Features play a crucial role in computer vision. Initially designed to detect salient elements by means of handcrafted algorithms, features now are often learned using different layers in convolutional neural networks (CNNs). This paper develops a generic computer vision system based on features extracted from trained CNNs. Multiple learned features are combined into a single structure to work on different image classification tasks. The proposed system was derived by testing several approaches for extracting features from the inner layers of CNNs and using them as inputs to support vector machines that are then combined by sum rule. Several dimensionality reduction techniques were tested for reducing the high dimensionality of the inner layers so that they can work with SVMs. The empirically derived generic vision system based on applying a discrete cosine transform (DCT) separately to each channel is shown to significantly boost the performance of standard CNNs across a large and diverse collection of image data sets. In addition, an ensemble of different topologies taking the same DCT approach and combined with global mean thresholding pooling obtained state-of-the-art results on a benchmark image virus data set.

For splice site recognition, one has to solve two classification problems: discriminating true from decoy splice sites for both acceptor and donor sites. Gene finding systems typically rely on Markov Chains to solve these tasks.

Alpha-helical transmembrane (TM) proteins are involved in a wide range of important biological processes such as cell signaling, transport of membrane-impermeable molecules, cell-cell communication, cell recognition and cell adhesion. Many are also prime drug targets, and it has been estimated that more than half of all drugs currently on the market target membrane proteins. However, due to the experimental difficulties involved in obtaining high quality crystals, this class of protein is severely under-represented in structural databases. In the absence of structural data, sequence-based prediction methods allow TM protein topology to be investigated.

Swallowing sounds from cervical auscultation include information related to the swallowing function. Several studies have been conducted on the screening tests of dysphagia. The literature shows a significant difference between the characteristics of swallowing sounds obtained from different subjects (e.g., healthy and dysphagic subjects; young and old adults). These studies demonstrate the usefulness of swallowing sounds during dysphagic screening. However, the degree of classification for dysphagia based on swallowing sounds has not been thoroughly studied. In this study, we investigate the use of machine learning for classifying swallowing sounds into various types, such as normal swallowing or mild, moderate, and severe dysphagia. In particular, swallowing sounds were recorded from patients with dysphagia. Support vector machines (SVMs) were trained using some features extracted from the obtained swallowing sounds. Moreover, the accuracy of the classification of swallowing sounds using the trained SVMs was evaluated via cross-validation techniques. In the two-class scenario, wherein the swallowing sounds were divided into two categories (viz. normal and dysphagic subjects), the maximum F-measure was 78.9%. In the four-class scenario, where the swallowing sounds were divided into four categories (viz. normal subject, and mild, moderate, and severe dysphagic subjects), the F-measure values for the classes were 65.6%, 53.1%, 51.1%, and 37.1%, respectively.

The goal of this study is to develop, test, and compare multinomial logistic regression (MLR) and support vector machines (SVM) in classifying preschool-aged children physical activity data acquired from an accelerometer. In this study, 69 children aged 3-5 years old were asked to participate in a supervised protocol of physical activities while wearing a triaxial accelerometer. Accelerometer counts, steps, and position were obtained from the device. We applied K-means clustering to determine the number of natural groupings presented by the data. We used MLR and SVM to classify the six activity types. Using direct observation as the criterion method, the 10-fold cross-validation (CV) error rate was used to compare MLR and SVM classifiers, with and without sleep. Altogether, 58 classification models based on combinations of the accelerometer output variables were developed. In general, the SVM classifiers have a smaller 10-fold CV error rate than their MLR counterparts. Including sleep, a SVM classifier provided the best performance with a 10-fold CV error rate of 24.70%. Without sleep, a SVM classifier-based triaxial accelerometer counts, vector magnitude, steps, position, and 1- and 2-min lag and lead values achieved a 10-fold CV error rate of 20.16% and an overall classification error rate of 15.56%. SVM supersedes the classical classifier MLR in categorizing physical activities in preschool-aged children. Using accelerometer data, SVM can be used to correctly classify physical activities typical of preschool-aged children with an acceptable classification error rate.

