CARTEpigenoQC: QC Toolkit for CAR-T Epigenomics

• CARTEpigenoQC is an R-based dual QC toolkit that evaluates both standard scQC metrics (e.g., FRiP, TSS enrichment) and CAR engineering evidence from single-cell epigenomic data.
• It integrates fragment-level analysis with user-defined CAR-target annotations for assays like scATAC-seq, scCUT&Tag, and scBS-seq in CAR-T settings.
• The toolkit produces self-contained HTML reports with PNG outputs, supporting exploratory analysis, translational research, and preclinical documentation.

Searching arXiv for the primary paper and closely related methodological context. CARTEpigenoQC is an R-based quality-control toolkit for single-cell epigenomic datasets generated from Chimeric Antigen Receptor (CAR)-engineered T cells. It was developed to combine standard epigenomic QC with transgene-aware QC, so that ordinary assay-quality assessment and direct evidence of CAR engineering can be evaluated within a single workflow. The toolkit is positioned for scATAC-seq and, more broadly, for scATAC-seq, scCUT&Tag, and scBS-seq in CAR-T settings; it supports both 10x Genomics and non-10x data formats and generates Markdown-based reports rendered to self-contained HTML with PNG summary outputs suited to exploratory analysis, regulatory-grade preclinical reporting, and “clinical handoff” style documentation (Lai, 30 Jul 2025).

1. Problem domain and conceptual motivation

CARTEpigenoQC addresses a specific deficiency in standard single-cell epigenomic QC for engineered cell therapies. Existing tools such as ArchR, Signac, and SnapATAC evaluate generic chromatin data quality, but they do not address whether a dataset contains convincing signal from the engineered CAR construct or its genomic insertion target. In CAR-T studies, that omission is consequential because a cell can pass generic QC yet remain biologically irrelevant if it does not carry the engineered construct, if editing is mosaic, or if expected vector-target signals are absent (Lai, 30 Jul 2025).

The toolkit therefore frames QC along two axes. The first is standard chromatin quality, including metrics such as FRiP and TSS enrichment. The second is CAR-related signal validity, assessed through direct interrogation of predefined CAR-related genomic regions such as EF1-a, PGK1, or the TRAC locus. This dual-QC formulation is the central organizing principle of the software. Its intended use extends across exploratory analysis, core-facility workflows, and translational immunology programs that require reproducible validation of engineered-cell identity in addition to technical assay quality.

This positioning also distinguishes CARTEpigenoQC from generic plotting utilities. The paper repeatedly emphasizes report generation, engineered-cell confirmation, and consistent documentation. A plausible implication is that the package is meant to serve as a secondary QC and reporting layer in preclinical CAR-T pipelines rather than as an ad hoc downstream visualization script.

2. Scope, assay types, and preprocessing assumptions

The stated scope of CARTEpigenoQC includes scATAC-seq, scCUT&Tag, and scBS-seq in CAR-T settings. At the same time, the strongest concrete examples are ATAC-like fragment-based workflows, because the implemented features described in detail—FRiP, fragment overlap with peaks, and per-cell insertion-site coverage—are most naturally formulated for fragment-level chromatin data. The paper therefore presents a broad intended scope but a practically strongest fit to fragment-centric assays (Lai, 30 Jul 2025).

CARTEpigenoQC operates after basic alignment and preprocessing have already been completed. It is not presented as a raw-read aligner or a complete primary processing pipeline. The workflow assumes that the user already has a fragment file, a set of peak regions, a BED file of CAR insertion coordinates or CAR-related target loci, cell barcodes embedded in the fragment file, and a consistent reference genome coordinate system across all genomic inputs. The strongest implied assumptions are that reads have already been aligned to a reference genome, duplicates and mapping status have already been resolved sufficiently to produce fragment calls, peaks are available from an upstream caller or standard pipeline, and CAR insertion or target loci have been curated into BED coordinates.

The package’s format flexibility is centered on the fragment table abstraction rather than a fixed vendor-specific preprocessing stack. The manuscript explicitly lists compatibility with 10x Genomics fragments.tsv.gz files, ArchR fragments, and BED-formatted CAR insertion site annotations. It also states support for 10x and non-10x data formats, which suggests that non-10x data are admissible when they can be represented in a fragment-like form with genomic intervals and cell barcodes.

3. Inputs, data model, and implemented workflow

The implemented workflow can be reconstructed from the feature list, example usage, and example outputs. The user supplies fragment-level single-cell epigenomic data, loads peak annotations for FRiP calculation, and loads a BED file of CAR insertion or target site annotations. The annotation examples shown in the paper include labels such as TRAC_CAR_Eyquem2017, AAVS1_CAR_Common, CCR5_CAR_Target, RAB11A_CAR_Target, and B2M_KO_CAR_Target, indicating that the annotation model is user-specified and label-based rather than a fixed built-in ontology (Lai, 30 Jul 2025).

