Towards the Discovery of Reliable Biomarkers from Gene-Expression Profiles: An Iterative Constraint Satisfaction Learning Approach

摘要

The article demonstrates the use of Multiple Iterative Constraint Satisfaction Learning (MICSL) process in inducing gene-markers from microarray gene-expression profiles. MICSL adopts a supervised learning from examples framework and proceeds by optimizing an evolving zero-one optimization model with constraints. After a data discretization pre-processing step, each example sample is transformed into a corresponding constraint. Extra constraints are added to guarantee mutual-exclusiveness between gene (feature) and assigned phenotype (class) values. The objective function corresponds to the learning outcome and strives to minimize use of genes by following an iterative constraint-satisfaction mode that finds solutions of increasing complexity. Standard (c4.5-like) pruning and rule-simplification processes are also incorporated. MICSL is applied on several well-known microarray datasets and exhibits very good performance that outperforms other established algorithms, providing evidence that the approach is suited for the discovery of biomarkers from microarray experiments. Implications of the approach in the biomedical informatics domain are also discussed.