Computational Learning Methods Research Articles

Abstract SY14-03: Inferring causality for oncology drug discovery and development

Abstract Decisions in drug development are made based upon determinations of cause and effect from experimental observations that span all development phases. Despite major advances in our powers of observation due to the advent of multi-omic technologies, the ability to determine disease mechanisms, biomarkers, and effective combination therapies from large-scale and diverse data sets in the context of ill-defined and complex biological systems continues to be a major bottleneck in the drug development process. Current methods of statistical analysis do not allow for causal inferences, leaving this impossible task to the human mind. This major bottleneck can only be overcome by utilizing automated computational learning methods that identify from compound data the circuits and connections between drug-affected molecular constituents and physiological observables. We will present several case studies in oncology drug development where we employ massively parallel machines to reverse engineer populations of models that predict the ‘gearing’ between drug, genetic, and genomic molecular components with respect to oncology phenotypic outcomes. Since a single topology or model cannot capture the reality of a biological system under investigation, Monte Carlo simulation strategies over the population of models reveal how the drug, genetic, gene expression and clinical variables work together to impact phenotypic outcomes. We will show how the marriage of multi-omic technologies within the network inference approach is helping to fill in the missing insights needed to improve drug development success rates. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 102nd Annual Meeting of the American Association for Cancer Research; 2011 Apr 2-6; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2011;71(8 Suppl):Abstract nr SY14-03. doi:10.1158/1538-7445.AM2011-SY14-03

Diagnosis of Several Diseases by Using Combined Kernels with Support Vector Machine

Machine learning techniques have gained increasing demand in biomedical research due to capability of extracting complex relationships and correlations among members of the large data sets. Thus, over the past few decades, scientists have been concerned about computer information technology to provide computational learning methods for solving the complex medical problems. Support Vector Machine is an efficient classifier that is widely applied to biomedical and other disciplines. In recent years, new opportunities have been developed on improving Support Vector Machines' classification efficiency by combining with any other statistical and computational methods. This study proposes a new method of Support Vector Machines for influential classification using combined kernel functions. The classification performance of the developed method, which is a type of non-linear classifier, was compared to the standart Support Vector Machine method by applying on seven different datasets of medical diseases. The results show that the new method provides a significant improvement in terms of the probability excess.

Creating an Agile ECE Learning Environment Through Engineering Clinics

To keep up with rapidly advancing technology, numerous innovations to the electrical and computer engineering (ECE) curriculum, learning methods and pedagogy have been envisioned, tested, and implemented. It is safe to say that no single approach will work for all of the diverse ECE technologies and every type of learner. However, a few key innovations appear useful in keeping undergraduate students motivated to learn, resilient to technology evolution, and oriented amid the overload of new information and ECE applications. Engineering clinics, similar to their medical clinic counterparts, provide project-based experiences within the core of an ECE education that enable transformation of the entire curriculum toward an outcomes-oriented, student-centered, total-quality environment. Clinics and project-based learning approaches build skills that give the students confidence and motivation to continuously self-learn and adapt as the technologies around them give way to new, more effective paradigms. Perhaps more importantly, engineering clinic experiences provide numerous opportunities for students to experience the holism of true engineering problem-solving approaches and the ranges of potential technology solutions. This paper reviews the clinic innovations that will enable ECE education to become more effective in the midst of the present plethora of information and technology. Assessment results are provided and are very encouraging. This paper concludes that agile learning environments, created to graduate engineers who can be rapidly productive in the professional and research worlds, are enhanced by clinic and/or project-based learning experiences in the ECE curriculum.

Computational Identification of Discriminating Features of Pathogenic and Symbiotic Type III Secreted Effector Proteins

Type III secretion systems (T3SS) deliver bacterial proteins, or “effectors”, into eukaryotic host cells, inducing physiological responses in the hosts. Effector proteins have been considered virulence factors of pathogenic bacteria, but T3SSs have now been found in symbiotic bacteria as well. Whether any physicochemical difference exists between the two types of effectors remains unknown. In this work, we combined computational statistical and machine learning methods to identify features that could be responsible for the difference. For computational statistical method we used generalized Bayesian information criterion and kernel logistic regression, and for machine learning method we used support vector machine. It was clearly shown that differences in amino acid composition exist between pathogenic and symbiotic effector proteins. All identified discriminating features were those of amino acid composition and average residue weight, and their classification performance could be nearly identical to that using all physicochemical features, with sensitivity and specificity of over 80%. Further analysis on the seven discriminating features by graphical modeling revealed three dominant features among them. Moreover, amino acid regions that were distinctive for the seven features were explored by sliding window analysis. This study provides a methodological basis and important insights into the functional differences between pathogenic and symbiotic T3SS effectors.

