Related Experiment Videos
On the statistical assessment of classifiers using DNA microarray data
N Ancona1, R Maglietta, A Piepoli
1lstituto di Studi sui Sistemi Intelligenti per I'Automazione-CNR, Via Amendola 122/D-l, 70126 Bari, Italy. ancona@ba.issia.cnr.it
BMC Bioinformatics
|August 22, 2006
Summary
Accurate colon cancer classification is achievable with minimal data. Statistical methods assess gene expression profiles, identifying key genes for reliable diagnosis and prognosis.
Area of Science:
- Bioinformatics
- Computational Biology
- Genomics
Background:
- Presents a novel statistical method for assessing cancer predictors using gene expression profiles.
- Utilizes a new dataset of microarray gene expression data from colon cancer patients (22 normal, 25 tumor specimens).
Purpose of the Study:
- To determine if available data is sufficient for accurate cancer classifiers.
- To assess the statistical significance of error rates in classification.
- To identify the number of genes crucial for accurate colon cancer classification.
Main Methods:
- Employs Leave-K-Out Cross Validation to estimate generalization error for three classification schemes.
- Uses permutation tests to measure the statistical significance of error rates.
- Analyzes gene frequencies and their correlation with colon cancer pathology.
Main Results:
- Weighted Voting Algorithm (WVA) achieved 21% error with 25 training examples (p=0.045).
- Regularized Least Squares (RLS) and Support Vector Machines (SVM) achieved 19% (p=0.035) and 18% (p=0.037) error with 15 training examples, respectively.
- RLS and SVM reached optimal performance with 35 training examples (14% and 11% error, respectively) and identified ~6000 significant genes (p<0.05).
Conclusions:
- The proposed method offers statistically significant insights for cancer diagnosis and prognosis.
- Effective colon cancer classifiers can be trained with as few as 15 examples.
- The number of genes required for reliable classification is dependent on the desired accuracy.