SqueezeCall

A nanopore sequencer never observes a base. It records the electrical current disturbance as a strand threads through a protein pore, and because five or so nucleotides occupy the pore at once, every measurement reflects a k-mer rather than a single letter. Turning that noisy trace into a sequence is basecalling, and it sets a ceiling on everything downstream — assembly, variant calling, metagenomic classification. SqueezeCall, from Zhongxu Zhu at Zhejiang Cancer Hospital and the Hangzhou Institute of Medicine, is an end-to-end basecaller that borrows its encoder from automatic speech recognition.

The architectural lineage is deliberate. Basecalling and speech recognition share a shape — a long, noisy one-dimensional signal mapped to a shorter discrete symbol sequence — and successive basecallers have tracked the ASR field, from recurrent networks through self-attention to convolution-augmented transformers. SqueezeCall adopts Squeezeformer, a refinement of Conformer that attacks the temporal redundancy in neighbouring frames by inserting a U-Net-style downsampling stage in the middle of the encoder and recovering resolution at the end. Two ideas from self-supervised speech modelling come with it: masking a fraction of the convolutional feature time steps during training, and combining a CTC-CRF objective with intermediate-layer supervision and a label-smoothed KL term.

The model was published in Gigabyte in 2025 after a January preprint. It differs in purpose from DeepSME, which bootstraps a basecaller for chemically modified DNA that standard models cannot read at all; SqueezeCall targets accuracy on canonical sequencing chemistry.

#Key Features

• Speech-recognition encoder, adapted: The Squeezeformer temporal U-Net halves the sampling rate through the middle blocks and upsamples at the end, spending less compute on redundant adjacent frames than a flat transformer stack.
• Masked feature training: A proportion of convolutional output time steps is replaced by a trained shared vector before the encoder, a regularizer borrowed from Wav2vec 2.0 that improves noise tolerance.
• Three complementary losses: CTC-CRF supervision at the output, the same objective at an intermediate block, and a label-smoothed KL divergence term each contribute independently to accuracy.
• Beam-search CTC decoding: The encoder emits per-position log probabilities over A, G, C, T, and blank, which a width-5 beam search resolves into a read.
• Two model sizes: An 8-layer and a 10-layer configuration trade accuracy against inference cost within one architecture.

#Technical Details

The encoder begins with three 1D convolution blocks — kernel sizes 5, 5, and 19, output channels 4, 16, and 512, each with layer normalization and a GELU activation — which downsample the raw signal and capture local dependencies. The mask module sits between the convolutions and the Squeezeformer stack. SqueezeCall-M uses 8 layers and 79 million parameters, SqueezeCall-L 10 layers and 95 million, against 27 million for Bonito-LSTM and 79 million for Bonito-Transformer. Training slices reads into non-overlapping 3,600-point chunks, uses Adam with a 1,000-step warmup to a peak learning rate of 0.0005, and runs on 4 GPUs at batch size 64. Labels come from Tombo resquiggle alignments, with poorly resquiggled reads discarded.

Evaluation spans 11 datasets: three NA12878 human reference runs, a lambda phage set, and nine bacterial species held out from a 50-genome training collection. SqueezeCall-L reaches the lowest error rate on every dataset — 8.45% on NA12878 against 9.72% for Bonito-Transformer and 9.91% for Bonito-LSTM, and 0.24% on lambda phage against 0.57% and 0.74% — and the highest average read identity at 93.97%, ahead of SqueezeCall-M at 93.50%, Bonito-Transformer at 92.79%, Bonito-LSTM at 92.25%, CATCaller at 91.2%, and SACall at 90.66%. Homopolymer accuracy holds up better as run length grows, reaching 0.76 at length 5. Ablations on the ONT chunk dataset trace the median match rate from 96.37% with CTC-CRF loss alone to 96.64% with intermediate loss and 96.93% with the KL term.

#Applications

Basecalling sits upstream of nearly every nanopore workflow, so accuracy gains propagate: fewer indels in homopolymer runs improve assembly contiguity, reduce spurious variant calls, and sharpen species assignment in environmental and clinical metagenomics. The multi-species evaluation is the practically relevant part — a basecaller trained on one organism and applied to another is the normal deployment case, and the bacterial test genomes were held out from training entirely. The author identifies direct modified-base calling as a natural extension of the same architecture.

#Impact

SqueezeCall is a focused demonstration that architectural transfer from speech recognition continues to pay off in basecalling, and that the gains come from encoder design and objective composition rather than scale — the larger variant is comparable in size to the vendor transformer it outperforms. The practical caveat is availability: the MIT-licensed repository documents a complete training pipeline and links demo data on Zenodo, but no pretrained checkpoint is distributed, so using the model means training it. The evaluation is also DNA-only; although the paper opens on nanopore's direct RNA capability, every reported experiment is on DNA reads.