BACKGROUND: Plant pathogens cause billions of dollars of crop loss every year and are a major threat to global food security. Effector proteins are the tools such pathogens use to infect the cell, predicting effectors de novo from sequence is difficult because of the heterogeneity of the sequences. We hypothesised that deep learning classifiers based on Convolutional Neural Networks would be able to identify effectors and deliver new insights. RESULTS: We created a training set of manually curated effector sequences from PHI-Base and used these to train a range of model architectures for classifying bacteria, fungal and oomycete sequences. The best performing classifiers had accuracies from 93 to 84%. The models were tested against popular effector detection software on our own test data and data provided with those models. We observed better performance from our models. Specifically our models showed greater accuracy and lower tendencies to call false positives on a secreted protein negative test set and a greater generalisability. We used GRAD-CAM activation map analysis to identify the sequences that activated our CNN-LSTM models and found short but distinct N-terminal regions in each taxon that was indicative of effector sequences. No motifs could be observed in these regions but an analysis of amino acid types indicated differing patterns of enrichment and depletion that varied between taxa. CONCLUSIONS: Small training sets can be used effectively to train highly accurate and sensitive deep learning models without need for the operator to know anything other than sequence and without arbitrary decisions made about what sequence features or physico-chemical properties are important. Biological insight on subsequences important for classification can be achieved by examining the activations in the model.

BACKGROUND: Traditional Map based Cloning approaches, used for the identification of desirable alleles, are extremely labour intensive and years can elapse between the mutagenesis and the detection of the polymorphism. High throughput sequencing based Mapping-by-sequencing approach requires an ordered genome assembly and cannot be used with fragmented, un-scaffolded draft genomes, limiting its application to model species and precluding many important organisms. RESULTS: We addressed this gap in resource and presented a computational method and software implementations called CHERIPIC (Computing Homozygosity Enriched Regions In genomes to Prioritise Identification of Candidate variants). We have successfully validated implementation of CHERIPIC using three different types of bulk segregant sequence data from Arabidopsis, maize and barley, respectively. CONCLUSIONS: CHERIPIC allows users to rapidly analyse bulk segregant sequence data and we have made it available as a pre-packaged binary with all dependencies for Linux and MacOS and as Galaxy tool.

SUMMARY: Rapid technological advances have led to an explosion of biomedical data in recent years. The pace of change has inspired new collaborative approaches for sharing materials and resources to help train life scientists both in the use of cutting-edge bioinformatics tools and databases and in how to analyse and interpret large datasets. A prototype platform for sharing such training resources was recently created by the Bioinformatics Training Network (BTN). Building on this work, we have created a centralized portal for sharing training materials and courses, including a catalogue of trainers and course organizers, and an announcement service for training events. For course organizers, the portal provides opportunities to promote their training events; for trainers, the portal offers an environment for sharing materials, for gaining visibility for their work and promoting their skills; for trainees, it offers a convenient one-stop shop for finding suitable training resources and identifying relevant training events and activities locally and worldwide. AVAILABILITY AND IMPLEMENTATION: http://mygoblet.org/training-portal.

Abstract Motivation: bio-samtools is a Ruby language interface to SAMtools, the highly popular library that provides utilities for manipulating high-throughput sequence alignments in the Sequence Alignment/Map format. Advances in Ruby, now allow us to improve the analysis capabilities and increase bio-samtools utility, allowing users to accomplish a large amount of analysis using a very small amount of code. bio-samtools can also be easily developed to include additional SAMtools methods and hence stay current with the latest SAMtools releases. Results: We have added new Ruby classes for the MPileup and Variant Call Format (VCF) data formats emitted by SAMtools and introduced more analysis methods for variant analysis, including alternative allele calculation and allele frequency calling for SNPs. Our new implementation of bio-samtools also ensures that all the functionality of the SAMtools library is now supported and that bio-samtools can be easily extended to include future changes in SAMtools. bio-samtools 2 also provides methods that allow the user to directly produce visualization of alignment data. Availability and implementation: bio-samtools is available as a BioGem from http://www.biogems.info or as source code from https://github.com/helios/bioruby-samtools under the MIT License. Contact: [email protected]

