Nanopore base calling on the edge

Abstract

Motivation: MinION is a portable nanopore sequencing device that can be easily operated in the field with features including monitoring of run progress and selective sequencing. To fully exploit these features, real-time base calling is required. Up to date, this has only been achieved at the cost of high computing requirements that pose limitations in terms of hardware availability in common laptops and energy consumption.

Results: We developed a new base caller DeepNano-coral for nanopore sequencing, which is optimized to run on the Coral Edge Tensor Processing Unit, a small USB-attached hardware accelerator. To achieve this goal, we have designed new versions of two key components used in convolutional neural networks for speech recognition and base calling. In our components, we propose a new way of factorization of a full convolution into smaller operations, which decreases memory access operations, memory access being a bottleneck on this device. DeepNano-coral achieves real-time base calling during sequencing with the accuracy slightly better than the fast mode of the Guppy base caller and is extremely energy efficient, using only 10 W of power.

Availability and implementation: https://github.com/fmfi-compbio/coral-basecaller.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Bonito CNN-based architecture. The architecture is composed of high-level blocks depicted as colored rectangles, where the output of the previous block serves as an input of the following block. The neural network is composed of three blocks of type C, five blocks of type B and a Decoder block. The construction of these block types is depicted on the right; each block type is composed of standard building blocks used in deep learning

Fig. 2.

Receptive fields for basic types of convolutions. For each convolution type, the two rows represent input and output data of the convolution. Multiple channels are stacked. The colored value in the output tensor is computed from the values of the same color in the input tensor

Fig. 3.

Comparison of different convolution factorizations. Individual filled squares represent input values convolved with a weight tensor to a single output or intermediate value. Our k-blueprint-separable convolutions introduce a novel combination of basic blocks. (a) $C_{\text{in}}=C_{\text{out}}=128$ and tensor size (4, 1668, 128) and (b) $C_{\text{in}}=C_{\text{out}}=256$ and tensor size (4, 556, 256)

Fig. 4.

Inference time of different convolutions on the Coral device. Note that pointwise corresponds to a full convolution with depth = 1

Fig. 5.

Residual block with depth-to-space compression (right) is a drop-in replacement for a regular Bonito B-type block (left). The first separable convolution is replaced by a compression block, which compresses the input tensor dimensions from (T, C) to $(T/x,Cy)$. Subsequent separable convolutions use depth reduced by a factor of x and channels increased by a factor of y. The final separable convolution is replaced by a decompression block

Fig. 6.

Speed-vers us-accuracy frontier for various architectures and depthwise kernel sizes. Each point represents the tradeoff for a given kernel size