Learning Classifiers for Assigning Protein Sequences to Gene Ontology Functional Families

Assigning putative functions to novel proteins and the discovery of sequence correlates of protein function are important challenges in bioinformatics. In this paper, we explore several machine learning approaches to data-driven construction of classifiers for assigning protein sequences to appropriate Gene Ontology (GO) function families using a class conditional probabilistic representation of amino acid sequences. Specifically, we represent protein sequences using class conditional probability distribution of amino acids (amino acid composition) or short (k-letter) subsequences (k-grams) of amino acids. We compare a model (NB k-grams) that ignores the statistical dependencies among overlapping k-grams with an alternative, NB(k), that uses an undirected probabilistic graphical model that captures the relevant dependencies. These two methods require only one pass through the training data during the learning phase, making them especially attractive in settings where there is a need to update the classifiers as new training data become available. We also explore a support vector machine (SVM) classifier, SVM k-grams, trained on the k-gram class conditional probability distributions of sequences. We report the performance of the resulting classifiers on three data sets of functional families from the Gene Ontology (GO) database. Our results show that NB(k) classifier outperforms NB k-grams in terms of accuracy of classification (as measured by cross-validation) by a few percentage points. SVM k-grams outperforms NB(k) in the majority of test cases. These results suggest the possibility of developing fully automated and computationally efficient approaches to construction of classifiers based on undirected graphical models of overlapping k-grams that can be easily updated as additional training data become available. Our results also show that further gains in accuracy of the classifiers are achievable (at the expense of increased computational demands and hence greater difficulty of frequent updates to the classifier as new training data become available) using SVM k-grams.