Comprehensive guide for epigenetics and transcriptomics data quality control

Summary

Host response to environmental exposures such as pathogens and chemicals can include modifications to the epigenome and transcriptome. Improved signature discovery, including the identification of the agent and timing of exposure, has been enabled by advancements in assaying techniques to detect RNA expression, DNA base modifications, histone modifications, and chromatin accessibility. The interrogation of the epigenome and transcriptome cascade requires analyzing disparate datasets from multiple assay types, often at single-cell resolution, derived from the same biospecimen. However, there remains a paucity of rigorous quality control standards of those datasets that reflect quality assurance of the underlying assay. This guide outlines a comprehensive suite of metrics that can be used to ensure quality from 11 different epigenetics and transcriptomics assays. Recommended mitigative actions to address failed metrics are provided. The workflow presented aims to improve benchwork protocols and dataset quality to enable accurate discovery of exposure signatures.

Early identification of low-quality datasets enables (1) avoidance of low-quality data into downstream analytics, (2) monitoring of experimental assays for early identification of possible issues, and (3) assay selections samples with limited cells or DNA available.

Introduction

An epigenome consists of the current state of the chemical modifications to DNA and to the histone proteins that determine the packaging of the DNA. The main epigenetic modification to DNA is the methylation of cytosine to 5-methylcytosine, which almost exclusively occurs on the cytosine of a CpG dinucleotide (i.e., when a cytosine follows a guanine). DNA methylation particularly at gene promotors can inhibit the binding of transcription factors that can initiate transcription.¹ DNA methylation is an epigenetic mark that is involved in regulating several cellular processes including gene transcription, genomic imprinting, and X chromosome inactivation.² DNA methylomes, the representation of methyl groups on a genome’s 5′-CpG-3′ dinucleotides sites added by methyltransferases, can serve as maps to environmental changes such as exposure to chemicals or pathogens to inform epigenetic signature development. Another main epigenetic modification is the post-translational modification (PTM) of histones, commonly the acetylation or methylation of lysine residues. PTMs to histones are able to dictate the accessibility of DNA to transcription initiating proteins by influencing the association of positively charged histones to negatively charged DNA. The addition of a negatively charged functional group (i.e., acetyl) neutralizes the histone’s positive charge and weakens the protein’s association with DNA, opening up access for transcriptional machinery to bind to regulatory regions and initiate transcription.³ The transcriptome represents all coding and non-coding cellular ribonucleic acid (RNA) transcripts. Transcriptional regulation is often controlled by non-coding RNA (ncRNA), whereby ncRNA transcripts will bind to complementary messenger RNA (mRNA) to repress expression.⁴

Some of these epigenetic and transcriptomic alterations can be transitory while other changes can persist for years to decades.⁵ Epigenetic exposure modifications, if distinctive, may provide potential patterns that can be combined together into exposure signatures.^6,7,8,9 These epigenetic signatures may be gene and cell-type specific.^6,7,8,9 Determining if an exposure signature is distinct can be difficult as exposures may only imprint subtle changes to the epigenome or transcriptome. These changes can be further masked if samples are analyzed in a pool containing poor quality samples. Without in-depth quality control (QC), features extracted from a dataset may be the result of experimental conditions (e.g., batch effects) in the sample preparation process rather than the result of an exposure. For example, if partial sample degradation occurs at any point of the sample workflow, amplification for library construction will only occur on the non-degraded sequences, resulting in a false representation of the original sample. Any sample processing step can leave compounding effects due to the potential of PCR amplification bias: the varying efficiency with which PCR amplifies sequences causing the end product after amplification to not accurately reflect the input DNA. Amplification bias often presents as the preferential amplification of sequences with high GC content and would be further exacerbated after any sample degradation.¹⁰ As library construction for most epigenomic and transcriptomic sequencing requires PCR, just slightly mishandled sample prep steps can exponentially skew sequencing results. Because QC metrics are a result of the characterization of a sequence, they can hold insights into which sample preparation steps are potentially introducing these biases. Thus, leveraging QC metrics is a powerful tool to develop accurate and reproducible sample preparation protocols for epigenomic and transcriptomic datasets.

QC of epigenomics and transcriptomics datasets is often an overlooked step. Common bioinformatics workflows include bulked preprocessing steps for datasets that remove outliers and low-quality reads based on their deviation from all samples of a dataset opposed to QC assessment at the sample level prior to downstream analysis. Omitting poor quality samples prior to downstream analysis is essential to distinguish truly expressed genes from artifacts, variations introduced by non-biological processes. If a transcriptomic dataset is analyzed with a mix of poor and high quality samples, aberrant or differential gene expression could be removed as an outlier in the bulk preprocessing step, resulting in the loss or weakening of the signature.¹¹ QC sample assessment at the individual level can also elucidate possible batch effects. Some common causes of batch effects could be a subset of samples relating to the sample source, sampling method, or storage conditions.

In this article, QC metrics for 11 epigenetics and transcriptomics assays from individual peripheral blood mononuclear cells (PBMCs) human samples are assessed. Herein, we review the potential cause for when reported metrics fall outside passing parameters and propose the best courses of mitigation.

The 11 assays selected include a wide variety of transcriptomics and epigenomics techniques including assay for transposase-accessible chromatin using sequencing (ATAC-seq),¹² single-cell ATAC-seq (scATAC-seq),¹³ ChIPmentation,¹⁴ Infinium 850K MethylationEPIC BeadChip,¹⁵ methylated DNA immunoprecipitation sequencing (MeDIPseq,¹⁶ multiplexed indexed T7 (Mint) chromatin immunoprecipitation (ChIP) sequencing (Mint-ChIP-seq),¹⁷ micro ribonucleic acid (RNA) sequencing (miRNAseq),¹⁸ 10X Genomics chromium single-cell multiome (scMultiome) scATAC-seq and single-cell RNA sequencing (scRNA-seq), bulk RNA sequencing (RNA-seq),¹⁹ scRNA-seq²⁰ or singular nuclear RNA-seq (snRNA-seq), and single-nucleus methylcytosine sequencing (snmC-seq) and (snmC-seq2)^21,22 (see Figure 1).

Figure 1.

Schematic overview of workflow

All epigenetic and transcriptomic signatures developed from PBMC samples from high quality sequencing. Different assays inferred different aspects of exposure. Epigenetics and transcriptomics assays/methods assess different aspects of gene regulation and expression, respectively.

(A) Chromatin accessibility is measured by ATAC-seq and scATAC-seq (sometimes referred to as snATAC-seq for single nucleus).

(B) Histone modifications are measured by MINT-ChIP-seq or ChIPmentation.

(C) Cytosine (C) base methylation assessed by snmC-seq2, MeDIPseq, or EPIC.

(D) RNA/transcript sequencing by RNA-seq, scRNA-seq, or miRNAseq. scMultiome assesses both scRNA-seq and scATAC-seq from the same cells.

Figure 1 depicts the schematic workflow throughout this study that aimed to develop exposure-specific signatures. Host response characterization to various exposures was achieved through different assays. Epigenomic assays reveal the underlying mechanisms controlling genomic expression while transcriptomic assays quantify actively expressed genes. Different assays were able to infer different aspects of exposure. RNA sequencing was able to identify pathogen exposure. Assays interrogating methylation patterns were able to predict time since exposure.²³ ATAC sequencing revealed changes in chromatin accessibility to be associated with symptom severity. Combining epigenomics and transcriptomic assays with multiome sequencing revealed gene regulatory networks and circuitry associated with host response.²⁴ Previously, QC metrics were determined for all assays at a good/pass/fail categorization. Singularity pipelines were developed for each assay type to output assay-specific QC metrics for each individual sample.²⁵ These QC metrics cover a variety of attributes relating to sample quality such as sequencing depth, fraction of uniquely mapped reads, percent of aligned reads, etc. Table 1 shows the metrics outputted for each sample’s assay from these pipelines . Single-cell assays include further granularity in QC, including metrics such as number of cells and median unique molecular identifiers (UMI) per cell. Table 1 highlights the most common and relevant QC metrics for each assay; however, additional QC metrics not encompassed in Table 1 are discussed in each section for each assay. Benchwork protocols used for assays are reviewed alongside sample QC checks that can be done throughout the wet lab sample preparation process. General QC metrics that apply to multiple assays are assessed. Further, assay-specific metrics and mitigations are reviewed.

Table 1.