Input category	Explicitly stated format or example	Role
Fragment data	fragments.tsv.gz, ArchR fragments	Per-cell fragment representation
Peak annotations	Peak regions	FRiP calculation
CAR-site annotations	BED-formatted CAR insertion sites	CAR-site overlap and coverage

After loading these inputs, the toolkit computes per-cell standard QC metrics and then overlaps fragments with CAR-site BED intervals. This yields per-cell read or fragment counts at each CAR-related site and total coverage per site across all cells. These quantities are then summarized as a heatmap of per-cell CAR insertion site coverage and a ranked table of top CAR sites by total supporting reads.

The workflow culminates in a report-generation step. The report is described as Markdown-based, rendered to self-contained HTML, and accompanied by PNG summary outputs. The paper also implies a simple software architecture consisting of an R-based QC engine, a top-level script called run_qc.R, and a report template that renders a self-contained report. The manuscript does not list package dependencies explicitly; it states only that the tool is lightweight, fast, and does not require complex dependencies.

4. QC metrics and transgene-aware coverage assessment

The paper explicitly discusses or names FRiP, TSS enrichment, duplication, peak overlap, and CAR insertion site coverage. Of these, FRiP is operationally described in sufficient detail to reconstruct directly. The paper summary gives the per-cell expression

$$\mathrm{FRiP}_i = \frac{N^{(i)}_{\mathrm{fragments\ in\ peaks}}}{N^{(i)}_{\mathrm{total\ fragments}}}$$

or, equivalently, if reads rather than fragments are counted,

$$\mathrm{FRiP}_i = \frac{N^{(i)}_{\mathrm{reads\ in\ peaks}}}{N^{(i)}_{\mathrm{total\ mapped\ reads}}}.$$

The example output table includes columns for Total, In Peaks, and FRiP, and the manuscript notes that high-quality scATAC-seq data are expected to show a FRiP distribution peaking around $0.2$–$0.4$, with cells satisfying $\mathrm{FRiP} < 0.2$ as plausible downstream filtering candidates (Lai, 30 Jul 2025).

TSS enrichment and duplication are mentioned as part of the conceptual standard QC layer, but the paper does not provide a detailed implementation section for either metric. Peak overlap is conceptually embedded within FRiP. The primary extension beyond generic chromatin QC is the engineered-locus-aware overlap analysis. A reconstructed expression in the paper synthesis for per-cell CAR-site coverage is

$$C_{i,s} = \sum_{f \in F_i} \mathbf{1}\{f \cap I_s \neq \emptyset\},$$

where $F_i$ is the set of fragments from cell $i$ and $I_s$ is the genomic interval for CAR site $s$. Total support for a site across all cells is then

$$\mathrm{FRiP}_i = \frac{N^{(i)}_{\mathrm{reads\ in\ peaks}}}{N^{(i)}_{\mathrm{total\ mapped\ reads}}}.$$0

This overlap-based design has a specific interpretive consequence: CARTEpigenoQC assesses coverage at expected genomic target regions, not de novo vector-genome junctions. It is therefore a target-locus support assessment rather than a structural variant caller or insertion breakpoint detector. This distinction is central to correct use. The paper does not provide sensitivity or specificity benchmarking, formal classifier cutoffs for declaring a cell “CAR-positive,” or numeric thresholds for CAR-site support. CAR-site signal is instead presented as evidence display and QC interpretation, to be read jointly with standard assay-quality metrics.

5. Outputs, visual summaries, and reporting uses

The report includes a histogram of FRiP score distribution across cells, a ranked table of top cells by FRiP, a heatmap of CAR insertion site coverage across individual cells, and a ranked summary table of top CAR insertion sites by total read count. The FRiP histogram is used to summarize the distribution of signal enrichment across the dataset; the manuscript notes that such distributions are often right-skewed and that a peak around $$\mathrm{FRiP}_i = \frac{N^{(i)}_{\mathrm{reads\ in\ peaks}}}{N^{(i)}_{\mathrm{total\ mapped\ reads}}}.$$1–$$\mathrm{FRiP}_i = \frac{N^{(i)}_{\mathrm{reads\ in\ peaks}}}{N^{(i)}_{\mathrm{total\ mapped\ reads}}}.$$2 is expected for high-quality scATAC-seq. The top-FRiP table contains barcode identifier, total reads or fragments, in-peaks reads or fragments, and FRiP score (Lai, 30 Jul 2025).

