Training nu-support vector classifiers: theory and algorithms

A minimum spanning forest based classification method for dedicated breast CT images

PURPOSE

To develop and test an automated algorithm to classify different types of tissue in dedicated breast CT images.

METHODS

Images of a single breast of five different patients were acquired with a dedicated breast CT clinical prototype. The breast CT images were processed by a multiscale bilateral filter to reduce noise while keeping edge information and were corrected to overcome cupping artifacts. As skin and glandular tissue have similar CT values on breast CT images, morphologic processing is used to identify the skin based on its position information. A support vector machine (SVM) is trained and the resulting model used to create a pixelwise classification map of fat and glandular tissue. By combining the results of the skin mask with the SVM results, the breast tissue is classified as skin, fat, and glandular tissue. This map is then used to identify markers for a minimum spanning forest that is grown to segment the image using spatial and intensity information. To evaluate the authors' classification method, they use DICE overlap ratios to compare the results of the automated classification to those obtained by manual segmentation on five patient images.

RESULTS

Comparison between the automatic and the manual segmentation shows that the minimum spanning forest based classification method was able to successfully classify dedicated breast CT image with average DICE ratios of 96.9%, 89.8%, and 89.5% for fat, glandular, and skin tissue, respectively.

CONCLUSIONS

A 2D minimum spanning forest based classification method was proposed and evaluated for classifying the fat, skin, and glandular tissue in dedicated breast CT images. The classification method can be used for dense breast tissue quantification, radiation dose assessment, and other applications in breast imaging.

Wild-type APC predicts poor prognosis in microsatellite-stable proximal colon cancer

Background:

APC mutations (APC-mt) occur in ∼70% of colorectal cancers (CRCs), but their relationship to prognosis is unclear.

Methods:

APC prognostic value was evaluated in 746 stage I–IV CRC patients, stratifying for tumour location and microsatellite instability (MSI). Microarrays were used to identify a gene signature that could classify APC mutation status, and classifier ability to predict prognosis was examined in an independent cohort.

Results:

Wild-type APC microsatellite stable (APC-wt/MSS) tumours from the proximal colon showed poorer overall and recurrence-free survival (OS, RFS) than APC-mt/MSS proximal, APC-wt/MSS distal and APC-mt/MSS distal tumours (OS HR⩾1.79, P⩽0.015; RFS HR⩾1.88, P⩽0.026). APC was a stronger prognostic indicator than BRAF, KRAS, PIK3CA, TP53, CpG island methylator phenotype or chromosomal instability status (P⩽0.036). Microarray analysis similarly revealed poorer survival in MSS proximal cancers with an APC-wt-like signature (P=0.019). APC status did not affect outcomes in MSI tumours. In a validation on 206 patients with proximal colon cancer, APC-wt-like signature MSS cases showed poorer survival than APC-mt-like signature MSS or MSI cases (OS HR⩾2.50, P⩽0.010; RFS HR⩾2.14, P⩽0.025). Poor prognosis APC-wt/MSS proximal tumours exhibited features of the sessile serrated neoplasia pathway (P⩽0.016).

Conclusions:

APC-wt status is a marker of poor prognosis in MSS proximal colon cancer.

Design of a Computer-Assisted System to Automatically Detect Cell Types Using ANA IIF Images for the Diagnosis of Autoimmune Diseases

Indirect immunofluorescence technique applied on HEp-2 cell substrates provides the major screening method to detect ANA patterns in the diagnosis of autoimmune diseases. Currently, the ANA patterns are mostly inspected by experienced physicians to identify abnormal cell patterns. The objective of this study is to design a computer-assisted system to automatically detect cell patterns of IIF images for the diagnosis of autoimmune diseases in the clinical setting. The system simulates the functions of modern flow cytometer and provides the diagnostic reports generated by the system to the technicians and physicians through the radar graphs, box-plots, and tables. The experimental results show that, among the IIF images collected from 17 patients, 6 were classified as coarse-speckled, 3 as diffused, 2 as discrete-speckled, 1 as fine-speckled, 2 as nucleolar, and 3 as peripheral patterns, which were consistent with the patterns determined by the physicians. In addition to recognition of cell patterns, the system also provides the function to automatically generate the report for each patient. The time needed for the whole procedure is less than 30 min, which is more efficient than the manual operation of the physician after inspecting the ANA IIF images. Besides, the system can be easily deployed on many desktop and laptop computers. In conclusion, the designed system, containing functions for automatic detection of ANA cell pattern and generation of diagnostic report, is effective and efficient to assist physicians to diagnose patients with autoimmune diseases. The limitations of the current developed system include (1) only a unique cell pattern was considered for the IIF images collected from a patient, and (2) the cells during the process of mitosis were not adopted for cell classification.

A generic support vector machine model for preoperative glioma survival associations

PURPOSE

To develop a generic support vector machine (SVM) model by using magnetic resonance (MR) imaging-based blood volume distribution data for preoperative glioma survival associations and to prospectively evaluate the diagnostic effectiveness of this model in autonomous patient data.

MATERIALS AND METHODS

Institutional and regional medical ethics committees approved the study, and all patients signed a consent form. Two hundred thirty-five preoperative adult patients from two institutions with a subsequent histologically confirmed diagnosis of glioma after surgery were included retrospectively. An SVM learning technique was applied to MR imaging-based whole-tumor relative cerebral blood volume (rCBV) histograms. SVM models with the highest diagnostic accuracy for 6-month and 1-, 2-, and 3-year survival associations were trained on 101 patients from the first institution. With Cox survival analysis, the diagnostic effectiveness of the SVM models was tested on independent data from 134 patients at the second institution.

