Semi-supervised learning improves gene expression-based prediction of cancer recurrence

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression profiling holds promise for cancer outcome prediction.
• Small sample sizes limit the accuracy of traditional supervised learning classifiers.
• Vast amounts of public microarray data with limited follow-up information are often underutilized.

Purpose of the Study:

• To investigate the utility of semi-supervised learning for cancer outcome prediction.
• To address the limitations of small sample sizes in gene expression-based classification.
• To leverage unlabeled data for improved predictive model robustness.

Main Methods:

• Employed a semi-supervised learning technique called low density separation (LDS).
• Applied LDS to predict recurrence risk in colorectal cancer patients.
• Validated the method on human breast cancer datasets.

Main Results:

• Semi-supervised classification using LDS significantly improved prediction accuracy compared to supervised methods like Support Vector Machines (SVM).
• Prediction performance increased proportionally with the number of unlabeled samples utilized.
• The LDS method demonstrated robustness irrespective of the number of input features and accommodated data from different institutions.

Conclusions:

• Semi-supervised learning, particularly LDS, offers a powerful approach to enhance gene expression-based cancer outcome prediction.
• This method effectively utilizes previously disregarded unlabeled data, improving classifier accuracy and robustness.
• LDS shows broad applicability across different cancer types, including colorectal and breast cancer.