Bacterial proteins dubbed virulence factors (VFs) are a highly diverse group of sequences, whose only obvious commonality is the very property of being, more or less directly, involved in virulence. It is therefore tempting to speculate whether their prediction, based on direct sequence similarity (seqsim) to known VFs, could be enhanced or even replaced by using machine-learning methods. Specifically, when trained on a large and diverse set of VFs, such may be able to detect putative, non-trivial characteristics shared by otherwise unrelated VF families and therefore better predict novel VFs with insignificant similarity to each individual family. We therefore first reassess the performance of dimer-based Support Vector Machines, as used in the widely used MP3 method, in light of seqsim-only and seqsim/dimer-hybrid classifiers. We then repeat the analysis with a novel, considerably more diverse data set, also addressing the important problem of negative data selection. Finally, we move on to the real-world use case of proteome-wide VF prediction, outlining different approaches to estimating specificity in this scenario. We find that direct seqsim is of unparalleled importance and therefore should always be exploited. Further, we observe strikingly low correlations between different feature and classifier types when ranking proteins by VF likeness. We therefore propose a 'best of each world' approach to prioritize proteins for experimental testing, focussing on the top predictions of each classifier. Further, classifiers for individual VF families should be developed.

MOTIVATION: We expect novel pathogens to arise due to their fast-paced evolution, and new species to be discovered thanks to advances in DNA sequencing and metagenomics. Moreover, recent developments in synthetic biology raise concerns that some strains of bacteria could be modified for malicious purposes. Traditional approaches to open-view pathogen detection depend on databases of known organisms, which limits their performance on unknown, unrecognized and unmapped sequences. In contrast, machine learning methods can infer pathogenic phenotypes from single NGS reads, even though the biological context is unavailable. RESULTS: We present DeePaC, a Deep Learning Approach to Pathogenicity Classification. It includes a flexible framework allowing easy evaluation of neural architectures with reverse-complement parameter sharing. We show that convolutional neural networks and LSTMs outperform the state-of-the-art based on both sequence homology and machine learning. Combining a deep learning approach with integrating the predictions for both mates in a read pair results in cutting the error rate almost in half in comparison to the previous state-of-the-art. AVAILABILITY AND IMPLEMENTATION: The code and the models are available at: https://gitlab.com/rki_bioinformatics/DeePaC. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

There is a growing interest in the mechanisms and the prediction of how flexible peptides bind proteins, often in a highly selective and conserved manner. While both existing small-molecule docking methods and custom protocols can be used, even short peptides make difficult targets owing to their high torsional flexibility. Any benchmarking should therefore start with those. We compiled a meta-data set of 47 complexes with peptides up to five residues, based on 11 related studies from the past decade. Although their highly varying strategies and constraints preclude direct, quantitative comparisons, we still provide a comprehensive overview of the reported results, using a simple yet stringent measure: the quality of the top-scoring peptide pose. Using the entire data set, this is augmented by our own benchmark of AutoDock Vina, a freely available, fast and widely used docking tool. It particularly addresses non-expert users and was therefore implemented in a highly integrated manner. Guidelines addressing important issues such as the amount of sampling required for result reproducibility are so far lacking. Using peptide docking as an example, this is the first study to address these issues in detail. Finally, to encourage further, standardized benchmarking efforts, the compiled data set is made available in an accessible, transparent and extendable manner.

Here we assessed the use of domain families for predicting the functions of whole proteins. These 'functional families' (FunFams) were derived using a protocol that combines sequence clustering with supervised cluster evaluation, relying on available high-quality Gene Ontology (GO) annotation data in the latter step. In essence, the protocol groups domain sequences belonging to the same superfamily into families based on the GO annotations of their parent proteins. An initial test based on enzyme sequences confirmed that the FunFams resemble enzyme (domain) families much better than do families produced by sequence clustering alone. For the CAFA 2011 experiment, we further associated the FunFams with GO terms probabilistically. All target proteins were first submitted to domain superfamily assignment, followed by FunFam assignment and, eventually, function assignment. The latter included an integration step for multi-domain target proteins. The CAFA results put our domain-based approach among the top ten of 31 competing groups and 56 prediction methods, confirming that it outperforms simple pairwise whole-protein sequence comparisons.