Understanding why only some breast cancers become metastatic, and predicting metastatic risks at earlier cancer stages, are two major goals of breast cancer research. These goals are clearly connected and synergistic, but are rarely integrated within a single study. Knowledge discovery-oriented research has identified critical biological pathways related to metastasis, but the interplay between these pathways remains elusive, resulting in models that lack prediction accuracy. Conversely, complex machine learning models achieve high prediction accuracy without detailed biological knowledge, making the explanation challenging. Here, we propose a novel computational framework that addresses both goals simultaneously. To support knowledge discovery, we construct gene regulatory networks (GRNs) to model the cellular states of metastatic and non-metastatic patients. To enable explainable metastasis prediction, we introduce a dysregulation score based on the GRN models. Experimental results demonstrate that our method not only identified significant metastasis-associated GRN changes, but also revealed the loss of co-regulation among key biological processes in metastatic patients. Leveraging the dysregulation score, our model-free classifier outperformed complex machine learning models under rigorous evaluation. This work bridges a significant gap between knowledge discovery and accurate, explainable prediction by employing carefully designed knowledge models and knowledge-based prediction, with potential applicability in other disease contexts.

Classification tasks in many real-world domains are exacerbated by class imbalance, relatively small sample sizes compared to high dimensionality, and measurement uncertainty. The problem of class imbalance has been extensively studied, and data augmentation methods based on interpolation of minority class instances have been proposed as a viable solution to mitigate imbalance. It remains to be seen whether augmentation can be applied to improve the overall performance while maintaining stability, especially with a limited number of samples. In this paper, we present a novel feature-space augmentation technique that can be applied to high-dimensional data for classification tasks and address these issues. Our method utilizes uniform random sampling and introduces synthetic instances by taking advantage of the local distributions of individual features in the observed instances. The core augmentation algorithm is class-invariant, which opens up an unexplored avenue of simultaneously improving and stabilizing performance by augmenting unlabeled instances. The proposed method is evaluated using a comprehensive performance analysis involving multiple classifiers and metrics. Comparative analysis with existing feature space augmentation methods strongly suggests that the proposed algorithm can result in improved classification performance while also increasing the overall reliability of the performance evaluation.

Interpretability of machine learning (ML) models represents the extent to which a model's decision-making process can be understood by model developers and/or end users. Transcriptomics-based cancer prognosis models, for example, while achieving good accuracy, are usually hard to interpret, due to the high-dimensional feature space and the complexity of models. As interpretability is critical for the transparency and fairness of ML models, several algorithms have been proposed to improve the interpretability of arbitrary classifiers. However, evaluation of these algorithms often requires substantial domain knowledge. Here, we propose a breast cancer metastasis prediction model using a very small number of biologically interpretable features, and a simple yet novel model interpretation approach that can provide personalized interpretations. In addition, we contributed, to the best of our knowledge, the first method to quantitatively compare different interpretation algorithms. Experimental results show that our model not only achieved competitive prediction accuracy, but also higher inter-classifier interpretation consistency than state-of-the-art interpretation methods. Importantly, our interpretation results can improve the generalizability of the prediction models. Overall, this work provides several novel ideas to construct and evaluate interpretable ML models that can be valuable to both the cancer machine learning community and related application domains.

