In a very large scale genetic study...
...discover genes of physiological interest or SNPs of predictive interest	...visualise population heterogeneity and discover sub-phenotypes	...investigate gene by gene interactions

The intent of this method is to detect, amplify, analyze and visualize signal and signal heterogeneity in high dimensional datasets (nVariables >> nObservations), such as whole-genome scans or complex datasets incorporating large number of clinical (sub-phenotypes) and non-clinical (genetics, genomics, metabolomics,...) variables of various types (discrete and continous).

'Taxonomy 3' provides a statistical framework to large scale or complex problems, and produces simple answers, visually and biologically meaningful..

The method reduces the complexity and the dimensionality of the data, reveals independent sets of correlated variables and meaningful sub-groups of observations. Technically, it is a single multivariate eigen analysis (without multiple testing) based on correlations or covariations of contrasts of empirically-derived log Bayes factors (LBFs). The LBF properties allow usage of prior knowledge for data aggregation.

Since a primary objective of 'Taxonomy 3' is to visualize a complex dataset, it does not produce impenetrable 'black box' solutions such as other multivariate methods (artificial neural networks or support vector machine).

The authors believe this method can address several industry and academic needs such as disease understanding, data integration (integrated biomarker strategies) and decision making in drug discovery and clinical development.

The main features of the method can be tested on-line with small datasets: see the on-line analysis page.

 

	- signal to noise detection	: detect genes of interest relevant to the disease investigated, target discovery
	- detection of subjects' heterogeneity	: discover human sub-phenotypes and population heterogeneity
	- detection of independent sets of correlated variables	: detect new gene ontologies and their associated sub-phenotypes
	- analysis of variable interactions	: investigate and detect important gene by gene interactions
	- usage of prior knowledge for data aggregation	: analyse at gene level or ontology level, using SNP to gene to ontology maps
	- tunable signal amplification	: discover either genes of physiological interest or SNP of predictive interest
	- multivariate predictor	: individual prediction, dataset predictive capabilities

A comprehensive overview of 'Taxonomy 3' can be found in this article (PDF available on request):

Visualizing gene determinants of disease in drug discovery. Delrieu O and Bowman C.
Pharmacogenomics. 2006 Apr;7(3):311-29.
PMID: 16610942 / Pharmacogenomics
    

• partnerships with clinical/genetic groups having access to high dimensional datasets willing to try this method out
• partnerships with mathematical/statistical/IT groups willing to develop and implement this method
• integration within an existing (genetic/stat) software suite.

Site map

top of page

• Introduction
  
  
• Methods
  • Background
  • Method outline
  • Input Dataset
  • Log Bayes Factors (LBF)
  • Aggregation of LBFs
  • Principal Component Analysis
  • Results and Interpretation
    
    
• Examples
  • Dataset 1: Genes of physiological interest, population heterogeneity and sub-phenotypes
  • Dataset 2: Genetic heterogeneity in controls due to sampling bias
  • Dataset 3: Gene by gene interactions and sub-phenotypes
  • Dataset 4: 650K SNP level analysis, SNPs of predictive interest, anecdotal quality control
    
    
• Publications, talks and citations
  
  
• Software
  • Overview
  • Multi Platform JAVA application
  • Sas Macros
  • On-line analysis : try-out the main features of the method on-line with dummy datasets, or your own data.
    
    
• Authors

top of page