Jennifer A. Smith

Integrating Thermodynamic and Observed-Frequency Data for Non-Coding RNA Gene Search

Jennifer Smith et al. — Wed, 10 Aug 2011 14:18:55 PDT

Among the most powerful and commonly used methods for finding new members of non-coding RNA gene families in genomic data are covariance models. The parameters of these models are estimated from the observed position-specific frequencies of insertions, deletions, and mutations in a multiple alignment of known non-coding RNA family members. Since the vast majority of positions in the multiple alignment have no observed changes, yet there is no reason to rule them out, some form of prior is applied to the estimate. Currently, observed-frequency priors are generated from non-family members based on model node type and child node type allowing for some differentiation between priors for loops versus helices and between internal segments of structures and edges of structures. In this work it is shown that parameter estimates might be improved when thermodynamic data is combined with the consensus structure/sequence and observed-frequency priors to create more realistic position-specific priors.

Efficient Non-Coding RNA Gene Searches Through Classical and Evolutionary Methods

Jennifer A. Smith — Wed, 10 Aug 2011 13:53:48 PDT

Successful non-coding RNA gene searching requires examination of long-range intramolecular base pairing possibilities. This results in search algorithms with extremely long run times such that large-scale use of the algorithms often becomes computationally infeasible. Methods for the efficient search of the solution space are examined. A review of the standard dynamic-programming covariance model search algorithm is given. An analysis of the statistically probable regions of the search space is undertaken and a method of limiting the traditional dynamic-programming algorithm to this region is shown. An alternative search method using a Genetic Algorithm (GA) which favours the probable region of the search space is also given.

Joint Loop End Modeling Improves Covariance Model Based Non-coding RNA Gene Search

Jennifer Smith — Wed, 10 Nov 2010 08:32:35 PST

The effect of more detailed modeling of the interface between stem and loop in non-coding RNA hairpin structures on efficacy of covariance-model-based non-coding RNA gene search is examined. Currently, the prior probabilities of the two stem nucleotides and two loop-end nucleotides at the interface are treated the same as any other stem and loop nucleotides respectively. Laboratory thermodynamic studies show that hairpin stability is dependent on the identities of these four nucleotides, but this is not taken into account in current covariance models. It is shown that separate estimation of emission priors for these nucleotides and joint treatment of substitution probabilities for the two loop-end nucleotides leads to improved non-coding RNA gene search.

Efficient Non-Coding RNA Gene Searches through Classical and Evolutionary Methods

Jennifer A. Smith — Tue, 26 Oct 2010 07:39:15 PDT

Computation Intelligence Method to Find Generic Non-Coding RNA Search Models

Jennifer A. Smith — Wed, 15 Sep 2010 09:51:04 PDT

Fairly effective methods exist for finding new noncoding RNA genes using search models based on known families of ncRNA genes (for example covariance models). However, these models only find new members of the existing families and are not useful in finding potential members of novel ncRNA families. Other problems with family-specific search include large processing requirements, ambiguity in defining which sequences form a family and lack of sufficient numbers of known sequences to properly estimate model parameters. An ncRNA search model is proposed which includes a collection of non-overlapping RNA hairpin structure covariance models. The hairpin models are chosen from a hairpin-model list compiled from many families in the Rfam non-coding RNA families database. The specific hairpin models included and the overall score threshold for the search model is determined through the use of a genetic algorithm.

RNA Search with Decision Trees and Partial Covariance Models

Jennifer A. Smith — Thu, 06 Aug 2009 17:49:34 PDT

The use of partial covariance models to search for RNA family members in genomic sequence databases is explored. The partial models are formed from contiguous subranges of the overall RNA family multiple alignment columns. A binary decision-tree framework is presented for choosing the order to apply the partial models and the score thresholds on which to make the decisions. The decision trees are chosen to minimize computation time subject to the constraint that all of the training sequences are passed to the full covariance model for final evaluation. Computational intelligence methods are suggested to select the decision tree since the tree can be quite complex and there is no obvious method to build the tree in these cases. Experimental results from seven RNA families shows execution times of 0.066-0.268 relative to using the full covariance model alone. Tests on the full sets of known sequences for each family show that at least 95 percent of these sequences are found for two families and 100 percent for five others. Since the full covariance model is run on all sequences accepted by the partial model decision tree, the false alarm rate is at least as low as that of the full model alone.

Improved Covariance Model Parameter Estimation Using RNA Thermodynamic Properties

Jennifer A. Smith et al. — Wed, 20 May 2009 06:35:51 PDT

Covariance models are a powerful description of non-coding RNA (ncRNA) families that can be used to search nucleotide databases for new members of these ncRNA families. Currently, estimation of the parameters of a covariance model (state transition and emission scores) is based only on the observed frequencies of mutations, insertions, and deletions in known ncRNA sequences. For families with very few known members, this can result in rather uninformative models where the consensus sequence has a good score and most deviations from consensus have a fairly uniform poor score. It is proposed here to combine the traditional observed-frequency information with known information about free energy changes in RNA helix formation and loop length changes. More thermodynamically probable deviations from the consensus sequence will then be favored in database search. The thermodynamic information may be incorporated into the models as informative priors that depend on neighboring consensus nucleotides and on loop lengths.

