Symbolic Discriminant Analysis for Mining Gene Expression Patterns

Created by W.Langdon from gp-bibliography.bib Revision:1.4524

  author =       "Jason H. Moore and Joel S. Parker and Lance W. Hahn",
  title =        "Symbolic Discriminant Analysis for Mining Gene
                 Expression Patterns",
  booktitle =    "Critical Assessment of Techniques for Microarray Data
                 Analysis (CAMDA00)",
  year =         "2000",
  address =      "Levine Science Research Building, Duke University,
                 Durham, N.C.",
  month =        "18-19 " # dec,
  note =         "submitted abstract",
  keywords =     "genetic algorithms, genetic programming, SDA",
  URL =          "",
  URL =          "",
  size =         "1 page",
  abstract =     "Linear discriminant analysis is a popular multivariate
                 statistical approach for classification of observations
                 into groups because the theory is well described and
                 the method is easy to implement and interpret. However,
                 an important limitation is that linear discriminant
                 functions need to be pre-specified. That is, specific
                 variables need to be selected and added linearly into
                 the model. Only the coefficients are estimated from the
                 data. To address this limitation, we developed symbolic
                 discriminant analysis (SDA) for the automatic selection
                 of gene expression variables and discriminant functions
                 that can take any form.

                 Our SDA approach is inspired by the symbolic regression
                 approach of Koza (1992). We begin by defining the
                 mathematical functions (e.g. +, -, /, *, log, sqrt,
                 etc.) and the list of gene expression variables that
                 could potentially be used as the building blocks for
                 discriminant functions. Symbolic discriminant functions
                 are evaluated by generating discriminant scores for
                 each observation to be classified. The overlap in
                 distributions of discriminant scores between groups is
                 an estimate of the classification error. Class
                 membership for new observations can be predicted from
                 the discriminant score that separates the
                 distributions. To identify optimal symbolic
                 discriminant functions from the near infinite model
                 space, we employed parallel genetic programming for
                 machine learning on a 110 processor Beowulf-style
                 parallel supercomputer.

                 We applied the SDA approach to identifying subsets of
                 gene expression variables and symbolic discriminant
                 functions that can correctly classify and predict types
                 of human acute leukemia. Using a leave-one-out
                 cross-validation strategy, we identified no fewer than
                 15 different combinations of gene expression variables
                 and symbolic discriminant functions that correctly
                 classified 38/38 observations in the first dataset and
                 correctly predicted 31/34 observations in the
                 independent dataset. The most common gene identified
                 across these models was the human synaptonemal complex
                 protein 1 (SCP1) gene that is expressed in solid tumors
                 and haematological malignancies.

                 We conclude that the SDA approach provides a powerful
                 alternative to traditional multivariate statistical
                 methods for identifying gene expression patterns. The
                 advantages of SDA include the ability to identify an
                 important subset of gene expression variables from
                 among thousands of candidates and the ability to
                 identify the most appropriate mathematical functions
                 relating the gene expression variables to a clinical
                 endpoint. We anticipate this will be an important
                 methodology to add to the repertoire of approaches for
                 mining gene expression patterns.",
  notes =        "Program in Human Genetics, Department of Molecular
                 Physiology and Biophysics, Vanderbilt University
                 Medical School, Nashville, TN 37232-0700",

Genetic Programming entries for Jason H Moore Joel S Parker Lance W Hahn