Deep learning-based variant classifier
The technology disclosed directly operates on sequencing data and derives its own feature filters. It processes a plurality of aligned reads that span a target base position. It combines elegant encoding of the reads with a lightweight analysis to produce good recall and precision using lightweight hardware. For instance, one million training examples of target base variant sites with 50 to 100 reads each can be trained on a single GPU card in less than 10 hours with good recall and precision. A single GPU card is desirable because it a computer with a single GPU is inexpensive, almost universally within reach for users looking at genetic data. It is readily available on could-based platforms.
1. A system for a trained variant classifier, the system including:
numerous processors operating in parallel and coupled to memory;
a convolutional neural network running on the numerous processors, trained on at least 50000 training examples of groups of reads spanning candidate variant sites labeled with true variant classifications of the groups of reads using a backpropagation-based gradient update technique that progressively matches outputs of the convolutional neural network with corresponding ground truth labels;
wherein each of the at least 50000 training examples used in the training includes a group of reads aligned to a reference read, each of the reads including a target base position flanked by or padded to at least 110 bases on each side, each of the at least 110 bases in the reads accompanied by
a corresponding reference base in the reference read,
a base call accuracy score of reading the base,
a strandedness of reading the base,
insertion count of changes adjoining a position of the base, and
deletion flag at the position of the base;
an input module of the convolutional neural network which runs on at least one of the numerous processors and feeds the group of reads for evaluation of the target base position; and
an output module of the convolutional neural network which runs on at least one of the numerous processors and translates analysis by the convolutional neural network into classification scores for likelihood that each candidate variant at the target base position is a true variant or a false variant.
2. The system of claim 1 , wherein each of the at least 110 bases in the reads is further accompanied by a mapping quality score of aligning a corresponding read that contains the base to the reference read.
3. The system of claim 1 , wherein the convolutional neural network has one or more convolution layers and one or more fully-connected layers.
4. The system of claim 3 , wherein the convolutional neural network processes the group of reads through the one or more convolution layers and concatenates output of the one or more convolution layers with corresponding empirical variant score (abbreviated EVS) features, and feeds result of the concatenation to the one or more fully-connected layers.
5. The system of claim 4 , wherein each convolution layer has convolution filters and each of the convolution filters has convolution kernels.
6. The system of claim 5 , wherein the convolution filters use depthwise separable convolutions.
7. The system of claim 1 , wherein the convolutional neural network has one or more max pooling layers and one or more batch normalization layers.
8. The system of claim 1 , wherein the convolutional neural network uses a softmax classification layer to produce the classification scores.
9. The system of claim 1 , wherein the convolutional neural network uses dropout.
10. The system of claim 1 , wherein the convolutional neural network uses flattening layers.
11. The system of claim 1 , wherein the convolutional neural network uses concatenation layers.
12. The system of claim 1 , wherein the convolutional neural network runs on a GPU and iterates evaluation of the at least 50000 training examples over five to fifty epochs, with one epoch taking one hour to complete.
13. The system of claim 1 , wherein the convolutional neural network is trained on 1000000 training examples.
14. A method of variant calling, including:
feeding an array of input features to a convolutional neural network and processing the array through the convolutional neural network;
wherein the convolutional neural network runs on numerous processors operating in parallel and coupled to memory, and is trained on at least 50000 training examples of groups of reads spanning candidate variant sites labeled with true variant classifications of the groups of reads using a backpropagation-based gradient update technique that progressively matches outputs of the convolutional neural network with corresponding ground truth labels;
wherein the array encodes a group of reads that are aligned to a reference read and include a target base position flanked by or padded to at least 30 bases on each side;
wherein each input feature in the array corresponds to a base in the reads and has a plurality of dimensions, including
a first dimension set identifying the base,
a second dimension set identifying a reference base aligned to the base,
a third dimension set identifying a base call accuracy score of the base,
a fourth dimension set identifying strandedness of the base,
a fifth dimension set identifying an insertion count of changes adjoining a position of the base, and
a sixth dimension set identifying a deletion flag at the position of the base; and
translating processing of the array by the convolutional neural network into classification scores for likelihood that each input feature at the target base position is a true variant or a false variant.
15. The method of claim 14 , wherein each input feature in the array further includes a seventh dimension set identifying a mapping quality score of aligning a corresponding read that contains the base to the reference read.
16. The method of claim 14 , wherein the convolutional neural network has one or more convolution layers and one or more fully-connected layers.
17. A system for a trained variant classifier, the system including:
numerous processors operating in parallel and coupled to memory;
a fully-connected neural network running on the numerous processors, trained on at least 50000 training examples of empirical variant score (abbreviated EVS) feature sets of candidate variant sites labeled with true variant classifications of the candidate variant sites using a backpropagation-based gradient update technique that progressively matches outputs of the fully-connected neural network with corresponding ground truth labels;
wherein each of the at least 50000 training examples used in the training includes an EVS feature set representing characteristics of a corresponding candidate variant site in a group of reads;
an input module of the fully-connected neural network which runs on at least one of the numerous processors and feeds the EVS feature set for evaluation of a target candidate variant site; and
an output module of the fully-connected neural network which runs on at least one of the numerous processors and translates analysis by the fully-connected neural network into classification scores for likelihood that at least one variant occurring at the target candidate variant site is a true variant or a false variant.
18. The system of claim 17 , wherein the fully-connected neural network has one or more max pooling layers and one or more batch normalization layers.
19. The system of claim 17 , wherein the fully-connected neural network uses dropout.
20. The system of claim 17 , wherein the fully-connected neural network uses a softmax classification layer to produce the classification scores.