Quality metrics for different assays

Assay	Metric	Below threshold	Pass threshold	High quality	Mitigations
ATAC-seq	sequencing depth	<25M	≥25M		remove sources of sample degradation; repeat library preparation
	percent of aligned reads	<50%	[50%, 75%]	≥75%	repeat nuclei extraction
	non-duplicate reads	<15M	[15M, 40M]	≥40M	repeat library preparation; increase initial cell input; concentrate sample for sequence input
	overlap FRIP (fraction of reads in peaks)	<0.05	[0.05, 0.1]	≥0.1	repeat transposition step; ensure cell viability as a population of cells with highly accessible DNA (e.g., activated granulocytes, dead or dying cells, unsupported organisms)
	nucleosome-free region (NFR) peak detected	no	yes		repeat library preparation; quantify library again to ensure fragment size distribution; remove sources of sample degradation; repeat initial sample preparation
	mononucleosomal peak detected	no	yes		repeat initial sample preparation; quantify library again to ensure fragment size distribution; check library has 200 bp peak
	TSS enrichment	<4	[4, 6]	≥6	consider pre-treating cells with DNase or using flow cytometry to sort viable cells; low score may indicate poor sample preparation, poor sample quality, a population of cells with highly accessible DNA (e.g., activated granulocytes, dead or dying cells, unsupported organisms)
ChIPmentation	uniquely mapped	<60%	[60%, 80%]	≥80%	increase initial cell numbers
	sequence length	<50 bp	≥50 bp		remove sources of sample degradation
MethylationEPIC	Percentage of failed probes	>10%	[10%, 1%]	≤1%	ensure optimal amount of input DNA for the bisulfite conversion kit; optimize PCR conditions in whole genome isothermal amplification
	beta value distribution	>2 peaks	2 peaks		remove unreliable probes; remove sources of background contamination
	percent of CpG methylation	<20% or >80%	[20%, 80%]		repeat bisulfite conversation and whole gnome amplification
MeDIP-seq	sequencing depth	<30M	≥30M		repeat DNA extraction and purification; repeat immunoprecipitation
	CpG coverage	<40%	[40%, 60%]	≥60%	if caused by nonspecific binding, consider magnetic beads for immunoprecipitation instead of agarose beads; ailiconized tubes can also mitigate the non-specific binding of DNA to tube walls; alter antibody and DNA incubation time
	sequence length	<50	≥50		remove sources of sample degradation; repeat DNA extraction and purification; repeat immunoprecipitation
MINT-ChIP	uniquely mapped	<60%	[60%, 80%]	≥80%	remove sources of sample degradation
	uniquely mapped reads	<2M	[2M, 3M]	≥3M	remove sources of sample degradation; remove sources of sample contamination
	reads in peaks	<35	[35, 50]	≥50	incorporate total Histone (H3) as control
	useful reads efficiency	<70%	≥70%		repeat IP reaction and amplification
	PCR duplicates	≥50%	<50%		repeat IP reaction and amplification
	sequence length	<50 bp	≥50 bp		remove sources of sample degradation
miRNA-seq	sequencing depth	<4M	[4M, 8M]	≥8M	increase starting material; remove sources of RNA contamination
	aligned reads	<2M	≥2M		repeat library construction; ensure proper adaptor ligation in library construction; use adapters with random sequences so that there is a greater probability of adapter ligating to a miRNA
	sequence length	<17 bp	≥17 bp		remove sources of RNA contamination; repeat size exclusion steps
Multiome	number of cells	<500	[500, 1,000]	≥1,000	check nuclei for proper quantification; ensure proper nuclei handling; remove background RNA and DNA
	median UMI per cell	<500	[500, 1,500]	≥1,500	possible improper emulsification, partially or nonhomogeneous GEM emulsification reactions can occur, which can be detected by visually inspecting pipette tips throughout steps; if this occurs, re-preparing the sample is the recommendation from 10x Genomics
	reads mapped to exons	<40%	[40%, 60%]	≥60%	remove sources of RNA degradation or contamination
	transcriptomic reads in cell	<40%	[40%, 70%]	≥70%	remove sources of RNA degradation or contamination; repeat sample preparation; repeat library preparation; if transcriptomic reads low due to a population of nuclei with low RNA content, rerun pipeline with appropriate cell calling parameters
	cells with mitochondrial reads	>50%	[20%, 50%]	≥20%	repeat tissue extraction avoiding necrotic cells; repeat cell isolation
	mapped reads	<50%	[50%, 75%]	≥75%	repeat library construction
	fragments in peaks	<5%	[5%, 10%]	≥10%	ensure nuclei integrity; repeat transposition step; repeat emulsification; repeat sample extraction as low values could be from population of cells with un-compacted DNA (activated granulocytes, dead or dying cells)
	GEX sequence length	<50 bp	≥50 bp		remove sources of RNA degradation or contamination; repeat library preparation
	ATAC sequence length	<37 bp	≥37 bp		remove sources of DNA degradation or contamination; repeat library preparation
RNA-seq	sequencing depth	<20M	[20M, 50M]	≥50M	increase sample input for sequencing
	GC%	<20% or >80%	[20%, 40%] or [60%, 80%]	[40%, 60%]	optimize PCR parameters for cDNA library construction
	percent of rRNA	>20%	[10%, 20%]	≤10%	enrichment for mRNA with polyA selection methods; rRNA depletion methods
	percent of globin RNA	>30%	[15%, 30%]	≤15%	consider globin depletion methods; repeat PBMC extraction ensuring exclusion of RBCs
	mRNA reads	<16M	[16M, 30M]	≥30M	enrichment for mRNA with polyA selection methods; consider rRNA removal kit; consider RRNA removal kit
	percent of uniquely mapped	<40%	[40%, 75%]	≥75%	increase library diversity by increasing/improving quality RNA input for cDNA library; optimize PCR parameters for cDNA library construction; remove sources of RNA contamination
	sequence length	<50 bp	≥50 bp		remove sources of RNA degradation or contamination
scATAC-seq	mapped reads	<50%	[50%, 75%]	≥75%	repeat library construction; assess lysis efficiency
	fragments in peaks	<5%	[5%, 10%]	≥10%	repeat transposition step; repeat sample extraction as low values could be from population of cells with un-compacted DNA (activated granulocytes, dead or dying cells); ensure nuclei integrity; assess lysis efficiency
	sequence length	<37 bp	≥37 bp		remove sources of DNA degradation or contamination; repeat library preparation
scRNA-seq	number of cells	<500	[500, 1,000]	≥1,000	repeat cell preparation; optimize cell preservation methods; repeat library construction
	median UMI per cell	<500	[500, 1,500]	≥1500	increase input material for cDNA synthesis; if sample quality is strong, increasing the PCR cycles by two extra cycles during cDNA amplification can bring transcripts above background levels
	reads mapped to exons	<40%	[40%, 60%]	≥60%	remove sources of RNA degradation or contamination; remove sources of cDNA degradation
	transcriptomic reads in cell	<40%	[40%, 70%]	≥70%	remove sources of RNA degradation or contamination; remove sources of cDNA degradation
	cells with mitochondrial reads	>50%	[20%, 50%]	≥20%	repeat tissue extraction avoiding necrotic cells; repeat cell isolation; implement measures to ensure cellular quality and prevent cell lysis
	sequence length	<50 bp	≥50 bp		remove sources of RNA degradation or contamination
snmC-seq	percent of genome	<1%	[1%, 5%]	≥5%	ensure sample quality
	CCC methylation rate	>3%	[2%, 3%]	≥2%	repeat bisulfite conversion and subsequent steps
	percent of uniquely mapped	<40%	[40%, 70%]	≥70	increase input material
	reads per cell	<10K	[10K, 1M]	≥1M	ensure sample quality
	cell level mCG rate	<0.5	[0.5, 0.7]	≥0.7	repeat bisulfite conversion and subsequent steps
	cell level mCH rate	≥0.08	<0.08		repeat bisulfite conversion and subsequent steps
	sequence length	<30 bp	≥30 bp		remove sources of sample degradation

Bounds are given for failing, passing, and excellent, with suggested mitigating procedures to improve quality. Adapted with permission from Ricke et al.²⁵

Methods

Wet lab methods

Bulk RNA-seq

The RNA content in both prokaryotic and eukaryotic cell consists of 80%–90% rRNA, 10%–15% transfer RNA (tRNA), and 3%–7% messenger RNA (mRNA) and regulatory ncRNA.²⁶ RNA sequencing (RNA-seq) is a next-generation sequencing (NGS) method that aims to quantify the mRNA transcripts in all cells from a given sample. Quantifying the transcriptome can reveal splice variants/isoforms, differentially expressed genes, novel transcripts, and gene fusions that can characterize phenotypic variation in both disease etiology and epigenetic responses.²⁷ When harmonized with epigenomic analysis, transcriptomic characterization can validate proposed epigenetic effects.

