Nanopore quality score resolution can be reduced with little effect on downstream analysis

Martin Rivara-Espasandín, Lucía Balestrazzi, Guillermo Dufort y Álvarez, Idoia Ochoa, G. Seroussi, Pablo Smircich, José Sotelo‐Silveira, Álvaro MartínbioRxiv 2022 · DOI

Abstract

We investigate the effect of quality score information loss on downstream analysis from nanopore sequencing FASTQ files. We polished de novo assemblies for a mock microbial community and a human genome, and we called variants on a human genome. We repeated these experiments using various pipelines, under various coverage level scenarios, and various quality score quantizers. In all cases we found that the quantization of quality scores cause little difference on (or even improves) the results obtained with the original (non-quantized) data. This suggests that the precision that is currently used for nanopore quality scores is unnecessarily high, and motivates the use of lossy compression algorithms for this kind of data. Moreover, we show that even a non-specialized compressor, like gzip, yields large storage space savings after quantization of quality scores.

Introduction

Nanopore sequencing data is notoriously bulky, largely due to high-precision quality scores that are difficult to compress. The standard FASTQ file format stores DNA sequences alongside "quality scores"—estimates of the probability that a base call is an error. Oxford Nanopore Technologies (ONT) currently utilizes a high-resolution scale of 94 possible values. While high precision sounds advantageous, it creates a massive data footprint. Because these scores are noisy and high-entropy, they are difficult to compress losslessly, leading to genomic datasets that scale to terabytes and create storage bottlenecks.

This study asks whether genomic analysis pipelines actually require this level of precision or if the current 94-value scale is simply storing expensive noise.

Methods

The researchers evaluated the impact of quantization—mapping the 94 original values into a much smaller "alphabet." Several quantizers were developed:

  • $Q_2, Q_4, Q_8$: Mapping original values into 2, 4, or 8 discrete levels.
  • $F_{10}$: A constant quantizer setting every quality score to 10.
  • Context-Aware Quantization ($\\langle Q_2, Q_8 \\rangle_\\delta$): A sophisticated approach maintaining higher resolution near repetitive DNA regions (homopolymers).

These quantized datasets were processed through standard bioinformatics pipelines for assembly polishing (Racon, Medaka, HELEN) and variant calling (PEPPER-Margin-DeepVariant). Tests were performed on three datasets: the ZymoBIOMICS Microbial Community Standard, Human Genome Sample HG003 (for variant calling), and Human Genome Sample HG00733 (for human assembly polishing).

Results

Assembly Polishing

For microbial assemblies, the impact of quantization was minimal. Surprisingly, in several instances, quantization actually improved the results, supporting the hypothesis that it may act as a "denoising" step by smoothing out insignificant fluctuations in the raw sequencing data.

Assembly polishing performance (mismatches per 100kb) across different tools and quantization levels.

As shown in the table above, the number of mismatches per 100kb remained remarkably stable across all quantization levels. For example, using the Medaka tool, the $Q_4$ quantized data (174.17 mismatches) actually outperformed the original high-precision data (175.56 mismatches).

Variant Calling

Variant calling performance on human data (HG003) showed similar resilience to information loss. At 90x coverage, the $Q_8$ quantizer resulted in SNP precision and recall metrics that differed from the original data by less than 0.0001.

Variant calling performance metrics for human genome HG003 at various coverage levels and quantization schemes.

A critical finding illustrated in the table above is that coverage trumps precision. High-coverage data (e.g., 90x) that is heavily quantized ($Q_4$ or $Q_2$) consistently yields significantly better F1 scores than lower-coverage data (e.g., 40x) at full original precision. This suggests that researchers can achieve better biological results by prioritizing more sequencing depth over high-precision quality scores.

Storage and Compression

Reducing the entropy of quality scores through quantization leads to massive space savings. Once quantized, standard tools like gzip become highly efficient:

  • $Q_4$ quantization reduced compressed file size by 48.5%.
  • $F_{10}$ quantization saved nearly 70% of storage space, amounting to over 175 GB for a single human genome dataset.

Conclusion

The current 94-value scale for nanopore quality scores is unnecessarily high. Reducing this resolution to as few as 4 or 8 levels has a negligible impact on downstream genomic analysis while drastically reducing storage requirements. Future work may involve retraining machine learning models (like DeepVariant) specifically on quantized data to further optimize performance. These findings motivate the immediate adoption of lossy compression strategies for nanopore sequencing data.

Summary updated

Abstract

Abstract We investigate the effect of quality score information loss on downstream analysis from nanopore sequencing FASTQ files. We polished denovo assemblies for a mock microbial community and a human genome, and we called variants on a human genome. We repeated these experiments using various pipelines, under various coverage level scenarios, and various quality score quantizers. In all cases we found that the quantization of quality scores cause little difference on (or even improves) the results obtained with the original (non-quantized) data. This suggests that the precision that is currently used for nanopore quality scores is unnecessarily high, and motivates the use of lossy compression algorithms for this kind of data. Moreover, we show that even a non-specialized compressor, like gzip, yields large storage space savings after quantization of quality scores.
BibTeX
@misc{rivaraespasandin2022nanopore,
  author = {Martin Rivara-Espasandín and Lucía Balestrazzi and Guillermo Dufort y Álvarez and Idoia Ochoa and G. Seroussi and Pablo Smircich and José Sotelo‐Silveira and Álvaro Martín},
  title = {{Nanopore quality score resolution can be reduced with little effect on downstream analysis}},
  year = {2022},
  doi = {10.1101/2022.03.03.482048},
  howpublished = {bioRxiv},
}
Figure