Robust SVD Method for Missing Value Estimation of DNA Microarrays

A majority of DNA microarray datasets contain missing or corrupt values and it is critical to estimate these values accurately. These missing values are most often attributed to insufficient experimental resolution or the presence of foreign objects on the experimental slide’s surface. To improve existing missing value estimation algorithms, this paper introduces and investigates the scalable singular value decomposition (SSVD) solver, which is an improvement upon the Jacobi singular value decomposition (SVD) approach. Experiments were conducted on a study comparing SSVD to the Jacobi and QR SVD methods against several legitimate microarray datasets. The robustness of SSVD is verified by subjecting it to several test cases containing 1–20% of missing values. For nearly all of the test cases across all configurations of missing value percentages, SSVD provides more accurate recovery results than Jacobi and SQ SVD. These numerical results strongly suggest SSVD is a robust and scalable solver.