Split architecture for artificial intelligence-based base caller

1 . A system, comprising:

a sequencing instrument;

at least one processor; and

a non-transitory computer-readable medium comprising instructions that, when executed by the at least one processor, cause the system to:

capture, by one or more light detectors of the sequencing instrument at imaging events of sequencing cycles of a sequencing run, per-cycle sequencing image sets comprising signals from nucleic-acid reaction sites;

process, through a spatial convolution network, a window of per-cycle sequencing image sets for a series of sequencing cycles of the sequencing cycles on a cycle-by-cycle basis by independently convolving intra-cycle data associated with each of the respective per-cycle sequencing image sets within the window of per-cycle sequencing image sets through discrete sequences of spatial convolution layers to generate, for each respective sequencing cycle within the series of sequencing cycles, respective per-cycle spatial feature map sets encoding alternative representations of the signals captured within the per-cycle sequencing image sets;

wherein the respective sequences of spatial convolution layers have respective sequences of spatial convolution filter banks, wherein trained coefficients of spatial convolution filters in spatial convolution filter banks of the respective sequences of spatial convolution filter banks vary between sequences of spatial convolution layers in the respective sequences of spatial convolution layers;

process, through a temporal convolution network, the per-cycle spatial feature map sets on a groupwise basis by convolving on respective overlapping groups of per-cycle spatial feature map sets in the per-cycle spatial feature map sets using respective temporal convolution filter banks of a first temporal convolution layer to generate respective per-group temporal feature map sets for the respective overlapping groups of per-cycle spatial feature map sets;

wherein trained coefficients of temporal convolution filters in the respective temporal convolution filter banks vary between temporal convolution filter banks in the respective temporal convolution filter banks; and

generate, for a respective sequencing cycle and based on the respective per-group temporal feature map sets, a base call prediction of a nucleotide type for corresponding signals captured in a respective per-cycle sequencing image set.

2 . The system of claim 1 , wherein the spatial convolution filters use intra-cycle segregated convolutions.

3 . The system of claim 1 , wherein the temporal convolution filters use inter-cycle combinatory convolutions.

4 . The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to separately convolve, utilizing a compression network, the respective per-cycle spatial feature map sets through respective compression convolution layers to generate respective per-cycle compressed spatial feature map sets for the respective sequencing cycles.

5 . The system of claim 4 , wherein trained coefficients of compression convolution filters in the respective compression convolution layers vary between compression convolution layers in the respective compression convolution layers.

6 . The system of claim 5 , further comprising instructions that, when executed by the at least one processor, cause the system to process, through the temporal convolution network, the per-group temporal feature map sets on the groupwise basis by convolving on respective overlapping groups of per-group temporal feature map sets in the per-group temporal feature map sets using respective temporal convolution filter banks of a second temporal convolution layer to generate respective further per-group temporal feature map sets for the respective overlapping groups of per-group temporal feature map sets.

7 . The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to process, through an output network, a final temporal feature map set generated by a final temporal convolution layer to generate a final output.

8 . The system of claim 7 , further comprising instructions that, when executed by the at least one processor, cause the system to generate, for respective sequencing cycles and based on the final output, the base call prediction and other base call predications of nucleotide types for corresponding signals captured in the respective per-cycle sequencing image sets.

9 . A system, comprising:

a sequencing instrument;

at least one processor; and

a non-transitory computer-readable medium comprising instructions that, when executed by the at least one processor, cause the system to:

capture, by one or more light detectors of the sequencing instrument at imaging events of sequencing cycles of a sequencing run, per-cycle sequencing image sets comprising signals from nucleic-acid reaction sites;

process, through a spatial convolution network, a window of per-cycle sequencing image sets for a series of sequencing cycles of the sequencing cycles on a cycle-by-cycle basis by independently convolving intra-cycle data associated with each of the respective per-cycle sequencing image sets in the window of per-cycle sequencing image sets through discrete sequences of spatial convolution layers to generate, for each respective sequencing cycle within the series of sequencing cycles, respective per-cycle spatial feature map sets encoding alternative representations of the signals captured within the per-cycle sequencing image sets;

process, through a temporal convolution network, the per-cycle spatial feature map sets on a groupwise basis by convolving on respective overlapping groups of per-cycle spatial feature map sets in the per-cycle spatial feature map sets using respective temporal convolution filter banks to generate respective per-group temporal feature map sets for the respective overlapping groups of per-cycle spatial feature map sets;

wherein trained coefficients of temporal convolution filters in the respective temporal convolution filter banks vary between temporal convolution filter banks in the respective temporal convolution filter banks; and

generate, for a respective sequencing cycle and based on the respective per-group temporal feature map sets, a base call prediction of a nucleotide type for corresponding signals captured in a respective per-cycle sequencing image set.

