Number Of False Positive Predictions Research Articles

Evaluating Risk Prediction with Data Collection Costs: Novel Estimation of Test Tradeoff Curves.

The test tradeoff curve helps investigators decide if collecting data for risk prediction is worthwhile when risk prediction is used for treatment decisions. At a given benefit-cost ratio (the number of false-positive predictions one would trade for a true positive prediction) or risk threshold (the probability of developing disease at indifference between treatment and no treatment), the test tradeoff is the minimum number of data collections per true positive to yield a positive maximum expected utility of risk prediction. For example, a test tradeoff of 3,000 invasive tests per true-positive prediction of cancer may suggest that risk prediction is not worthwhile. A test tradeoff curve plots test tradeoff versus benefit-cost ratio or risk threshold. The test tradeoff curve evaluates risk prediction at the optimal risk score cutpoint for treatment, which is the cutpoint of the risk score (the estimated risk of developing disease) that maximizes the expected utility of risk prediction when the receiver-operating characteristic (ROC) curve is concave. Previous methods for estimating the test tradeoff required grouping risk scores. Using individual risk scores, the new method estimates a concave ROC curve by constructing a concave envelope of ROC points, taking a slope-based moving average, minimizing a sum of squared errors, and connecting successive ROC points with line segments. The estimated concave ROC curve yields an estimated test tradeoff curve. Analyses of 2 synthetic data sets illustrate the method. Estimating the test tradeoff curve based on individual risk scores is straightforward to implement and more appealing than previous estimation methods that required grouping risk scores. The test tradeoff curve helps investigators decide if collecting data for risk prediction is worthwhile when risk prediction is used for treatment decisions.At a given benefit-cost ratio or risk threshold, the test tradeoff is the minimum number of data collections per true positive to yield a positive maximum expected utility of risk prediction.Unlike previous estimation methods that grouped risk scores, the method uses individual risk scores to estimate a concave ROC curve, which yields an estimated test tradeoff curve.

β-LacFamPred: An online tool for prediction and classification of β-lactamase class, subclass, and family.

β-Lactams are a broad class of antimicrobial agents with a high safety profile, making them the most widely used class in clinical, agricultural, and veterinary setups. The widespread use of β-lactams has induced the extensive spread of β-lactamase hydrolyzing enzymes known as β-lactamases (BLs). To neutralize the effect of β-lactamases, newer generations of β-lactams have been developed, which ultimately led to the evolution of a highly diverse family of BLs. Based on sequence homology, BLs are categorized into four classes: A-D in Ambler's classification system. Further, each class is subdivided into families. Class B is first divided into subclasses B1-B3, and then each subclass is divided into families. The class to which a BL belongs gives a lot of insight into its hydrolytic profile. Traditional methods of determining the hydrolytic profile of BLs and their classification are time-consuming and require resources. Hence we developed a machine-learning-based in silico method, named as β-LacFamPred, for the prediction and annotation of Ambler's class, subclass, and 96 families of BLs. During leave-one-out cross-validation, except one all β-LacFamPred model HMMs showed 100% accuracy. Benchmarking with other BL family prediction methods showed β-LacFamPred to be the most accurate. Out of 60 penicillin-binding proteins (PBPs) and 57 glyoxalase II proteins, β-LacFamPred correctly predicted 56 PBPs and none of the glyoxalase II sequences as non-BLs. Proteome-wide annotation of BLs by β-LacFamPred showed a very less number of false-positive predictions in comparison to the recently developed BL class prediction tool DeepBL. β-LacFamPred is available both as a web-server and standalone tool at http://proteininformatics.org/mkumar/blacfampred and GitHub repository https://github.com/mkubiophysics/B-LacFamPred respectively.

Drug Target Identification with Machine Learning: How to Choose Negative Examples.

Identification of the protein targets of hit molecules is essential in the drug discovery process. Target prediction with machine learning algorithms can help accelerate this search, limiting the number of required experiments. However, Drug-Target Interactions databases used for training present high statistical bias, leading to a high number of false positives, thus increasing time and cost of experimental validation campaigns. To minimize the number of false positives among predicted targets, we propose a new scheme for choosing negative examples, so that each protein and each drug appears an equal number of times in positive and negative examples. We artificially reproduce the process of target identification for three specific drugs, and more globally for 200 approved drugs. For the detailed three drug examples, and for the larger set of 200 drugs, training with the proposed scheme for the choice of negative examples improved target prediction results: the average number of false positives among the top ranked predicted targets decreased, and overall, the rank of the true targets was improved.Our method corrects databases’ statistical bias and reduces the number of false positive predictions, and therefore the number of useless experiments potentially undertaken.