RESULTS

were adjusted for known survival predictors, including patient age, tumor size, neurologic status, and postsurgery treatment, and were compared with survival associations from an expert reader.

RESULTS

Compared with total qualitative assessment by an expert reader, the whole-tumor rCBV-based SVM model was the strongest parameter associated with 6-month and 1-, 2-, and 3-year survival in the independent patient data (area under the receiver operating characteristic curve, 0.794-0.851; hazard ratio, 5.4-21.2).

DISCUSSION

Machine learning by means of SVM in combination with whole-tumor rCBV histogram analysis can be used to identify early patient survival in aggressive gliomas. The SVM model returned higher diagnostic accuracy values than an expert reader, and the model appears to be insensitive to patient, observer, and institutional variations.

Imaging tumour heterogeneity of the consequences of a PKCα-substrate interaction in breast cancer patients

Breast cancer heterogeneity demands that prognostic models must be biologically driven and recent clinical evidence indicates that future prognostic signatures need evaluation in the context of early compared with late metastatic risk prediction. In pre-clinical studies, we and others have shown that various protein-protein interactions, pertaining to the actin microfilament-associated proteins, ezrin and cofilin, mediate breast cancer cell migration, a prerequisite for cancer metastasis. Moreover, as a direct substrate for protein kinase Cα, ezrin has been shown to be a determinant of cancer metastasis for a variety of tumour types, besides breast cancer; and has been described as a pivotal regulator of metastasis by linking the plasma membrane to the actin cytoskeleton. In the present article, we demonstrate that our tissue imaging-derived parameters that pertain to or are a consequence of the PKC-ezrin interaction can be used for breast cancer prognostication, with inter-cohort reproducibility. The application of fluorescence lifetime imaging microscopy (FLIM) in formalin-fixed paraffin-embedded patient samples to probe protein proximity within the typically <10 nm range to address the oncological challenge of tumour heterogeneity, is discussed.

Semi-automatic segmentation and classification of Pap smear cells

Cytologic screening has been widely used for detecting the cervical cancers. In this study, a semiautomatic PC-based cellular image analysis system was developed for segmenting nuclear and cytoplasmic contours and for computing morphometric and textual features to train support vector machine (SVM) classifiers to classify four different types of cells and to discriminate dysplastic from normal cells. A software program incorporating function, including image reviewing and standardized denomination of file names, was also designed to facilitate and standardize the workflow of cell analyses. Two experiments were conducted to verify the classification performance. The cross-validation results of the first experiment showed that average accuracies of 97.16% and 98.83%, respectively, for differentiating four different types of cells and in discriminating dysplastic from normal cells have been achieved using salient features (8 for four-cluster and 7 for two-cluster classifiers) selected with SVM recursive feature addition. In the second experiment, 70% (837) of the cell images were used for training and 30% (361) for testing, achieving an accuracy of 96.12% and 98.61% for four-cluster and two-cluster classifiers, respectively. The proposed system provides a feasible and effective tool in evaluating cytologic specimens.

Identification of biomarkers that distinguish chemical contaminants based on gene expression profiles

Background

High throughput transcriptomics profiles such as those generated using microarrays have been useful in identifying biomarkers for different classification and toxicity prediction purposes. Here, we investigated the use of microarrays to predict chemical toxicants and their possible mechanisms of action.

Results

In this study, in vitro cultures of primary rat hepatocytes were exposed to 105 chemicals and vehicle controls, representing 14 compound classes. We comprehensively compared various normalization of gene expression profiles, feature selection and classification algorithms for the classification of these 105 chemicals into14 compound classes. We found that normalization had little effect on the averaged classification accuracy. Two support vector machine (SVM) methods, LibSVM and sequential minimal optimization, had better classification performance than other methods. SVM recursive feature selection (SVM-RFE) had the highest overfitting rate when an independent dataset was used for a prediction. Therefore, we developed a new feature selection algorithm called gradient method that had a relatively high training classification as well as prediction accuracy with the lowest overfitting rate of the methods tested. Analysis of biomarkers that distinguished the 14 classes of compounds identified a group of genes principally involved in cell cycle function that were significantly downregulated by metal and inflammatory compounds, but were induced by anti-microbial, cancer related drugs, pesticides, and PXR mediators.

Conclusions

Our results indicate that using microarrays and a supervised machine learning approach to predict chemical toxicants, their potential toxicity and mechanisms of action is practical and efficient. Choosing the right feature and classification algorithms for this multiple category classification and prediction is critical.

Automated quantitative histology reveals vascular morphodynamics during Arabidopsis hypocotyl secondary growth

Among various advantages, their small size makes model organisms preferred subjects of investigation. Yet, even in model systems detailed analysis of numerous developmental processes at cellular level is severely hampered by their scale. For instance, secondary growth of Arabidopsis hypocotyls creates a radial pattern of highly specialized tissues that comprises several thousand cells starting from a few dozen. This dynamic process is difficult to follow because of its scale and because it can only be investigated invasively, precluding comprehensive understanding of the cell proliferation, differentiation, and patterning events involved. To overcome such limitation, we established an automated quantitative histology approach. We acquired hypocotyl cross-sections from tiled high-resolution images and extracted their information content using custom high-throughput image processing and segmentation. Coupled with automated cell type recognition through machine learning, we could establish a cellular resolution atlas that reveals vascular morphodynamics during secondary growth, for example equidistant phloem pole formation.