BACKGROUND: The abundance of molecular profiling of breast cancer tissues entailed active research on molecular marker-based early diagnosis of metastasis. Recently there is a surging interest in combining gene expression with gene networks such as protein-protein interaction (PPI) network, gene co-expression (CE) network and pathway information to identify robust and accurate biomarkers for metastasis prediction, reflecting the common belief that cancer is a systems biology disease. However, controversy exists in the literature regarding whether network markers are indeed better features than genes alone for predicting as well as understanding metastasis. We believe much of the existing results may have been biased by the overly complicated prediction algorithms, unfair evaluation, and lack of rigorous statistics. In this study, we propose a simple approach to use network edges as features, based on two types of networks respectively, and compared their prediction power using three classification algorithms and rigorous statistical procedure on one of the largest datasets available. To detect biomarkers that are significant for the prediction and to compare the robustness of different feature types, we propose an unbiased and novel procedure to measure feature importance that eliminates the potential bias from factors such as different sample size, number of features, as well as class distribution. RESULTS: Experimental results reveal that edge-based feature types consistently outperformed gene-based feature type in random forest and logistic regression models under all performance evaluation metrics, while the prediction accuracy of edge-based support vector machine (SVM) model was poorer, due to the larger number of edge features compared to gene features and the lack of feature selection in SVM model. Experimental results also show that edge features are much more robust than gene features and the top biomarkers from edge feature types are statistically more significantly enriched in the biological processes that are well known to be related to breast cancer metastasis. CONCLUSIONS: Overall, this study validates the utility of edge features as biomarkers but also highlights the importance of carefully designed experimental procedures in order to achieve statistically reliable comparison results.

Multiple classifier system (MCS) has become a successful alternative for improving classification performance. However, studies have shown inconsistent results for different MCSs, and it is often difficult to predict which MCS algorithm works the best on a particular problem. We believe that the two crucial steps of MCS - base classifier generation and multiple classifier combination, need to be designed coordinately to produce robust results. In this work, we show that for different testing instances, better classifiers may be trained from different subdomains of training instances including, for example, neighboring instances of the testing instance, or even instances far away from the testing instance. To utilize this intuition, we propose Individualized Classifier Ensemble (ICE). ICE groups training data into overlapping clusters, builds a classifier for each cluster, and then associates each training instance to the top-performing models while taking into account model types and frequency. In testing, ICE finds the k most similar training instances for a testing instance, then predicts class label of the testing instance by averaging the prediction from models associated with these training instances. Evaluation results on 49 benchmarks show that ICE has a stable improvement on a significant proportion of datasets over existing MCS methods. ICE provides a novel choice of utilizing internal patterns among instances to improve classification, and can be easily combined with various classification models and applied to many application domains.

Multi-label feature selection focuses on constructing a reduced feature space for discriminating multi-label instances. In consideration of the complex structures of label and feature spaces, a critical issue that explicitly determines selection performance is how to induce consistent information from both spaces to steer feature selection. Existing approaches tackle this issue in various spatial-aware views, without sufficient consideration of the negative effects of irrelevant features and imbalanced neighbors on inferring space structure. Inspired by the superiority of margin theory in assessing reliable space structure, we approach multi-label feature selection in the learning framework of preserving label-feature space consistency through probabilistic margin in this paper. In contrast to existing approaches, our model assesses the weighted margin based on the probabilistic nearest neighbors, and preserves consistent margin information in label and feature spaces. In this manner, label-feature space consistency is elegantly achieved, which conduces to effectively capturing discriminative features suitable for multi-label learning tasks and eliminating noisy features. Experimental results on multi-label data sets demonstrate the encouraging performance of the proposed model.

Semi-supervised learning and multi-label learning pose different challenges for feature selection, which is one of the core techniques for dimension reduction, and the exploration of reducing feature space for multi-label learning with incomplete label information is far from satisfactory. Existing feature selection approaches devote attention to either of two issues, namely, alleviating negative effects of imperfectly predicted labels and quantitatively evaluating label correlations, exclusively for semi-supervised or multi-label scenarios. A unified framework to extract label correlation information with incomplete prior knowledge and embed this information in feature selection however, is rarely touched. In this paper, we propose a space consistency-based feature selection model to address this issue. Specifically, correlation information in feature space is learned based on the probabilistic neighborhood similarities, and correlation information in label space is optimized by preserving feature-label space consistency. This mechanism contributes to appropriately extracting label information in semi-supervised multi-label learning scenario and effectively employing this information to select discriminative features. An extensive experimental evaluation on real-world data shows the superiority of the proposed approach under various evaluation metrics.