For RNA-seq, RNA was extracted from cells (1,000 to 100,000 cells) (Direct-zol RNA MiniPrep kit). RNA concentration was assessed with fluorimetry using RiboGreen and RNA quality/integrity (RIN) and by capillary gel electrophoresis using a Fragment Analyzer. At this stage, poor-quality RNA samples, which either have a concentration below 8 ng/ul or a RIN below 7, are excluded as they may cause artifacts. If RIN is below 7, repeating the RNA extraction may be necessary. For samples meeting RNA quality requirements, complimentary DNA (cDNA) libraries were prepared using the universal Plus mRNA-Seq with NuQuant kit (Tecan Genomics) from poly(A) selected RNA. Sequencing library construction involved reverse transcription to generate cDNA with random and poly(T) primers followed by the addition of sequencing adaptors to each end of the cDNA fragment.²⁸ Library concentration and the quality and extent of RNA degradation were evaluated by fluorimetry using Qubit and by capillary gel electrophoresis using Fragment Analyzer, respectively. Initially, shallow sequencing of the pooled libraries using a Mi-Seq system (Illumina) allows the confirmation of library quality and adjustment of the pools for deep sequencing. Deep sequencing of the libraries for up to 30 million reads per sample was performed on a Novaseq sequencer (Illumina).

scRNA-seq

Single-cell RNA sequencing (scRNA-seq) is used to quantify coding and noncoding RNA transcripts of individual cells. scRNA-seq reveals the transcriptional cellular heterogeneity that is masked by bulk RNA-seq methods. This delineation can refine an exposure-based signature to be cell or tissue-type specific.

Coupling microfluidic and nanodroplet-based systems with NGS enables the genome-wide expression profiling of thousands of cells.²⁹ The scRNA-seq samples were characterized with the 10x Chromium Single Cell 3′ Gene Expression assay kit. This method uses a microfluidic device to encapsulate single cells in a droplet that holds reverse transcription reagents. In each droplet, the cell is lysed and a micro-reaction of reverse transcription occurs on polyadenylated RNAs, creating cDNA with capture sequences, UMI (unique molecular identifier) to distinguish separate RNA transcripts, and a unique cell barcode causing all cDNAs from the same cell to have the same barcode. The emulsion is then broken and pooled cDNAs undergo PCR amplification. The quality of the amplified cDNA was assessed with microcapillary-based electrophoresis using a Bioanalyzer. If low-quality RNA is indicated at this point, the sample may have initially been of poor sample quality (<80% cell viability) or the RNA may have been degraded and new cell preparation or nuclei extraction is likely the best solution. Initial cell viability can be assessed by flow cytometry. Clogs or wetting failures in the microfluidic device can also lead to low-quality RNA. cDNA is used for library construction. Adapter and sample indices are incorporated into libraries for next-generation short read sequence compatibility. Libraries were checked on a Bioanalyzer for a distribution of peaks between 300–-1000 base pair (bp). If a significant portion of library falls under the integrals of peaks around and below 200 bp, this could be an indication of carryover of adapter/primer dimers for which size selection cleanup is recommended.

ATAC-seq

Assay for transposase-accessible chromatin sequencing (ATAC-seq) is a fast and high-throughput method that has been widely used to profile chromatin accessibility across the genome.^13,30,31,32 ATAC-seq identifies open chromatin regions (OCRs) by sequencing the 20–90 bp sequences that connect nucleosomes. These open chromatin regions are often regulatory elements like promoters, enhancers, insulators, and silencers that transcription factors bind to influence and control gene expression.^33,34 ATAC-seq offers valuable insight as most RNA polymerase activity to initiate transcription co-occurs with OCRs.³⁵ By coupling the same biological sample with ATAC-seq and downstream sequencing applications, we inferred underlying gene regulatory mechanisms by correlating differential chromatin accessibility with differential gene expression.

Nuclei extraction was performed, using an input of 5,000–100,000 cells. Nuclei were lysed following standard protocols, and the genomic DNA was exposed to Tn5, a highly active transposase that targets open chromatin sites and simultaneously cuts the ends of open chromatin regions and ligates sequencing adaptors to them to create fragments that can be PCR amplified. Library preparation was done with Nextera DNA Flex Library Prep. Amplified libraries were purified. Next, library quality was assessed by microcapillary-based electrophoresis using a Bioanalyzer to ensure that their profiles exhibited the expected fragment size distribution characteristic to nucleosomal-free region (NFR) lengths. Fragment size distribution of the amplified libraries is expected to show a characteristic nucleosomal laddering pattern with a periodicity of 200 bp, corresponding to DNA fragments from nucleosome-free regions (NFRs). High-quality libraries were subjected to paired-end sequencing on a Nova-Seq instrument.

scATAC-seq and snATAC-seq

Single-nucleus ATAC sequencing (snATAC-seq) and single-cell ATAC sequencing (scATAC-seq) generate profiles of chromatin accessibility at a single-cell resolution.³⁶ Further delineating the OCRs of a sample to single-cell granularity enables a more comprehensive view of a sample’s regulatory landscape.³⁷

Similar to ATAC-seq, scATAC-seq nuclei are isolated and treated with a transposase to cut open chromatin sites and ligate barcodes to both ends of the sequence for each individual nuclei. As for scMultiome described below, a microfluidic chip is then used to generate Gel-in-Emulsion droplets that isolated individual nuclei into individual droplets in the presence of a barcoded bead to tag individual nuclei. The 10x Genomics protocol used for scATAC is very similar after that stage to any 10x Genomics protocols. Next, the emulsion is broken and the sample is processed for library construction before being QCed and sequenced.

scMultiome

Single-cell multiome (scMultiome) sequencing pairs scATAC-seq and scRNA-seq to simultaneously generate a cell-type-specific chromatin accessibility and transcriptional profiles in the same cells.³⁸ These concurrently made profiles can identify the variations in chromatin accessibility for the regulatory regions of a certain gene that result in the same transcriptional phenotype. The ability to take multi-omic measurements in parallel allows scMultiome to provide unambiguous inferences on genetic heterogeneity.^39,40,41

The 10x Genomics Chromium Next GEM Multiome ATAC GEX (Gene Expression) kit was used to generate single-cell libraries. First, nuclei are isolated and are bulk treated with Tn5 to generate chromatin-accessible fragments. A microfluidic chip is then used to partition nuclei into individual emulsion droplets where UMI barcodes are added to the transposed DNA and mRNA. Droplets are pooled for preamplification PCR to ensure maximum recovery of barcoded transposed DNA and cDNA fragments. This PCR product is then used as input for ATAC library construction and cDNA amplification for gene expression library construction. Libraries were sequenced on a NovaSeq instrument.

MethylationEPIC

The cell-free methylome can be determined by methylationEPIC analysis by inputted cell-free DNA (cfDNA) to detect abnormalities in methylation patterns indicative of cancer.⁴² MethylationEPIC analysis via the Infinium MethylationEPIC v2.0 BeadChip provides quantitative genome-wide CpG methylation screening at a single-nucleotide resolution for over 850,000 CpG sites.¹⁵ Probes are only designed to interrogate known CpG sites that have been demonstrated to be enhancers and regions associated with tumors.⁴³

Input DNA first underwent bisulfite conversion in a thermocycling process causing unmethylated cytosine bases to deaminate to uracil; methyl-cytosine bases remain as cytosine. Bisulfite-converted DNA then underwent whole genome isothermal amplification and were enzymatically fragmented. DNA was purified and added to bead-bound probes where the DNA hybridizes to the CpG-locus-specific oligomers on individual beads. Bead oligomers will either correspond to the CpG locus of the methylated (C) or the unmethylated (T) state. Next, allele-specific single base extension of the probes was conducted with either biotin-labeled C and G nucleotides or dinitrophenyl-labeled A and T nucleotides. Nucleotide labels were incorporated to improve signal-to-noise ratios. The array was fluorescently stained and scanned and the intensities of the unmethylated and methylated bead types at each locus were measured to determine the relative degree of locus methylation.

