Search

Automated Protein Function Prediction is the task of automatically predicting functional annotations for a protein based on gold-standard annotations derived from experimental assays. These experiment-based annotations accumulate over time: proteins without annotations get annotated, and new functions of already annotated proteins are discovered. Therefore, function prediction can be considered a combination of two sub-tasks: making predictions on annotated proteins and making predictions on previously unannotated proteins. In previous work, we analyzed the performance of several protein function prediction methods in these two scenarios. Our results showed that GOstruct, which is based on the structured output framework, had lower accuracy in the task of predicting annotations for proteins with existing annotations, while its performance on un-annotated proteins was similar to the performance in cross-validation. In this work, we present GOstruct 2.0 which includes improvements that allow the model to make use of information of a protein's current annotations to better handle the task of predicting novel annotations for previously annotated proteins. This is highly important for model organisms where most proteins have some level of annotations. Experimental results on human data show that GOstruct 2.0 outperforms the original GOstruct in this task, demonstrating the effectiveness of the proposed improvements. This is the first study that focuses on adapting the structured output framework for applications in which labels are incomplete by nature.

Predicting the structure of genes from RNA-Seq data remains a significant challenge in bioinformatics. Although the amount of data available for analysis is growing at an accelerating rate, the capability to leverage these data to construct complete gene models remains elusive. In addition, the tools that predict novel transcripts exhibit poor accuracy. We present a novel approach to predicting splice graphs from RNA-Seq data that uses patterns of acceptor and donor sites to recognize when novel exons can be predicted unequivocally. This simple approach achieves much higher precision and higher recall than methods like Cufflinks or IsoLasso when predicting novel exons from real and simulated data. The ambiguities that arise from RNA-Seq data can preclude making decisive predictions, so we use a realignment procedure that can predict additional novel exons while maintaining high precision. We show that these accurate splice graph predictions provide a suitable basis for making accurate transcript predictions using tools such as IsoLasso and PSGInfer. Using both real and simulated data, we show that this integrated method predicts transcripts with higher recall and precision than using these other tools alone, and in comparison to Cufflinks. SpliceGrapherXT is available from the SpliceGrapher web page at http://SpliceGrapher.sf.net.

Calmodulin (CaM) is a calcium-binding protein that is involved in a variety of cellular processes, interacting with many proteins. Since many CaM interactions are calcium-dependent, they are difficult to detect using high-throughput methods like yeast-two-hybrid. Furthermore, detection of CaM binding sites requires a significant experimental effort. Using a collection of CaM binding sites extracted from the Calmodulin Target Database we trained SVM-based classifiers to detect CaM binding sites using a variety of sequence features; our best classifier achieved an area under the ROC curve of 0.89 for detecting binding site locations at the amino acid level. We apply our classifiers to the problem of detecting CaM binding proteins in Arabidopsis; at a false-positive level of 0.05 we detected 638 novel putative CaM binding proteins. These proteins share overrepresented Gene Ontology terms associated with the functions of known CaM binders.

The problem of predicting protein function using Gene Ontology terms is a hierarchical classification problem. There are a variety of genomic data that are relevant to a protein's function: its sequence, its interactions with other proteins, expression of its gene, etc. Some of these sources (interactions and expression) are species-specific, while protein sequence is comparable across species, which complicates the task of integrating labeled data from a target species with labeled data from other species. We address this problem using the methodology of structured output learning, present a framework based on multi-view learning that is naturally suited for combining both types of data, and demonstrate its effectiveness in making predictions for proteins in S. cerevisiae and M. musculus. The code for our framework is available at http://strut.sourceforge.net.