Comparison of Machine Learning Algorithms for Identifying Cancer Types

This coauthored 2014 poster compares neural networks, support-vector machines, and random forests for classifying cancer types from gene expression. It was shown at the MCBIOS conference.

The poster by Garima Saxena, Joseph Helsing, Omar Costilla Reyes, and Rajeev K. Azad of the University of North Texas was shown at the MidSouth Computational Biology and Bioinformatics Society (MCBIOS) annual conference in 2014, and the University of North Texas Digital Library archives it. Dr. Costilla Reyes contributed during his master’s studies at the university, where his coursework included bioinformatics alongside signal processing and parallel computing.

Microarray technology measures the activity of thousands of genes in a tissue sample at once, which helps in studying a disease such as cancer, where many genes interact and tumor types differ. The resulting tables usually contain far more measured genes than samples, so classifying a sample correctly is a demanding statistical task. The poster follows an earlier study that used random forests, an ensemble of many decision trees that vote on a class, and asks how artificial neural networks and support-vector machines compare with them and how changing each algorithm’s settings affects performance.

The authors classified samples in a range of cancer gene-expression datasets shared by that earlier study, covering tumor types such as breast, colon, leukemia, lymphoma, and prostate. They used the WEKA software package and varied settings such as the number of trees, the layers and learning rates of the neural networks, and the kernels of the support-vector machines. According to the poster, random forests were the strongest overall, leading on most of the datasets, with neural networks leading on the rest; support-vector machines outperformed neural networks in some cases, and tuning the settings raised accuracy for some datasets. The comparison shows how the choice of algorithm and its settings shape the way gene-expression measurements are turned into cancer-type predictions, which is useful guidance for computational biologists who analyze molecular data. It belongs with Dr. Costilla Reyes’s wider work applying machine learning to biomedical and healthcare data.