DOI:http://dx.doi.org/10.7554/eLife.01567.001

Our understanding of the living world has been advanced greatly by studies of ‘model organisms’, such as mice, zebrafish, and fruit flies. Studying these creatures has been crucial to uncovering the genes that control how our bodies develop and grow, and also to discover the genetic basis of diseases such as cancer.

Thale cress—or Arabidopsis thaliana to give its formal name—is the model organism of choice for many plant biologists. This tiny weed has been widely studied because it can complete its lifecycle, from seed to seed, in about 6 weeks, and because its relatively small genome simplifies the search for genes that control specific traits. However, as with other much-studied model systems, understanding the changes that underpin the development of some of the more complex tissues in Arabidopsis has been severely hampered by the shear number of cells involved.

After it has emerged from the seed, the plant’s first stem will develop from a few dozen cells in width to several thousand cells with highly specialized tissues arranged in a complex pattern of concentric circles. Although this stem thickening process represents a major developmental change in many plants—from Arabidopsis to oak trees—it has been under-researched. This is partly because it involves so many different cells, and also because it can only be observed in thin sections cut out of the plant’s stem.

Now Sankar, Nieminen, Ragni et al. have developed a novel approach, termed ‘automated quantitative histology’, to overcome these problems. This strategy involves ‘teaching’ a computer to automatically recognize different plant cells and to measure their important features in high-resolution images of tissue sections. The resulting ‘map’ of the developing stem—which required over 800 hr of computing time to complete—reveals the changes to cells and tissues as they develop that allow the transport of water, sugars and nutrients between the above- and below-ground organs. Sankar, Nieminen, Ragni et al. suggest that their novel approach could, in the future, also be applied to study the development of other tissues and organisms, including animals.

DOI:http://dx.doi.org/10.7554/eLife.01567.002

Application of Hybrid Functional Groups to Predict ATP Binding Proteins

The ATP binding proteins exist as a hybrid of proteins with Walker A motif and universal stress proteins (USPs) having an alternative motif for binding ATP. There is an urgent need to find a reliable and comprehensive hybrid predictor for ATP binding proteins using whole sequence information. In this paper the open source LIBSVM toolbox was used to build a classifier at 10-fold cross-validation. The best hybrid model was the combination of amino acid and dipeptide composition with an accuracy of 84.57% and Mathews correlation coefficient (MCC) value of 0.693. This classifier proves to be better than many classical ATP binding protein predictors. The general trend observed is that combinations of descriptors performed better and improved the overall performances of individual descriptors, particularly when combined with amino acid composition. The work developed a comprehensive model for predicting ATP binding proteins irrespective of their functional motifs. This model provides a high probability of success for molecular biologists in predicting and selecting diverse groups of ATP binding proteins irrespective of their functional motifs.

Using single-trial EEG to predict and analyze subsequent memory

We show that it is possible to successfully predict subsequent memory performance based on single-trial EEG activity before and during item presentation in the study phase. Two-class classification was conducted to predict subsequently remembered vs. forgotten trials based on subjects' responses in the recognition phase. The overall accuracy across 18 subjects was 59.6% by combining pre- and during-stimulus information. The single-trial classification analysis provides a dimensionality reduction method to project the high-dimensional EEG data onto a discriminative space. These projections revealed novel findings in the pre- and during-stimulus periods related to levels of encoding. It was observed that the pre-stimulus information (specifically oscillatory activity between 25 and 35Hz) -300 to 0ms before stimulus presentation and during-stimulus alpha (7-12Hz) information between 1000 and 1400ms after stimulus onset distinguished between recollection and familiarity while the during-stimulus alpha information and temporal information between 400 and 800ms after stimulus onset mapped these two states to similar values.

Feasibility and finite convergence analysis for accurate on-line ν-support vector machine

The ν-support vector machine ( ν-SVM) for classification has the advantage of using a parameter ν on controlling the number of support vectors and margin errors. Recently, an interesting accurate on-line algorithm accurate on-line ν-SVM algorithm (AONSVM) is proposed for training ν-SVM. AONSVM can be viewed as a special case of parametric quadratic programming techniques. It is demonstrated that AONSVM avoids the infeasible updating path as far as possible, and successfully converges to the optimal solution based on experimental analysis. However, because of the differences between AONSVM and classical parametric quadratic programming techniques, there is no theoretical justification for these conclusions. In this paper, we prove the feasibility and finite convergence of AONSVM under two assumptions. The main results of feasibility analysis include: 1) the inverses of the two key matrices in AONSVM always exist; 2) the rules for updating the two key inverse matrices are reliable; 3) the variable ζ can control the adjustment of the sum of all the weights efficiently; and 4) a sample cannot migrate back and forth in successive adjustment steps among the set of margin support vectors, the set of error support vectors, and the set of the remaining vectors. Moreover, the analyses of AONSVM also provide the proofs of the feasibility and finite convergence for accurate on-line C-SVM learning directly.