MOTIVATION: The study of transcriptional regulation is still difficult yet fundamental in molecular biology research. While the development of both in vivo and in vitro profiling techniques have significantly enhanced our knowledge of transcription factor (TF)-DNA interactions, computational models of TF-DNA interactions are relatively simple and may not reveal sufficient biological insight. In particular, supervised learning based models for TF-DNA interactions attempt to map sequence-level features ( k -mers) to binding event but usually ignore the location of k -mers, which can cause data fragmentation and consequently inferior model performance. RESULTS: Here, we propose a novel algorithm based on the so-called multiple-instance learning (MIL) paradigm. MIL breaks each DNA sequence into multiple overlapping subsequences and models each subsequence separately, therefore implicitly takes into consideration binding site locations, resulting in both higher accuracy and better interpretability of the models. The result from both in vivo and in vitro TF-DNA interaction data show that our approach significantly outperform conventional single-instance learning based algorithms. Importantly, the models learned from in vitro data using our approach can predict in vivo binding with very good accuracy. In addition, the location information obtained by our method provides additional insight for motif finding results from ChIP-Seq data. Finally, our approach can be easily combined with other state-of-the-art TF-DNA interaction modeling methods. AVAILABILITY AND IMPLEMENTATION: http://www.cs.utsa.edu/∼jruan/MIL/. CONTACT: [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The 2016 International Conference on Intelligent Biology and Medicine (ICIBM 2016) was held on December 8-10, 2016 in Houston, Texas, USA. ICIBM included eight scientific sessions, four tutorials, one poster session, four highlighted talks and four keynotes that covered topics on 3D genomics structural analysis, next generation sequencing (NGS) analysis, computational drug discovery, medical informatics, cancer genomics, and systems biology. Here, we present a summary of the nine research articles selected from ICIBM 2016 program for publishing in BMC Bioinformatics.

MOTIVATION: Metastasis prediction is a well-known problem in breast cancer research. As breast cancer is a complex and heterogeneous disease with many molecular subtypes, predictive models trained for one cohort often perform poorly on other cohorts, and a combined model may be suboptimal for individual patients. Furthermore, attempting to develop subtype-specific models is hindered by the ambiguity and stereotypical definitions of subtypes. RESULTS: Here, we propose a personalized approach by relaxing the definition of breast cancer subtypes. We assume that each patient belongs to a distinct subtype, defined implicitly by a set of patients with similar molecular characteristics, and construct a different predictive model for each patient, using as training data, only the patients defining the subtype. To increase robustness, we also develop a committee-based prediction method by pooling together multiple personalized models. Using both intra- and inter-dataset validations, we show that our approach can significantly improve the prediction accuracy of breast cancer metastasis compared with several popular approaches, especially on those hard-to-learn cases. Furthermore, we find that breast cancer patients belonging to different canonical subtypes tend to have different predictive models and gene signatures, suggesting that metastasis in different canonical subtypes are likely governed by different molecular mechanisms. AVAILABILITY AND IMPLEMENTATION: Source code implemented in MATLAB and Java available at www.cs.utsa.edu/∼jruan/PCC/.

prediction of drug side-effects in early stage of drug development is becoming more popular now days, which not only reduces the time for drug design but also reduces the drug development costs. In this article we propose an ensemble approach to predict drug side-effects of drug molecules based on their chemical structure. Our idea originates from the observation that similar drugs have similar side-effects. Based on this observation we design an ensemble approach that combine the results from different classification models where each model is generated by a different set of similar drugs. We applied our approach to 1385 side-effects in the SIDER database for 888 drugs. Results show that our approach outperformed previously published approaches and standard classifiers. Furthermore, we applied our method to a number of uncharacterized drug molecules in DrugBank database and predict their side-effect profiles for future usage. Results from various sources confirm that our method is able to predict the side-effects for uncharacterized drugs and more importantly able to predict rare side-effects which are often ignored by other approaches. The method described in this article can be useful to predict side-effects in drug design in an early stage to reduce experimental cost and time.

