Logo des Repositoriums

Auflistung nach Autor:in "Keilwagen, Jens"

1 - 5 von 5
  • Konferenzbeitrag
    A general approach for discriminative de novo motif discovery from high-throughput data
    (German conference on bioinformatics 2014, 2014) Grau, Jan; Posch, Stefan; Grosse, Ivo; Keilwagen, Jens
    High-throughput techniques like ChIP-seq, ChIP-exo, and protein binding microarrays (PBMs) demand for novel de novo motif discovery approaches that focus on accuracy and runtime on large data sets. While specialized algorithms have been designed for discovering motifs in in-vivo ChIP-seq/ChIP-exo or in in-vitro PBM data, none of these works equally well for all these high-throughput techniques. Here, we present Dimont, a general approach for fast and accurate de-novo motif discovery from high-throughput data, which achieves a competitive performance on both ChIP-seq and PBM data compared to recent approaches specifically designed for either technique. Hence, Dimont allows for investigating differences between in-vitro and in-vivo binding in an unbiased manner using a unified approach. For most transcription factors, Dimont discovers similar motifs from in-vivo and in-vitro data, but we also find notable exceptions. Scrutinizing the benefit of modeling dependencies between binding site positions, we find that more complex motif models often increase prediction performance and, hence, are a worthwhile field of research. Original paper: doi: 10.1093/nar/gkt831
  • Konferenzbeitrag
    Predicting miRNA targets utilizing an extended profile HMM
    (German Conference on Bioinformatics 2010, 2010) Grau, Jan; Arend, Daniel; Grosse, Ivo; Hatzigeorgiou, Artemis G.; Keilwagen, Jens; Maragkakis, Manolis; Weinholdt, Claus; Posch, Stefan
    The regulation of many cellular processes is influenced by miRNAs, and bioinformatics approaches for predicting miRNA targets evolve rapidly. Here, we propose conditional profile HMMs that learn rules of miRNA-target site interaction automatically from data. We demonstrate that conditional profile HMMs detect the rules implemented into existing approaches from their predictions. And we show that a simple UTR model utilizing conditional profile HMMs predicts target genes of miR- NAs with a precision that is competitive compared to leading approaches, although it does not exploit cross-species conservation.
  • Konferenzbeitrag
    Supervised posteriors for DNA-motif classification
    (German conference on bioinformatics – GCB 2007, 2007) Grau, Jan; Keilwagen, Jens; Kel, Alexander; Grosse, Ivo; Posch, Stefan
    Markov models have been proposed for the classification of DNA-motifs using generative approaches for parameter learning. Here, we propose to apply the discriminative paradigm for this problem and study two different priors to facilitate parameter estimation using the maximum supervised posterior. Considering seven sets of eukaryotic transcription factor binding sites we find this approach to be superior employing area under the ROC curve and false positive rate as performance criterion, and better in general using sensitivity. In addition, we discuss potential reasons for the improved performance.
  • Konferenzbeitrag
    Utilizing promoter pair orientations for HMM-based analysis of ChIP-chip data
    (German Conference on Bioinformatics, 2008) Seifert, Michael; Keilwagen, Jens; Strickert, Marc; Grosse, Ivo
    Array-based analysis of chromatin immunoprecipitation data (ChIP-chip) is a powerful technique for identifying DNA target regions of individual transcription factors. Here, we present three approaches, a standard log-fold-change analysis (LFC), a basic method based on a Hidden Markov Model (HMM), and an ex- tension of the HMM approach to an HMM with scaled transition matrices (SHMM) to incorporate different promoter pair orientations. We compare the prediction of ABI3 target genes for the three methods and evaluate these genes using Geneves- tigator expression profiles and transient assays. We find that the application of the SHMM leads to a superior identification of ABI3 target genes. The software and the ChIP-chip data set used in our case study can be downloaded from http://dig.ipk- gatersleben.de/SHMMs/ChIPchip/ChIPchip.html.
  • Textdokument
    Vorhersage von DNA-Bindungsstellen mit generativen, diskriminativen und hybriden Lernverfahren
    (Ausgezeichnete Informatikdissertationen 2010, 2011) Keilwagen, Jens
    Probabilistische Modelle werden heutzutage aufgrund ihrer Flexibilität in vielen Bereichen zur Modellierung und Klassifikation von anfallenden Daten genutzt. Von entscheidender Bedeutung ist neben der Wahl des entsprechenden Modells auch die Wahl des Lernverfahrens, welches die Modellparameter aus gegebenen Daten inferriert. Häufig wird dieser Aspekt völlig aus den Augen gelassen, obwohl er sehr viel Potenzial birgt. In der vorgelegten Dissertation wird u.a. ein verallgemeinertes Lernverfahren vorgestellt und auf biologische Daten angewendet. Die objektorientierte und quelloffene Implementierung ermöglicht eine Vielzahl weiterer Anwendungen und Erweiterungen.