Integer quantized multilayer perceptron neural network for lossless compression of nanopore signal

MOTIVATION: Nanopore sequencing is indispensable for advanced genomic analyses, but the raw signal output creates substantial storage and bandwidth demands. Specialized model-free compressors such as VBZ and ex-zd have been proposed to address this challenge. They deliver high throughput and strong compression performance, yet they cannot fully exploit higher-order statistical dependencies, which limits further gains. Learning-based compressors can achieve better compression ratios, but they often introduce floating-point non-determinism and are too slow for large-scale nanopore datasets, reducing their practicality. RESULTS: We introduce SigComp, a learned lossless compressor for nanopore signal data. SigComp employs a compact multilayer perceptron network with a fixed context window to capture higher-order dependencies within the dominant one-byte signal deltas. By integrating post-training quantization and a fully GPU-native range coder, SigComp eliminates GPU-CPU synchronization overhead in the autoregressive coding loop while ensuring strict bit-exactness across diverse execution settings. Evaluations on diverse species and sequencing chemistries (DNA R10.4.1 and RNA004) demonstrate that SigComp consistently reduces compressed file sizes by 6.00%-7.09% and 3.86%-5.34% over VBZ and ex-zd, respectively, albeit with a 10× lower processing throughput. Compared to the leading neural compressor, SigComp achieves a comparable compression ratio, sacrificing merely 0.4%-1.9%, while delivering 46-74× higher throughput. Ablation studies confirm that SigComp's integer-only, GPU-native approach improves throughput by 7.75-8.30× overall over hybrid GPU-CPU pipelines without model quantization, and the default 64-symbol context window effectively captures the vast majority of exploitable redundancy. SigComp can also be extended to lossy compression by eliminating the least significant bits, and achieve compression ratios of 4.047-4.868, outperforming lossy ex-zd by 9.18%-14.38%. Although processing remains slower than specialized codecs, SigComp offers an excellent trade-off for long-term cold archival storage, particularly when scaled via multi-GPU environments. AVAILABILITY AND IMPLEMENTATION: SigComp is freely available on GitHub at: https://github.com/BioinfoSZU/SigComp. It supports CPU-only, NVIDIA, and AMD GPU execution backends with seamless cross-platform archive interoperability, as well as multi-GPU scaling. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Authors

Institutions

Publication Details

Journal
Bioinformatics
Published
2026-11-12
DOI
https://doi.org/10.1093/bioinformatics/btag835
Primary Topic
Algorithms and Data Compression
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Integer quantized multilayer perceptron neural network for lossless compression of nanopore signal

Yew Soon Ong, Shumin Li, Shaohui Xie, Lulu Ding et al.
Bioinformatics
Algorithms and Data Compression
article

Integer quantized multilayer perceptron neural network for lossless compression of nanopore signal

Yew Soon Ong, Shumin Li, Shaohui Xie, Lulu Ding, Ling Liu, Zexuan Zhu
article en

Abstract

MOTIVATION: Nanopore sequencing is indispensable for advanced genomic analyses, but the raw signal output creates substantial storage and bandwidth demands. Specialized model-free compressors such as VBZ and ex-zd have been proposed to address this challenge. They deliver high throughput and strong compression performance, yet they cannot fully exploit higher-order statistical dependencies, which limits further gains. Learning-based compressors can achieve better compression ratios, but they often introduce floating-point non-determinism and are too slow for large-scale nanopore datasets, reducing their practicality. RESULTS: We introduce SigComp, a learned lossless compressor for nanopore signal data. SigComp employs a compact multilayer perceptron network with a fixed context window to capture higher-order dependencies within the dominant one-byte signal deltas. By integrating post-training quantization and a fully GPU-native range coder, SigComp eliminates GPU-CPU synchronization overhead in the autoregressive coding loop while ensuring strict bit-exactness across diverse execution settings. Evaluations on diverse species and sequencing chemistries (DNA R10.4.1 and RNA004) demonstrate that SigComp consistently reduces compressed file sizes by 6.00%-7.09% and 3.86%-5.34% over VBZ and ex-zd, respectively, albeit with a 10× lower processing throughput. Compared to the leading neural compressor, SigComp achieves a comparable compression ratio, sacrificing merely 0.4%-1.9%, while delivering 46-74× higher throughput. Ablation studies confirm that SigComp's integer-only, GPU-native approach improves throughput by 7.75-8.30× overall over hybrid GPU-CPU pipelines without model quantization, and the default 64-symbol context window effectively captures the vast majority of exploitable redundancy. SigComp can also be extended to lossy compression by eliminating the least significant bits, and achieve compression ratios of 4.047-4.868, outperforming lossy ex-zd by 9.18%-14.38%. Although processing remains slower than specialized codecs, SigComp offers an excellent trade-off for long-term cold archival storage, particularly when scaled via multi-GPU environments. AVAILABILITY AND IMPLEMENTATION: SigComp is freely available on GitHub at: https://github.com/BioinfoSZU/SigComp. It supports CPU-only, NVIDIA, and AMD GPU execution backends with seamless cross-platform archive interoperability, as well as multi-GPU scaling. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Bioinformatics
Xidian University (CN), Nanyang Technological University (SG), Shenzhen University (CN), University Town of Shenzhen (CN), Shenzhen Research Institute of Big Data (CN), Guangzhou Institute of Technology (CN)
Algorithms and Data Compression
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon