an analysis that is based on assumptions of linearity such as multiple linear regression (or basic QSAR), principal component regression (PCR), or partial least squares (PLS). Various implementations of RP exist but none have been adapted to the specific problem of generating SAR. The present invention features a new computer program, Statistical Classification of Molecules using recursive partitioning (SCAM), to analyze large numbers of binary descriptors (which are concerned only with the presence or absence of a particular feature) and to interactively partition a data set into active classes.
SUMMARY OF THE INVENTION
In brief summary, the invention is a computer-based method of encoding features of mixtures, whether the features be of individual data objects in a mixture or features of mixtures themselves, and of identifying and correlating those individual features to a response characteristic that is a trait of interest of the individual data object or of the mixture. The method is applicable to data objects in those types of data sets that are characterized in being a mixture of data object classes, each data object class containing one or more of the data objects, and wherein multiple data objects present a same trait of interest, but classes of data objects produce the response characteristic that is a trait of interest through different underlying mechanisms. The method comprises the steps of: assembling a set of descriptors and converting said set of descriptors into the form of a bit string such that each descriptor reflects the presence or absence of a potentially useful feature in a data object of interest; examining each data object for presence or absence of each of said descriptors; assembling the results of looking for descriptors into a vector for each data object, noting the presence or absence of each feature in said data object; assembling all vectors thus generated into a matrix; dividing the data in said matrix into two daughter sets on the basis of presence or absence of a given descriptor from said set of descriptors; and iteratively repeating this step until each member of said mixture has been classified into a group. The method is applicable to three broad situations. Firstly, those situations in which data objects are unique, but the data set is a mixture in the sense that the data objects act in different ways, e.g. a population of human patients having different biological genotypes that nonetheless lead to a phenotypically identical clinical disease diagnosis. Secondly, those situations in which the data objects are themselves mixtures, e.g. a mixture of k chemical compounds tested together in a high throughput screen, or a mixture of different structural modes of a compound, and those data objects that show a given activity of interest do so in the same fashion or through the same underlying mechanism of action. And thirdly, those situations in which the data objects are mixtures and the active elements in the mixtures produce the same activity, but are acting through different mechanisms, for example, where k chemical compounds are screened together for activity and two of the compounds bind to a biological receptor, but bind to it in different places or in different conformations. Each of these three types of situations can be addressed whether they are planned or inadvertent mixtures. A planned mixture occurs where the fact of being a mixture is capable of manual control as is the case with carrying out a combinatorial synthesis, or where a high throughput screening is carried out with, for example, 20 compounds test together. An inadvertent mixture is said to be present whenever it is inherent in the situation, for example where there are multiple structural conformations of a chemical compound, or where a data set contains compounds producing the same chemical result but acting by different mechanisms, or where a data set contains compounds producing the same biochemical result, but binding to different receptor sites or places, or where the data set is a human population having
Farmen Mark W.
Lambert Christophe Gerard
Rusinko, III Andrew R.
Young Sidney Stanley
Black Thomas
Grassler Frank P.
Holmes Michael B.
SmithKline Beecham Corporation
LandOfFree
Statistical deconvoluting of mixtures does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Statistical deconvoluting of mixtures, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Statistical deconvoluting of mixtures will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2913741