Deep learning-based aberrant splicing detection
The technology disclosed relates to constructing a convolutional neural network-based classifier for variant classification. In particular, it relates to training a convolutional neural network-based classifier on training data using a backpropagation-based gradient update technique that progressively match outputs of the convolutional neural network-based classifier with corresponding ground truth labels. The convolutional neural network-based classifier comprises groups of residual blocks, each group of residual blocks is parameterized by a number of convolution filters in the residual blocks, a convolution window size of the residual blocks, and an atrous convolution rate of the residual blocks, the size of convolution window varies between groups of residual blocks, the atrous convolution rate varies between groups of residual blocks. The training data includes benign training examples and pathogenic training examples of translated sequence pairs generated from benign variants and pathogenic variants.
1. A system for aberrant splicing determination, including at least one processor coupled to memory, the memory loaded with instructions that, when executed by the at least one processor, cause the system to perform operations comprising:
a trained atrous convolutional neural network, running on the at least one processor, that processes pre-mRNA sequences, including:
an input layer that
receives a variant sequence with target nucleotides flanked on each side by flanking nucleotides; and
accesses a reference sequence corresponding to the variant sequence;
convolutional layers that perform atrous convolutions on the target nucleotides in the variant sequence and corresponding reference nucleotides in the reference sequence, and generate, for each of the target nucleotides and the corresponding reference nucleotides, a triplet splice site score comprising a donor site probability, an acceptor site probability, and a non-splicing site probability; and
an output layer that determines, from position-wise differences in respective triplet splice site scores of the target nucleotides and the corresponding reference nucleotides, whether a variant in the variant sequence causes aberrant splicing at any of the target nucleotides and is therefore pathogenic.
2. The system of claim 1 , wherein the trained atrous convolutional neural network comprises groups of residual blocks.
3. The system of claim 2 , wherein each group of residual blocks is parameterized by a number of convolution filters in residual blocks of a group, a convolution window size of the residual blocks of the group, and an atrous convolution rate of the residual blocks of the group.
4. The system of claim 2 , wherein the trained atrous convolutional neural network is parameterized by a number of residual blocks, a number of skip connections, and a number of residual connections.
5. The system of claim 2 , wherein each group of residual blocks produces an intermediate output by processing a preceding input, wherein dimensionality of the intermediate output is (I−[{(W−1)*D}*A])×N, where:
I is a dimensionality of the preceding input;
W is a convolution window size of residual blocks of a group;
D is an atrous convolution rate of the residual blocks of the group;
A is a number of atrous convolution layers in the group; and
N is a number of convolution filters in the residual blocks of the group.
6. The system of claim 5 , wherein the atrous convolution rate progresses non-exponentially from a lower residual block group to a higher residual block group.
7. The system of claim 6 , wherein atrous convolutions conserve partial convolution calculations for reuse as adjacent bases are processed.
8. The system of claim 7 , wherein atrous convolutions conserve partial convolution calculations for reuse as adjacent bases are processed.
9. The system of claim 1 , wherein the position-wise differences in the respective triplet splice site scores are determined position-wise between the target nucleotides in the variant sequence and the corresponding reference nucleotides in the reference sequence.
10. The system of claim 1 , wherein dimensionality of input received at the input layer is (C u +L+C d )×4, where:
C u is a number of upstream context bases;
C d is a number of downstream context bases; and
L is a number of bases in a target base sequence.
11. The system of claim 1 , wherein the trained atrous convolutional neural network comprises dimensionality altering convolution layers that reshape spatial and feature dimensions of a preceding input.
12. The system of claim 2 , wherein each residual block comprises at least one batch normalization layer, at least one rectified linear unit (abbreviated ReLU) layer, at least one atrous convolution layer, and at least one residual connection.
13. The system of claim 12 , wherein each residual block comprises two batch normalization layers, two ReLU non-linearity layers, two atrous convolution layers, and one residual connection.
14. The system of claim 1 , wherein the trained atrous convolutional neural network batch-wise evaluates training examples during an epoch.
15. The system of claim 14 , wherein the training examples are randomly sampled into batches, wherein each batch has a predetermined batch size.
16. The system of claim 14 , wherein the trained atrous convolutional neural network iterates evaluation of the training examples over ten epochs.
17. A computer-implemented method, including:
receiving a variant sequence with target nucleotides flanked on each side by flanking nucleotides;
accessing a reference sequence corresponding to the variant sequence;
processing the target nucleotides in the variant sequence and corresponding reference nucleotides in the reference sequence through a trained atrous convolutional neural network to generate, for each of the target nucleotides and the corresponding reference nucleotides, a triplet splice site score comprising a donor site probability, an acceptor site probability, and a non-splicing site probability; and
determining, from position-wise differences in respective triplet splice site scores of the target nucleotides and the corresponding reference nucleotides, whether a variant in the variant sequence causes aberrant splicing at any of the target nucleotides and is therefore pathogenic.
18. The computer-implemented method of claim 17 , wherein the trained atrous convolutional neural network comprises groups of residual blocks.
19. A non-transitory computer readable medium storing executable instructions, that, when executed by at least one processor, cause the at least one processor to perform operations comprising:
receiving a variant sequence with target nucleotides flanked on each side by flanking nucleotides;
accessing a reference sequence corresponding to the variant sequence;
processing the target nucleotides in the variant sequence and corresponding reference nucleotides in the reference sequence through a trained atrous convolutional neural network to generate, for each of the target nucleotides and the corresponding reference nucleotides, a triplet splice site score comprising a donor site probability, an acceptor site probability, or a non-splicing site probability; and
determining, from position-wise differences in respective triplet splice site scores of the target nucleotides and the corresponding reference nucleotides, whether a variant in the variant sequence causes aberrant splicing at any of the target nucleotides and is therefore pathogenic.
20. The non-transitory computer readable medium of claim 19 , wherein the trained atrous convolutional neural network comprises groups of residual blocks.