Multi-proteins similarity-based sampling to select representative genomes from large databases.

Background: Genome sequence databases are growing exponentially, but with high redundancy and uneven data quality. For these reasons, selecting representative subsets of genomes is an essential step for almost all studies. However, most current sampling approaches are biased and unable to process large datasets in a reasonable time.

Methods: Here we present MPS-Sampling (Multiple-Protein Similarity-based Sampling), a fast, scalable, and efficient method for selecting reliable and representative samples of genomes from very large datasets. Using families of homologous proteins as input, MPS-Sampling delineates homogeneous groups of genomes through two successive clustering steps. Representative genomes are then selected within these groups according to predefined or user-defined priority criteria.

Results: MPS-Sampling was applied to a dataset of 48 ribosomal protein families from 178,203 bacterial genomes to generate representative genome sets of various size, corresponding to a sampling of 32.17% down to 0.3% of the complete dataset. An in-depth analysis shows that the selected genomes are both taxonomically and phylogenetically representative of the complete dataset, demonstrating the relevance of the approach.

Conclusion: MPS-Sampling provides an efficient, fast and scalable way to sample large collections of genomes in an acceptable computational time. MPS-Sampling does not rely on taxonomic information and does not require the inference of phylogenetic trees, thus avoiding the biases inherent in these approaches. As such, MPS-Sampling meets the needs of a growing number of users.