Mint-ChIP-seq

Mint-ChIP-seq is a high-throughput method for genome-wide mapping of nucleosomes with specific histone modifications, such as H3K27ac and H3K4me3; H3K27ac is an acetyl-modified lysine residue of histone H3 and H3K4me3 is three methyl residues added to lysine residue 4 of histone H3. In this method, cross-linked chromatin is first pulled down using an antibody against the histone modification of interest and the corresponding genomic region is sequenced.¹⁷ These histone modifications can alter chromatin structure and influence the accessibility of genomic regions to regulatory proteins and transcription factors. Histone-specific signatures from Mint-ChIP have been used in characterizing disease mechanisms.⁴⁴

Traditional ChIP-seq assays require a large number of input cells for sample preparation.⁴⁵ Mint-ChIP-seq uses a method that simultaneously fragments and digests chromatin using micrococcal nuclease (MNase) from a small number of cells that have been optimized for minimizing sample loss and conducting chromatin immunoprecipitation (IP). Double-stranded barcoded adapters are ligated to the 5′ end of the fragmented nucleosome DNA. These adapters, referred to as Index #1, contain the following:

• (1)T7 promoter
• (2)Illumina SBS3 PCR priming sequence
• (3)Barcode that is unique for each sample
• (4)5′ C3 spacer to prevent self-ligation

The T7 promoter is used for DNA amplification in a transcription reaction where multiple copies of RNA for each nucleosomal DNA are generated, which greatly reduces the amount of input cells required. Reverse transcription using primers that contain a PCR priming sequence on both ends generates cDNA that is PCR amplified for downstream library generation.

Throughput in ChIP-seq assays is low because each sample is processed individually. The Mint-ChIP assay yields a higher throughput by using a pool-and-split multiplexing approach where samples are pooled together using the sample barcode in the Index #1 adapter. The pooled samples are then split into individual ChIP assays, one for each histone modification mark and total H3 that serves as the control. Each split sample undergoes an immunoprecipitation reaction using an antibody specific to the histone mark or H3 histone (control) in a parallel reaction, which increases the throughput (H3, H3K4me1, H3K4me3, H3K9me3, H3K27ac, H3K27me3, and H3K36me3).

A quantitative comparison of histone modification across samples in traditional ChIP-seq assay is difficult because separate IP reactions are required for the samples, and they are sensitive to the amount of starting material. Some strategies that have recently been used include incorporating exogenous DNA or synthetic histone spike-in control, but they don’t work well for small numbers of cells. The pool-and-split approach, which allows for processing multiple samples in the same assay, allows for quantitative normalization and comparison across samples.

After the samples are split, one for each histone mark, a second barcode, Index #2, is added, which can identify the individual histone modification (ChIP assay). The libraries are then combined, which now contain different samples, represented by Index #1, and different histone marks, represented by Index #2, and sequenced using pair-end sequencing. The FASTQ files generated from the sequencing are then demultiplexed for each histone mark by Index #2 followed by demultiplexing for each sample using Index #1.

MicroRNA-seq

MicroRNA sequencing (microRNA-seq) sequences miRNA, a type of RNA that is 21–24 nt long that regulates gene expression at the transcriptional or post-transcriptional level by binding to complementary mRNAs. Abnormal miRNA expression has been implicated in many human diseases, including infectious disease. Due to their conserved function in gene expression, an understanding of the changes in miRNA expression between healthy and disease states or before and after an infection or exposure to chemicals can provide insights into mechanistic pathways and novel biomarkers and enable development of predictive signatures of infection or exposure. Pre-processing steps for samples include sample acquisition, total RNA extraction (Norgen total RNA extraction kit), enrichment of sRNA, and enzymatic modifications. There are several methods that can be used for extraction of total RNA and enrichment of sRNA, each of which has potential to introduce bias. Due to the high abundance of other RNA species in total RNA, most of which are greater in length than miRNA, it is important to maintain the integrity of the total RNA to avoid degradation into smaller-length RNA, which could confound extraction of miRNA based on their size range. A total size distribution from the read alignment should show distinct peaks in the 21–24 nt range, which would otherwise indicate problems with the input material or library preparation. In addition, mapping of reads to known miRNA loci can be done using a database such as miRBase to ensure that the majority of the reads are in known miRNA genomic regions. Based on these considerations, sequencing length should be greater than 17 bp. Sequences under 17 bp may be indicative of degradation and sequences over this may require repeating the size exclusion steps.

Libraries were prepared using the QIAseq miRNAseq Library kit with 100 ng total RNA input. miRNAs can have modifications at their 5′ and 3′ ends, which require appropriate enzymatic modifications to enable proper adapter ligation during the library preparation step. If proper enzymatic modifications are not done, it can introduce errors. Once the samples are ready for library preparation, one of several ways to prepare a library can be used, each of which could have its own specific biases. In general, the steps in library preparation include possible polyA tailing, ligation of sequencing adapter, PCR primer and sample barcode for multiplexing, cDNA synthesis by reverse transcription, and PCR amplification. Improper adapter ligation is usually the greatest source of bias that can lower the percentage of aligned reads and misidentification of differentially expressed miRNAs. One possible solution is to use adapters with random sequences so that there is a greater probability of adapter ligating to a miRNA. In this strategy, all the different adapter sequences should be trimmed in the downstream bioinformatics analysis steps before proceeding with the data analysis. Sequencing was performed using the Illumina NextSeq 500 High-Output 74bp single-end read with 5–10M (million) read clusters per sample.

snmC-seq

Single-nucleus methylcytosine sequencing (snmC-seq) characterizes nucleotide modifications at a single-nucleosome resolution level by measuring the occurrence of 5-methyl- and 5-hydroxymethylcytosines. Profiling single-cell methylomes enables cell state delineation and provides insight to the cell cycle stage, transcriptional activity, mitotic age, and proliferation potential.^46,47 This single-cell epigenomics assay has been used to identify and characterize diverse neuronal types and offers scalability to characterize cellular populations traditionally difficult to identify.⁴⁸

Isolated nuclei were sorted by flow cytometry. Single-nucleus methylome library preparation began with bisulfite conversion of single nuclei (Zymo EZ-96 DNA Methylation-Direct TM kit). Four indexed random primers were used to synthesize bisulfite-converted single nuclei in a 96-well plate. Samples were bead purified (DynaMag-96 Side Magnet, ThermoFisher). Samples were pooled for an adaptase reaction to attach adapters for sequencing. (Accel-NGS Adaptase Module for Single Cell Methyl-Seq Library Preparation, Swift Biosciences). Samples were sequenced on an Illumina HiSeq 4000.

MeDIP-seq

Methylated DNA immunoprecipitation sequencing (MeDIP-seq) couples a magnetic-bead-based antibody capturing method with NGS to agnostically assess genomic methylation. MeDIP-seq assesses genomic methylation by characterizing the modifications of cytosines: 5mc (5-methylcytosine) or 5hmc (5-hydroxymethylcytosine). MeDIP-seq has been used to demonstrate environmentally induced epigenetic transgenerational inheritance of disease and phenotypic variation.⁴⁹

Genomic DNA from PBMCs was extracted and purified (Promega Genomic DNA Purification kit) followed by sonication (Covaris 300bp protocol). DNA was heat denatured into single-stranded DNA to improve antibody binding affinity. DNA was incubated with anti-5mC (5-methylcytosine) antibodies (monoclonal mouse anti 5-methyl cytidine; Diagenode #C15200006). The DNA-antibody complex was captured by magnetic beads (Dynabeads M-280 Sheep antiMouse IgG; 11201 D) and unbound DNA was removed. Samples were treated with proteinase K to digest the antibody and eluted DNA was prepared for NGS (NEBNextVR UltraTM RNA Library Prep Kit for Illumina).

ChIPmentation

Chromatin immunoprecipitation with on-bead tagmentation (ChIPmentation) introduces an expedited library preparation method for ChIP-seq. In tagmentation, a hyperactive Tn5 transposase simultaneously fragments DNA and ligates adaptors directly to bead-bound immunoprecipitated chromatin. In combination with next-generation sequencing, ChIpmentation maps histones and their post-translational modifications of transcription factors throughout the genome. This method offers efficiency, reduces cost, and greatly reduces input requirements.⁵⁰

Bioinformatics methods

General quality control metrics

Although different assays have unique sample preparation methods, there are many general QC metrics that apply to any assay that ends with next-generation sequencing due to commonalities in methods to create the libraries to be sequenced. Any assay culminating in NGS requires PCR amplification to construct a library for sequencing. The need to utilize PCR in sample preparation arises from the initial sample not having an adequate amount of DNA or cDNA molecules for sequencing. However, PCR can stochastically introduce errors and embed artifacts into samples that propagate to later cycles.⁵¹

For all assays that utilize PCR for sample and library preparation, careful consideration is given to metrics that can point to biases potentially introduced from PCR steps.

Experimental datasets can vary from high to low quality. Several assays (ATAC-seq, ChIPmentation, MeDIP-seq, Mint-ChIPseq, RNA-seq, and others) that rely upon amplification of target sequences with PCR share PCR-associated QC metrics. PCR during library preparation has been shown to be a principal source of bias. When source target nucleic acids are limited or experimental techniques are in need of improvement, excessive amplification of duplicate target sequences can result. This is because PCR can introduce artifacts to samples through over-amplification or under-amplification of PCR products during cDNA and library preparation steps. Sequence-dependent bias can also be introduced from PCR from differences in GC content. GC-rich regions have a higher melting temperature and thus are more resistant to denaturation during PCR amplification, potentially causing these regions to be under amplified and less represented in a final library.