MOTIVATION: Single nucleotide polymorphism (SNP) discovery is an important preliminary for understanding genetic variation. With current sequencing methods, we can sample genomes comprehensively. SNPs are found by aligning sequence reads against longer assembled references. De Bruijn graphs are efficient data structures that can deal with the vast amount of data from modern technologies. Recent work has shown that the topology of these graphs captures enough information to allow the detection and characterization of genetic variants, offering an alternative to alignment-based methods. Such methods rely on depth-first walks of the graph to identify closing bifurcations. These methods are conservative or generate many false-positive results, particularly when traversing highly inter-connected (complex) regions of the graph or in regions of very high coverage. RESULTS: We devised an algorithm that calls SNPs in converted De Bruijn graphs by enumerating 2k + 2 cycles. We evaluated the accuracy of predicted SNPs by comparison with SNP lists from alignment-based methods. We tested accuracy of the SNP calling using sequence data from 16 ecotypes of Arabidopsis thaliana and found that accuracy was high. We found that SNP calling was even across the genome and genomic feature types. Using sequence-based attributes of the graph to train a decision tree allowed us to increase accuracy of SNP calls further. Together these results indicate that our algorithm is capable of finding SNPs accurately in complex sub-graphs and potentially comprehensively from whole genome graphs. AVAILABILITY AND IMPLEMENTATION: The source code for a C++ implementation of our algorithm is available under the GNU Public Licence v3 at: https://github.com/danmaclean/2kplus2. The datasets used in this study are available at the European Nucleotide Archive, reference ERP00565, http://www.ebi.ac.uk/ena/data/view/ERP000565.

MOTIVATION: Drawing genomic features in attractive and informative ways is a key task in visualization of genomics data. Scalable Vector Graphics (SVG) format is a modern and flexible open standard that provides advanced features including modular graphic design, advanced web interactivity and animation within a suitable client. SVGs do not suffer from loss of image quality on re-scaling and provide the ability to edit individual elements of a graphic on the whole object level independent of the whole image. These features make SVG a potentially useful format for the preparation of publication quality figures including genomic objects such as genes or sequencing coverage and for web applications that require rich user-interaction with the graphical elements. RESULTS: SVGenes is a Ruby-language library that uses SVG primitives to render typical genomic glyphs through a simple and flexible Ruby interface. The library implements a simple Page object that spaces and contains horizontal Track objects that in turn style, colour and positions features within them. Tracks are the level at which visual information is supplied providing the full styling capability of the SVG standard. Genomic entities like genes, transcripts and histograms are modelled in Glyph objects that are attached to a track and take advantage of SVG primitives to render the genomic features in a track as any of a selection of defined glyphs. The feature model within SVGenes is simple but flexible and not dependent on particular existing gene feature formats meaning graphics for any existing datasets can easily be created without need for conversion. AVAILABILITY: The library is provided as a Ruby Gem from https://rubygems.org/gems/bio-svgenes under the MIT license, and open source code is available at https://github.com/danmaclean/bioruby-svgenes also under the MIT License. CONTACT: [email protected].

SUMMARY: Biogem provides a software development environment for the Ruby programming language, which encourages community-based software development for bioinformatics while lowering the barrier to entry and encouraging best practices. Biogem, with its targeted modular and decentralized approach, software generator, tools and tight web integration, is an improved general model for scaling up collaborative open source software development in bioinformatics. AVAILABILITY: Biogem and modules are free and are OSS. Biogem runs on all systems that support recent versions of Ruby, including Linux, Mac OS X and Windows. Further information at http://www.biogems.info. A tutorial is available at http://www.biogems.info/howto.html CONTACT: [email protected].

