Section: Research Program

Axis 2: Algorithms

The main goal of the GenScale team is to develop optimized tools dedicated to NGS processing. Optimization can be seen both in terms of space (low memory footprint) and in terms of time (fast execution time). The first point is mainly related to advanced data structures as presented in the previous section (axis 1). The second point relies on new algorithms and, when possible implementation on parallel structures (axis 3).

We do not have the ambition to cover the vast panel of software related to NGS needs. We particularly focused on the following areas:

• NGS data Compression De-Bruijn graphs are de facto a compressed representation of the NGS information from which very efficient and specific compressors can be designed. Furthermore, compressing the data using smart structures may speed up some downstream graph-based analyses since a graph structure is already built [1].

• Genome assembly This task remains very complicated, especially for large and complex genomes, such as plant genomes with polyploid and highly repeated structures. We worked both on the generation of contigs [3] and on the scaffolding step [26].

• Detection of variants This is often the first information we want to extract from billions of reads. Variant structures range from SNPs or short indels to large insertions/deletions and long inversions over the chromosomes. We developed original methods to find variants without any reference genome [7]. We also worked on the detection of structural variants using approaches of local assembly [6].

• Metagenomics We focussed our research on comparative metagenomics by providing methods able to compare hundreds of metagenomic samples together. This is achieved by combining very low memory data structures and efficient implementation and parallelization on large clusters [2].

• Genome Wide Association Study (GWAS) We tackle this problem with algorithms commonly used in data mining. From two cohorts of individuals (case and control) we can exhibit statistically significant patterns spaning over full genomes.

In addition, we also proposed new algorithmic solutions for analyzing third generation sequencing data, in order to benefit from their larger read size while taking into account their higher sequencing error rate [16].