Section: New Results

Algorithms & Methods

Short Read Correction

Participants : Antoine Limasset, Pierre Peterlongo.

We proposed a new method to correct short reads using de Bruijn graphs, and we implemented it as a tool called Bcool. As a first step, Bcool constructs a corrected compacted de Bruijn graph from the reads. This graph is then used as a reference and the reads are corrected according to their mapping on the graph. We showed that this approach yields a better correction than $k$mer-spectrum techniques, while being scalable, making it possible to apply it to human-size genomic datasets and beyond [41].

Long transcriptomic read clustering

Participants : Camille Marchet, Pierre Peterlongo.

This contribution tackles the problem of clustering RNA reads in clusters representing all variants of each gene, in a de novo way i.e. without any reference sequences. Such problem is not new as is, but the latest, Third Generation Sequencing (TGS) data redefine it. Reads can now span full-length transcripts but at the price of very high error rates, mostly insertions and deletions. This makes difficult or impossible to use tools designed for previous sequencing data. Still, the property to obtain whole RNA molecules through reads is very promising to better describe a transcriptome. In this work, we targeted the need to extract relevant information from a TGS transcriptome, even when no reference is available. In collaboration with Jacques Nicolas from the Inria/IRISA Dyliss team, we propose a novel algorithm in the community detection framework, based on the clustering coefficient. In addition we propose an implementation of this algorithm in the tool CARNAC-LR and a pipeline for the processing of transcriptome data. We validated our tool on real data from mouse and showed that it could be accurate and precise even for lowly expressed genes. We showed that our approach can be complementary to a mapping in the case a reference exists, and that a straightforward use of CARNAC-LR enables to quickly assess the genes'e expression levels [42].

Statistically Significant Discriminative Patterns Search

Participants : Hoang Son Pham, Dominique Lavenier.

Identifying multiple SNPs combinations associated with diseases such as cancers or diabetes is a central goal of human genetics. Recently, discriminative pattern mining algorithms have been investigated to tackle genome-wide association studies (GWAS). We designed an algorithm, called SSDPS, to discover groups of items which have significant difference of frequency in case-control datasets. The algorithm directly uses relative risk measures such as risk ratio, odds ratio and absolute risk reduction combined with confidence intervals as anti-monotonic properties to efficiently prune the search space. The algorithm discovers a complete set of discriminative patterns with regard to given thresholds or applies heuristic strategies to extract the largest statistically significant discriminative patterns in a given dataset. Experimental results on both synthetic datasets and three real variant datasets (Age-Related Macular Degeneration, Breast Cancer and Type 2 Diabetes) demonstrate that the SSDPS algorithm effectively detects multiple SNPs combinations in an acceptable execution time.

Reference free SNP detection in RAD-seq data

Participants : Jeremy Gauthier, Claire Lemaitre, Pierre Peterlongo.

We developed an original method for reference-free variant calling from Restriction site associated DNA Sequencing (RAD-Seq) data. RAD-seq is a technique widely employed in the evolutionary biology field. Based on the variant caller DiscoSnp, DiscoSnp-RAD explores the De Bruijn Graph built from all the read datasets to detect SNP and Indels. Tested on simulated and real datasets, DiscoSnp-RAD identifies thousands of variants suitable for different population genomics analyses. Furthermore, DiscoSnp-RAD stands out from other tools due to his completely different principle, making it significantly faster, in particular on large datasets [39].

Global Optimization for Scaffolding and Completing Genome Assemblies

Participants : Sebastien Francois, Rumen Andonov, Dominique Lavenier.

We developed a method for solving genome scaffolding as a problem of finding the longest simple path in a graph defined by the contigs that satisfies a maximal number of additional constraints encoding the insert-size information [26]. Then we solved the resulting mixed integer linear program to optimality using the Gurobi solver. We tested our algorithm on a benchmark of chloroplast genomes and showed that it outperforms other widely-used assembly solvers by the accuracy of the results.

Identification and characterization of long non-coding RNA

Participant : Fabrice Legeai.

We participated in the development and validation of the tool FeelNC (collaboration with IGDR group). This is a tool allowing the identification of long non coding RNA (lncRNA) from RNASeq reads with or without a reference genome. Contrary to other tools, it does not depend on the comparison with protein databanks, which usually require lots of computations, but used a machine learning approach based on a Random Forest model trained with general features such as multi k-mer frequencies and relaxed open reading frames. We delivered a module that allows to characterize the relationships of each long non coding RNA with the other genes in its genomics close environment, giving insights about the putative impact of the lncRNAs to the regulation of these genes [23], [24].

Characterizing repeat-associated subgraphs in de Bruijn graphs

Participant : Camille Marchet.

The main problem in genome assembly, namely repeats, is also present in transcriptomic data. They are dealt with using various heuristics in the de Bruijn Graph framework (dBG). In this work, we introduce a formal model for representing high copy-number and low-divergence repeats in RNA-seq data in dBG and infer the definition of repeat-associated subgraphs. We show that the problem of identifying such subgraphs in a dBG is NP-complete. Then we place ourselves in the case of local assembly of alternative splicing and show that such subgraphs can be avoided implicitly. Thus, more alternative splicing events can be enumerated than with previous approaches. Finally we show that this exploration of DBG explorations can improve de novo transcriptome evaluation methods [16].