10 . The system of claim 9 , wherein the respective sequences of spatial convolution layers have respective sequences of spatial convolution filter banks, wherein trained coefficients of spatial convolution filters in spatial convolution filter banks of the respective sequences of spatial convolution filter banks are shared between sequences of spatial convolution layers in the respective sequences of spatial convolution layers.

11 . The system of claim 9 , further comprising instructions that, when executed by the at least one processor, cause the system to separately convolve, utilizing a compression network, the respective per-cycle spatial feature map sets through respective compression convolution layers to generate respective per-cycle compressed spatial feature map sets for the respective sequencing cycles, wherein trained coefficients of compression convolution filters in the respective compression convolution layers vary between compression convolution layers in the respective compression convolution layers.

12 . An artificial intelligence-based method of base calling, the method including:

capturing, by one or more light detectors of a sequencing instrument at imaging events of sequencing cycles of a sequencing run, per-cycle sequencing image sets comprising signals from nucleic-acid reaction sites;

processing, through a spatial convolution network, a window of per-cycle sequencing image sets for a series of sequencing cycles of the sequencing cycles on a cycle-by-cycle basis by independently convolving intra-cycle data associated with each of the respective per-cycle sequencing image sets within the window of per-cycle sequencing image sets through discrete sequences of spatial convolution layers, and generating, for each respective sequencing cycle within the series of sequencing cycles, respective per-cycle spatial feature map sets encoding alternative representations of the signals captured within the per-cycle sequencing image sets;

wherein the respective sequences of spatial convolution layers have respective sequences of spatial convolution filter banks, wherein trained coefficients of spatial convolution filters in spatial convolution filter banks of the respective sequences of spatial convolution filter banks vary between sequences of spatial convolution layers in the respective sequences of spatial convolution layers;

processing, through a temporal convolution network, the per-cycle spatial feature map sets on a groupwise basis by convolving on respective overlapping groups of per-cycle spatial feature map sets in the per-cycle spatial feature map sets using respective temporal convolution filter banks of a first temporal convolution layer, and generating respective per-group temporal feature map sets for the respective overlapping groups of per-cycle spatial feature map sets;

wherein trained coefficients of temporal convolution filters in the respective temporal convolution filter banks vary between temporal convolution filter banks in the respective temporal convolution filter banks; and

generating, for a respective sequencing cycle and based on the respective per-group temporal feature map sets, a base call prediction of a nucleotide type for corresponding signals captured in a respective per-cycle sequencing image set.

13 . The artificial intelligence-based method of claim 12 , further including separately convolving the respective per-cycle spatial feature map sets through respective compression convolution layers of a compression network and generating respective per-cycle compressed spatial feature map sets for the respective sequencing cycles.

14 . The artificial intelligence-based method of claim 13 , wherein trained coefficients of compression convolution filters in the respective compression convolution layers vary between compression convolution layers in the respective compression convolution layers.

15 . The artificial intelligence-based method of claim 14 , further including processing, through the temporal convolution network, the per-group temporal feature map sets on the groupwise basis by convolving on respective overlapping groups of per-group temporal feature map sets in the per-group temporal feature map sets using respective temporal convolution filter banks of a second temporal convolution layer, and generating respective further per-group temporal feature map sets for the respective overlapping groups of per-group temporal feature map sets.

16 . The artificial intelligence-based method of claim 12 , further including processing, through an output network, a final temporal feature map set generated by a final temporal convolution layer, and generating a final output.

17 . The artificial intelligence-based method of claim 16 , further including generating, for respective sequencing cycles and based on the final output, the base call prediction and other base call predications of nucleotide types for corresponding signals captured in the respective per-cycle sequencing image sets.

18 . The artificial intelligence-based method of claim 12 , wherein the spatial convolution filters use intra-cycle segregated convolutions.

19 . The artificial intelligence-based method of claim 12 , wherein the temporal convolution filters use inter-cycle combinatory convolutions.

20 . The artificial intelligence-based method of claim 12 , wherein generating the respective per-cycle spatial feature map sets comprises generating a number of spatial feature map sets corresponding to a number of the spatial convolution filters within a spatial convolution layer.