RNA Gene Finding with Biased Mutation Operators

Jennifer A. Smith — Wed, 20 May 2009 06:35:51 PDT

The use of genetic algorithms for non-coding RNA gene finding has previously been investigated and found to be a potentially viable method for accelerating covariance-model-based database search relative to full dynamic-programming methods. The mutation operators in previous work chose new alignment insertion and deletion locations uniformly over the length of the model consensus sequence. Since the covariance models are estimated from multiple known members of a non-coding RNA family, information is available as to the likelihood of insertions or deletions at the individual model positions. This information is implicit in the state-transition parameters of the estimated covariance models. In the current work, the use of mutation operators which are biased toward selection of insertions and deletions at model positions with low insertion or deletion penalties is examined in hopes of speeding up convergence. The performance of the biased and unbiased mutation operators is compared. Both biased and unbiased genetic algorithms are also compared to a steepest-descent algorithm, which is a comparison lacking in prior work.

Searching for Protein Classification Features

Jennifer A. Smith — Wed, 20 May 2009 06:35:50 PDT

A genetic algorithm is used to search for a set of classification features for a protein superfamily which is as unique as possible to the superfamily. These features may then be used for very fast classification of a query sequence into a protein superfamily. The features are based on windows onto modified consensus sequences of multiple aligned members of a training set for the protein superfamily. The efficacy of the method is demonstrated using receiver operating characteristic (ROC) values and the performance of resulting algorithm is compared with other database search algorithms.

Protein Family Classification Using Structural and Sequence Information

Jennifer A. Smith — Wed, 20 May 2009 06:35:50 PDT

Protein family classification usually relies on sequence information (as in the case of hidden Markov models and position-specific scoring matrices) or on structural information where some sort of average positional error between the atomic locations is used. The positional error method requires that the structure of all the proteins to be classified is known. Sequence methods have the advantage that a much larger number of proteins can be classified (since far more sequences are know than structures). However, sequence methods discard a large amount of useful information contained in the structures of the subset of proteins in the family for which structures are known. A protein family classification system is presented which uses both structural and sequence information and combines this information in a way consistent with fuzzy systems theory. The non-linear fuzzy-theory-based method is found to perform better than either an equally-weighted linear combination of the sequence and structural information or the sequence information alone.

Accelerated Non-coding RNA Searches with Covariance Model Approximations

Jennifer A. Smith — Wed, 20 May 2009 06:35:49 PDT

Covariance models (CMs) are a very sensitive tool for finding non-coding RNA (ncRNA) genes in DNA sequence data. However, CMs are extremely slow. One reason why CMs are so slow is that they allow all possible combinations of insertions and deletions relative to the consensus model even though the vast majority of these are never seen in practice. In this paper we examine reduction in the number of states in covariance models. A simplified CM with reduced states which can be scored much faster is introduced. A comparison of the results of a full CM versus a reduced-state model found using a genetic algorithm is given for the let7 ncRNA family.

A Genetic Algorithms Approach to Non-coding RNA Gene Searches

Jennifer A. Smith — Wed, 20 May 2009 06:35:48 PDT

A genetic algorithm is proposed as an alternative to the traditional linear programming method for scoring covariance models in non-coding RNA (ncRNA) gene searches. The standard method is guaranteed to find the best score, but it is too slow for general use. The observation that most of the search space investigated by the linear programming method does not even remotely resemble any observed sequence in real sequence data can be used to motivate the use of genetic algorithms (GAs) to quickly reject regions of the search space. A search space with many local minima makes gradient decent an unattractive alternative. It is shown that a fixed-length representation for alignment of two sequences taken from the protein threading literature can be adapted for use with covariance models.

Truncated Profile Hidden Markov Models

Jennifer A. Smith — Wed, 20 May 2009 06:35:48 PDT

The profile hidden Markov model (HMM) is a powerful method for remote homolog database search. However, evaluating the score of each database sequence against a profile HMM is computationally demanding. The computation time required for score evaluation is proportional to the number of states in the profile HMM. This paper examines whether the number of states can be truncated without reducing the ability of the HMM to find proteins containing members of a protein domain family. A genetic algorithm (GA) is presented which finds a good truncation of the HMM states. The results of using truncation on searches of the yeast, E. coli, and pig genomes for several different protein domain families is shown.

An Asynchronous GALS Interface with Applications

Jennifer A. Smith — Wed, 20 May 2009 06:35:46 PDT

A low-latency asynchronous interface for use in globally-asynchronous locally-synchronous (GALS) integrated circuits is presented. The interface is compact and does not alter the local clocks of the interfaced local clock domains in any way (unlike many existing GALS interfaces). Two applications of the interface to GALS systems are shown. The first is a single-chip shared-memory multiprocessor for generic supercomputing use. The second is an application-specific coprocessor for hardware acceleration of the Smith-Waterman algorithm. This is a bioinformatics algorithm used for sequence alignment (similarity searching) between DNA or amino acid (protein) sequences and sequence databases such as the recently completed human genome database.

Covariance Searches for ncRNA Gene Finding

Jennifer A. Smith — Wed, 20 May 2009 06:35:45 PDT

The use of covariance models for non-coding RNA gene finding is extremely powerful and also extremely computationally demanding. A major reason for the high computational burden of this algorithm is that the search proceeds through every possible start position in the database and every possible sequence length between zero and a user-defined maximum length at every one of these start positions. Furthermore, for every start position and sequence length, all possible combinations of insertions and deletions leading to the given sequence length are searched. It has been previously shown that a large portion of this search space is nowhere near any database match observed in practice and that the search space can be limited significantly with little change in expected search results. In this work a different approach is taken in which the space of starting positions, sequence lengths, and insertion/deletion patterns is searched using a genetic algorithm.