Identification of microRNAs (miRNAs) is an important step toward understanding post-transcriptional gene regulation and miRNA-related pathology. Difficulties in identifying miRNAs through experimental techniques combined with the huge amount of data from new sequencing technologies have made in silico discrimination of bona fide miRNA precursors from non-miRNA hairpin-like structures an important topic in bioinformatics. Among various techniques developed for this classification problem, machine learning approaches have proved to be the most promising. However these approaches require the use of training data, which is problematic due to an imbalance in the number of miRNAs (positive data) and non-miRNAs (negative data), which leads to a degradation of their performance. In order to address this issue, we present an ensemble method that uses a boosting technique with support vector machine components to deal with imbalanced training data. Classification is performed following a feature selection on 187 novel and existing features. The algorithm, miRBoost, performed better in comparison with state-of-the-art methods on imbalanced human and cross-species data. It also showed the highest ability among the tested methods for discovering novel miRNA precursors. In addition, miRBoost was over 1400 times faster than the second most accurate tool tested and was significantly faster than most of the other tools. miRBoost thus provides a good compromise between prediction efficiency and execution time, making it highly suitable for use in genome-wide miRNA precursor prediction. The software miRBoost is available on our web server http://EvryRNA.ibisc.univ-evry.fr.

One of the major goals in gene and protein expression profiling of cancer is to identify biomarkers and build classification models for prediction of disease prognosis or treatment response. Many traditional statistical methods, based on microarray gene expression data alone and individual genes' discriminatory power, often fail to identify biologically meaningful biomarkers thus resulting in poor prediction performance across data sets. Nonetheless, the variables in multivariable classifiers should synergistically interact to produce more effective classifiers than individual biomarkers.

The increasing size of datasets in drug discovery makes it challenging to build robust and accurate predictive models within a reasonable amount of time. In order to investigate the effect of dataset sizes on predictive performance and modelling time, ligand-based regression models were trained on open datasets of varying sizes of up to 1.2 million chemical structures. For modelling, two implementations of support vector machines (SVM) were used. Chemical structures were described by the signatures molecular descriptor. Results showed that for the larger datasets, the LIBLINEAR SVM implementation performed on par with the well-established libsvm with a radial basis function kernel, but with dramatically less time for model building even on modest computer resources. Using a non-linear kernel proved to be infeasible for large data sizes, even with substantial computational resources on a computer cluster. To deploy the resulting models, we extended the Bioclipse decision support framework to support models from LIBLINEAR and made our models of logD and solubility available from within Bioclipse.

While the genomes of hundreds of organisms have been sequenced and good approaches exist for finding protein encoding genes, an important remaining challenge is predicting the functions of the large fraction of genes for which there is no annotation. Large gene expression datasets from microarray experiments already exist and many of these can be used to help assign potential functions to these genes. We have applied Support Vector Machines (SVM), a sigmoid fitting function and a stratified cross-validation approach to analyze a large microarray experiment dataset from Drosophila melanogaster in order to predict possible functions for previously un-annotated genes. A total of approximately 5043 different genes, or about one-third of the predicted genes in the D. melanogaster genome, are represented in the dataset and 1854 (or 37%) of these genes are un-annotated.

The current progress in sequencing projects calls for rapid, reliable and accurate function assignments of gene products. A variety of methods has been designed to annotate sequences on a large scale. However, these methods can either only be applied for specific subsets, or their results are not formalised, or they do not provide precise confidence estimates for their predictions.