If it is suspected that library diversity is being compromised from under amplification of GC-rich regions, GC-rich regions could be rescued by adding 2M betaine and extending denaturation times. Entire GC% spectrums can also be rescued by switching polymerases and further fine tuning the thermoprofile.^52,53

QC metrics that indicate low library diversity include low sequencing depth, low aligned or mapped reads, low number of cells identified, or low number of transcriptomic reads. For many assays, a repeat of library preparation or any steps involving PCR is recommended when QC metrics point to low library diversity. Spiking in PhiX can alleviate low diversity libraries that have a skewed proportion of A,T,C,G for sequence positions.

If primer/adapter dimers are suspected in a library through DNA quantification methods, general PCR mitigation steps for primer dimers can be taken (fine tuning thermoprofile by increasing annealing temperature or increasing temperature of denaturation, decreasing primer/adapter concentration, or adding dimethyl sulfoxide [DMSO] if hairpins are suspected) before a library size selection cleanup. Reducing primer/adapter dimers is important because a high amount of primer/adapter dimers taking up library composition causes low library diversity and reduces the sequencing and quantification of relevant transcripts.

PCR duplication rate is also a concern for many assays. PCR duplication rate refers to the percentage of reads that are made from the same initial cDNA molecule from PCR. A high PCR duplication rate can be the result of low library diversity. Incorporating UMIs to initial sample sequences can alleviate bias introduced from PCR duplicates. By tagging the initial sample with a UMI, PCR duplicates can be identified and removed from downstream analysis to increase accuracy of frequency quantification. PCR duplication can also be alleviated by lowering PCR cycles in amplification steps.⁵⁴

Early identification of low-quality datasets enables 1) avoidance of low-quality data into downstream analytics, 2) monitoring of experimental assays for early identification of possible issues, and 3) assay selections samples with limited cells or DNA available.

Assay-specific quality control metrics

Bulk RNA-seq

QC checks are applied to ensure data quality that involves the analysis of raw reads, read alignment, transcript quantification, and data reproducibility.⁵⁵ Low sequencing depth may be from low sample input. Low uniquely mapped reads (i.e., reads that align directly to only one motif of the reference genome) below 75% can be the result of low library diversity. Low library diversity may be from low sample input to cDNA library construction or from PCR amplification bias. Optimizing PCR conditions for cDNA library construction by reducing cycles or tweaking heat steps can improve library diversity and can help achieve an optimal end GC %. Integration of additional kits to workflows can be considered for sub-optimal ribosomal RNA(rRNA)% or %globinRNA. As rRNA is the predominant form of RNA, proper isolation of mRNA must be achieved for accurate transcript quantification. rRNA removal can enable higher sequencing depth of mRNA, leading to better detection of mRNA transcripts as well as uniquely mapped reads.⁵⁶ For PBMC samples, globin depletion prior to sequencing has been shown to improve the quality of RNA-seq data.^57,58 Proper isolation of human peripheral blood mononuclear cells (PBMCs) (i.e., lymphocytes, monocytes, natural killer cells [NK cells], or dendritic cells) is essential to capture the RNA expression variations relevant to diseases and conditions in different cell types. Hemoglobin (Hgb) RNAs originate from red blood cells (RBCs), which can make up to 45% of the total volume of blood and 99% of the cellular volume. Thus, globin mRNA can compromise the detection of PBMC mRNAs by reducing the sensitivity to detect lower-level transcripts in the blood transcriptome.⁵⁹

Quality distribution

In addition to QC as part of the amplification and sequencing processes, evaluation of the outputs of any secondary analysis pipelines is imperative for understanding the quality of the algorithmic processing around assembly, alignment, quantifications, and any variant calling steps.

One example of this in RNA-seq and WES is the use of Picard,⁶⁰ a tool from the Broad Institute, or the graphical wrapper, FastQC,⁶¹ that enables the manipulation of high-throughput sequencing and contains functions to derive metrics around the quality of your alignment. Using functions such as Picard’s CollectAlignmentSummaryMetrics, CollectBaseDistributionByCycle, CollectRnaSeqMetrics, CollectHsMetrics, CollectInsertSizeMetrics, MeanQualityByCycle, or QualityScoreDistribution, we can evaluate the quality of the sequencing inputs, but also the alignment performance of the sample. This is important in the assessment of the bioinformatics process that occurs after the assembly of reads through the secondary analysis steps.

The quality distribution metrics from Picard’s QualityScoreDistribution function provide a count of reads by Phred quality score. This allows for the identification of quality issues where there may be significant proportion of low-quality reads that will affect subsequent analyses. Phred quality scores represent base-calling error probabilities using the following equation:

$$Q=-10{log}_{10}P\text{,}$$

where P is the base-calling error probability and Q is the resulting Phred score. For example, a base-calling accuracy of 99% (or 1/100 error probability) would be represented with a Phred score of 20.

In Table 2, this example shows that 3.4% of the bases have a quality score of 11, which may be problematic, depending on the sensitivity of the research use case in question.

Table 2.

Example of the quality distribution output from Picard’s QualityScoreDistribution function

QUALITY	COUNT_OF_Q
11	457812700
25	708452935
37	12137945809

Furthermore, as part of the sequencing process, base quality should not vary significantly between cycles. Picard’s MeanQualityByCycle function also provides a breakdown of mean read quality by cycle. This view helps to identify if there is a significant change in read quality over time throughout the sequencing process. In Table 3, the example output shows that the mean remains stable (and high, >20) through the first several cycles shown.

Table 3.

Example of the mean quality output by cycle from Picard’s MeanQualityByCycle function

CYCLE	MEAN_QUALITY
1	35.40155
2	35.58375
3	36.02574
4	36.13188
5	36.21469
6	36.53705
7	36.54161
8	36.57361
...	...

Alignment metrics

Once the quality of sequencing reads has been deemed sufficient and an alignment has been performed, an additional QC step includes the evaluation of alignment performance. Picard’s AlignmentSummaryMetrics, RnaSeqMetrics, and HsMetrics functions provide convenient breakdowns of the reads and their alignment.

AlignmentSummaryMetrics shows the total reads versus the number of reads that were aligned along with metrics such as mismatch rate, mean read length, and indel rate. It is useful to confirm that the total reads aligned is on par with the reported number of reads from the sequencing process. Plus, this output can help identify if there is a high number of bad cycles, an unacceptable error rate, or an unacceptably high number of reads that were not aligned in pairs.

In Table 4, note how the mismatch and error rates are quite low. Plus, the high-quality (HQ) aligned reads are a significant majority of all aligned reads (85,211,166 / 102,759,046 ≈ 83%) and 100% of the reads are aligned in pairs. If there were issues with the alignment quality, some of these metrics may help indicate the source of the issue(s). Furthermore, investigating the distribution of the bases within the transcripts of an RNA-seq experiment, we can better understand the makeup of the sample that was processed through an RNA-seq secondary analysis pipeline. For example, in Table 5, we can see that 0 reads were ignored and that there were a very small number of bases that belong to rRNA and 95.8% of the bases were mRNA (messenger RNA), which is desired. Also, given that 25% of the human genome is introns⁶² (32) and that, in theory, RNA-seq samples should contain minimal intronic regions of the genome after downstream analyses, we can adjudicate this notion by seeing that <4% of the bases in the sample shown above are intronic bases.

Table 4.

Example of the alignment metrics from Picard’s AlignmentSummaryMetrics function

CATEGORY	FIRST_OF_PAIR	SECOND_OF_PAIR	PAIR
TOTAL_READS	54929952	54929952	109859904
PF_READS	54929952	54929952	109859904
PCT_PF_READS	1	1	1
PF_NOISE_READS	0	0	0
PF_READS_ALIGNED	51379523	51379523	102759046
PCT_PF_READS_ALIGNED	0.91668	0.91668	0.91668
PF_ALIGNED_BASES	5128423636	5125843525	10254267161
PF_HQ_ALIGNED_READS	42605583	42605583	85211166
PF_HQ_ALIGNED_BASES	4256380666	4255199751	8511580417
PF_HQ_ALIGNED_Q20_BASES	4220745549	4199744152	8420489701
PF_HQ_MEDIAN_MISMATCHES	0	0	0
PF_MISMATCH_RATE	0.002022	0.002443	0.002232
PF_HQ_ERROR_RATE	0.001804	0.002239	0.002022
PF_INDEL_RATE	0.000146	0.000147	0.000146
MEAN_READ_LENGTH	98.212138	98.182183	98.19716
READS_ALIGNED_IN_PAIRS	51379523	51379523	102759046
PCT_READS_ALIGNED_IN_PAIRS	1	1	1
PF_READS_IMPROPER_PAIRS	0	0	0
PCT_PF_READS_IMPROPER_PAIRS	0	0	0
BAD_CYCLES	0	0	0
STRAND_BALANCE	0.473177	0.506847	0.490012
PCT_CHIMERAS	0.005451	0.005451	0.005451
PCT_ADAPTER	0	0.000001	0

Table 5.

