An Automated Combination of Kernels for Predicting Protein Subcellular Localization

Abstract

Abstract. Protein subcellular localization is a crucial ingredient to many important inferences about cellular processes, including prediction of protein function and protein interactions. While many predictive computational tools have been proposed, they tend to have complicated architectures and require many design decisions from the developer. Here we utilize the multiclass support vector machine (m-SVM) method to directly solve protein subcellular localization without resorting to the common approach of splitting the problem into several binary classification problems. We further propose a general class of protein sequence kernels which considers all motifs, including motifs with gaps. Instead of heuristically selecting one or a few kernels from this family, we utilize a recent extension of SVMs that optimizes over multiple kernels simultaneously. This way, we automatically search over families of possible amino acid motifs. We compare our automated approach to three other predictors on four different datasets, and show that we perform better than the current state of the art. Further, our method provides some insights as to which sequence motifs are most useful for determining subcellular localization, which are in agreement with biological reasoning. Data files, kernel matrices and open source software are available at http://www.fml.mpg.de/raetsch/projects/protsubloc.

Non-sparse Multiple Kernel Learning for Fisher Discriminant Analysis

“…• In Section 3.6 we further compare ℓ p MK-FDA and ℓ p MK-SVM on the protein subcellular localisation problem studied in Zien and Ong (2007) and Ong and Zien (2008). On this dataset ℓ p MK-SVM outperforms ℓ p MK-FDA by a small margin, and the results suggest that given the same set of base kernels, the two MKL algorithms may favour slightly different norms.…”

“…For the first five datasets, due to the kernel functions used, the kernel matrices are by definition spherically normalised: all data points lie on the unit hypersphere in the feature space. For the protein localisation dataset, the kernels are multiplicatively normalised following Ong and Zien (2008) and Kloft et al (2011) to allow comparison with Kloft et al (2011). After normalisation, the kernels are then centred in the feature spaces, as required by ℓ p MK-FDA.…”

“…In this section, we further compare ℓ p MK-FDA and ℓ p MK-SVM on a computational biology problem, namely, the prediction of subcellular localisation of proteins (Zien and Ong, 2007;Ong and Zien, 2008 The protein subcellular localisation problem contains 4 datasets, corresponding to 4 different sets of organisms: plant (plant), non-plant eukaryotes (nonpl), Gram-positive (psortPos) and Gramnegative bacteria (psortNeg). Each of the 4 datasets can be considered as a multiclass classification problem, with the number of classes ranging between 3 and 5.…”

Multiclass multiple kernel learning

“…When comparing our predictor to current state of the art methods, we perform substantially better. Figure 3 summarizes the results in [17] on three datasets.…”

Prediction of protein submitochondria locations based on data fusion of various features of sequences

“…Fruitful knowledge on the functions of proteins can be provided by the information of their subcellular locations. Various effective computational systems have been developed for prediction of protein subcellular locations during the last decade (Aarti and Raghava, 2008;Blum et al, 2009;Chou and Shen, 2008;Fyshe et al, 2008;Gardy et al, 2004;Guda and Subramaniam, 2005;Huang et al, 2008;Lee et al, 2008;Ong and Zien, 2008;Su et al, 2006).The recent progresses about qthis problem were comprehensively and usefully described in Chou and Shen (2007). Another area which has received little attention in the literature is the prediction of protein localization in the organelles of the cell, such as nucleus and mitochondria.…”