RIKEN's FANTOM project has revealed many previously unknown coding sequences, as well as an unexpected degree of variation in transcripts resulting from alternative promoter usage and splicing. Ever more transcripts that do not code for proteins have been identified by transcriptome studies, in general. Increasing evidence points to the important cellular roles of such non-coding RNAs (ncRNAs). The distinction of protein-coding RNA transcripts from ncRNA transcripts is therefore an important problem in understanding the transcriptome and carrying out its annotation. Very few in silico methods have specifically addressed this problem. Here, we introduce CONC (for "coding or non-coding"), a novel method based on support vector machines that classifies transcripts according to features they would have if they were coding for proteins. These features include peptide length, amino acid composition, predicted secondary structure content, predicted percentage of exposed residues, compositional entropy, number of homologs from database searches, and alignment entropy. Nucleotide frequencies are also incorporated into the method. Confirmed coding cDNAs for eukaryotic proteins from the Swiss-Prot database constituted the set of true positives, ncRNAs from RNAdb and NONCODE the true negatives. Ten-fold cross-validation suggested that CONC distinguished coding RNAs from ncRNAs at about 97% specificity and 98% sensitivity. Applied to 102,801 mouse cDNAs from the FANTOM3 dataset, our method reliably identified over 14,000 ncRNAs and estimated the total number of ncRNAs to be about 28,000.

Conotoxin has been proven to be effective in drug design and could be used to treat various disorders such as schizophrenia, neuromuscular disorders and chronic pain. With the rapidly growing interest in conotoxin, accurate conotoxin superfamily classification tools are desirable to systematize the increasing number of newly discovered sequences and structures. However, despite the significance and extensive experimental investigations on conotoxin, those tools have not been intensively explored.

Alzheimer's disease (AD) is a neurodegenerative disease that mainly affects older adults. Currently, AD is associated with certain hypometabolic biomarkers, beta-amyloid peptides, hyperphosphorylated tau protein, and changes in brain morphology. Accurate diagnosis of AD, as well as mild cognitive impairment (MCI) (prodromal stage of AD), is essential for early care of the disease. As a result, machine learning techniques have been used in recent years for the diagnosis of AD. In this research, we propose a novel methodology to generate a multivariate model that combines different types of features for the detection of AD. In order to obtain a robust biomarker, ADNI baseline data, clinical and neuropsychological assessments (1024 features) of 106 patients were used. The data were normalized, and a genetic algorithm was implemented for the selection of the most significant features. Subsequently, for the development and validation of the multivariate classification model, a support vector machine model was created, and a five-fold cross-validation with an AUC of 87.63% was used to measure model performance. Lastly, an independent blind test of our final model, using 20 patients not considered during the model construction, yielded an AUC of 100%.

Protein-protein interactions are critically dependent on just a few 'hot spot' residues at the interface. Hot spots make a dominant contribution to the free energy of binding and they can disrupt the interaction if mutated to alanine. Here, we present HSPred, a support vector machine(SVM)-based method to predict hot spot residues, given the structure of a complex. HSPred represents an improvement over a previously described approach (Lise et al, BMC Bioinformatics 2009, 10:365). It achieves higher accuracy by treating separately predictions involving either an arginine or a glutamic acid residue. These are the amino acid types on which the original model did not perform well. We have therefore developed two additional SVM classifiers, specifically optimised for these cases. HSPred reaches an overall precision and recall respectively of 61% and 69%, which roughly corresponds to a 10% improvement. An implementation of the described method is available as a web server at http://bioinf.cs.ucl.ac.uk/hspred. It is free to non-commercial users.

Identifying trauma patients at risk of imminent hemorrhagic shock is a challenging task in intraoperative and battlefield settings given the variability of traditional vital signs, such as heart rate and blood pressure, and their inability to detect blood loss at an early stage. To this end, we acquired N = 58 photoplethysmographic (PPG) recordings from both trauma patients with suspected hemorrhage admitted to the hospital, and healthy volunteers subjected to blood withdrawal of 0.9 L. We propose four features to characterize each recording: goodness of fit (r2), the slope of the trend line, percentage change, and the absolute change between amplitude estimates in the heart rate frequency range at the first and last time points. Also, we propose a machine learning algorithm to distinguish between blood loss and no blood loss. The optimal overall accuracy of discriminating between hypovolemia and euvolemia was 88.38%, while sensitivity and specificity were 88.86% and 87.90%, respectively. In addition, the proposed features and algorithm performed well even when moderate blood volume was withdrawn. The results suggest that the proposed features and algorithm are suitable for the automatic discrimination between hypovolemia and euvolemia, and can be beneficial and applicable in both intraoperative/emergency and combat casualty care.