Convergence Properties of a Computational Learning Model for Unknown Markov Chains

The increasing complexity of engineering systems has motivated continuing research on computational learning methods toward making autonomous intelligent systems that can learn how to improve their performance over time while interacting with their environment. These systems need not only to sense their environment, but also to integrate information from the environment into all decision-makings. The evolution of such systems is modeled as an unknown controlled Markov chain. In a previous research, the predictive optimal decision-making (POD) model was developed, aiming to learn in real time the unknown transition probabilities and associated costs over a varying finite time horizon. In this paper, the convergence of the POD to the stationary distribution of a Markov chain is proven, thus establishing the POD as a robust model for making autonomous intelligent systems. This paper provides the conditions that the POD can be valid, and be an interpretation of its underlying structure.

Computer learning method for scattering junction identification

Acoustic reflectometry is a non-invasive, time-domain method of identifying the geometry of an acoustical space. A sound pulse is injected into a space and the resulting impulse response details particular changes of impedance, which is a result of a cross-sectional area change or an elbow/T- intersection. Each cause of reflection, known as a scattering junction, has a distinct reflection contour. Previous works were able to identify these scattering junctions via algorithms that attempt to extract particular contours from the impulse response. In the present study, the prominent reflections of the space are observed, isolated, and then compared to a training database of all possible scattering junctions. This method eliminates the necessity to create a contour identification algorithm, as scattering junctions are defined based on its most similar neighbors in the training database. Preliminary results suggest that this computer-learning algorithm can successfully identify reflection contours of a space with varying cross-sectional areas from those that were stored in the training database, which suggests that this method could be a more efficient and versatile alternative to previous identification processes.

MiRFinder: an improved approach and software implementation for genome-wide fast microRNA precursor scans.

BackgroundMicroRNAs (miRNAs) are recognized as one of the most important families of non-coding RNAs that serve as important sequence-specific post-transcriptional regulators of gene expression. Identification of miRNAs is an important requirement for understanding the mechanisms of post-transcriptional regulation. Hundreds of miRNAs have been identified by direct cloning and computational approaches in several species. However, there are still many miRNAs that remain to be identified due to lack of either sequence features or robust algorithms to efficiently identify them.ResultsWe have evaluated features valuable for pre-miRNA prediction, such as the local secondary structure differences of the stem region of miRNA and non-miRNA hairpins. We have also established correlations between different types of mutations and the secondary structures of pre-miRNAs. Utilizing these features and combining some improvements of the current pre-miRNA prediction methods, we implemented a computational learning method SVM (support vector machine) to build a high throughput and good performance computational pre-miRNA prediction tool called MiRFinder. The tool was designed for genome-wise, pair-wise sequences from two related species. The method built into the tool consisted of two major steps: 1) genome wide search for hairpin candidates and 2) exclusion of the non-robust structures based on analysis of 18 parameters by the SVM method. Results from applying the tool for chicken/human and D. melanogaster/D. pseudoobscura pair-wise genome alignments showed that the tool can be used for genome wide pre-miRNA predictions.ConclusionThe MiRFinder can be a good alternative to current miRNA discovery software. This tool is available at .

A parallel genetic algorithm for single class pattern classification and its application for gene expression profiling in Streptomyces coelicolor

BackgroundIdentification of coordinately regulated genes according to the level of their expression during the time course of a process allows for discovering functional relationships among genes involved in the process.ResultsWe present a single class classification method for the identification of genes of similar function from a gene expression time series. It is based on a parallel genetic algorithm which is a supervised computer learning method exploiting prior knowledge of gene function to identify unknown genes of similar function from expression data. The algorithm was tested with a set of randomly generated patterns; the results were compared with seven other classification algorithms including support vector machines. The algorithm avoids several problems associated with unsupervised clustering methods, and it shows better performance then the other algorithms. The algorithm was applied to the identification of secondary metabolite gene clusters of the antibiotic-producing eubacterium Streptomyces coelicolor. The algorithm also identified pathways associated with transport of the secondary metabolites out of the cell. We used the method for the prediction of the functional role of particular ORFs based on the expression data.ConclusionThrough analysis of a time series of gene expression, the algorithm identifies pathways which are directly or indirectly associated with genes of interest, and which are active during the time course of the experiment.