Finding out the associations between an input gene set, such as genes associated with a certain phenotype, and annotated gene sets, such as known pathways, are a very important problem in modern molecular biology. The existing approaches mainly focus on the overlap between the two, and may miss important but subtle relationships between genes. In this paper, we propose a method, NetPEA, by combining the known pathways and high-throughput networks. Our method not only considers the shared genes, but also takes the gene interactions into account. It utilizes a protein-protein interaction network and a random walk procedure to identify hidden relationships between gene sets, and uses a randomization strategy to evaluate the significance for pathways to achieve such similarity scores. Compared with the over-representation based method, our method can identify more relationships. Compared with a state of the art network-based method, EnrichNet, our method not only provides a ranked list of pathways, but also provides the statistical significant information. Importantly, through independent tests, we show that our method likely has a higher sensitivity in revealing the true casual pathways, while at the same time achieve a higher specificity. Literature review of selected results indicates that some of the novel pathways reported by our method are biologically relevant and important.

MOTIVATION: Recent advances in technology have dramatically increased the availability of protein-protein interaction (PPI) data and stimulated the development of many methods for improving the systems level understanding the cell. However, those efforts have been significantly hindered by the high level of noise, sparseness and highly skewed degree distribution of PPI networks. Here, we present a novel algorithm to reduce the noise present in PPI networks. The key idea of our algorithm is that two proteins sharing some higher-order topological similarities, measured by a novel random walk-based procedure, are likely interacting with each other and may belong to the same protein complex. RESULTS: Applying our algorithm to a yeast PPI network, we found that the edges in the reconstructed network have higher biological relevance than in the original network, assessed by multiple types of information, including gene ontology, gene expression, essentiality, conservation between species and known protein complexes. Comparison with existing methods shows that the network reconstructed by our method has the highest quality. Using two independent graph clustering algorithms, we found that the reconstructed network has resulted in significantly improved prediction accuracy of protein complexes. Furthermore, our method is applicable to PPI networks obtained with different experimental systems, such as affinity purification, yeast two-hybrid (Y2H) and protein-fragment complementation assay (PCA), and evidence shows that the predicted edges are likely bona fide physical interactions. Finally, an application to a human PPI network increased the coverage of the network by at least 100%. AVAILABILITY: www.cs.utsa.edu/∼jruan/RWS/.

Recent advances in high-throughput technology have dramatically increased the quantity of available protein-protein interaction (PPI) data and stimulated the development of many methods for predicting protein complexes, which are important in understanding the functional organization of protein-protein interaction networks in different biological processes. However, automated protein complex prediction from PPI data alone is significantly hindered by the high level of noise, sparseness, and highly skewed degree distribution of PPI networks. Here we present a novel network topology-based algorithm to remove spurious interactions and recover missing ones by computational predictions, and to increase the accuracy of protein complex prediction by reducing the impact of hub nodes. The key idea of our algorithm is that two proteins sharing some high-order topological similarities, which are measured by a novel random walk-based procedure, are likely interacting with each other and may belong to the same protein complex. Applying our algorithm to a yeast protein-protein interaction network, we found that the interactions in the reconstructed PPI network have more significant biological relevance than the original network, assessed by multiple types of information, including gene ontology, gene expression, essentiality, conservation between species, and known protein complexes. Comparison with several existing methods show that the network reconstructed by our method has the highest quality. Finally, using two independent graph clustering algorithms, we found that the reconstructed network has resulted in significantly improved prediction accuracy of protein complexes.