One of the important tasks in cancer research is to identify biomarkers and build classification models for clinical outcome prediction. In this paper, we develop a CyNetSVM software package, implemented in Java and integrated with Cytoscape as an app, to identify network biomarkers using network-constrained support vector machines (NetSVM). The Cytoscape app of NetSVM is specifically designed to improve the usability of NetSVM with the following enhancements: (1) user-friendly graphical user interface (GUI), (2) computationally efficient core program and (3) convenient network visualization capability. The CyNetSVM app has been used to analyze breast cancer data to identify network genes associated with breast cancer recurrence. The biological function of these network genes is enriched in signaling pathways associated with breast cancer progression, showing the effectiveness of CyNetSVM for cancer biomarker identification. The CyNetSVM package is available at Cytoscape App Store and http://sourceforge.net/projects/netsvmjava; a sample data set is also provided at sourceforge.net.

G-protein coupled receptors (GPCRs) represent one of the most important classes of drug targets for pharmaceutical industry and play important roles in cellular signal transduction. Predicting the coupling specificity of GPCRs to G-proteins is vital for further understanding the mechanism of signal transduction and the function of the receptors within a cell, which can provide new clues for pharmaceutical research and development. In this study, the features of amino acid compositions and physiochemical properties of the full-length GPCR sequences have been analyzed and extracted. Based on these features, classifiers have been developed to predict the coupling specificity of GPCRs to G-proteins using support vector machines. The testing results show that this method could obtain better prediction accuracy.

The cancerlectin plays an important role in the initiation, survival, growth, metastasis, and spread of cancer. Therefore, to study the function of cancerlectin is greatly significant because it can help to identify tumor markers and tumor prevention, treatment, and prognosis. However, plenty of studies have generated a large amount of protein data. Traditional prediction methods have been unable to meet the needs of analysis. Developing powerful computational models based on these data to discriminate cancerlectins and non-cancerlectins on a large scale has been treated as one of the most important topics. In this study, we developed a feature extraction method to identify cancerlectins based on fusion of g-gap dipeptides. The analysis of variance was used to select the optimal feature set and a support vector machine was used to classify the data. The rigorous nested 10-fold cross-validation results, demonstrated that our method obtained the prediction accuracy of 83.91% and sensitivity of 83.15%. At the same time, in order to evaluate the performance of the classification model constructed in this work, we constructed a new data set. The prediction accuracy of the new data set reaches 83.3%. Experimental results show that the performance of our method is better than the state-of-the-art methods.

Although many computational methods have been developed to predict protein subcellular localization, most of the methods are limited to the prediction of single-location proteins. Multi-location proteins are either not considered or assumed not existing. However, proteins with multiple locations are particularly interesting because they may have special biological functions, which are essential to both basic research and drug discovery.

The chemical modification of histones at specific DNA regulatory elements is linked to the activation, inactivation and poising of genes. A number of tools exist to predict enhancers from chromatin modification maps, but their practical application is limited because they either (i) consider a smaller number of marks than those necessary to define the various enhancer classes or (ii) work with an excessive number of marks, which is experimentally unviable. We have developed a method for chromatin state detection using support vector machines in combination with genetic algorithm optimization, called ChromaGenSVM. ChromaGenSVM selects optimum combinations of specific histone epigenetic marks to predict enhancers. In an independent test, ChromaGenSVM recovered 88% of the experimentally supported enhancers in the pilot ENCODE region of interferon gamma-treated HeLa cells. Furthermore, ChromaGenSVM successfully combined the profiles of only five distinct methylation and acetylation marks from ChIP-seq libraries done in human CD4(+) T cells to predict ∼21,000 experimentally supported enhancers within 1.0 kb regions and with a precision of ∼90%, thereby improving previous predictions on the same dataset by 21%. The combined results indicate that ChromaGenSVM comfortably outperforms previously published methods and that enhancers are best predicted by specific combinations of histone methylation and acetylation marks.