Example of the RNA-seq metrics from Picard’s CollectRnaSeqMetrics function

CATEGORY	METRIC
PF_BASES	12103166420
PF_ALIGNED_BASES	11274633239
RIBOSOMAL_BASES	4910
CODING_BASES	9300635422
UTR_BASES	1499010311
INTRONIC_BASES	370755932
INTERGENIC_BASES	104226665
IGNORED_READS	0
CORRECT_STRAND_READS	8479813
INCORRECT_STRAND_READS	285381
NUM_R1_TRANSCRIPT_STRAND_READS	121327
NUM_R2_TRANSCRIPT_STRAND_READS	4059650
NUM_UNEXPLAINED_READS	220590
PCT_R1_TRANSCRIPT_STRAND_READS	0.029019
PCT_R2_TRANSCRIPT_STRAND_READS	0.970981
PCT_RIBOSOMAL_BASES	0
PCT_CODING_BASES	0.824917
PCT_UTR_BASES	0.132954
PCT_INTRONIC_BASES	0.032884
PCT_INTERGENIC_BASES	0.009244
PCT_MRNA_BASES	0.957871
PCT_USABLE_BASES	0.892299
PCT_CORRECT_STRAND_READS	0.967442
MEDIAN_CV_COVERAGE	0.919178
MEDIAN_5PRIME_BIAS	0.231197
MEDIAN_3PRIME_BIAS	0.286529
MEDIAN_5PRIME_TO_3PRIME_BIAS	0.392208

It is also useful to note the percentage of the bases in your sample that were able to be aligned. In the example above, 93% (11,274,633,239 / 12,103,166,420) of the bases were aligned and the percentage of usable bases in the sample was 89%.

These aforementioned Picard outputs are just some of many metrics that may help to assess the quality of the results from any bioinformatics secondary analyses. While the quality of the input sequences may be high, issues with alignment may persist due to sample impurity, off target amplification, and other factors.

scRNA-seq

The quality and viability of cells at the start of any single-cell experiment is critical to achieve good QC metrics downstream. Cryopreservation and sample preparation technique discrepancies can cause variation in sample integrity and input quantity. As RNA degradation is closely related to sample preservation technique, ensuring proper storage and extraction methods can mitigate initiating the sequencing process with low- and poor-quality input samples. If performing whole-cell scRNA-seq, cell viability is assessed and should preferentially be above 70%. If the starting material are nuclei, assessing nuclei quantity using a fluorescence-based cell counter is recommended.^48,63,64

High mitochondrial reads in cells can be an indicator of damaged or dying cells. This is because in stressed cellular environments, mitochondria can release proteins to induce apoptosis.⁶⁵ This can be caused by either dissection of necrotic tissue or by the stress caused on cells throughout the sample prep process disassociating cells from the tissue sample, which can introduce artifacts to the sample. Also, if a cell is lysed, cytoplasmic mRNA can leak out through a broken membrane resulting in only mitochondrial RNA to be conserved and in turn over-amplified and sequenced.⁶⁶ If samples have high amounts of cells with mitochondrial reads, cell isolation techniques should be reviewed for optimization.⁴⁶

RNA degradation could be the culprit for any failed metric. Working on ice and working quickly will alleviate RNA degradation. RNA degradation prior to cDNA synthesis can result in low cDNA yield and can cause the total number of cells detected to be low. Low cDNA synthesis yields will in turn cause low inputs for cDNA library construction. Low input into library construction can increase the chances of amplification bias in library construction resulting in low diversity libraries, which can also cause a low number of cells to be detected in sequencing. Partial sample degradation of the cDNA library can result in low UMI per cell. However, median UMI counts may be skewed low if a sample contains a large proportion of neutrophils and granulocytes as they have a relatively low RNA content and high levels of ribonucleases (RNases), which results in fewer sequencing reads per neutrophil and granulocyte cell. If a low median UMI is not due to poor sample quality, optimal neutrophil and granulocyte sequencing reads may be achieved by increasing the PCR cycles by two during cDNA amplification. Deeper sequencing through increased library size may also improve QC metrics as more transcripts can be detected leading to more precise transcript quantification.

General gene expression

As mentioned above, the input cell-level quality is imperative to the quality of downstream analyses. The metrics of the number of cells per sample and fraction of reads in cells can be combined to estimate the number of dead cells per sample. This can be used to discern if the issue is a sequencing issue or sample integrity issue. One example program that outputs these metrics is Cell Ranger, which includes a data preparation, exploration, and QC toolkit.⁶⁷ Cell Ranger’s count function provides count-level metrics per sample of the scRNA-seq output. In the example output in Table 6, note the comparatively low number of cells found in the fourth sample along with its low fraction of reads that were found in cells. This may indicate an issue in the quality or integrity of the cells found in that sample.

Table 6.

Example Cell Ranger targeted gene expression counts output

Sample	30-1006397_202401.SC-C1	30-1006342_202401.SC-D1	30-1006401_202401.SC-A2	30-1006422_202401.SC-C2	30-1006425_202401.SC-B1
Estimated number of cells	7,520	15,316	10,752	677∗	15,002
Mean reads per cell	79,382	34,812	45,758	537,410	24,610
Median genes per cell	1,312	3,624	870	609	134
Number of reads	612,744,323	521,800,171	514,464,287	372,162,063	376,655,791
Valid barcodes	98.50%	98.50%	98.50%	97.50%	97.50%
Valid UMIs	100.00%	99.50%	100.00%	100.00%	99.99%
Sequencing saturation	93.50%	51.50%∗	91.50%	97.50%	97.50%
Q30 bases in barcode	96.20%	96.30%	96.20%	95.60%	95.60%
Q30 bases in RNA read	95.10%	95.40%	95.60%	94.20%	94.40%
...	...	...	...	...	...
Fraction of reads in cells	81.66%	93.60%	71.82%	18.10%∗	81.23%
Total genes detected	14,393	12,152	17,219	17,794	16,560
Median UMI counts per cell	1,441	691	1,408	259	727

Failed samples are highlighted with an asterisk (∗).

Furthermore, this table shows that the percentage of bases that are high quality (Q30 or above) is >90% across all the samples, as is the validity of the samples’ barcodes (>90%). So, the issue for the problematic sample shown above is not in the quality of the sequencing, but rather in the cells in that sample. This type of output may also show sample impurity if the total genes detected is lower than expected or the mean reads per cell are also quite low.

Common issues in scRNA sequencing may arise from issues in sample integrity. For example, cells may have apoptosed or have been inadvertently lysed as part of the sample preparation process. This issue may be detected when looking at the cell-level QC output from Cell Ranger or other tools. For example, Table 7 shows that the percentage of cells that are predicted to be dead are quite high in the fourth sample from the above example.

Table 7.

Example Cell Ranger cell-level QC predicting the number and percentage of dead cells

Sample	Pct. predicted dead	N predicted dead
30-1006397_202401.SC-C1	0.1%	8
30-1006342_202401.SC-D1	2.6%	398
30-1006401_202401.SC-A2	0.0%	0
30-1006422_202401.SC-C2	68.9%∗	466
30-1006425_202401.SC-B1	12.7%	1905

Failed samples are highlighted with an asterisk (∗).

It is expected for some cells to be dead as part of any input sample. However, sequencing performed on older samples or those that have been frozen and thawed improperly may increase the quantity of dead cells. Then, when a sequencing process results in few viable cells and a low percentage of in-cell reads, the presence of a high percentage of dead cells further hinders the samples’ viability for downstream processing and usage.

Cell classification

A common goal of scRNA-seq may be to characterize the variation in gene expression by cell type across samples. For a given study, cells can be classified into a number of different cell types based on an input gene signature matrix. For example, using an R library such as SingleR, cells can be mapped to a variety of cell types based on their gene expression. In the example data shown in Table 8, there are varying quantities and percentages of the few cell types that are defined.

Table 8.

