Semi-Supervised Projective Non-Negative Matrix Factorization for Cancer Classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• DNA microarray technology enables gene expression profiling for cancer identification.
• Traditional supervised methods require expensive labeled clinical data, limiting practical application.
• Unlabeled data is abundant but underutilized in current cancer classification models.

Purpose of the Study:

• To develop a semi-supervised learning method for cancer classification using gene expression profiles.
• To leverage both labeled and unlabeled samples to enhance classifier performance.
• To propose a novel Semi-PNMF algorithm for more robust cancer identification.

Main Methods:

• A semi-supervised projective non-negative matrix factorization (Semi-PNMF) approach is proposed.
• The method jointly learns a non-negative subspace from labeled and unlabeled samples.
• A multiplicative update rule (MUR) is developed and its convergence is proven for optimization.

Main Results:

• Semi-PNMF learns more representative subspaces by incorporating information from unlabeled samples.
• The method demonstrates improved cancer classification performance compared to existing techniques.
• Experimental validation on two multiclass cancer gene expression datasets confirms superior results.

Conclusions:

• Semi-PNMF offers an effective solution for cancer classification by integrating labeled and unlabeled data.
• The proposed method enhances classification accuracy and subspace representation.
• This approach addresses the limitations of expensive labeling in clinical cancer research.