We present a statistical model of genes in DNA. A Generalized Hidden Markov Model (GHMM) provides the framework for describing the grammar of a legal parse of a DNA sequence (Stormo & Haussler 1994). Probabilities are assigned to transitions between states in the GHMM and to the generation of each nucleotide base given a particular state. Machine learning techniques are applied to optimize these probabilities using a standardized training set. Given a new candidate sequence, the best parse is deduced from the model using a dynamic programming algorithm to identify the path through the model with maximum probability. The GHMM is flexible and modular, so new sensors and additional states can be inserted easily. In addition, it provides simple solutions for integrating cardinality constraints, reading frame constraints, "indels", and homology searching. The description and results of an implementation of such a gene-finding model, called Genie, is presented. The exon sensor is a codon fre...

We present an improved splice site predictor for the genefinding program Genie. Genie is based on a generalized Hidden Markov Model (GHMM) that describes the grammar of a legal parse of a multi-exon gene in a DNA sequence. In Genie, probabilities are estimated for gene features by using dynamic programming to combine information from multiple content and signal sensors, including sensors that integrate matches to homologous sequences from a database. One of the hardest problems in genefinding is to determine the complete gene structure correctly. The splice site sensors are the key signal sensors that address this problem. We replaced the existing splice site sensors in Genie with two novel neural networks based on dinucleotide frequencies. Using these novel sensors, Genie shows significant improvements in the sensitivity and specificity of gene structure identification. Experimental results in tests using a standard set of annotated genes showed that Genie identified 86% of coding nuc...

this paper, we present a generalized form of SVG, which supports additional biologically-relevant operations by going beyond homomorphisms, instead uniformly applying substitutions in either a forward or reverse direction (see Definition 2.1) to bindings of logic variables. We give a constructive proof of our conjecture [26] that the languages describable by SVG are contained in the indexed languages, and furthermore show that the containment is proper, thus refining the position of an important class of biological sequences in the hierarchy of languages. We also describe a simple grammar translator, give a number of examples of mathematical and biological languages, discuss the distinctions between SVG, DG, TAG, and RPDAs, and suggest extensions well-suited to the overlapping languages of genes. Finally, we describe a large-scale implementation of a domain-specific parser called GenLang which incorporates a practical version of these ideas, and which has been successful in parsing several types of genes from DNA sequence data [9, 30], in a form of pattern-matching search termed syntactic pattern recognition [10]. 6 2. STRING VARIABLE GRAMMAR

Research into new methods to identify genes in anonymous genomic sequences has been going on for more than 15 years. Over this period of time, the field has evolved from the designing of programs to identify protein coding regions in compact mitochondrial or bacterial genomes, to the challenge of predicting the detailed organization of multi-exon vertebrate genes. The best program currently available perfectly locates more than 80 % of the internal coding exons, and only 5 % of the predictions do not overlap a real exon. Given such accuracy, computational methods are indeed very useful; however, they do not alleviate the need for experimental validation. If the performances are satisfactory for the identification of the coding moiety of genes (internal coding exons), the determination of the full extent of the transcript (5 ′ and 3 ′ extremities of the gene) and the location of promoter regions are still unreliable. As the human and mouse genome sequencing projects enter a production mode, the fully automated annotation of megabase-long anonymous genomic sequences is the next big challenge in bioinformatics.

The analysis of the massive data sets collected by scientific instruments demands automation as a pre-requisite to analysis. There is an urgent need to create an intermediate level at which scientists can operate effectively; isolating them from the massive sizes and harnessing human analysis capabilities to focus on tasks in which machines do not even remotely approach humans---namely, creative data analysis, theory and hypothesis formation, and drawing insights into underlying phenomena. We give an overview of the main issues in the exploitation of scientific datasets, present five case studies where KDD tools play important and enabling roles, and conclude with future challenges for data mining and KDD techniques in science data analysis. keywords: Applications in Science, Data Analysis, overview article, large databases, automated analysis, scietific data sets, scientific discovery. To appear: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining...

Molecular biologists use algorithms that compare and otherwise analyze sequences that represent genetic and protein molecules. Most of these algorithms, however, operate on the basic sequence and do not incorporate the additional information that is often known about the molecule and its pieces. This research describes schemes to combinatorially annotate this information onto sequences so that it can be analyzed in tandem with the sequence; the overall result would thus reflect both types of information about the sequence. These annotation schemes include adding colours and arcs to the sequence. Colouring a sequence would produce a same-length sequence of colours or other symbols that highlight or label parts of the sequence. Arcs can be used to link sequence symbols (or coloured substrings) to indicate molecular bonds or other relationships. Adding these annotations to sequence analysis problems such as sequence alignment or finding the longest common subsequence can make the problem more complex, often depending on the complexity of the annotation scheme. This research examines the different annotation schemes and the corresponding problems of verifying annotations, creating annotations, and finding the longest common subsequence of pairs of sequences with annotations. This work involves both the conventional complexity framework and parameterized complexity, and includes algorithms and hardness results for both frameworks. Automata and transducers are created for some annotation verification and creation problems. Different restrictions on layered substring and arc annotation are considered to de- iii termine what properties an annotation scheme must have to make its incorporation feasible. Extensions to the algorithms that use weighting schemes are explored. Examin...

A general probabilistic model of the gene structural and compositional properties of human genomic DNA is introduced and applied to the problem of identifying genes in unannotated human genomic sequences. The model uses a \Hidden semi-Markov" or semi-Markov source architecture which incorporates probabilistic descriptions of fundamental transcriptional, translational and splicing signals, as well as length distri-butions and compositional features of exons, introns and intergenic regions. Distinct sets of model parameters are derived which account for many of the substantial di er-ences in gene density and structure observed in distinct C+G compositional regions (\isochores") of the human genome. A novel model building procedure, termed Max-imal Dependence Decomposition, is introduced which captures potentially important dependencies between non-adjacent aswell as adjacent positions in a biological signal. Application of this model to the donor splice signal not only gives better discrimina-tion of potential donor sites than previous probabilistic models, but also reveals subtle properties of this signal which suggest aspects of its biochemical function. Acceptor

Article cited in:

Given a sequence of DNA nucleotide bases, the task of gene prediction is to find subsequences of bases that encode proteins. Reasonable performance on this task has been achieved using generatively trained sequence models, such as hidden Markov models. We propose instead the use of a discriminitively trained sequence model, the conditional random field (CRF). CRFs can naturally incorporate arbitrary, non-independent features of the input without making conditional independence assumptions among the features. This can be particularly important for gene finding, where including evidence from protein databases, EST data, or tiling arrays may improve accuracy. We evaluate our model on human genomic data, and show that CRFs perform better than HMM-based models at incorporating homology evidence from protein databases, achieving a 10 % reduction in base-level errors. 1

Biological macromolecules can be viewed, at one level, as strings of symbols. Collections of such molecules can thus be considered to be sets of strings, i.e. formal languages. This article reviews languagetheoretic approaches to describing intramolecular and intermolecular structural interactions within these molecules, and evolutionary relationships between them. 1 Introduction The author has for some time been investigating the application of formal language theory to biological macromolecules, primarily nucleic acids because of the relative simplicity of the biochemical structures and interactions. After introducing the very simple mathematical foundations for these investigations, this article will review three major lines of research. These can largely be found in more fully developed form in referenced publications, though some new material is also included in each case. The sections below will deal with the use of formal grammars to describe intramolecular interactions [17, 18...