Example cell type mapping from scRNA-seq analysis

Samples	Pct. CD8 T cells	CD8 T cells	Pct. macrophages	Macrophages	Pct. memory B cells	Memory B cells	Pct. naïve B cells	Pct. naïve B cells
30-1006397_202401.SC-C1	2.1%	158	3.6%	271	0.3%	19	0.0%	0
30-1006342_202401.SC-D1	0.4%	61	0.2%	31	0.5%	77	0.1%	15
30-1006401_202401.SC-A2	37.8%	4064	25.7%	2763	21.1%	2269	0.2%	22
30-1006422_202401.SC-C2	3.0%	20	3.0%	20	0.0%	0	1.1%	7
30-1006425_202401.SC-B1	0.2%	30	74.3%	11146	2.2%	330	0.0%	0

In the example above, the third sample seems to have a much larger ratio of memory B cells and CD8+ T cells as compared to the others. While this may not be problematic, it may be indicative that the samples are quite different. This is important in studies where the samples (that are perhaps from different study participants) need to be of the same tissue types to get as close to a one-to-one comparison. In this example set, it becomes clear that the input tissues may not have been the same or the state of those tissues certainly varies. Also, it is important to pre-define the cell types of interest as this changes the required reference input for the bioinformatics mapping process. In the example above, only four immunorelated cell types were used. However, it completely misses other cells that may have been of interest in the samples (e.g., erythrocytes, plasma cells, keratinocytes, etc.).

ATAC-seq

If sequencing depth is below 25 million, repeating the transposase reaction and library prep may be required to achieve proper sequencing depth. Over 40 million non-duplicate reads is indicative of strong library complexity. A high duplicate rate (>30%) may be an indicator of a low library complexity, for which increasing initial cell input, repeating library preparation, or adjusting the volumes of low concentration samples when making a new pool for sequencing may resolve the issue. Although ATAC-seq is intended to primarily sequence regions in between nucleosomes, a high-quality ATAC-seq sample will also include a sequence peak for one nucleosomal sequence, which enables proper alignment of the sequences.

Signal-to-noise metrics are essential to look at in ATAC-seq since highly accessible DNA is a characteristic for cells of poor viability like activated granulates or dead and dying cells. Key signal-to-noise metrics are FRiP (fraction of reads in peaks) and TSS (transcriptional start sites) enrichment scores. Peaks are determined based on a 10- to 30-fold enrichment compared to background.⁶⁸ High FRiP scores indicate that the experiment has a high signal-to-noise ratio as a majority of the reads are in peaks. A low FRiP score indicates the experiment to have had a low specificity as a majority of the reads are in non-specific regions. TSS enrichment scores are calculated by the ENCODE metrics.⁶⁹ TSS enrichment score is assessed as an aggregate distribution of reads centered on TSSs and extending to 2,000 bp in either direction. TSS regions should be enriched as they are typically in accessible chromatin regions. To ensure strong signal-to-noise ratios, cell viability and sample integrity must be ensured. Repeating the transposition step ensuring assay integrity would be necessary if the sample indicated non-specific sequencing.

scATAC-seq and snATAC-seq

The same mitigation steps for ATACseq from the same failed metrics can be taken for scATAC-seq. If the protocol is not followed and the nuclei are overtransposed, this may lead to shorter sequences.^38,70,71

scMultiome

As for scATAC, entering the assay with high-quality nuclei is critical for assay success. Extracellular debris in the sample can clog the microfluidic channels that partition nuclei into emulsions, which can result in a low number of cells detected by sequencing. Overlysed nuclei can stick together and form clumps leading to clogs. Additionally, overlysis can cause nuclei content to leak out of nuclei, resulting in a high level of background sequencing. Following the manufacturers’ recommendation and established protocols is critical. Improper emulsification from debris can also increase the chances of doublet formation, when two or more cells are in one emulsion droplet resulting in transcripts from more than one cell receiving the same cellular barcode. Minimizing the chances of doublet formation is critical for transcripts to be mapped to the proper cell type to enable overall accurate cellular annotation. Low quality nuclei can also cause nonhomogeneous emulsification reactions, which can result in low median UMI detected per cell. Low GEX and ATAC sequence lengths can be indications of sample degradation.

QC of the scRNA sequencing from a scMultiome sample is visualized in Figure 2. Plots are made from Scanpy (https://pubmed.ncbi.nlm.nih.gov/29409532/). The clusters of different cell types that are identifiable by the distribution of cell-specific marker genes.⁷² The failed metrics for the poor-quality sample were a suboptimal percent of transcriptomic reads in cells (49.35%) and suboptimal median UMI per cell (544). All other metrics for both samples had passing thresholds.

Figure 2.

QC visualization of scRNA-seq data from PBMC samples from individuals exposed to pentaerythritol tetranitrate (PETN), a nitrate ester explosive

Each point represents a cell. Clustering is driven by transcriptomic profiles. Leiden clustering and distribution of marker genes for NK cells (NKG7), CD4 T cells (MAL), and B cells (MS4A1) shown for a passed sample (A) and failed sample (B). Cell clustering should compositionally be representative of the different cell types from a PBMC sample. Clustering is driven by transcriptomic profiles that are different for each cell type. Similar profiles of upregulated genes drive cluster formation and are used to ID cell type.

MethylationEPIC

The EPIC array data quality checks include the following:

• (1)Seventeen control metrics are defined by the manufacturer.⁷³ The Infinium MethylationEPIC v2.0 BeadChip assay comes equipped with internal controls to assess important QC metrics. These manufacture-defined QC metrics are reported by the BeadArray Controls Reporter software for the internal sample controls. Metrics are calculated based on the intensity of the bead associated with the individual control probe and each scanner’s varying levels of intensity. These internal sample controls are assessed by the manufacturer for formalin-fixed paraffin-embedded (FFPE) DNA restoration, staining efficiency of red and green channels, extension efficiency of probes, hybridization performance, efficiency of the stripping step after the extension reaction, bisulfite conversion efficiency, nonspecific primer extension, and overall assay performance from amplification to detection by querying a particular base in a nonpolymorphic region of the genome. Notably, Bisulfite Conversion II, which monitors successful bisulfite conversion samples with a Bisulfite Conversion II metric below 1, might be partially converted, leading to inaccurate estimates of methylation levels, which are important to assess. Also, samples with too many undetected probes or low overall fluorescence intensity are excluded. As fluorescence intensities of a SNP locus are calculated from two probes targeting either the wild type or the common mutant variant, the distribution of the proportion of methylated strands of a SNP locus (beta value) falls into three disjunct clusters: the heterozygous and the two homozygous genotypes. Beta value deviation from the ideal trimodal distribution can serve as a means of outlier identification and a metric for poor technical performance.⁷³
• (2)A sex check to detect mislabeled sex-discordant samples.
• (3)An identity check for fingerprinting sample donors.
• (4)A measure of sample contamination based on probes querying high-frequency SNPs.

Additionally, undetected probes should be filtered out from analysis to reduce spurious values.⁷⁴

Additional quality checks consist in evaluating sample clustering, background noise, probe variability, methylation call rate, beta value distribution, detection p value distribution, and principal component analysis (PCA), as described below:

• (1)Sample clustering: The sample clustering should clearly separate samples into appropriate groups, with minimal overlap between groups.
• (2)Background noise: The background noise levels should be low and consistent across all arrays, with values less than 2% of the average signal intensity.
• (3)Probes with high variability: The proportion of probes with high variability (CV > 30%) should be low, typically less than 10%.
• (4)Methylation call rate: The methylation call rate should be high, with a minimum of 95% of probes with a valid methylation call in each sample.
• (5)Beta value distribution: The beta value represents the proportion of methylated strands for each CpG site. The beta value distribution should be within the expected range (0–1), with a minimum balance of probes with extreme values (less than 0 or greater than 1).
• (6)Detection p value distribution: A low signal-to-noise ratio of fluorescence intensities results in probes with high detection p values. Such probes should be removed as they are unreliable. The detection p values should have a minimal proportion of probes with p values greater than 0.05, typically less than 5%.
• (7)PCA: The first two principal components should explain a significant proportion of the sample variation (typically >50%), with minimal overlap between samples of different groups.

Input DNA quantity and quality influence assay performance

Failed probe detection can be the result of a high background. Lowering background may be alleviated by ensuring optimal amount of input DNA for the bisulfite conversion kit (being between 200–500 ng). High levels of background are also alleviated by optimizing PCR conditions in whole genome isothermal amplification—optimal primers and high annealing temperatures between 55°C–60°C coupled with hot start polymerases can help since the template is AT rich due to the non-methylated cytosine conversion to uracil in the PCR template. Proper reaction times for the initial bisulfite reaction could be also important; although methylated cytosine base deaminates at about two orders of magnitude slower, overextending the bisulfite reaction time could lead to methylated cytosine to undergo the deamination, yielding inaccurate results.

