Rami N. Mahdi

Reduced HyperBF Networks: Regularization by Explicit Complexity Reduction and Scaled Rprop-Based Training

Hyper basis function (HyperBF) networks are generalized radial basis function neural networks (where the activation function is a radial function of a weighted distance. Such generalization provides HyperBF networks with high capacity to learn complex functions, which in turn make them susceptible to overfitting and poor generalization. Moreover, training a HyperBF network demands the weights, centers, and local scaling factors to be optimized simultaneously. In the case of a relatively large dataset with a large network structure, such optimization becomes computationally challenging. In this paper, a new regularization method that performs soft local dimension reduction in addition to weight decay is proposed. The regularized HyperBF network is shown to provide classification accuracy competitive to a support vector machine while requiring a significantly smaller network structure. Furthermore, a practical training to construct HyperBF networks is presented. Hierarchal clustering is used to initialize neurons followed by a gradient optimization using a scaled version of the Rprop algorithm with a localized partial backtracking step. Experimental results on seven datasets show that the proposed training provides faster and smoother convergence than the regular Rprop algorithm.

Empirical Bayes conditional independence graphs for regulatory network recovery

MOTIVATION Computational inference methods that make use of graphical models to extract regulatory networks from gene expression data can have difficulty reconstructing dense regions of a network, a consequence of both computational complexity and unreliable parameter estimation when sample size is small. As a result, identification of hub genes is of special difficulty for these methods. METHODS We present a new algorithm, Empirical Light Mutual Min (ELMM), for large network reconstruction that has properties well suited for recovery of graphs with high-degree nodes. ELMM reconstructs the undirected graph of a regulatory network using empirical Bayes conditional independence testing with a heuristic relaxation of independence constraints in dense areas of the graph. This relaxation allows only one gene of a pair with a putative relation to be aware of the network connection, an approach that is aimed at easing multiple testing problems associated with recovering densely connected structures. RESULTS Using in silico data, we show that ELMM has better performance than commonly used network inference algorithms including GeneNet, ARACNE, FOCI, GENIE3 and GLASSO. We also apply ELMM to reconstruct a network among 5492 genes expressed in human lung airway epithelium of healthy non-smokers, healthy smokers and individuals with chronic obstructive pulmonary disease assayed using microarrays. The analysis identifies dense sub-networks that are consistent with known regulatory relationships in the lung airway and also suggests novel hub regulatory relationships among a number of genes that play roles in oxidative stress and secretion. AVAILABILITY AND IMPLEMENTATION Software for running ELMM is made available at http://mezeylab.cb.bscb.cornell.edu/Software.aspx. CONTACT ramimahdi@yahoo.com or jgm45@cornell.edu SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.

Semi-Supervised Clustering and Feature Discrimination with Instance-Level Constraints

We propose a Semi-Supervised Clustering and Attribute Discrimination (S-SCAD) algorithm that performs fuzzy clustering and coarse feature weighting simultaneously. The supervision information in S-SCAD consists of a small set of constraints on which instances should or should not reside in the same cluster. The feature set is divided into logical subsets of features, and a degree of relevance is dynamically assigned to each subset based on its partial degree of dissimilarity. These weights have two advantages. First, they help in partitioning the data set into more meaningful clusters. Second, they can be used as part of a more complex learning system to enhance its learning behavior. We show that the partial supervision can guide the algorithm in learning the prototype parameters and the feature relevance weights, and thus, improve the final partition. The performance of the proposed algorithm is illustrated by using it to categorize a collection of color images. We use four feature subsets that encode color, structure, and texture information. The results are compared to other similar algorithms.

Buffered codons in human transcriptional units

Background Codon usage is well established for a number of different species. Multiple models have been proposed to show codon bias as a balance between mutation and selection. Most of these models emphasize controlling the speed of protein translation from the mRNA and increasing the accuracy where this bias is dependent on the abundance of the available tRNA. We show codon usage bias from a different angle based on a new hypothesis where selection is expected to act in a direction to favor codons that are more buffered, or protected, from mutation than those sensitive to mutation. It is anticipated that the more buffered the original coding sequence, the higher the survival chance for the whole organism since the resulting protein sequence remains unchanged. Two different complementary measures are developed to compute the average buffering capacity in a given sequence. We show that the buffering capacity of coding sequences is in general higher than that of randomly generated sequences and that of shifted reading frames. Highly expressed genes are shown to have an even higher buffering capacity than non-housekeeping genes.