BACKGROUND: Several large-scale gene co-expression networks have been constructed successfully for predicting gene functional modules and cis-regulatory elements in Arabidopsis (Arabidopsis thaliana). However, these networks are usually constructed and analyzed in an ad hoc manner. In this study, we propose a completely parameter-free and systematic method for constructing gene co-expression networks and predicting functional modules as well as cis-regulatory elements. RESULTS: Our novel method consists of an automated network construction algorithm, a parameter-free procedure to predict functional modules, and a strategy for finding known cis-regulatory elements that is suitable for consensus scanning without prior knowledge of the allowed extent of degeneracy of the motif. We apply the method to study a large collection of gene expression microarray data in Arabidopsis. We estimate that our co-expression network has ~94% of accuracy, and has topological properties similar to other biological networks, such as being scale-free and having a high clustering coefficient. Remarkably, among the ~300 predicted modules whose sizes are at least 20, 88% have at least one significantly enriched functions, including a few extremely significant ones (ribosome, p < 1E-300, photosynthetic membrane, p < 1.3E-137, proteasome complex, p < 5.9E-126). In addition, we are able to predict cis-regulatory elements for 66.7% of the modules, and the association between the enriched cis-regulatory elements and the enriched functional terms can often be confirmed by the literature. Overall, our results are much more significant than those reported by several previous studies on similar data sets. Finally, we utilize the co-expression network to dissect the promoters of 19 Arabidopsis genes involved in the metabolism and signaling of the important plant hormone gibberellin, and achieved promising results that reveal interesting insight into the biosynthesis and signaling of gibberellin. CONCLUSIONS: The results show that our method is highly effective in finding functional modules from real microarray data. Our application on Arabidopsis leads to the discovery of the largest number of annotated Arabidopsis functional modules in the literature. Given the high statistical significance of functional enrichment and the agreement between cis-regulatory and functional annotations, we believe our Arabidopsis gene modules can be used to predict the functions of unknown genes in Arabidopsis, and to understand the regulatory mechanisms of many genes.

Identifying modules, or natural communities, in large complex networks is fundamental in many fields, including social sciences, biological sciences and engineering. Recently several methods have been developed to automatically identify communities from complex networks by optimizing the modularity function. The advantage of this type of approaches is that the algorithm does not require any parameter to be tuned. However, the modularity-based methods for community discovery assume that the network structure is given explicitly and is correct. In addition, these methods work best if the network is unweighted and/or sparse. In reality, networks are often not directly defined, or may be given as an affinity matrix. In the first case, each node of the network is defined as a point in a high dimensional space and different networks can be obtained with different network construction methods, resulting in different community structures. In the second case, an affinity matrix may define a dense weighted graph, for which modularity-based methods do not perform well. In this work, we propose a very simple algorithm to automatically identify community structures from these two types of data. Our approach utilizes a k-nearest-neighbor network construction method to capture the topology embedded in high dimensional data, and applies a modularity-based algorithm to identify the optimal community structure. A key to our approach is that the network construction is incorporated with the community identification process and is totally parameter-free. Furthermore, our method can suggest appropriate pre-processing/ normalization of the data to improve the results of community identification. We tested our methods on several synthetic and real data sets, and evaluated its performance by internal or external accuracy indices. Compared with several existing approaches, our method is not only fully automatic, but also has the best accuracy overall.

BACKGROUND: The osteocyte is a type of cell that appears to be one of the key endocrine regulators of bone metabolism and a key responder to initiate bone formation and remodeling. Identifying the regulatory networks in osteocytes may lead to new therapies for osteoporosis and loss of bone. RESULTS: Using microarray, we identified 269 genes over-expressed in osteocyte, many of which have known functions in bone and muscle differentiation and contractility. We determined the evolutionarily conserved and enriched TF binding sites in the 5 kb promoter regions of these genes. Using this data, a transcriptional regulatory network was constructed and subsequently partitioned to identify cis-regulatory modules. CONCLUSION: Our results show that many osteocyte-specific genes, including two well-known osteocyte markers DMP1 and Sost, have highly conserved clustering of muscle-related cis-regulatory modules, thus supporting the concept that a muscle-related gene network is important in osteocyte biology and may play a role in contractility and dynamic movements of the osteocyte.

