TL;DR
This paper systematically evaluates how lossy compression of base quality values in RNA-seq data impacts downstream analysis pipelines. While lossy compression offers significant file size reduction, its effects vary greatly depending on the analysis task and the specific compression algorithm used, necessitating careful selection to avoid compromising results.
What Problem Does This Paper Address?
High-throughput sequencing technologies, particularly RNA-seq, generate massive amounts of data. Storing, transmitting, and processing this data efficiently is a growing challenge. Lossy compression, which sacrifices some data fidelity for greater compression ratios, is a promising solution, especially for base quality values that are highly entropic. While lossy compression has been explored for DNA sequencing, its impact on RNA-seq data analysis pipelines remains largely unevaluated. This paper aims to fill that gap by assessing the performance of various lossy compression algorithms across common RNA-seq analysis tasks.
Methods
The researchers evaluated eight different lossy quality value compression algorithms: LEON, DSRC, Quartz, CALQ, P-block, R-block, Crumble, and ScaleQC. They applied these algorithms to raw RNA-seq data, then decompressed the quality values and merged them back with the original sequence data to create new FASTQ or BAM files. These processed files were then fed into standard RNA-seq analysis pipelines, including:
- Alignment: Using STAR and HISAT2.
- Gene Expression Quantification: Using HISAT2 + featureCounts.
- Transcriptome Assembly: Using StringTie and Scallop.
- Short Variants Detection: Using GATK HC.
The performance of these pipelines with lossy compressed data was compared against pipelines using original, uncompressed quality values, which served as the ground truth. Statistical methods, including one-way ANOVA, were used to assess the significance of any observed differences.
Datasets
The study utilized four RNA-seq datasets from different species and sequencing platforms:
- Human cell line: SRR10509596 (HiSeq2000)
- Mouse: SRR10433000 (NovaSeq 6000)
- Rice: SRR8499098 (HiSeq2500)
- Arabidopsis: SRX4122949 (HiSeq X Ten)
Each dataset was split into three technical replicates to allow for robust statistical analysis of compression impacts.
Key Insights and Ideas
- Not all lossy compression is equal: The study highlights that the effectiveness and impact of lossy compression are highly dependent on the specific algorithm and its parameters. Some algorithms offer substantial compression with minimal downstream effects, while others can introduce significant errors.
- Pipeline sensitivity varies: RNA-seq analysis pipelines exhibit different sensitivities to lossy quality compression. Tools that do not rely on base quality values (like STAR for alignment) are generally unaffected, whereas those that do (like HISAT2 for alignment or GATK for variant calling) can be more susceptible.
- Quality values are key to compression: The paper reinforces that base quality values, due to their high entropy, are the primary target for achieving significant data size reduction through compression.
- Species and platform matter: The performance of compression algorithms can vary across different species and sequencing platforms, suggesting that a one-size-fits-all approach may not be optimal.
- Trade-offs are inevitable: There's a clear trade-off between compression ratio and potential impact on analysis results. The "best" algorithm depends on the specific downstream application and its tolerance for error.
Key Results
- Compression Performance: Lossy compression algorithms achieved file size reductions of up to 1.2-3 times compared to lossless methods. Crumble and ScaleQC, in particular, showed strong performance across datasets.
- Alignment: STAR-based alignment was unaffected by lossy compression. HISAT2, which uses quality values, showed minor but noticeable degradation in alignment rates with some lossy compressors, particularly CALQ.
- Gene Expression Quantification: For HISAT2 + featureCounts, higher compression ratios generally led to more differentially expressed (DE) genes. ScaleQC and LEON showed minimal impact on DE gene counts even at high compression levels.
- Transcriptome Assembly: Lossy compression affected isoform reconstruction, with impacts being larger for Scallop than StringTie when using HISAT2 as the aligner. CALQ showed a notably greater difference in reconstructed isoforms.
- Short Variants Detection: Lossy compression generally led to lower F1 scores for SNP and Indel calling, with HISAT2 showing more pronounced effects than STAR. ScaleQC and LEON demonstrated relatively small impacts.
- Computational Performance: Run-time and memory usage varied significantly among compression algorithms. Some, like P-Block and R-Block, were very fast, while CALQ was the slowest. The impact of lossy compression on the run-time and memory consumption of downstream analysis tools was found to be negligible.
Limitations and Future Work
The authors acknowledge that the impact of lossy compression is pipeline-dependent and can be influenced by the specific characteristics of the RNA-seq data (e.g., species, sequencing platform). They also note that some algorithms, like Quartz, are designed with specific reference genomes in mind, limiting their applicability to other species.
Future work could involve developing RNA-seq specific compression algorithms that are optimized for the unique characteristics of this data type. The authors suggest that established RNA-seq data processing pipelines should be considered in the development of future compression tools to ensure their broad applicability. The study provides a valuable framework for evaluating such tools.
Summary updated