Output	Reported purpose	Format
FRiP score distribution histogram	Dataset-wide assay-quality overview	HTML, PNG
Top cells by FRiP	Ranked high-signal barcodes	HTML table
CAR insertion site coverage heatmap	Per-cell engineered-locus visualization	HTML, PNG
Top CAR insertion sites by total coverage	Sample-level site ranking	HTML table

The CAR insertion heatmap is the distinctive visualization of the package. One axis corresponds to cell barcodes, the other to CAR insertion or target sites, and color intensity reflects read count or coverage. The paper states that this plot enables visual identification of cells likely to carry vector insertion signal and can help reveal mosaicism in editing. The ranked site table, by contrast, operates at sample level and identifies which loci are most represented across all cells.

The report-centric design is fundamental to the toolkit’s intended placement in preclinical workflows. Because the outputs are self-contained HTML plus PNG summaries, the package is positioned for both interactive review and archival documentation. The paper explicitly links these outputs to exploratory analysis, formal reporting, and regulatory-grade preclinical documentation.

6. Relation to adjacent sparse single-cell QC frameworks

CARTEpigenoQC is specialized for CAR-T epigenomic datasets, but its conceptual structure has affinities with earlier work on sparse single-cell molecular measurements. McDavid et al. developed a statistical framework for single-cell qPCR in which non-detection is modeled as a true zero state and positive measurements are treated separately from detection itself; their QC framework removes null wells, uses robust detection-rate outlier scores, and distinguishes technical failure from biologically meaningful zeros (McDavid et al., 2012). This suggests a broader methodological context in which CARTEpigenoQC’s separation between ordinary assay quality and CAR-site signal can be understood as a domain-specific instance of multi-component QC for sparse single-cell assays.

A related connection arises on the methylation side. The Bayesian epiallele detection framework for tumour RRBS models read-pattern heterogeneity through latent epialleles, an explicit per-CpG noise parameter $$\mathrm{FRiP}_i = \frac{N^{(i)}_{\mathrm{reads\ in\ peaks}}}{N^{(i)}_{\mathrm{total\ mapped\ reads}}}.$$3, posterior assignment uncertainty, and contamination-aware mixture modeling (Barrett et al., 2017). CARTEpigenoQC does not implement that model, but the comparison is instructive because the paper explicitly positions CARTEpigenoQC for scBS-seq as well as fragment-based assays. A plausible implication is that future CAR-T epigenomic QC could combine CAR-site-aware reporting with read-pattern-level methylation coherence metrics in settings where single-cell methylome data are sufficiently phased and covered.

These connections should not be overstated. CARTEpigenoQC, as described, is a lightweight R-based QC and reporting layer centered on fragment overlaps, standard chromatin QC summaries, and user-defined CAR-site coverage. It is not a hurdle-modeling framework in the style of single-cell qPCR inference, nor a Bayesian epiallele caller. The value of these adjacent methods lies primarily in clarifying the broader analytical problem: sparse single-cell data often require QC procedures that preserve biologically meaningful zeros while still identifying globally poor cells or poorly interpretable molecular patterns.

7. Limitations, assumptions, and interpretive boundaries

Several limitations are explicit or strongly implied. First, the CAR-specific analysis depends on user-supplied BED annotations for insertion or target loci. If the BED file is incomplete, incorrect, or based on the wrong genome build, the CAR-site QC layer becomes misleading. Second, the toolkit appears to quantify support at predefined target regions rather than direct vector-genome breakpoint structure. It therefore does not substitute for integration-site sequencing, structural variant analysis, or chimeric read or junction detection (Lai, 30 Jul 2025).

Third, although the package is positioned for scATAC-seq, scCUT&Tag, and scBS-seq, the implemented examples are most natural for fragment-based accessibility data. For assays such as scBS-seq, the paper indicates intended compatibility provided the user can supply the relevant genomic intervals and read or fragment representations, but it does not provide assay-specific benchmarking across platforms. Fourth, the paper does not report runtime benchmarks, precision or recall for CAR-positive detection, or formal comparative evaluation against ArchR, Signac, or SnapATAC. The validation is therefore best described as problem-solution justification plus illustrative outputs rather than rigorous benchmarking.

Interpretation of CAR-site coverage is correspondingly qualitative. The manuscript recommends inspecting per-cell heatmaps and total site counts jointly with ordinary QC metrics and expected target loci. In edge cases—such as low-depth datasets, poorly called peak sets, mixed engineered and non-engineered products, or targeted insertion into loci with weak local chromatin accessibility—a technically good but biologically ambiguous profile may occur. This suggests that CARTEpigenoQC is most effective when used as a structured evidence layer within a larger experimental and preprocessing context, rather than as an autonomous classifier of engineering success.