Automatic discovery of community structures in complex networks is a fundamental task in many disciplines, including social science, engineering, and biology. A quantitative measure called modularity (Q) has been proposed to effectively assess the quality of community structures. Several community discovery algorithms have since been developed based on the optimization of Q. However, this optimization problem is NP-hard, and the existing algorithms have a low accuracy or are computationally expensive. In this paper, we present an efficient spectral algorithm for modularity optimization. When tested on a large number of synthetic or real-world networks, and compared to the existing algorithms, our method is efficient and and has a high accuracy. In addition, we have successfully applied our algorithm to detect interesting and meaningful community structures from real-world networks in different domains, including biology, medicine and social science. Due to space limitation, results of these applications are presented in a complete version of the paper available on our Website (http://cse .wustl.edu/~jruan/).

MicroRNAs are short, noncoding RNAs that play important roles in post-transcriptional gene regulation. Although many functions of microRNAs in plants and animals have been revealed in recent years, the transcriptional mechanism of microRNA genes is not well-understood. To elucidate the transcriptional regulation of microRNA genes, we study and characterize, in a genome scale, the promoters of intergenic microRNA genes in Caenorhabditis elegans, Homo sapiens, Arabidopsis thaliana, and Oryza sativa. We show that most known microRNA genes in these four species have the same type of promoters as protein-coding genes have. To further characterize the promoters of microRNA genes, we developed a novel promoter prediction method, called common query voting (CoVote), which is more effective than available promoter prediction methods. Using this new method, we identify putative core promoters of most known microRNA genes in the four model species. Moreover, we characterize the promoters of microRNA genes in these four species. We discover many significant, characteristic sequence motifs in these core promoters, several of which match or resemble the known cis-acting elements for transcription initiation. Among these motifs, some are conserved across different species while some are specific to microRNA genes of individual species.

MOTIVATION: The transcriptional regulation of a gene depends on the binding of cis-regulatory elements on its promoter to some transcription factors and the expression levels of the transcription factors. Most existing approaches to studying transcriptional regulation model these dependencies separately, i.e. either from promoters to gene expression or from the expression levels of transcription factors to the expression levels of genes. Little effort has been devoted to a single model for integrating both dependencies. RESULTS: We propose a novel method to model gene expression using both promoter sequences and the expression levels of putative regulators. The proposed method, called bi-dimensional regression tree (BDTree), extends a multivariate regression tree approach by applying it simultaneously to both genes and conditions of an expression matrix. The method produces hypotheses about the condition-specific binding motifs and regulators for each gene. As a side-product, the method also partitions the expression matrix into small submatrices in a way similar to bi-clustering. We propose and compare several splitting functions for building the tree. When applied to two microarray datasets of the yeast Saccharomyces cerevisiae, BDTree successfully identifies most motifs and regulators that are known to regulate the biological processes underlying the datasets. Comparing with an existing algorithm, BDTree provides a higher prediction accuracy in cross-validations.

MOTIVATION: A fundamental problem of computational genomics is identifying the genes that respond to certain endogenous cues and environmental stimuli. This problem can be referred to as targeted gene finding. Since gene regulation is mainly determined by the binding of transcription factors and cis-regulatory DNA sequences, most existing gene annotation methods, which exploit the conservation of open reading frames, are not effective in finding target genes. RESULTS: A viable approach to targeted gene finding is to exploit the cis-regulatory elements that are known to be responsible for the transcription of target genes. Given such cis-elements, putative target genes whose promoters contain the elements can be identified. As a case study, we apply the above approach to predict the genes in model plant Arabidopsis thaliana which are inducible by a phytohormone, abscisic acid (ABA), and abiotic stress, such as drought, cold and salinity. We first construct and analyze two ABA specific cis-elements, ABA-responsive element (ABRE) and its coupling element (CE), in A.thaliana, based on their conservation in rice and other cereal plants. We then use the ABRE-CE module to identify putative ABA-responsive genes in A.thaliana. Based on RT-PCR verification and the results from literature, this method has an accuracy rate of 67.5% for the top 40 predictions. The cis-element based targeted gene finding approach is expected to be widely applicable since a large number of cis-elements in many species are available.