Soft margin multiple kernel learning

Multiple kernel learning (MKL) has been proposed for kernel methods by learning the optimal kernel from a set of predefined base kernels. However, the traditional L1MKL method often achieves worse results than the simplest method using the average of base kernels (i.e., average kernel) in some practical applications. In order to improve the effectiveness of MKL, this paper presents a novel soft margin perspective for MKL. Specifically, we introduce an additional slack variable called kernel slack variable to each quadratic constraint of MKL, which corresponds to one support vector machine model using a single base kernel. We first show that L1MKL can be deemed as hard margin MKL, and then we propose a novel soft margin framework for MKL. Three commonly used loss functions, including the hinge loss, the square hinge loss, and the square loss, can be readily incorporated into this framework, leading to the new soft margin MKL objective functions. Many existing MKL methods can be shown as special cases under our soft margin framework. For example, the hinge loss soft margin MKL leads to a new box constraint for kernel combination coefficients. Using different hyper-parameter values for this formulation, we can inherently bridge the method using average kernel, L1MKL, and the hinge loss soft margin MKL. The square hinge loss soft margin MKL unifies the family of elastic net constraint/regularizer based approaches; and the square loss soft margin MKL incorporates L2MKL naturally. Moreover, we also develop efficient algorithms for solving both the hinge loss and square hinge loss soft margin MKL. Comprehensive experimental studies for various MKL algorithms on several benchmark data sets and two real world applications, including video action recognition and event recognition demonstrate that our proposed algorithms can efficiently achieve an effective yet sparse solution for MKL.

Random generalized linear model: a highly accurate and interpretable ensemble predictor

BACKGROUND

Ensemble predictors such as the random forest are known to have superior accuracy but their black-box predictions are difficult to interpret. In contrast, a generalized linear model (GLM) is very interpretable especially when forward feature selection is used to construct the model. However, forward feature selection tends to overfit the data and leads to low predictive accuracy. Therefore, it remains an important research goal to combine the advantages of ensemble predictors (high accuracy) with the advantages of forward regression modeling (interpretability). To address this goal several articles have explored GLM based ensemble predictors. Since limited evaluations suggested that these ensemble predictors were less accurate than alternative predictors, they have found little attention in the literature.

RESULTS

Comprehensive evaluations involving hundreds of genomic data sets, the UCI machine learning benchmark data, and simulations are used to give GLM based ensemble predictors a new and careful look. A novel bootstrap aggregated (bagged) GLM predictor that incorporates several elements of randomness and instability (random subspace method, optional interaction terms, forward variable selection) often outperforms a host of alternative prediction methods including random forests and penalized regression models (ridge regression, elastic net, lasso). This random generalized linear model (RGLM) predictor provides variable importance measures that can be used to define a "thinned" ensemble predictor (involving few features) that retains excellent predictive accuracy.

CONCLUSION

RGLM is a state of the art predictor that shares the advantages of a random forest (excellent predictive accuracy, feature importance measures, out-of-bag estimates of accuracy) with those of a forward selected generalized linear model (interpretability). These methods are implemented in the freely available R software package randomGLM.

A unified classification model based on robust optimization

A wide variety of machine learning algorithms such as the support vector machine (SVM), minimax probability machine (MPM), and Fisher discriminant analysis (FDA) exist for binary classification. The purpose of this letter is to provide a unified classification model that includes these models through a robust optimization approach. This unified model has several benefits. One is that the extensions and improvements intended for SVMs become applicable to MPM and FDA, and vice versa. For example, we can obtain nonconvex variants of MPM and FDA by mimicking Perez-Cruz, Weston, Hermann, and Schölkopf's (2003) extension from convex ν-SVM to nonconvex Eν-SVM. Another benefit is to provide theoretical results concerning these learning methods at once by dealing with the unified model. We give a statistical interpretation of the unified classification model and prove that the model is a good approximation for the worst-case minimization of an expected loss with respect to the uncertain probability distribution. We also propose a nonconvex optimization algorithm that can be applied to nonconvex variants of existing learning methods and show promising numerical results.

Predicting atrial fibrillation inducibility in a canine model by multi-threshold spectra of the recurrence complex network

The purpose of this study is to predict atrial fibrillation (AF) from epicardial signals by investigating the recurrence property of atrial activity dynamic system before AF. A novel scheme is proposed to predict AF by using multi-threshold spectra of the recurrence complex network. Firstly, epicardial signals are transformed into the recurrence complex network to quantify structural properties of the recurrence in the phase space. Spectral parameters with multi-threshold are used to characterize the global structure of the network. Then the feature sequential forward searching algorithm and mutual information based Maximum Relevance Minimum Redundancy criterion are used to find the optimal feature set. Finally, a support vector machine is used to predict the occurrence of AF. This method is assessed on the pre-AF epicardial signals of canine which includes the normal group A (no further AF will happen), the mild group B (the following AF time is less than 180s) and the severe group C (the following AF time is more than 180s). 25 optimal features are selected out of 180 features from each sample. With these features, sensitivity, specificity and accuracy are 99.40%, 99.70% and 99.60%, respectively, which are the best among the recurrence based methods. The results suggest that the proposed method can predict AF accurately and thus can be prospectively used in the postoperative evaluation.

PKMiner: a database for exploring type II polyketide synthases

Background

Bacterial aromatic polyketides are a pharmacologically important group of natural products synthesized by type II polyketide synthases (type II PKSs) in actinobacteria. Isolation of novel aromatic polyketides from microbial sources is currently impeded because of the lack of knowledge about prolific taxa for polyketide synthesis and the difficulties in finding and optimizing target microorganisms. Comprehensive analysis of type II PKSs and the prediction of possible polyketide chemotypes in various actinobacterial genomes will thus enable the discovery or synthesis of novel polyketides in the most plausible microorganisms.

