Protein Secondary Structure Prediction Using Machine Learning

Protein structure prediction is an important component in understanding protein structures and functions. Accurate prediction of protein secondary structure helps in understanding protein folding. In many applications such as drug discovery it is required to predict the secondary structure of unknown proteins. In this paper we report our first attempt to secondary structure predication, and approach it as a sequence classification problem, where the task is equivalent to assigning a sequence of labels (i.e. helix, sheet, and coil) to the given protein sequence. We propose an ensemble technique that is based on two stochastic supervised machine learning algorithms, namely Maximum Entropy Markov Model (MEMM) and Conditional Random Field (CRF). We identify and implement a set of features that mostly deal with the contextual information. The proposed approach is evaluated with a benchmark dataset, and it yields encouraging performance to explore it further. We obtain the highest predictive accuracy of 61.26% and segment overlap score (SOY) of 52.30%.