Combining dissimilarities in a Hyper Reproducing Kernel Hilbert Space for complex human cancer prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Oncology

Background:

• DNA microarrays generate gene expression profiles crucial for cancer prediction.
• Support Vector Machines (SVM) are widely used for cancer sample classification but have limitations with Euclidean distances.
• Euclidean distances may not accurately capture sample profile similarities, leading to potential misclassification errors.

Purpose of the Study:

• To enhance cancer sample classification accuracy by incorporating non-Euclidean dissimilarities into the nu-SVM algorithm.
• To develop a method that better reflects the proximities among sample profiles in gene expression data.
• To reduce misclassification errors in the analysis of human cancer problems.

Main Methods:

• Incorporation of a linear combination of non-Euclidean dissimilarities into the nu-SVM algorithm.
• Learning combination weights within a Hyper Reproducing Kernel Hilbert Space (HRKHS).
• Utilizing a Semidefinite Programming algorithm for weight optimization and incorporating a smoothing term to prevent overfitting.

Main Results:

• The proposed method effectively integrates non-Euclidean dissimilarities to improve classification.
• The approach successfully reduces misclassification errors in several human cancer datasets.
• The incorporation of a smoothing term in HRKHS helps penalize distance family complexity and avoid overfitting.

Conclusions:

• The novel nu-SVM approach utilizing non-Euclidean dissimilarities offers improved accuracy in cancer prediction.
• This method provides a more robust way to analyze gene expression data for cancer classification.
• The findings suggest a significant reduction in misclassification errors for human cancer problems using this enhanced SVM technique.