Methods and systems for sequence calling
The present disclosure provides methods, systems, and media for accurate and efficient estimation of a genome of a genus.
1. A computer-implemented method for generating a training set, the method comprising:
obtaining a first trained algorithm comprising a first mapping between actual reference sequencing signals and trusted reference sequencing signals, wherein the actual reference sequencing signals and the trusted reference sequencing signals comprise analog signals generated at least in part by sequencing nucleic acids using a high-throughput sequencer, wherein the actual reference sequencing signals and the trusted reference sequencing signals represent parts of a reference genome of a first genus;
obtaining actual sequencing signals corresponding to a second genome of a second genus that differs from the first genus, wherein the reference genome is smaller than the second genome; and
generating the training set for training a second trained algorithm comprising a second mapping between the actual sequencing signals corresponding to the second genome and trusted sequencing signals corresponding to the second genome, wherein the training set is generated based at least in part on applying the first mapping to the actual sequencing signals corresponding to the second genome.
2. The method of claim 1 , wherein the second genome is at least a portion of a human genome.
3. The method of claim 2 , wherein the first trained algorithm comprises a first neural network, and wherein the second trained algorithm comprises a second neural network.
4. The method of claim 3 , further comprising training the second neural network, using the training set, to map the actual sequencing signals corresponding to the second genome to the trusted sequencing signals corresponding to the second genome.
5. The method of claim 3 , wherein generating the training set comprises aligning the actual sequencing signals corresponding to the second genome to the trusted reference sequencing signals.
6. The method of claim 3 , further comprising training the first neural network at least in part by aligning the actual reference sequencing signals to the trusted reference sequencing signals.
7. The method of claim 6 , wherein training the first neural network comprises aligning, using a first alignment process, the actual reference sequencing signals to the trusted reference sequencing signals; wherein generating the training set comprises aligning, using a second alignment process, the actual sequencing signals to the trusted reference sequencing signals; and wherein the first alignment process consumes less resources than the second alignment process.
8. The method of claim 7 , wherein the first alignment process comprises determining correlations between the actual reference sequencing signals and different parts of the trusted reference sequencing signals.
9. The method of claim 7 , wherein the second alignment process comprises using a hash-based search.
10. The method of claim 3 , further comprising training the first neural network at least in part by performing one or more iterations of:
selecting a portion of the actual reference sequencing signals and a portion of the trusted reference sequencing signals associated therewith;
using the first neural network to process the selected portion of the actual reference sequencing signals to produce first neural network output signals;
determining an error that represents a difference between the first neural network output signals and the selected portion of the trusted reference sequencing signals; and
configuring the first neural network by backpropagating the error.
11. The method of claim 10 , wherein the first neural network comprises a regression network.
12. The method of claim 11 , wherein the regression network comprises a fully connected regression network.
13. The method of claim 11 , wherein the regression network comprises an input layer that comprises one neuron per value of the actual reference sequencing signals, and a plurality of intermediate layers that are larger than the input layer.
14. The method of claim 3 , wherein generating the training set comprises truncating the actual sequencing signals and the trusted reference sequencing signals, and aligning the truncated actual sequencing signals to the truncated trusted reference sequencing signals.
15. The method of claim 3 , further comprising using the second neural network to process the actual sequencing signals and additional information of a type that differs from the actual sequencing signals.
16. The method of claim 15 , wherein the additional information comprises information regarding background noise of the actual sequencing signals.
17. The method of claim 15 , wherein the additional information comprises sequencing signals obtained from a preamble of the sequencing.
18. The method of claim 15 , wherein the additional information comprises local information corresponding to a vicinity of the actual sequencing signals.
19. The method of claim 15 , wherein the additional information comprises flow information indicative of at least one of a flow base and a flow position of the actual sequencing signals.
20. A method for estimating, based at least in part on a reference genome of a first genus, a second genome of a second genus, the method comprising:
for each of a plurality of parts of the second genome:
obtaining actual sequencing signals that represent the part of the second genome; and
estimating the part of the second genome based at least in part on the actual sequencing signals;
wherein the estimating comprises applying a machine learning classifier to the actual sequencing signals;
wherein the machine learning classifier is trained to provide a second mapping between the actual sequencing signals and trusted sequencing signals corresponding to the second genome, wherein the actual sequencing signals and the trusted sequencing signals comprise analog signals generated at least in part by sequencing nucleic acids using a high-throughput sequencer;
wherein the second mapping is generated based at least in part on a first mapping between actual reference sequencing signals and trusted reference sequencing signals corresponding to the reference genome; and
wherein the actual reference sequencing signals and the trusted reference sequencing signals represent parts of the reference genome of the first genus that differs from the second genus, wherein the reference genome is smaller than the second genome.