Classification of a Diverse Set of Tetrahymena pyriformis Toxicity Chemical Compounds from Molecular Descriptors by Statistical Learning Methods

Toxicity of various compounds has been measured in many studies by their toxic effects against Tetrahymena pyriformis. Efforts have also been made to use computational quantitative structure-activity relationship (QSAR) and statistical learning methods (SLMs) for predicting Tetrahymena pyriformis toxicity (TPT) at impressive accuracies. Because of the diversity of compounds and toxicity mechanisms, it is desirable to explore additional methods and to examine if these methods are applicable to more diverse sets of compounds. We tested several SLMs (logistic regression, C4.5 decision tree, k-nearest neighbor, probabilistic neural network, support vector machines) for their capability in predicting TPT by using 1129 compounds (841 TPT and 288 non-TPT agents) which are more diverse than those in other studies. A feature selection method was used for improving prediction performance and selecting molecular descriptors responsible for distinguishing TPT and non-TPT agents. The prediction accuracies are 86.9% approximately 94.2% for TPT and 71.2% approximately 87.5% for non-TPT agents based on 5-fold cross-validation studies, which are comparable to some of earlier studies despite the use of more diverse sets of compounds. The selected molecular descriptors are consistent with those used in other studies and experimental findings. These suggest that SLMs are useful for predicting TPT potential of diverse sets of compounds and for characterizing the molecular descriptors associated with TPT.

Robo-Erectus: a low-cost autonomous humanoid soccer robot

(2004). Robo-Erectus: a low-cost autonomous humanoid soccer robot. Advanced Robotics: Vol. 18, No. 7, pp. 717-720.

Application of Pyramidal Networks to the Search for New Electro-Optical Inorganic Materials

The problem of predicting new inorganic material systems with predefined properties on the basis of elemental properties remains largely unsolved. Within the context of this problem, a computer learning method (concept formation), based on semantic networks of a special kind - growing pyramidal networks (GPNs), was used in the process of searching for novel compounds with electro-optical (EO) properties. The basic chemical hypothesis of this investigation was that only compounds with acentric crystal space groups possess EO properties. It is expected that new EO compounds will be similar in composition and crystal structure to already known compounds. The crystal chemical types used to build the GPN were selected on the basis of known acousto-optical, electro-optical, and nonlinear-optical materials. Hundreds of new ternary and quaternary EO compounds are predicted using the resultant GPN.

Knowledge-based analysis of microarray gene expression data by using support vector machines.

We introduce a method of functionally classifying genes by using gene expression data from DNA microarray hybridization experiments. The method is based on the theory of support vector machines (SVMs). SVMs are considered a supervised computer learning method because they exploit prior knowledge of gene function to identify unknown genes of similar function from expression data. SVMs avoid several problems associated with unsupervised clustering methods, such as hierarchical clustering and self-organizing maps. SVMs have many mathematical features that make them attractive for gene expression analysis, including their flexibility in choosing a similarity function, sparseness of solution when dealing with large data sets, the ability to handle large feature spaces, and the ability to identify outliers. We test several SVMs that use different similarity metrics, as well as some other supervised learning methods, and find that the SVMs best identify sets of genes with a common function using expression data. Finally, we use SVMs to predict functional roles for uncharacterized yeast ORFs based on their expression data.

An Empirical Study of Computer System Learning: Comparison of Co-Discovery and Self-Discovery Methods

This paper reports a study that examined two types of exploratory computer learning methods: self-discovery vs. co-discovery, the latter of which involves two users working together to learn a system. An experiment was conducted to compare these two methods and the results were interpreted within a mental model framework. Co-discovery subjects were better than self-discovery subjects at making inferences about the capability and extended functions of the system. Furthermore, while working by themselves after an initial period of learning, they performed better in a similar, though more complex task than the one they encountered at the learning phase. Process tracing analysis showed that self-discovery subjects focused more on surface structures, such as detailed physical actions, for implementing the task. On the other hand, co-discovery groups focused more on relating lower level actions to higher level goals. Therefore, co-discovery subjects had a better understanding of the relationships between the physical actions and goals, and hence formed mental models with higher inference potential than self-discovery subjects.

The ClusNet algorithm and time series prediction.

This paper describes a novel neural network architecture named ClusNet. This network is designed to study the trade-offs between the simplicity of instance-based methods and the accuracy of the more computational intensive learning methods. The features that make this network different from existing learning algorithms are outlined. A simple proof of convergence of the ClusNet algorithm is given. Experimental results showing the convergence of the algorithm on a specific problem is also presented. In this paper, ClusNet is applied to predict the temporal continuation of the Mackey-Glass chaotic time series. A comparison between the results obtained with ClusNet and other neural network algorithms is made. For example, ClusNet requires one-tenth the computing resources of the instance-based local linear method for this application while achieving comparable accuracy in this task. The sensitivity of ClusNet prediction accuracies on specific clustering algorithms is examined for an application. The simplicity and fast convergence of ClusNet makes it ideal as a rapid prototyping tool for applications where on-line learning is required.

Comparative study of methods of identification of alkylbenzenes from their mass spectra

Different computer learning methods, identification flowcharts, and library searching methods are investigated to determine their abilities to identify an unknown alkylbenzene from its mass spectrum. It is concluded that the best results are obtained by using different methods in a complementary manner.