SUMMARY: Scientists now use high-throughput sequencing technologies and short-read assembly methods to create draft genome assemblies in just days. Tools and pipelines like the assembler, and the workflow management environments make it easy for a non-specialist to implement complicated pipelines to produce genome assemblies and annotations very quickly. Such accessibility results in a proliferation of assemblies and associated files, often for many organisms. These assemblies get used as a working reference by lots of different workers, from a bioinformatician doing gene prediction or a bench scientist designing primers for PCR. Here we describe Gee Fu, a database tool for genomic assembly and feature data, including next-generation sequence alignments. Gee Fu is an instance of a Ruby-On-Rails web application on a feature database that provides web and console interfaces for input, visualization of feature data via AnnoJ, access to data through a web-service interface, an API for direct data access by Ruby scripts and access to feature data stored in BAM files. Gee Fu provides a platform for storing and sharing different versions of an assembly and associated features that can be accessed and updated by bench biologists and bioinformaticians in ways that are easy and useful for each. AVAILABILITY: http://tinyurl.com/geefu CONTACT: [email protected].

BACKGROUND: Next-generation sequencing technologies allow researchers to obtain millions of sequence reads in a single experiment. One important use of the technology is the sequencing of small non-coding regulatory RNAs and the identification of the genomic locales from which they originate. Currently, there is a paucity of methods for finding small RNA generative locales. RESULTS: We describe and implement an algorithm that can determine small RNA generative locales from high-throughput sequencing data. The algorithm creates a network, or graph, of the small RNAs by creating links between them depending on their proximity on the target genome. For each of the sub-networks in the resulting graph the clustering coefficient, a measure of the interconnectedness of the subnetwork, is used to identify the generative locales. We test the algorithm over a wide range of parameters using RFAM sequences as positive controls and demonstrate that the algorithm has good sensitivity and specificity in a range of Arabidopsis and mouse small RNA sequence sets and that the locales it generates are robust to differences in the choice of parameters. CONCLUSIONS: NiBLS is a fast, reliable and sensitive method for determining small RNA locales in high-throughput sequence data that is generally applicable to all classes of small RNA.

This work describes a parallel implementation of a distinctly different concept in self-organizing maps (SOMs)—processing of the input as a whole, in parallel, via neurons attracted to cover regions of the input space rather than single input instances. The algorithm is analyzed from the perspective of its parallelism and a simple multi-machine concept is utilized to demonstrate the parallel nature of the neural network architecture. The nature of the classic SOM (Ozdzynski, P., Lin, A., Liljeholm, M. and Beatty, J. (2002) A parallel implementation of Kohonen's self-organizing map algorithm: performance and scalability. Neurocomputing, 44–46, 567–571) algorithm displays some parallel characteristics, but contains several bottlenecks, which prevent a true parallel execution. Most SOM parallel implementations either deal with expensive hardware that requires considerable algorithm adaptation, or run multiple instances of SOM on multiple machines and choose the most relevant, or adopt controller–worker architecture with the controller machine presenting a significant bottleneck in the parallelism. Our algorithm (Valova, I., Szer, D., Gueorguieva, G. and Buer, A. (2005) A parallel growing architecture for self-organizing maps with unsupervised learning. Neurocomputing, 68C, 177–195), which we eventually named ParaSOM, reveals no need for significant adaptation to be ported to parallelism and the analysis and implementations we discuss illustrate this quality. This work presents two designs for parallelizing our algorithm and analyzes both while demonstrating the easiness in the transition to a parallel paradigm. We have illustrated the performance gain of the proposed implementation and compared it with other similar methods, from the perspectives of both parallelism and nature of the mapping algorithm.

