Granted Patent B2

US 12,640,235 · App. 17/874,158 · Granted May 26, 2026

Inventors: Kishore Jaganathan (San Francisco, CA); Kai-how Farh (San Mateo, CA); Jeremy Francis McRae (Hayward, CA); Sofia Kyriazopoulou Panagiotopoulou (Redwood City, CA)

Assignee: Illumina, Inc.

G16B40/20G06N3/0464G06N3/048G06N3/084G16B20/00G16B30/00G16B40/00G16B50/00G06F18/24

Abstract

Claims (31)

1 . A system comprising a splice site predictor and at least one processor coupled to memory for running the splice site predictor, the system further comprising:

a convolutional neural network trained on training examples of donor splice sites, acceptor splice sites, and non-splicing sites;

an input stage of the convolutional neural network that feeds an input sequence of nucleotides for evaluation of target nucleotides in the input sequence; and

an output stage of the convolutional neural network that translates convolutions of the input sequence performed by the convolutional neural network into classification scores indicating respective probabilities of each of the target nucleotides being a donor splice site, an acceptor splice site, and a non-splicing site.

2 . The system of claim 1 , wherein the convolutional neural network is parameterized by a number of convolution layers, a number of convolution filters, and a number of subsampling layers.

3 . The system of claim 1 , wherein the convolutional neural network includes one or more fully-connected layers and a terminal classification layer.

4 . The system of claim 1 , wherein the convolutional neural network includes dimensionality altering layers that reshape spatial and feature dimensions of a preceding input.

5 . The system of claim 1 , wherein the convolutional neural network is parameterized by a number of residual blocks, a number of skip connections, and a number of residual connections, wherein each residual block comprises at least one batch normalization layer, at least one rectified linear unit (ReLU) layer, at least one dimensionality altering layer, and at least one residual connection, and wherein each residual block comprises two batch normalization layers, two ReLU non-linearity layers, two dimensionality altering layers, and one residual connection.

6 . The system of claim 1 , wherein the training examples include at least 50,000 training examples of donor sites, at least 50,000 training examples of acceptor sites, and at least 100,000 training examples of non-occurrence sites.

7 . The system of claim 1 , wherein each of the training examples is a target base sequence having at least one target base flanked by at least 20 bases on each side.

8 . The system of claim 1 , wherein the convolutional neural network batch-wise evaluates the training examples during an epoch.

9 . The system of claim 1 , wherein the training examples are randomly sampled into batches, wherein each batch has a predetermined batch size.

10 . The system of claim 1 , wherein the convolutional neural network iterates evaluation of the training examples over 10 epochs.

11 . A system comprising Aan aberrant splicing detector and at least one processor coupled to memory for running the aberrant splicing detector, the system further comprising:

a convolutional neural network trained to classify target nucleotides in an input sequence and assign, based on convolutions of the input sequence, splice site scores indicating respective probabilities of each of the target nucleotides being a donor splice site, an acceptor splice site, and a non-splicing site; and

a classifier configured to process a reference sequence and a variant sequence through convolutions of the convolutional neural network to (i) produce splice site scores indicating respective probabilities of each target nucleotide in the reference sequence and in the variant sequence being a donor splice site, an acceptor splice site, and a non-splicing site, and (ii) determine, based on the splice site scores for the reference sequence and the variant sequence, whether a variant that generated the variant sequence causes aberrant splicing and is therefore pathogenic.

12 . The system of claim 11 , wherein the classifier determines whether the variant is pathogenic based on differences between splice site scores of corresponding target nucleotides in the reference sequence and in the variant sequence.

13 . The system of claim 12 , wherein the differences in the splice site scores are determined position-wise between the corresponding target nucleotides in the reference sequence and the variant sequence.

14 . The system of claim 11 , wherein, for at least one target nucleotide position, a global maximum difference in the splice site scores is above a predetermined threshold, further configured to classify the variant as causing aberrant splicing and therefore pathogenic.

15 . The system of claim 11 , wherein, for at least one target nucleotide position, a global maximum difference in the splice site scores is below a predetermined threshold, further configured to classify the variant as not causing aberrant splicing and therefore benign.

16 . The system of claim 11 , further configured to identify variants that cause autism spectrum disorder.

17 . The system of claim 11 , further configured to identify variants that cause developmental delay disorder.

18 . A computer-implemented method of detecting genomic variants that cause aberrant splicing, comprising:

processing, by at least one processor, a reference sequence through convolutions of a convolutional neural network trained to detect differential splicing patterns in a target sub-sequence of an input sequence by generating classification scores indicating respective probabilities of each nucleotide in the target sub-sequence being a donor splice site, an acceptor splice site, and a non-splicing site;

based on the processing of the reference sequence, detecting a first differential splicing pattern in a reference target sub-sequence by generating classification scores indicating respective probabilities of each nucleotide in the reference target sub-sequence being a donor splice site, an acceptor splice site, and a non-splicing site;

processing, by the at least one processor, a variant sequence through convolutions of the convolutional neural network, wherein the variant sequence and the reference sequence differ by at least one variant nucleotide located in a variant target sub-sequence;

based on the processing of the variant sequence, detecting a second differential splicing pattern in the variant target sub-sequence by generating classification scores indicating respective probabilities of each nucleotide in the variant target sub-sequence being a donor splice site, an acceptor splice site, and a non-splicing site;

determining, by the at least one processor, a difference between the first differential splicing pattern and the second differential splicing pattern by comparing, on a nucleotide-by-nucleotide basis, classification scores of the reference target sub-sequence and the variant target sub-sequence; and

based on the difference being above a predetermined threshold, classifying the at least one variant nucleotide as causing aberrant splicing and therefore pathogenic.

19 . The computer-implemented method of claim 18 , wherein a differential splicing pattern identifies positional distribution of occurrence of splicing events in a target sub-sequence.

20 . The computer-implemented method of claim 19 , wherein the splicing events include at least one of cryptic splicing, exon skipping, mutually exclusive exons, alternative donor site, alternative acceptor site, and intron retention.

Assignments (2)

Continuity (6)

References Cited (202)