Description

We performed a comprehensive computational analysis of type II PKSs and their gene clusters in actinobacterial genomes. By identifying type II PKS subclasses from the sequence analysis of 280 known type II PKSs, we developed highly accurate domain classifiers for these subclasses and derived prediction rules for aromatic polyketide chemotypes generated by different combinations of type II PKS domains. Using 319 available actinobacterial genomes, we predicted 231 type II PKSs from 40 PKS gene clusters in 25 actinobacterial genomes, and polyketide chemotypes corresponding to 22 novel PKS gene clusters in 16 genomes. These results showed that the microorganisms capable of producing aromatic polyketides are specifically distributed within a certain suborder of Actinomycetales such as Catenulisporineae, Frankineae, Micrococcineae, Micromonosporineae, Pseudonocardineae, Streptomycineae, and Streptosporangineae.

Conclusions

We could identify the novel candidates of type II PKS gene clusters and their polyketide chemotypes in actinobacterial genomes by comprehensive analysis of type II PKSs and prediction of aromatic polyketides. The genome analysis results indicated that the specific suborders in actinomycetes could be used as prolific taxa for polyketide synthesis. The chemotype-prediction rules with the suggested type II PKS modules derived using this resource can be used further for microbial engineering to produce various aromatic polyketides. All these resources, together with the results of the analysis, are organized into an easy-to-use database PKMiner, which is accessible at the following URL: http://pks.kaist.ac.kr/pkminer. We believe that this web-based tool would be useful for research in the discovery of novel bacterial aromatic polyketides.

Regularization path for ν-support vector classification

The v-support vector classification (v-SVC) proposed by Schölkopf has the advantage of using a regularization parameter v for controlling the number of support vectors and margin errors. However, compared to C-SVC, its formulation is more complicated, and to date there are no effective methods for computing its regularization path. In this paper, we propose a new regularization path algorithm, which is designed on the basis of a modified formulation of v-SVC and traces the solution path with respect to the parameter v. Through theoretical analysis and confirmatory experiments, we show that our algorithm can avoid the infeasible updating path under several assumptions (i.e., Assumptions 1 and 2), and fit the entire solution path in a finite number of steps. When the regularization path of v-SVC is available, a novel approach proposed by Yang and Ong can be applied to obtain the global optimal solution of common validation functions for v-SVC, and the computation for the whole process is minimal. Numerical experiments show that it is more efficient than various kinds of grid search methods for selecting the optimal regularization parameter v.

Support vector regression scoring of receptor-ligand complexes for rank-ordering and virtual screening of chemical libraries

The community structure-activity resource (CSAR) data sets are used to develop and test a support vector machine-based scoring function in regression mode (SVR). Two scoring functions (SVR-KB and SVR-EP) are derived with the objective of reproducing the trend of the experimental binding affinities provided within the two CSAR data sets. The features used to train SVR-KB are knowledge-based pairwise potentials, while SVR-EP is based on physicochemical properties. SVR-KB and SVR-EP were compared to seven other widely used scoring functions, including Glide, X-score, GoldScore, ChemScore, Vina, Dock, and PMF. Results showed that SVR-KB trained with features obtained from three-dimensional complexes of the PDBbind data set outperformed all other scoring functions, including best performing X-score, by nearly 0.1 using three correlation coefficients, namely Pearson, Spearman, and Kendall. It was interesting that higher performance in rank ordering did not translate into greater enrichment in virtual screening assessed using the 40 targets of the Directory of Useful Decoys (DUD). To remedy this situation, a variant of SVR-KB (SVR-KBD) was developed by following a target-specific tailoring strategy that we had previously employed to derive SVM-SP. SVR-KBD showed a much higher enrichment, outperforming all other scoring functions tested, and was comparable in performance to our previously derived scoring function SVM-SP.

LigSeeSVM: ligand-based virtual screening using support vector machines and data fusion

Ligand-based in silico drug screening is useful for lead discovery, in particular for those targets without structures. Here, we have developed LigSeeSVM, a ligand-based screening tool using data fusion and Support Vector Machines (SVMs). We used Atom Pair (AP) structure descriptors and Physicochemical (PC) descriptors of compounds to generate SVM-AP and SVM-PC models. Sequentially, the two models were combined using rank-based data fusion to create LigSeeSVM model. LigSeeSVM was evaluated on five data sets. Experimental results show that the performance of LigSeeSVM is better than other ligand-based virtual screening approaches. We believe that LigSeeSVM is useful for lead compounds.

A new disease-specific machine learning approach for the prediction of cancer-causing missense variants

High-throughput genotyping and sequencing techniques are rapidly and inexpensively providing large amounts of human genetic variation data. Single Nucleotide Polymorphisms (SNPs) are an important source of human genome variability and have been implicated in several human diseases, including cancer. Amino acid mutations resulting from non-synonymous SNPs in coding regions may generate protein functional changes that affect cell proliferation. In this study, we developed a machine learning approach to predict cancer-causing missense variants. We present a Support Vector Machine (SVM) classifier trained on a set of 3163 cancer-causing variants and an equal number of neutral polymorphisms. The method achieve 93% overall accuracy, a correlation coefficient of 0.86, and area under ROC curve of 0.98. When compared with other previously developed algorithms such as SIFT and CHASM our method results in higher prediction accuracy and correlation coefficient in identifying cancer-causing variants.