RBF-TSS: Identification of Transcription Start Site in Human Using Radial Basis Functions Network and Oligonucleotide Positional Frequencies

Accurate identification of promoter regions and transcription start sites (TSS) in genomic DNA allows for a more complete understanding of the structure of genes and gene regulation within a given genome. Many recently published methods have achieved high identification accuracy of TSS. However, models providing more accurate modeling of promoters and TSS are needed. A novel identification method for identifying transcription start sites that improves the accuracy of TSS recognition for recently published methods is proposed. This method incorporates a metric feature based on oligonucleotide positional frequencies, taking into account the nature of promoters. A radial basis function neural network for identifying transcription start sites (RBF-TSS) is proposed and employed as a classification algorithm. Using non-overlapping chunks (windows) of size 50 and 500 on the human genome, the proposed method achieves an area under the Receiver Operator Characteristic curve (auROC) of 94.75% and 95.08% respectively, providing increased performance over existing TSS prediction methods.

Model Based Unsupervised Learning Guided by Abundant Background Samples

Many data sets contain an abundance of background data or samples belonging to classes not currently under consideration. We present a new unsupervised learning method based on fuzzy c-means to learn sub models of a class using background samples to guide cluster split and merge operations. The proposed method demonstrates how background samples can be used to guide and improve the clustering process. The proposed method results in more accurate clusters and helps to escape locally minimum solutions. In addition, the number of clusters is determined for the class under consideration. The method demonstrates remarkable performance on both synthetic 2D and real world data from the MNIST dataset of hand written digits.

Codon usage bias as a function of generation time and life expectancy

It has recently been demonstrated that human natural codon usage bias is optimized towards a higher buffering capacity to mutations (measured as the tendency of single point mutations in a DNA sequence to yield the same or similar amino acids) compared to random sequences. In this work, we investigate this phenomenon further by analyzing the natural DNA of four different species (human, mouse, zebrafish and fruit fly) to determine whether such a tolerance to mutations is correlated with the life span and age of sexual maturation for the corresponding organisms. We also propose a new measure to quantify the buffering capacity of a DNA sequence to mutations that takes into account the observed mutation rates within every genome and the effect of the corresponding mutation. Our results suggest there is a propensity for tolerance to mutations that is positively correlated with the life expectancy of the considered organisms. Moreover, random sequences that are constrained to produce the same protein as the naturally occurring sequences are found to be more buffered than completely random sequences while being less buffered than the natural sequences. These results suggest that optimization toward protective mechanisms tolerant to mutations is correlated with both life expectancy and age to sexual maturity at both the levels of codon usage bias and the bias of the natural sequence of codons itself.

Evidence of bias towards buffered codons in human transcripts

Codon usage bias is well established in different species from bacteria to mammals. A number of models have been proposed to show this bias as a balance between mutation and selection. Most of these models emphasize controlling the speed of protein translation from the mRNA and increasing the accuracy where this bias is dependent on the abundance and properties of the available tRNA. In this work, codon usage bias in general is considered from a different angle based on a new hypothesis where selection is expected to act in a direction to favor codons that are more buffered, or protected, from mutation than those more sensitive to mutation. It is anticipated that the more buffered the original coding sequence, the higher the survival chance for the whole organism since the resulting protein sequence remains unchanged. Two different complementary measures are developed to compute the average buffering capacity in a given sequence. We show that the buffering capacity of coding sequences in humans is in general higher than that of randomly generated sequences and that of shifted reading frames. Highly expressed genes are shown to have an even higher buffering capacity than non-housekeeping genes. This result is to be expected due to the necessity of housekeeping genes.

Combining feedback and image database categorization in CBIR

We propose a content-based image retrieval prototype that combines the advantages of relevance feedback and image database categorization. Our approach is based on an algorithm that performs clustering and feature weighting simultaneously and can incorporate partial supervision information. This information, extracted from the userpsilas feedback through visual exploration and interaction, is used to refine the clusterspsila distributions and their feature relevance weights in the vicinity of the query image. The cluster dependent feature weights are used in the retrieval phase to adapt the similarity to the different categories. The feedback information is encoded as a set of constraints on which instances should or should not reside in the same cluster. Thus, it does not depend on the query image explicitly. Consequently, partial supervision information from different query sessions could be saved, accumulated, and used to continuously refine the image categories and their feature weights.