ACME: pan-specific peptide-MHC class I binding prediction through attention-based deep neural networks.

Prediction of peptide binding to the major histocompatibility complex (MHC) plays a vital role in the development of therapeutic vaccines for the treatment of cancer. Algorithms with improved correlations between predicted and actual binding affinities are needed to increase precision and reduce the number of false positive predictions. We present ACME (Attention-based Convolutional neural networks for MHC Epitope binding prediction), a new pan-specific algorithm to accurately predict the binding affinities between peptides and MHC class I molecules, even for those new alleles that are not seen in the training data. Extensive tests have demonstrated that ACME can significantly outperform other state-of-the-art prediction methods with an increase of the Pearson correlation coefficient between predicted and measured binding affinities by up to 23 percentage points. In addition, its ability to identify strong-binding peptides has been experimentally validated. Moreover, by integrating the convolutional neural network with attention mechanism, ACME is able to extract interpretable patterns that can provide useful and detailed insights into the binding preferences between peptides and their MHC partners. All these results have demonstrated that ACME can provide a powerful and practically useful tool for the studies of peptide-MHC class I interactions. ACME is available as an open source software at https://github.com/HYsxe/ACME. Supplementary data are available at Bioinformatics online.

DeePromoter: Robust Promoter Predictor Using Deep Learning.

The promoter region is located near the transcription start sites and regulates transcription initiation of the gene by controlling the binding of RNA polymerase. Thus, promoter region recognition is an important area of interest in the field of bioinformatics. Numerous tools for promoter prediction were proposed. However, the reliability of these tools still needs to be improved. In this work, we propose a robust deep learning model, called DeePromoter, to analyze the characteristics of the short eukaryotic promoter sequences, and accurately recognize the human and mouse promoter sequences. DeePromoter combines a convolutional neural network (CNN) and a long short-term memory (LSTM). Additionally, instead of using non-promoter regions of the genome as a negative set, we derive a more challenging negative set from the promoter sequences. The proposed negative set reconstruction method improves the discrimination ability and significantly reduces the number of false positive predictions. Consequently, DeePromoter outperforms the previously proposed promoter prediction tools. In addition, a web-server for promoter prediction is developed based on the proposed methods and made available at https://home.jbnu.ac.kr/NSCL/deepromoter.htm.

An SVM-based method for assessment of transcription factor-DNA complex models

BackgroundAtomic details of protein-DNA complexes can provide insightful information for better understanding of the function and binding specificity of DNA binding proteins. In addition to experimental methods for solving protein-DNA complex structures, protein-DNA docking can be used to predict native or near-native complex models. A docking program typically generates a large number of complex conformations and predicts the complex model(s) based on interaction energies between protein and DNA. However, the prediction accuracy is hampered by current approaches to model assessment, especially when docking simulations fail to produce any near-native models.ResultsWe present here a Support Vector Machine (SVM)-based approach for quality assessment of the predicted transcription factor (TF)-DNA complex models. Besides a knowledge-based protein-DNA interaction potential DDNA3, we applied several structural features that have been shown to play important roles in binding specificity between transcription factors and DNA molecules to quality assessment of complex models. To address the issue of unbalanced positive and negative cases in the training dataset, we applied hard-negative mining, an iterative training process that selects an initial training dataset by combining all of the positive cases and a random sample from the negative cases. Results show that the SVM model greatly improves prediction accuracy (84.2%) over two knowledge-based protein-DNA interaction potentials, orientation potential (60.8%) and DDNA3 (68.4%). The improvement is achieved through reducing the number of false positive predictions, especially for the hard docking cases, in which a docking algorithm fails to produce any near-native complex models.ConclusionsA learning-based SVM scoring model with structural features for specific protein-DNA binding and an atomic-level protein-DNA interaction potential DDNA3 significantly improves prediction accuracy of complex models by successfully identifying cases without near-native structural models.

Improving eQTL Analysis Using a Machine Learning Approach for Data Integration: A Logistic Model Tree Solution.