Improving the prediction of disease-related variants using protein three-dimensional structure

Background

Single Nucleotide Polymorphisms (SNPs) are an important source of human genome variability. Non-synonymous SNPs occurring in coding regions result in single amino acid polymorphisms (SAPs) that may affect protein function and lead to pathology. Several methods attempt to estimate the impact of SAPs using different sources of information. Although sequence-based predictors have shown good performance, the quality of these predictions can be further improved by introducing new features derived from three-dimensional protein structures.

Results

In this paper, we present a structure-based machine learning approach for predicting disease-related SAPs. We have trained a Support Vector Machine (SVM) on a set of 3,342 disease-related mutations and 1,644 neutral polymorphisms from 784 protein chains. We use SVM input features derived from the protein’s sequence, structure, and function. After dataset balancing, the structure-based method (SVM-3D) reaches an overall accuracy of 85%, a correlation coefficient of 0.70, and an area under the receiving operating characteristic curve (AUC) of 0.92. When compared with a similar sequence-based predictor, SVM-3D results in an increase of the overall accuracy and AUC by 3%, and correlation coefficient by 0.06. The robustness of this improvement has been tested on different datasets and in all the cases SVM-3D performs better than previously developed methods even when compared with PolyPhen2, which explicitly considers in input protein structure information.

Conclusion

This work demonstrates that structural information can increase the accuracy of disease-related SAPs identification. Our results also quantify the magnitude of improvement on a large dataset. This improvement is in agreement with previously observed results, where structure information enhanced the prediction of protein stability changes upon mutation. Although the structural information contained in the Protein Data Bank is limiting the application and the performance of our structure-based method, we expect that SVM-3D will result in higher accuracy when more structural date become available.

Enabling fast brain-computer interaction by single-trial extraction of visual evoked potentials

This paper investigates the challenging issue of enabling fast brain-computer interaction to construct a mental speller. Exploiting visual evoked potentials as communication carriers, an online paradigm called "imitating-human-natural-reading" is realized. In this online paradigm, single-trial estimation with the intrinsically real-time feature should be used instead of grand average that is traditionally used in the cognitive or clinical experiments. By the use of several montages of component features from four channels with parameter optimization, we explored the support vector machines-based single-trial estimation of evoked potentials. The results on a human-subject show the advantages of the inducing paradigm used in our mental speller with a high classification rate.

Photoacoustic spectroscopy of ovarian normal, benign, and malignant tissues: a pilot study

Photoacoustic spectra of normal, benign, and malignant ovarian tissues are recorded using 325-nm pulsed laser excitation in vitro. A total of 102 (34 normal, 38 benign, and 30 malignant) spectra are obtained from 22 samples belonging to normal, benign, and malignant subjects. Applying multi-algorithm approach, comprised of methods such as, principal component analysis (PCA) based k-nearest neighbor (k-NN) analysis, artificial neural network (ANN) analysis, and support vector machine (SVM) analysis, classification of the data has been carried out. For PCA, first the calibration set is formed by pooling 45 spectra, 15 belonging to each of pathologically certified normal, benign, and malignant samples. PCA is then performed on the data matrix, comprised of the six spectral features extracted from each of 45 calibration samples, and three principal components (PCs) containing maximum diagnostic information are selected. The scores of the selected PCs are used to train the k-NN, ANN, and SVM classifiers. The ANN used is a classical multilayer feed forward network with back propagation algorithm for its training. For k-NN, the Euclidean distance based algorithm is used and for SVM, one-versus-rest multiclass kernel-radial basis function is used. The performance evaluation of the classification results are obtained by calculating statistical parameters like specificity and sensitivity. ANN and k-NN techniques showed identical performance with specificity and sensitivity values of 100 and 86.76%, whereas SVM had these values at 100 and 80.18%, respectively. In order to determine the relative diagnostic performance of the techniques, receiver operating characteristics analysis is also performed.

Enriching for correct prediction of biological processes using a combination of diverse classifiers

Background

Machine learning models (classifiers) for classifying genes to biological processes each have their own unique characteristics in what genes can be classified and to what biological processes. No single learning model is qualitatively superior to any other model and overall precision for each model tends to be low. The classification results for each classifier can be complementary and synergistic suggesting the benefit of a combination of algorithms, but often the prediction probability outputs of various learning models are neither comparable nor compatible for combining. A means to compare outputs regardless of the model and data used and combine the results into an improved comprehensive model is needed.

Results

Gene expression patterns from NCI's panel of 60 cell lines were used to train a Random Forest, a Support Vector Machine and a Neural Network model, plus two over-sampled models for classifying genes to biological processes. Each model produced unique characteristics in the classification results. We introduce the Precision Index measure (PIN) from the maximum posterior probability that allows assessing, comparing and combining multiple classifiers. The class specific precision measure (PIC) is introduced and used to select a subset of predictions across all classes and all classifiers with high precision. We developed a single classifier that combines the PINs from these five models in prediction and found that the PIN Combined Classifier (PINCom) significantly increased the number of correctly predicted genes over any single classifier. The PINCom applied to test genes that were not used in training also showed substantial improvement over any single model.

Conclusions

This paper introduces novel and effective ways of assessing predictions by their precision and recall plus a method that combines several machine learning models and capitalizes on synergy and complementation in class selection, resulting in higher precision and recall. Different machine learning models yielded incongruent results each of which were successfully combined into one superior model using the PIN measure we developed. Validation of the boosted predictions for gene functions showed the genes to be accurately predicted.

