Predicting Protein Function by Multi-Label Correlated Semi-Supervised Learning

Abstract

Assigning biological functions to uncharacterized proteins is a fundamental problem in the postgenomic era. The increasing availability of large amounts of data on protein-protein interactions (PPIs) has led to the emergence of a considerable number of computational methods for determining protein function in the context of a network. These algorithms, however, treat each functional class in isolation and thereby often suffer from the difficulty of the scarcity of labeled data. In reality, different functional classes are naturally dependent on one another. We propose a new algorithm, Multi-label Correlated Semi-supervised Learning (MCSL), to incorporate the intrinsic correlations among functional classes into protein function prediction by leveraging the relationships provided by the PPI network and the functional class network. The guiding intuition is that the classification function should be sufficiently smooth on subgraphs where the respective topologies of these two networks are a good match. We encode this intuition as regularized learning with intraclass and interclass consistency, which can be understood as an extension of the graph-based learning with local and global consistency (LGC) method. Cross validation on the yeast proteome illustrates that MCSL consistently outperforms several state-of-the-art methods. Most notably, it effectively overcomes the problem associated with scarcity of label data. The supplementary files are freely available at http://sites.google.com/site/csaijiang/MCSL.

Protein function prediction using weak-label learning

“…, yn] be the original label set with y ik = 1 if protein i has the k-th function, and y ik = 0 otherwise. It is important to incorporate function correlation in protein function prediction [8,13,23]. Various methods are proposed to measure the function correlation.…”

“…Various methods are proposed to measure the function correlation. We define a function correlation matrix C ∈ R K×K based on cosine similarity (also used in [8]) as follows:…”

“…The functions in FunCat are organized in a tree structure. We use the most informative functions as defined in [8] and [23]. Informative functions are the ones that have at least 30 proteins as members, and within the tree structure these functions do no not have a particular descendant node with more than 30 proteins.…”

“…In particular, some techniques organize the classifiers trained for each function according to the function catalogue hierarchical structure [1,15] and then annotate proteins [2,5,21]. Another class of protein function prediction methods incorporate the correlations between the functions (labels) to improve the multi-label prediction accuracy [8,23].…”

“…This type of multilabel prediction problem is referred to as the 'weak label' or 'incomplete class assignment' problem [4,18]. In this paper, unlike traditional multi-label learning methods [2,8,23], we develop a method, called Protein Function Prediction with W eak-label Learning (ProWL), which can annotate proteins with incomplete function assignment in the training set. The approach proposed in [4,18] for multi-label learning with weak labels considers the specified labels of an instance as relevant labels, and all the unspecified labels of the instance as candidates for relevant labels.…”