BACKGROUND: Many classification approaches have been applied to analyzing transcriptional regulation of gene expressions. These methods build models that can explain a gene's expression level from the regulatory elements (features) on its promoter sequence. Different types of features, such as experimentally verified binding motifs, motifs discovered by computer programs, or transcription factor binding data measured with Chromatin Immunoprecipitation (ChIP) assays, have been used towards this goal. Each type of features has been shown successful in modeling gene transcriptional regulation under certain conditions. However, no comparison has been made to evaluate the relative merit of these features. Furthermore, most publicly available classification tools were not designed specifically for modeling transcriptional regulation, and do not allow the user to combine different types of features. RESULTS: In this study, we use a specific classification method, decision trees, to model transcriptional regulation in yeast with features based on predefined motifs, automatically identified motifs, ChlP-chip data, or their combinations. We compare the accuracies and stability of these models, and analyze their capabilities in identifying functionally related genes. Furthermore, we design and implement a user-friendly web server called CAGER (Classification Analysis of Gene Expression Regulation) that integrates several software components for automated analysis of transcriptional regulation using decision trees. Finally, we use CAGER to study the transcriptional regulation of Arabidopsis genes in response to abscisic acid, and report some interesting new results. CONCLUSION: Models built with ChlP-chip data suffer from low accuracies when the condition under which gene expressions are measured is significantly different from the condition under which the ChIP experiment is conducted. Models built with automatically identified motifs can sometimes discover new features, but their modeling accuracies may have been over-estimated in previous studies. Furthermore, models built with automatically identified motifs are not stable with respect to noises. A combination of ChlP-chip data and predefined motifs can substantially improve modeling accuracies, and is effective in identifying true regulons. The CAGER web server, which is freely available at http://cic.cs.wustl.edu/CAGER/, allows the user to select combinations of different feature types for building decision trees, and interact with the models graphically. We believe that it will be a useful tool to facilitate the discovery of gene transcriptional regulatory networks.

MOTIVATION: Pseudoknots have generally been excluded from the prediction of RNA secondary structures due to its difficulty in modeling. Although, several dynamic programming algorithms exist for the prediction of pseudoknots using thermodynamic approaches, they are neither reliable nor efficient. On the other hand, comparative methods are more reliable, but are often done in an ad hoc manner and require expert intervention. Maximum weighted matching, an algorithm for pseudoknot prediction with comparative analysis, suffers from low-prediction accuracy in many cases. RESULTS: Here we present an algorithm, iterated loop matching, for reliably and efficiently predicting RNA secondary structures including pseudoknots. The method can utilize either thermodynamic or comparative information or both, thus is able to predict pseudoknots for both aligned and individual sequences. We have tested the algorithm on a number of RNA families. Using 8-12 homologous sequences, the algorithm correctly identifies more than 90% of base-pairs for short sequences and 80% overall. It correctly predicts nearly all pseudoknots and produces very few spurious base-pairs for sequences without pseudoknots. Comparisons show that our algorithm is both more sensitive and more specific than the maximum weighted matching method. In addition, our algorithm has high-prediction accuracy on individual sequences, comparable with the PKNOTS algorithm, while using much less computational resources. AVAILABILITY: The program has been implemented in ANSI C and is freely available for academic use at http://www.cse.wustl.edu/~zhang/projects/rna/ilm/ SUPPLEMENTARY INFORMATION: http://www.cse.wustl.edu/~zhang/projects/rna/ilm/