MOTIVATION: Machine learning may be the most popular computational tool in molecular biology. Providing sustained performance estimates is challenging. The standard cross-validation protocols usually fail in biology. Park and Marcotte found that even refined protocols fail for protein-protein interactions (PPIs). RESULTS: Here, we sketch additional problems for the prediction of PPIs from sequence alone. First, it not only matters whether proteins A or B of a target interaction A-B are similar to proteins of training interactions (positives), but also whether A or B are similar to proteins of non-interactions (negatives). Second, training on multiple interaction partners per protein did not improve performance for new proteins (not used to train). In contrary, a strictly non-redundant training that ignored good data slightly improved the prediction of difficult cases. Third, which prediction method appears to be best crucially depends on the sequence similarity between the test and the training set, how many true interactions should be found and the expected ratio of negatives to positives. The correct assessment of performance is the most complicated task in the development of prediction methods. Our analyses suggest that PPIs square the challenge for this task.

MOTIVATION: Many methods predict the physical interaction between two proteins (protein-protein interactions; PPIs) from sequence alone. Their performance drops substantially for proteins not used for training. RESULTS: Here, we introduce a new approach to predict PPIs from sequence alone which is based on evolutionary profiles and profile-kernel support vector machines. It improved over the state-of-the-art, in particular for proteins that are sequence-dissimilar to proteins with known interaction partners. Filtering by gene expression data increased accuracy further for the few, most reliably predicted interactions (low recall). The overall improvement was so substantial that we compiled a list of the most reliably predicted PPIs in human. Our method makes a significant difference for biology because it improves most for the majority of proteins without experimental annotations. AVAILABILITY AND IMPLEMENTATION: Implementation and most reliably predicted human PPIs available at https://rostlab.org/owiki/index.php/Profppikernel.

BACKGROUND: Any method that de novo predicts protein function should do better than random. More challenging, it also ought to outperform simple homology-based inference. METHODS: Here, we describe a few methods that predict protein function exclusively through homology. Together, they set the bar or lower limit for future improvements. RESULTS AND CONCLUSIONS: During the development of these methods, we faced two surprises. Firstly, our most successful implementation for the baseline ranked very high at CAFA1. In fact, our best combination of homology-based methods fared only slightly worse than the top-of-the-line prediction method from the Jones group. Secondly, although the concept of homology-based inference is simple, this work revealed that the precise details of the implementation are crucial: not only did the methods span from top to bottom performers at CAFA, but also the reasons for these differences were unexpected. In this work, we also propose a new rigorous measure to compare predicted and experimental annotations. It puts more emphasis on the details of protein function than the other measures employed by CAFA and may best reflect the expectations of users. Clearly, the definition of proper goals remains one major objective for CAFA.

MOTIVATION: Subcellular localization is one aspect of protein function. Despite advances in high-throughput imaging, localization maps remain incomplete. Several methods accurately predict localization, but many challenges remain to be tackled. RESULTS: In this study, we introduced a framework to predict localization in life's three domains, including globular and membrane proteins (3 classes for archaea; 6 for bacteria and 18 for eukaryota). The resulting method, LocTree2, works well even for protein fragments. It uses a hierarchical system of support vector machines that imitates the cascading mechanism of cellular sorting. The method reaches high levels of sustained performance (eukaryota: Q18=65%, bacteria: Q6=84%). LocTree2 also accurately distinguishes membrane and non-membrane proteins. In our hands, it compared favorably with top methods when tested on new data. AVAILABILITY: Online through PredictProtein (predictprotein.org); as standalone version at http://www.rostlab.org/services/loctree2. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The intricate molecular details of protein-protein interactions (PPIs) are crucial for function. Therefore, measuring the same interacting protein pair again, we expect the same result. This work measured the similarity in the molecular details of interaction for the same and for homologous protein pairs between different experiments. All scores analyzed suggested that different experiments often find exceptions in the interfaces of similar PPIs: up to 22% of all comparisons revealed some differences even for sequence-identical pairs of proteins. The corresponding number for pairs of close homologs reached 68%. Conversely, the interfaces differed entirely for 12-29% of all comparisons. All these estimates were calculated after redundancy reduction. The magnitude of interface differences ranged from subtle to the extreme, as illustrated by a few examples. An extreme case was a change of the interacting domains between two observations of the same biological interaction. One reason for different interfaces was the number of copies of an interaction in the same complex: the probability of observing alternative binding modes increases with the number of copies. Even after removing the special cases with alternative hetero-interfaces to the same homomer, a substantial variability remained. Our results strongly support the surprising notion that there are many alternative solutions to make the intricate molecular details of PPIs crucial for function.

MOTIVATION: The slow growth of expert-curated databases compared to experimental databases makes it necessary to build upon highly accurate automated processing pipelines to make the most of the data until curation becomes available. We address this problem in the context of protein structures and their classification into structural and functional classes, more specifically, the structural classification of proteins (SCOP). Structural alignment methods like Vorolign already provide good classification results, but effectively work in a 1-Nearest Neighbor mode. Model-based (in contrast to instance-based) approaches so far have been shown to be of limited values due to small classes arising in such classification schemes. RESULTS: In this article, we describe how kernels defined in terms of Vorolign scores can be used in SVM learning, and explore variants of combined instance-based and model-based learning, up to exclusively model-based learning. Our results suggest that kernels based on Vorolign scores are effective and that model-based learning can yield highly competitive classification results for the prediction of SCOP families. AVAILABILITY: The code is made available at: http://wwwkramer.in.tum.de/research/applications/vorolign-kernel.