Expression quantitative trait loci (eQTL) analysis is an emerging method for establishing the impact of genetic variations (such as single nucleotide polymorphisms) on the expression levels of genes. Although different methods for evaluating the impact of these variations are proposed in the literature, the results obtained are mostly in disagreement, entailing a considerable number of false-positive predictions. For this reason, we propose an approach based on Logistic Model Trees that integrates the predictions of different eQTL mapping tools to produce more reliable results. More precisely, we employ a machine learning-based method using logistic functions to perform a linear regression able to classify the predictions of three eQTL analysis tools (namely, R/qtl, MatrixEQTL, and mRMR). Given the lack of a reference dataset and that computational predictions are not so easy to test experimentally, the performance of our approach is assessed using data from the DREAM5 challenge. The results show the quality of the aggregated prediction is better than that obtained by each single tool in terms of both precision and recall. We also performed a test on real data, employing genotypes and microRNA expression profiles from Caenorhabditis elegans, which proved that we were able to correctly classify all the experimentally validated eQTLs. These good results come both from the integration of the different predictions, and from the ability of this machine learning algorithm to find the best cutoff thresholds for each tool. This combination makes our integration approach suitable for improving eQTL predictions for testing in a laboratory, reducing the number of false-positive results.

Machine Learning Methods for Rapid Prediction of Thermostabilizing Mutants of G‐protein‐coupled receptors in Detergents

G protein‐coupled receptors (GPCRs) are highly dynamic and often denature when extracted in detergents. Deriving thermostable mutants has been a successful strategy to stabilize GPCRs in detergents, but this process is experimentally tedious. We have developed a computational method that scans the entire receptor for alanine or leucine mutations and calculates the stability score for mutation of each position in the GPCR structural model. The method involves generating homology models of the receptors of varying accuracies and an ensemble of conformations by sampling the rigid body degrees of freedom of transmembrane helices. Then, an all‐atom force field function is used to calculate the enthalpy gain, known as the “stability score” upon mutation of every residue, in these receptor structures, to alanine. Inclusion of conformational sampling to account for structural perturbations improves the thermostability predictions compared to using a single structure. We have validated the method against experimentally measured thermostability data for single mutants of the β1‐adrenergic receptor (β1AR), adenosine A2A receptor (A2AR) and neurotensin receptor 1 (NTSR1). We will present some of our recent results on blind predictions of thermostabilizing mutations on NTSR1.Using the computed energy components in our dataset, we have recently tested several machine learning algorithms for rapid and reliable prediction of thermostable and non‐thermostable mutants. We used Matlab to build a binary classification model (thermostable vs. non‐thermostable) trained on data from an ensemble of three proteins (25% of dataset used for holdout cross‐validation), and tested against a fourth protein omitted from the training set. Initial dataset was heavily imbalanced to the “non‐thermostable” class. This was balanced by random replication of the undersampled “thermostable” class. Models trained by decision tree algorithms performed best in our study compared to k‐nearest neighbor (k‐NN) algorithms. Our precision and recall statistics for the best performing ensemble dataset were as follows: Boosted Trees – 62% precision, 34% recall, Simple Trees – 30% precision, 90% recall, Complex Trees – 36% precision, 73% recall, Medium k‐NN – 28% precision, 34% recall.These results are promising for predicting true thermostable mutants, and we are working to improve our methods for reducing the number of false positive predictions. This requires testing different training algorithms such as Support Vector Machines and other k‐NN and Decision Tree models, and potentially ensemble training models combining multiples of these algorithms. Our method is the first step toward a computational method for rapid prediction of thermostable mutants of GPCRs. These tools are critical for reliable and quick prediction of thermostabilizing mutations of GPCRs for drug discovery.Support or Funding InformationFunding for this work was from NIH‐RO1GM097261 to N.V.This abstract is from the Experimental Biology 2018 Meeting. There is no full text article associated with this abstract published in The FASEB Journal.