LIBSVM

LIBSVM is a library for Support Vector Machines (SVMs). We have been actively developing this package since the year 2000. The goal is to help users to easily apply SVM to their applications. LIBSVM has gained wide popularity in machine learning and many other areas. In this article, we present all implementation details of LIBSVM. Issues such as solving SVM optimization problems theoretical convergence multiclass classification probability estimates and parameter selection are discussed in detail.

Whole genome siRNA cell-based screen links mitochondria to Akt signaling network through uncoupling of electron transport chain

Akt activation sequesters FOXO1a away from its target genes and serves as an endpoint of a complex signaling network. A cell-based RNAi screen reveals an extensive network of genes, including UCP5, which directs nuclear localization of FOXO1a. Silencing of UCP5 disrupts the mitochondria and induces JNK1, creating a link to the Akt signaling network.

Forkhead transcription factors (FOXOs) alter a diverse array of cellular processes including the cell cycle, oxidative stress resistance, and aging. Insulin/Akt activation directs phosphorylation and cytoplasmic sequestration of FOXO away from its target genes and serves as an endpoint of a complex signaling network. Using a human genome small interfering RNA (siRNA) library in a cell-based assay, we identified an extensive network of proteins involved in nuclear export, focal adhesion, and mitochondrial respiration not previously implicated in FOXO localization. Furthermore, a detailed examination of mitochondrial factors revealed that loss of uncoupling protein 5 (UCP5) modifies the energy balance and increases free radicals through up-regulation of uncoupling protein 3 (UCP3). The increased superoxide content induces c-Jun N-terminal kinase 1 (JNK1) kinase activity, which in turn affects FOXO localization through a compensatory dephosphorylation of Akt. The resulting nuclear FOXO increases expression of target genes, including mitochondrial superoxide dismutase. By connecting free radical defense and mitochondrial uncoupling to Akt/FOXO signaling, these results have implications in obesity and type 2 diabetes development and the potential for therapeutic intervention.

Natural evolution of neural support vector machines

Two different neural implementations of support vector machines are described and applied to one-shot trainable pattern recognition. The first model is based on oscillating associative memory and is mapped to the olfactory system. The second model is founded on competitive queuing memory originally employed for generating motor action sequences in the brain. Both models include forward pathways where a stream of support vectors is evoked from memory and merges with sensory input to produce support vector machine classifications. Misclassified events are imprinted as new support vector candidates. Support vector machine weights are tuned by virtual experimentation in sleep. Recalled training examples masquerade as sensor input and feedback from the classification process drives a learning process where support vector weights are optimized. For both support vector machine models it is demonstrated that there is a plausible evolutionary path from a simple hard-wired pattern recognizer to a full implementation of a biological kernel machine. Simple and individually beneficial modifications are accumulated in each step along this path. Neural support vector machines can apparently emerge by natural processes.

Tuning support vector machines for minimax and Neyman-Pearson classification

This paper studies the training of support vector machine (SVM) classifiers with respect to the minimax and Neyman-Pearson criteria. In principle, these criteria can be optimized in a straightforward way using a cost-sensitive SVM. In practice, however, because these criteria require especially accurate error estimation, standard techniques for tuning SVM parameters, such as cross-validation, can lead to poor classifier performance. To address this issue, we first prove that the usual cost-sensitive SVM, here called the 2C-SVM, is equivalent to another formulation called the 2nu-SVM. We then exploit a characterization of the 2nu-SVM parameter space to develop a simple yet powerful approach to error estimation based on smoothing. In an extensive experimental study, we demonstrate that smoothing significantly improves the accuracy of cross-validation error estimates, leading to dramatic performance gains. Furthermore, we propose coordinate descent strategies that offer significant gains in computational efficiency, with little to no loss in performance.

Gene expression profile of BRCAness that correlates with responsiveness to chemotherapy and with outcome in patients with epithelial ovarian cancer

PURPOSE

To define a gene expression profile of BRCAness that correlates with chemotherapy response and outcome in epithelial ovarian cancer (EOC).

METHODS

A publicly available microarray data set including 61 patients with EOC with either sporadic disease or BRCA(1/2) germline mutations was used for development of the BRCAness profile. Correlation with platinum responsiveness was assessed in platinum-sensitive and platinum-resistant tumor biopsy specimens from six patients with BRCA germline mutations. Association with poly-ADP ribose polymerase (PARP) inhibitor responsiveness and with radiation-induced RAD51 foci formation (a surrogate of homologous recombination) was assessed in Capan-1 cell line clones. The BRCAness profile was validated in 70 patients enriched for sporadic disease to assess its association with outcome.

RESULTS

The BRCAness profile accurately predicted platinum responsiveness in eight out of 10 patient-derived tumor specimens, and between PARP-inhibitor sensitivity and resistance in four out of four Capan-1 clones. [corrected] When applied to the 70 patients with sporadic disease, patients with the BRCA-like (BL) profile had improved disease-free survival (34 months v 15 months; log-rank P = .013) and overall survival (72 months v 41 months; log-rank P = .006) compared with patients with a non-BRCA-like (NBL) profile, respectively. The BRCAness profile maintained independent prognostic value in multivariate analysis, which controlled for other known clinical prognostic factors.

CONCLUSION

The BRCAness profile correlates with responsiveness to platinum and PARP inhibitors and identifies a subset of sporadic patients with improved outcome. Additional evaluation of this profile as a predictive tool in patients with sporadic EOC is warranted.

