Iterative Variable Gene Discovery from Whole Genome Sequencing with a Bootstrapped Multiresolution Algorithm

In jawed vertebrates, variable (V) genes code for antigen-binding regions of B and T lymphocyte receptors, which generate a specific response to foreign pathogens. Obtaining the detailed repertoire of these genes across the jawed vertebrate kingdom would help to understand their evolution and function. However, annotations of V-genes are known for only a few model species since their extraction is not amenable to standard gene finding algorithms. Also, the more distant evolution of a taxon is from such model species, and there is less homology between their V-gene sequences. Here, we present an iterative supervised machine learning algorithm that begins by training a small set of known and verified V-gene sequences. The algorithm successively discovers homologous unaligned V-exons from a larger set of whole genome shotgun (WGS) datasets from many taxa. Upon each iteration, newly uncovered V-genes are added to the training set for the next predictions. This iterative learning/discovery process terminates when the number of new sequences discovered is negligible. This process is akin to "online" or reinforcement learning and is proven to be useful for discovering homologous V-genes from successively more distant taxa from the original set. Results are demonstrated for 14 primate WGS datasets and validated against Ensembl annotations. This algorithm is implemented in the Python programming language and is freely available at http://vgenerepertoire.org.