SPLICING SITES CLASSIFICATION BASED ON DEEP LEARNING Russian patent published in 2022 - IPC G16B40/30 

FIELD: biotechnology.

SUBSTANCE: computer-implemented method for prediction of the likelihood of splicing sites in pre-mRNA genomic sequences is described. The method includes: obtaining pre-mRNA genomic sequences by sequencing pre-mRNA transcripts, and training a sparse convolutional neural network (hereinafter – ACNN) based on training examples of pre-mRNA nucleotide sequences, including at least 50,000 training examples of donor splicing sites, at least 50,000 training examples of acceptor splicing sites, and at least 100,000 training examples of sites not related to splicing, and the trained ACNN generates triple assessments to assess the likelihood of that each nucleotide in target nucleotides is a donor splicing site, an acceptor splicing site, or a site not related to splicing. In this case, the specified training includes: input of training examples of nucleotide sequences, encoded by encoding with one active state, wherein each nucleotide sequence contains at least 401 nucleotides, with at least one target nucleotide and a context of at least 200 flanking nucleotides on each side, in the 5’ direction and in the 3’ direction from the target nucleotide; and correction, by reverse distribution, of filter parameters in the specified ACNN to predict the likelihood assessments of that each target nucleotide in the specified nucleotide sequence is a donor splicing site, an acceptor splicing site, or a site not related to splicing; wherein the trained ACNN receives as an input a pre-mRNA nucleotide sequence of at least 401 nucleotides, which is encoded by encoding with one active state, and which includes at least one target nucleotide and a context of at least 200 flanking nucleotides on each side. A system for prediction of the likelihood of splicing sites in pre-mRNA genomic sequences is also described, including one or more processors related to memory, wherein computer commands are loaded into memory, which, when executed on the specified processors, implement actions including: training a sparse convolutional neural network (ACNN) based on training examples of pre-mRNA nucleotide sequences, including at least 50,000 training examples of donor splicing sites, at least 50,000 training examples of acceptor splicing sites, and at least 100,000 training examples of sites not related to splicing, and the trained ACNN generates triple assessments to assess the likelihood of that each nucleotide in target nucleotides is a donor splicing site, an acceptor splicing site, or a site not related to splicing.

EFFECT: invention expands the range of tools for learning (training) deep convolutional neural networks.

41 cl, 59 dwg, 3 tbl