SigSeeker: a peak-calling ensemble approach for constructing epigenetic signatures

Epigenetic data are invaluable when determining the regulatory programs governing a cell. Based on use of next-generation sequencing data for characterizing epigenetic marks and transcription factor binding, numerous peak-calling approaches have been developed to determine sites of genomic significance in these data. Such analyses can produce a large number of false positive predictions, suggesting that sites supported by multiple algorithms provide a stronger foundation for inferring and characterizing regulatory programs associated with the epigenetic data. Few methodologies integrate epigenetic based predictions of multiple approaches when combining profiles generated by different tools. The SigSeeker peak-calling ensemble uses multiple tools to identify peaks, and with user-defined thresholds for peak overlap and signal strength it retains only those peaks that are concordant across multiple tools. Peaks predicted to be co-localized by only a very small number of tools, discovered to be only marginally overlapping, or found to represent significant outliers to the approximation model are removed from the results, providing concise and high quality epigenetic datasets. SigSeeker has been validated using established benchmarks for transcription factor binding and histone modification ChIP-Seq data. These comparisons indicate that the quality of our ensemble technique exceeds that of single tool approaches, enhances existing peak-calling ensembles, and results in epigenetic profiles of higher confidence. http://sigseeker.org. lichtenbergj@mail.nih.gov. Supplementary data are available at Bioinformatics online.

Machine learning classifier for identification of damaging missense mutations exclusive to human mitochondrial DNA-encoded polypeptides

BackgroundSeveral methods have been developed to predict the pathogenicity of missense mutations but none has been specifically designed for classification of variants in mtDNA-encoded polypeptides. Moreover, there is not available curated dataset of neutral and damaging mtDNA missense variants to test the accuracy of predictors. Because mtDNA sequencing of patients suffering mitochondrial diseases is revealing many missense mutations, it is needed to prioritize candidate substitutions for further confirmation. Predictors can be useful as screening tools but their performance must be improved.ResultsWe have developed a SVM classifier (Mitoclass.1) specific for mtDNA missense variants. Training and validation of the model was executed with 2,835 mtDNA damaging and neutral amino acid substitutions, previously curated by a set of rigorous pathogenicity criteria with high specificity. Each instance is described by a set of three attributes based on evolutionary conservation in Eukaryota of wildtype and mutant amino acids as well as coevolution and a novel evolutionary analysis of specific substitutions belonging to the same domain of mitochondrial polypeptides. Our classifier has performed better than other web-available tested predictors.We checked performance of three broadly used predictors with the total mutations of our curated dataset. PolyPhen-2 showed the best results for a screening proposal with a good sensitivity. Nevertheless, the number of false positive predictions was too high. Our method has an improved sensitivity and better specificity in relation to PolyPhen-2. We also publish predictions for the complete set of 24,201 possible missense variants in the 13 human mtDNA-encoded polypeptides.ConclusionsMitoclass.1 allows a better selection of candidate damaging missense variants from mtDNA. A careful search of discriminatory attributes and a training step based on a curated dataset of amino acid substitutions belonging exclusively to human mtDNA genes allows an improved performance. Mitoclass.1 accuracy could be improved in the future when more mtDNA missense substitutions will be available for updating the attributes and retraining the model.

Computational and Experimental Identification of Tissue-Specific MicroRNA Targets.

In this chapter we discuss computational methods for the prediction of microRNA (miRNA) targets. More specifically, we consider machine learning-based approaches and explain why these methods have been relatively unsuccessful in reducing the number of false positive predictions. Further we suggest approaches designed to improve their performance by considering tissue-specific target regulation. We argue that the miRNA targetome differs depending on the tissue type and introduce a novel algorithm that predicts miRNA targets specifically for colorectal cancer. We discuss features of miRNAs and target sites that affect target recognition, and how next-generation sequencing data can support the identification of novel miRNAs, differentially expressed miRNAs and their tissue-specific mRNA targets. In addition, we introduce some experimental approaches for the validation of miRNA targets as well as web-based resources sharing predicted and validated miRNA target interactions.

