| The protein family classification problem, which consists of determining the family memberships of given unknown protein sequences, is very important for a biologist for many practical reasons, such as drug discovery, prediction of molecular functions and medical diagnosis. Neural networks and Bayesian methods have performed well on the protein classification problem, achieving accuracy ranging from 90% to 98% while running relatively slowly in the learning stage. In this thesis, we present a principal component null space analysis (PCNSA) linear classifier to the problem and report excellent results compared to those of neural networks and support vector machines. The two main parameters of PCNSA are linked to the high dimensionality of the dataset used, and were optimized in an exhaustive manner to maximize accuracy. |