One drawback of MethylationEPIC is the background noise resulting from the off-target binding of probes due to their sequence homology.⁷⁵ Another source of noise in the data is the cross-talk between the green and red fluorophores used to differentiate the nucleotides that are incorporated in the single-base extension step. Beta value variations can occur depending on which instrument is used in quantifying signal intensities (Illumina iScan, NextSeq 550, or NextSeq 550Dx systems).

Mint-ChIP-seq

The Mint-ChIP-seq assay involves additional steps compared to the traditional ChIP-seq assay, which has the potential to introduce bias. Quality control steps therefore become critical before Mint-ChIP data can be analyzed. As with most sequencing experiments, some QC parameters that should be checked for each sample and histone mark as an indication of a high-quality library include the following:

• (1)Number of reads that uniquely map to the genome, which should be over 2 million but ideally greater than 3 million.
• (2)Percentage of uniquely mapped reads that map to the genome (alignment percentage), which at minimum should be greater than 60%, but for a high-quality library is expected to be greater than 85%.

The key objective of a ChIP-seq assay is the identification of peaks, which are regions of the genome where the histone modification is expected and is identified by the enrichment of alignment of reads in these regions. The presence of reads indicates the presence of the histone mark. FRIP is used to assess signal-to-noise ratio. Ideally, all the aligned reads should fall within peaks, but a value of greater than 40% is considered acceptable due to presence of background reads from regions other than the peaks. In traditional ChIP-seq assay, the background signals are controlled using histone spike-in or by incorporating exogenous DNA. This approach is not appropriate for the Mint-ChIP assay due to low number of cells but is overcome by pooling samples in the same IP reaction and using total Histone (H3) as control.

ChIP-seq assays suffer from high PCR duplication rate due to the low amount of immunoprecipitated DNA, which requires higher PCR amplification. In Mint-ChIP assays, this is further exacerbated due to the two-step amplification, which leads to higher PCR duplicate rate and low complexity, especially for histone marks that are not commonly found on the genome, such as H3K27ac. The PCR duplicate rate for these marks is high and difficult to control. We defined a new parameter called “useful reads efficiency,” which could better capture the usefulness of the deduplicated reads. Useful reads efficiency is calculated by dividing useful reads by all reads. Useful reads are those that uniquely map to the genome after removing the PCR duplication. The useful reads efficiency should be greater than 70% for optimal quality. Typically, if the PCR duplication rate is too high, the useful reads efficiency will decrease. Due to the expected variation in the PCR duplication rate, it is suggested to set a threshold based on specific marks. Recommended rates are listed in Table 9. A lower PCR duplication rate is better and leads to higher useful reads efficiency.

Table 9.

Mint-ChIP suboptimal PCR duplication rate for specific histone marks

Histone mark	PCR duplication rate (max)
H3K4me1	70%
H3K4me3	40%
H3K27me3	40%
H3K36me3	60%
H3K9me3	60%
H3	70%
H3K27ac	40%

microRNA-seq

High-quality miRNA sequences should be at least 17 bp. The overall length distribution can also be looked at to assess quality. A high portion of sequences lower than 17 bp could be an indication of sample degradation. A narrow peak at 22 bp should be seen for all samples; broad peaks could indicate sample degradation. The length distribution around 22 bp is informative if sequenced samples are derived from miRNA because they represent natural variations in miRNA length. Skews or deviation from a narrow peak at 22 bp could mean sequences were not derived from miRNAs.⁷⁶ Samples should have at least 2M aligned reads. Putative contamination could be assessed by comparing the percent of mapped reads to human miRNA databases to bacterial and viral genomes.⁷⁷ Sequence depth should be at least 4M reads. Another consideration for miRNA-seq is PCR duplication rate. The primary cause of higher PCR duplication rate is low starting material or mistakes in the steps preceding the library preparation. High PCR duplication can result in libraries with low complexity, which can lead to missing miRNA species in the data analysis step and lowering the percentage of reads mapped and aligned. Input sample quantity and quality in addition to the pre-processing steps should be optimized for minimizing PCR duplication rate.

snmC-seq

Optimal samples provide at least 1% of genome coverage. The methylation rate of the trinucleotide sequence CCC is assessed as a metric as an estimate for the rate of bisulfite non-conversion. Low input material can be the cause of low and failed uniquely mapped reads metric; specifically, the Smart-Seq2 protocol purification step can lead to a significant loss of material. Methylation of CG sites (mCG) rates and methylation of non-CG (mCH) rates are metrics indicative of both quality and accuracy of the assay. mCG and mCH rates should reflect a genomic methylation state: high mCG occurrence as 60% −90% of all CpG sites in mammals are methylated, and very low mCH rates as methylation outside of CpG sites is rare.⁷⁸ Repeating bisulfite conversion and subsequent steps would be necessary if mCG and mCH rates were significantly out of passing range.

MeDIP-seq

An optimized immunoprecipitation reaction with high-quality DNA input is required to achieve passing metrics. CpG coverage percent can be affected from non-specific binding in immunoprecipitation. Because antibody-based selection is biased toward hypermethylated regions, screening could exclude CpG islands if they are hypomethylated. Poor CpG coverage can be caused from nonspecific binding. Antibody and DNA incubation time can also alter binding results, which can influence sequenced motifs.

ChIPmentation

ChIPmentation assays can suffer from high PCR duplication rate due to the low immunoprecipitated DNA yield that requires a high number of PCR amplification cycles to generate a sufficient library for sequencing. This decreased complexity of input material for sequencing can cause high PCR duplication rates. Bias can also arise from Tn5 insertion frequency being higher in positions surrounding the center of nucleosomes where the DNA is most accessible.⁵⁰ Samples should have a uniquely mapped reads percentage over 80%. Increasing initial cell numbers can increase the number of uniquely mapped reads.⁶³ Optimal ChIPmentation samples should have a sequence length of over 50 bp after trimming. Any sample degradation leading to protein degradation will innately lead to poor sample sequencing, as chromatin proteins are protecting the DNA during the tagmentation process and are the link to accurately relate the DNA sequence to the antibody binding to the proteins around it.

Conclusion

Outlining high quality assay metrics for 11 different epigenomics and transcriptomics datasets alongside mitigation strategies to assist in optimizing protocols to yield high quality results will alleviate the issues in complex datasets that are difficult to navigate due to dynamic research teams that exhibit a multitude of research objectives. On projects where there may be numerous different handlers of a sample from collection to sequencing, it can be difficult to pinpoint the source of a sample’s problem or contamination. QC should be a prioritized step in all computational epigenomic and transcriptomic workflows. Although many workflows incorporate QC steps, the advantages of looking at individual sample quality within a given dataset is often overlooked. The ability to assess datasets for QC issues at the sample level prior to downstream analysis prevents the misinterpretation of results and serves as a strong indicator of bench protocols that yield high quality and reproducible results. Eliminating analysis on low quality data can optimize time and reduce the probability of developing erroneous results.

The assessed pipelines were containerized in Singularity and designed for a high-performance computing (HPC) environment, which offers great advantages in computing speed and reproducibility. The HPC environment is well suited for the QC of epigenetic and transcriptomic datasets because it enables the parallelization of pipeline steps to achieve rapid QC metric reporting. Singularity also offers the ability to run containers in internet-starved environments and on a shared environment without root access. Packaging software offers large research teams from different institutions and projects that span a long period of time to achieve reproducible results. All of these advantages offer great interoperability for research teams involving multiple collaborators, and reproducibility of results to ensure datasets from different time points are being analyzed with the identical tools, algorithms, and standardized set of metrics.

The previously developed pipelines are available on Github (https://github.com/mitll/Omics_QC_pipelines ), and interested users can also reach out to the corresponding author of that manuscript for additional help.²⁵

Acknowledgments

Funding and guidance were provided by the Defense Advanced Research Projects Agency; Epigenetic CHaracterization and Observation (ECHO) program. The authors acknowledge the MIT SuperCloud and Lincoln Laboratory Supercomputing Center for providing (HPC, database, and consultation) resources that have contributed to the research results reported within this paper, and Catherine Cabrera, PhD (MIT LL), Joseph Ecker, PhD (Salk Institute), Thomas Thomou, PhD, and Eric Van Gieson, PhD (DARPA).

Portions of this work by Salk Institute used the Anvil HPC cluster at Purdue University through allocation MCB130189 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.

Approved for public release. Distribution is unlimited. This material is based upon work supported by the Defense Advanced Research Projects Agency under Air Force contract no. FA8702-15-D-0001. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency.

Declaration of interests

C.T.F. is the owner of Tuple, LLC, a biotechnology consulting firm. S.C.S. is a consultant, equity owner, and interim chief scientific officer at GNOMX Corp.