Abstract 3762: The landscape of therapeutic targetable fusions

Abstract Gene fusions are an important class of cancer-contributing somatic alteration, and have an important significance as a tumor-initiating event and as a molecular therapeutic target for specific tumors. We analyzed RNA sequencing, DNA copy number and gene mutation data from 4,366 primary tumor samples to comprehensively detect fusion events in 13 tumor types. To reduce the number of false-positive predictions, we developed strict quality criteria on the basis of gene homology, transcript allele fraction (TAF), and partner gene variety. In addition, we used fusions detected in RNAseq data from 364 normal tissue samples to extract tumor-specific fusions. In total, 7,887 fusion transcripts with high confidence were detected across 13 tumor types. Our fusion prediction was validated by supporting evidence for a genomic rearrangement for 78 of 79 fusions in 48 glioma samples where whole genome sequencing data was available. Fusion transcript frequency was positively correlated with the levels of genomic instability in cancers, whereas tumor samples harboring fusions contained statistically significantly fewer driver gene mutations. We identified 221 fusion transcripts involving a tumor suppressor gene and these were wildtype in majority (98.2%), suggesting the fusion has a disruptive effect and a role in tumorigenesis. We identified at least one in-frame protein kinase fusion in 324 of 4,366 samples (7.4%). Potentially therapeutic targetable kinase fusions such as ALK, ROS, RET, NTRK, and FGFR gene families were detected in thyroid carcinoma (8.7%), glioblastoma (4.4%), bladder carcinoma (3.3%), lung squamous cell carcinoma (2.3%), lung adenocarcinoma (1.6%), lower grade glioma (1.5%), and head and neck cancer (1.0%), suggesting a potential for application of kinase inhibitors for multiple tumor types. In-frame fusion transcripts involving histone methyltransferase or histone demethylase genes, which may also be considered as therapeutic targets, were detected in 111 samples (2.5%). In summary, the landscape of transcript fusions detected across a large number of tumor samples revealed fusion events with clinical relevance that have not been previously recognized. Our findings support the concept of basket clinical trials where patients are matched with experimental therapies based on their genomic profile rather than the tissue where the tumor originated. Citation Format: Kosuke Yoshihara, Qianghu Wang, Wandaliz Torres-Garcia, Siyuan Zheng, Rahulsimham Vegesna, Hoon Kim, Roel GW Verhaak. The landscape of therapeutic targetable fusions. [abstract]. In: Proceedings of the 106th Annual Meeting of the American Association for Cancer Research; 2015 Apr 18-22; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2015;75(15 Suppl):Abstract nr 3762. doi:10.1158/1538-7445.AM2015-3762

HITS-CLIP reveals key regulators of nuclear receptor signaling in breast cancer.

miRNAs regulate the expression of genes in both normal physiology and disease. While miRNAs have been demonstrated to play a pivotal role in aspects of cancer biology, these reports have generally focused on the regulation of single genes. Such single-gene approaches have significant limitations, relying on miRNA expression levels and heuristic predictions of mRNA-binding sites. This results in only circumstantial evidence of miRNA-target interaction and typically leads to large numbers of false positive predictions. Here, we used a genome-wide approach (high-throughput sequencing of RNA isolated by crosslinking immunoprecipitation, HITS-CLIP) to define direct miRNA-mRNA interactions in three breast cancer subtypes (estrogen receptor positive, Her2 amplified, and triple negative). Focusing on steroid receptor signaling, we identified two novel regulators of the ER pathway (miR-9-5p and miR-193a/b-3p), which together target multiple genes involved in ER signaling. Moreover, this approach enabled the definition of miR-9-5p as a global regulator of steroid receptor signaling in breast cancer. We show that miRNA targets and networks defined by HITS-CLIP under physiologic conditions are predictive of patient outcomes and provide global insight into miRNA regulation in breast cancer.