TurSOM is a novel self-organizing map algorithm with the capability of connection reorganization, not just neuron reorganization. This behavior facilitates the ability to map distinct patterns in a given input space. Multiple networks exist, and operate independently. This work presents an application driven approach, based on the theoretical and empirical work of previous TurSOM experiments. TurSOM is a highly robust algorithm, designed to eliminate the need for post processing methods of cluster identification using SOM algorithms. One of the applications TurSOM is suitable for, but obviously not limited to, is image segmentation, as it is demonstrated in this work.

TurSOM, short for Turing self-organizing map, introduces new concepts, responsibilities and mechanisms to the traditional SOM algorithm. It draws its inspiration from Turing unorganized machines, competitive learning techniques, and SOM algorithms. Turing's unorganized machines (TUM) were one of the first computational concepts of modeling the cortex. Turing also described these machines as having self-organizing behaviors. The primary difference between Turing's self-organization description, and more traditional models we are familiar with (Grossberg, Kohonen), are that connections, rather than neurons, self-organize. TurSOM adheres to unsupervised, competitive learning techniques, wherein all neurons, and all connections between them are self-organizing and competing. As such, it presents a novel self-organizing neural network algorithm that eliminates the need for post-processing methods for cluster identification.

This work presents an application of TurSOM for high-dimensional segmentation and cluster identification. TurSOM is a variation algorithm of the SOM algorithm, which introduces a new mechanism of self-organization: connection reorganization. We have theoretically presented TurSOM in very recent previous work, however, the applicability of the novel architecture is expanding as we explore it numerous advantages and possibilities. The intent of these experiments, and TurSOM itself, is to be able to segment and identify various distinct objects in color data (red, green, blue).

The self-organizing map (SOM) is an effective method for topologically mapping datasets. By adapting the neurons to the inputs, the network can conform to the data and form clusters. However, with the classical SOM and growing architectures such as growing cells and growing grid, the neurons are simply points in space and do not cover entire regions of the input space. Therefore, inputs that are introduced after the network is trained need to have cluster membership determined by proximity to the trained neurons. The ParaSOM, being a different SOM architecture, where each neuron actually covers a region of the input space, opens up possibilities for different approaches to clustering and classification. An algorithm has been proposed to take advantage of the unique characteristics of the ParaSOM. The neighbors of each neuron are evaluated by distance to determine cluster separation. Once the clusters have successfully been identified, new inputs can be introduced to effectively determine which, if any, cluster each belongs to.

UNLABELLED: Recent developments in high-throughput sequencing technologies have generated considerable demand for tools to analyse large datasets of small RNA sequences. Here, we describe a suite of web-based tools for processing plant small RNA datasets. Our tools can be used to identify micro RNAs and their targets, compare expression levels in sRNA loci, and find putative trans-acting siRNA loci. AVAILABILITY: The tools are freely available for use at http://srna-tools.cmp.uea.ac.uk.

Genetic algorithms are an effective search technique to utilize when the search space of a problem is very large and an unintelligent brute-force search is too time-consuming. One such problem that would benefit from a genetic algorithm is the optimization of the ParaGSOM, a Self-Organizing Map that processes the input space in parallel. The ParaGSOM has several parameters that can be configured with a wide range of possible values. Each of these parameters can significantly change the behavior of the ParaGSOM, depending on the value. These behavioral changes will affect the ParaGSOM's ability to adapt to the input space, leading to anything from a fast convergence to a slow convergence to no convergence at all. Applying a genetic algorithm to determine the optimal parameters to use for fast, accurate convergence in the ParaGSOM yields results much faster than testing each parameter combination individually. A genetic algorithm gives insight about how particular parameter combinations affect the network and shows how their relationships can be exploited for maximum efficiency of the ParaGSOM.

Self-organizing maps can discover topological and multidimensional patterns using a variety of methods. We apply a parallel algorithm proposed by the authors (ParaSOM), which yields closer and denser approximations than other methods in a fraction of iterations, to a two-dimensional pattern in a parallel environment to demonstrate a high degree of neuron independence. In a second implementation, pieces of a two-dimensional input space are distributed over a network and processed by independent ParaSOM algorithms.