Packages for a methylation–fragmentomics disease fingerprint

A pragmatic open-source stack for EM-seq, fragment features, pathway analysis, integration, and future progression models

Best means fit for purpose

No single package fingerprints disease progression. The strongest stack separates read processing, methylation inference, fragment feature extraction, biological annotation, statistical modeling, and validation. For this project, transparent Bioconductor and workflow tools are preferable to a large black-box model.

A reproducible stack carries nucleosome-protected cfDNA through methylation sequencing to interpretable features.

Recommended core stack

Job	Preferred tool	Why it fits	Important caution
Reproducible orchestration	Nextflow + containers	Records dependencies and scales from workstation to compute cluster	Pin container digests and reference files
Read QC	fastp + MultiQC	Adapter/quality processing plus cohort-wide QC review	Do not hide group- or batch-specific QC failures
EM-seq alignment	bwa-meth	Established bisulfite-aware alignment route used in the planned pipeline	Validate compatibility and conversion metrics for the exact EM-seq protocol
Methylation extraction	MethylDackel	Produces per-cytosine methylation calls from aligned reads	Review overlap, duplicate, strand, and bias handling
DML/DMR inference	DSS	Coverage-aware beta-binomial modeling and multi-factor designs	Small samples still limit covariate complexity
Methylation objects/QC	bsseq	Standard count/coverage representation, smoothing, and visualization	Smoothing choices affect region calls
Alternative exploratory methylation workflow	methylKit	Accessible filtering, QC, differential methylation, and tabix-backed data	Pre-specify why it is used alongside or instead of DSS
Fragment ends	FrEIA	Open pipeline for fragment-end sequence abundance and diversity	Built around cfDNA WGS; benchmark carefully on EM-seq
Generic BAM features	samtools + deepTools + R/Bioconductor	Transparent insert-size, coverage, window, and TSS summaries	Define proper-pair, duplicate, MAPQ, and blacklist rules
Genomic annotation	GenomicRanges + annotatr	Reproducible overlap of CpGs/DMRs with genes and regulatory regions	Record genome build and many-to-many mappings
Pathway analysis	clusterProfiler	KEGG/GO workflows with explicit gene universes	Use the tested-region background, not every human gene
Penalized fingerprint model	glmnet via tidymodels	Elastic net is interpretable and suitable for correlated predictors	Tune and select features inside nested resampling
Model interpretation	DALEX or vip	Auditable variable-importance and diagnostics	Importance does not establish causality

Recommended architecture for this cohort

flowchart TD
  A["FASTQ"] --> B["fastp · bwa-meth<br/>MethylDackel"]
  B --> C["DSS / bsseq<br/>CpGs + DMRs"]
  B --> D["samtools · deepTools<br/>FrEIA benchmark"]
  C --> E["GenomicRanges<br/>clusterProfiler"]
  D --> F["Locked fragment<br/>feature table"]
  E --> G["Small pathway/DMR<br/>feature table"]
  H["sIgA + inflammation<br/>clinical data"] --> I["Participant-level<br/>analysis table"]
  F --> I
  G --> I
  I --> J["glmnet + tidymodels<br/>nested resampling"]

What to use now versus later

Use now

• DSS as the primary coverage-aware methylation model.
• bsseq/GenomicRanges for storage, visualization, and annotation.
• Simple BAM-derived fragment metrics with explicit definitions.
• clusterProfiler only after defining the correct tested-feature universe.
• glmnet only for a deliberately small exploratory integrated signature.

Benchmark before adopting

• FrEIA, because its assumptions and preprocessing were developed for cfDNA WGS rather than enzymatically converted libraries.
• Tissue-deconvolution tools such as EpiDISH, because available reference matrices may be platform-, tissue-, or array-specific and may not transfer to plasma EM-seq.
• FinaleMe/FRAGMA implementations, because the cited methods demonstrate important principles but are not automatic replacements for direct EM-seq methylation calls.

Add when longitudinal outcomes exist

• lme4 or glmmTMB for repeated continuous or generalized outcomes;
• survival for time-to-event baselines;
• JMbayes2 for joint longitudinal–survival models; and
• rsample with participant/group-aware splits for honest temporal or external validation.

Non-negotiable safeguards

1. Keep participant IDs grouped across all resampling splits.
2. Never select DMRs on the full dataset before cross-validation.
3. Fit normalization, imputation, and scaling on training folds only.
4. Preserve raw counts and coverage; do not model methylation percentages without depth awareness.
5. Record software versions, reference genome checksum, annotations, pathway database date, and parameters.
6. Report uncertainty, calibration, and feature stability alongside discrimination.

NotePractical recommendation

Start with a containerized Nextflow workflow producing one reviewed participant-level feature table. Use DSS for the primary methylation inference, transparent BAM summaries for first-pass fragmentomics, and elastic net only after reducing the feature space through pre-specified biology—not outcome-driven screening.

Follow the molecular pipeline → · See how the software stack evolved → · Understand the statistical procedures →