Daily prediction of marine beach water quality in Hong Kong

Bacterial concentration (Escherichia coli) is generally adopted as a key indicator of beach water quality. Currently the beach management system in Hong Kong relies on past water quality data sampled at intervals between 3 and 14 days. Beach advisories are issued when the geometric mean E. coli level of the past five samples exceeds the beach water quality objective (WQO) of 180 counts/100 mL. When the E. coli level varies dynamically, the system is not able to track the daily bacterial variation. And yet worldwide there does not exist a generally accepted method to predict beach water quality in a marine environment, which is influenced by hydro-meteorological variables, catchment characteristics, as well as complicated tidal currents and wave effects.A comprehensive study of beach water quality prediction has been carried out for four representative beaches in Hong Kong: Big Wave Bay (BW), Deep Water Bay (DW), New Cafeteria (NC) and Silvermine Bay (SIL). Statistical analysis of the extensive regular monitoring data was carried out for two periods before and after the commissioning of the Harbour Area Treatment Scheme (HATS): (1990–1997) and (2002–2006) respectively. The data analysis shows that E. coli is strongly correlated with seven hydro-environmental variables: rainfall, solar radiation, wind speed, tide level, salinity, water temperature and past E. coli concentration. The relative importance of the parameters is beach-specific, and depends on the local geographical and hydrographical characteristics as well as location of nearby pollution sources.Multiple Linear Regression (MLR) and Artificial Neural Network (ANN) models are developed from the sparsely sampled regular monitoring data (2002–2006) to predict the next-day E. coli concentration using the key hydro-environmental variables as input parameters. The models are validated against daily monitoring data in the bathing seasons of 2007 and 2008. The models are able to track the dynamic changes in E. coli concentration and predict WQO compliance/exceedance with an overall accuracy of 70–96%. Both the MLR and ANN models are superior to the current beach advisories in capturing water quality variations, and in predicting WQO exceedances. For example, the models predict around 80% and 50% of the exceedances at BW and NC respectively in June–July 2007, as compared to 0% and 14% based purely on past data. Similarly, observed exceedances are predicted with success rates of 71%, 42%, and 53% at BW, NC, and SIL respectively during July–October 2008, as compared with 0%, 0%, and 6% using the current water quality assessment criterion. The MLR and ANN models have similar performances; ANN model tends to be better in predicting the high-end concentrations, with however a greater number of false positive predictions (false alarms).This work demonstrates the practical feasibility of predicting bacterial concentration based on the critical hydro-environmental variables, and paves the way for developing a real time water quality forecast and management system for Hong Kong.

Assessment of a probabilistic framework for combining structure- and ligand-based virtual screening

A wide variety of structure- and ligand-based virtual screening approaches have been developed that aim at finding potential leads to initiate drug discovery efforts. Since each method has its strengths and weakness, combining the outcome of different structure- and ligand-based approaches can be expected to decrease the number of false positive predictions. However, a reliable fusion of information from different methods is challenging. This holds true in particular for new target structures, where target specific performance experiences are missing. Here, we assess the performance of a probilistic framework approach [1] that combines structure- and ligand-based information in a meaningful way by assigning probabilities that any two molecules are active. The approach is validated using two popular docking methods (GOLD and AutoDock) and an in-house ligand-based screening approach (ElectroShape [2]). Results of similarity search and docking calculations for the Directory of Useful Decoys (DUD) [3] are combined through rank fusion as well as a probabilistic framework approach. The study will be used to answer questions such as: How far do the virtual screening-approaches used provide complementary or redundant hit lists? Does the fusion of structure- and ligand-based approaches consistently outperform any single screening metric? Using a probabilistic framework approach, is it possible to obtain a quantification of the confidence that any molecule will be active?

Finding Transcription Factor Binding Motifs for Coregulated Genes by Combining Sequence Overrepresentation with Cross-Species Conservation

Novel computational methods for finding transcription factor binding motifs have long been sought due to tedious work of experimentally identifying them. However, the current prevailing methods yield a large number of false positive predictions due to the short, variable nature of transcriptional factor binding sites (TFBSs). We proposed here a method that combines sequence overrepresentation and cross-species sequence conservation to detect TFBSs in upstream regions of a given set of coregulated genes. We applied the method to 35S. cerevisiaetranscriptional factors with known DNA binding motifs (with the support of orthologous sequences from genomes ofS. mikatae,S. bayanus, andS. paradoxus), and the proposed method outperformed the single-genome-based motif finding methodsMEMEandAlignACEas well as the multiple-genome-based methodsPHYMEandFootprinterfor the majority of these transcriptional factors. Compared with the prevailing motif finding software, our method has some advantages in finding transcriptional factor binding motifs for potential coregulated genes if the gene upstream sequences of multiple closely related species are available. Although we used yeast genomes to assess our method in this study, it might also be applied to other organisms if suitable related species are available and the upstream sequences of coregulated genes can be obtained for the multiple closely related species.