Towards a user-friendly brain-computer interface: initial tests in ALS and PLS patients

OBJECTIVE

Patients usually require long-term training for effective EEG-based brain-computer interface (BCI) control due to fatigue caused by the demands for focused attention during prolonged BCI operation. We intended to develop a user-friendly BCI requiring minimal training and less mental load.

METHODS

Testing of BCI performance was investigated in three patients with amyotrophic lateral sclerosis (ALS) and three patients with primary lateral sclerosis (PLS), who had no previous BCI experience. All patients performed binary control of cursor movement. One ALS patient and one PLS patient performed four-directional cursor control in a two-dimensional domain under a BCI paradigm associated with human natural motor behavior using motor execution and motor imagery. Subjects practiced for 5-10min and then participated in a multi-session study of either binary control or four-directional control including online BCI game over 1.5-2h in a single visit.

RESULTS

Event-related desynchronization and event-related synchronization in the beta band were observed in all patients during the production of voluntary movement either by motor execution or motor imagery. The online binary control of cursor movement was achieved with an average accuracy about 82.1+/-8.2% with motor execution and about 80% with motor imagery, whereas offline accuracy was achieved with 91.4+/-3.4% with motor execution and 83.3+/-8.9% with motor imagery after optimization. In addition, four-directional cursor control was achieved with an accuracy of 50-60% with motor execution and motor imagery.

CONCLUSION

Patients with ALS or PLS may achieve BCI control without extended training, and fatigue might be reduced during operation of a BCI associated with human natural motor behavior.

SIGNIFICANCE

The development of a user-friendly BCI will promote practical BCI applications in paralyzed patients.

A neural support vector machine

Support vector machines are state-of-the-art pattern recognition algorithms that are well founded in optimization and generalization theory but not obviously applicable to the brain. This paper presents Bio-SVM, a biologically feasible support vector machine. An unstable associative memory oscillates between support vectors and interacts with a feed-forward classification pathway. Kernel neurons blend support vectors and sensory input. Downstream temporal integration generates the classification. Instant learning of surprising events and off-line tuning of support vector weights trains the system. Emotion-based learning, forgetting trivia, sleep and brain oscillations are phenomena that agree with the Bio-SVM model. A mapping to the olfactory system is suggested.

Functional annotations improve the predictive score of human disease-related mutations in proteins

Single nucleotide polymorphisms (SNPs) are the simplest and most frequent form of human DNA variation, also valuable as genetic markers of disease susceptibility. The most investigated SNPs are missense mutations resulting in residue substitutions in the protein. Here we propose SNPs&GO, an accurate method that, starting from a protein sequence, can predict whether a mutation is disease related or not by exploiting the protein functional annotation. The scoring efficiency of SNPs&GO is as high as 82%, with a Matthews correlation coefficient equal to 0.63 over a wide set of annotated nonsynonymous mutations in proteins, including 16,330 disease-related and 17,432 neutral polymorphisms. SNPs&GO collects in unique framework information derived from protein sequence, evolutionary information, and function as encoded in the Gene Ontology terms, and outperforms other available predictive methods.

Protein local 3D structure prediction by Super Granule Support Vector Machines (Super GSVM)

BACKGROUND

Understanding the relationship between the protein sequence and the 3D structure is a major research area in bioinformatics. The prediction of complete protein tertiary structure based only on sequence information is still an impractical work. This paper aims at revealing the hidden knowledge of the sequence motifs and the local tertiary structure.

RESULTS

In this paper, we propose a Super Granule Support Vector Machine (Super GSVM) model to obtain the high quality protein sequence motifs and to predict local tertiary structure information based on purely sequence information.

CONCLUSION

The proposed model overcomes the innate shortcoming of using the SVM on such a large data set, which is the inherent computational complexity involved in training support vectors for huge datasets including half million of samples. The satisfactory prediction results show the Super GSVM model generates decent protein sequence clusters and has the ability to capture the hidden sequence-to-structure information. This model also has a strong potential in the application of SVMs on other research areas with huge datasets.

Decoding human motor activity from EEG single trials for a discrete two-dimensional cursor control

This study aims to explore whether human intentions to move or cease to move right and left hands can be decoded from spatiotemporal features in non-invasive EEG in order to control a discrete two-dimensional cursor movement for a potential multidimensional brain-computer interface (BCI). Five naïve subjects performed either sustaining or stopping a motor task with time locking to a predefined time window by using motor execution with physical movement or motor imagery. Spatial filtering, temporal filtering, feature selection and classification methods were explored. The performance of the proposed BCI was evaluated by both offline classification and online two-dimensional cursor control. Event-related desynchronization (ERD) and post-movement event-related synchronization (ERS) were observed on the contralateral hemisphere to the hand moved for both motor execution and motor imagery. Feature analysis showed that EEG beta band activity in the contralateral hemisphere over the motor cortex provided the best detection of either sustained or ceased movement of the right or left hand. The offline classification of four motor tasks (sustain or cease to move right or left hand) provided 10-fold cross-validation accuracy as high as 88% for motor execution and 73% for motor imagery. The subjects participating in experiments with physical movement were able to complete the online game with motor execution at an average accuracy of 85.5 +/- 4.65%; the subjects participating in motor imagery study also completed the game successfully. The proposed BCI provides a new practical multidimensional method by noninvasive EEG signal associated with human natural behavior, which does not need long-term training.