SiTaR: a novel tool for transcription factor binding site prediction

Prediction of transcription factor binding sites (TFBSs) is crucial for promoter modeling and network inference. Quality of the predictions is spoiled by numerous false positives, which persist as the main problem for all presently available TFBS search methods. We suggest a novel approach, which is alternative to widely used position weight matrices (PWMs) and Hidden Markov Models. Each motif of the input set is used as a search template to scan a query sequence. Found motifs are assigned scores depending on the non-randomness of the motif's occurrence, the number of matching searching motifs and the number of mismatches. The non-randomness is estimated by comparison of observed numbers of matching motifs with those predicted to occur by chance. The latter can be calculated given the base compositions of the motif and the query sequence. The method does not require preliminary alignment of the input motifs, hence avoiding uncertainties introduced by the alignment procedure. In comparison with PWM-based tools, our method demonstrates higher precision by the same sensitivity and specificity. It also tends to outperform methods combining pattern and PWM search. Most important, it allows reducing the number of false positive predictions significantly. The method is implemented in a tool called SiTaR (Site Tracking and Recognition) and is available at http://sbi.hki-jena.de/sitar/index.php. ekaterina.shelest@hki-jena.de Supplementary data are available at Bioinformatics online.

In Silico Studies of the Relationship Between Chemical Structure and Drug Induced Phospholipidosis.

Drug-induced phospholipidosis (PLD) is a side effect of the administration of cationic amphiphilic drugs (CADs). It is desirable to identify and screen compounds with the potential to induce PLD as early as possible in drug development. Recently, a number of in silico methods have been developed to predict PLD. These models are low-cost and high-throughput strategies; however, they produce a high number of false positive predictions. The aim of this study was to assess the predictive performance of existing in silico approaches and to develop new strategies for the rapid identification of the potential PLD-inducers. Studies on 450 chemicals confirmed the high false positive rate of prediction of models based only on log P and pKa values. Modification of the methods by incorporating structural information gave moderate improvements in the prediction performance. Therefore, a new strategy, based on molecular fragments captured by SMARTS strings was developed. These structural fragments were able to identify potential PLD-inducers and achieved a high sensitivity of 85 %. The results showed that the phospholipidosis is linked directly to the molecular structure of chemical; therefore the SMARTS pattern methodology could be used as a first line of screening of PLD potential during the drug discovery process.

Identifying Fraud from the Language of Financial Reports

We compare the effectiveness of financial statement variables and pre-determined lists of suspicious words with an alternative approach to fraud detection. Using the text of the Management Discussion and Analysis section from 10-Q and 10-K reports, we allow the data to identify words most strongly associated with financial misrepresentation. Based on the identified words, we assign a probability that a particular financial report is truthful within the time series of reports for the same firm. We compare this probability of truthfulness to predictions of fraud based on the F-Score from Dechow et al. (2011) and pre-defined word lists designed to capture deception, negativity, uncertainty and litigious activity. We establish that the data-generated word list can be a useful complement to detection methods based on financial statement variables by dramatically reducing the number of false positive predictions. In addition, our approach produces higher correct classification rates than textual analysis using alternative pre-determined word lists. Particularly interesting is our finding that word lists designed to reflect conscious deception have virtually no ability to correctly identify financial misrepresentation. This suggests that many of the individuals involved in drafting financial reports, including employees, auditors, and legal counsel may be completely unaware that the fraud is occurring. The data-generated word list continues to outperform alternative detection methods in a second sample representing a broad cross-section of firms.

Identifying translation initiation sites in prokaryotes using support vector machine

Motivation: Gene identification in genomes has been a fundamental and long-standing task in bioinformatics and computational biology. Many computational methods have been developed to predict genes in prokaryote genomes by identifying translation initiation site (TIS) in transcript data. However, the pseudo-TISs at the genome level make these methods suffer from a high number of false positive predictions. In addition, most of the existing tools use an unsupervised learning framework, whose predictive accuracy may depend on the choice of specific organism. Results: In this paper, we present a supervised learning method, support vector machine (SVM), to identify translation initiation site at the genome level. The features are extracted from the sequence data by modeling the sequence segment around predicted TISs as a position specific weight matrix (PSWM). We train the parameters of our SVM through well constructed positive and negative TIS datasets. Then we apply the method to recognize translation initiation sites in E. coli, B. subtilis, and validate our method on two GC-rich bacteria genomes: Pseudomonas aeruginosa and Burkholderia pseudomallei K96243. We show that translation initiation sites can be recognized accurately at the genome level by our method, irrespective of their GC content. Furthermore, we compare our method with four existing methods and demonstrate that our method outperform these methods by obtaining better performance in all the four organisms.