Parallel & PRISM: Dual cfRNA Fragmentomics Pipelines
In cell-free RNA (cfRNA) fragmentomics, circulating molecules in plasma or extracellular vesicles carry information not only in their steady-state abundance, but in their cleavage coordinates, fragment length distributions, and protection states.
Standard bulk RNA-seq workflows treat reads as homogenous count matrices and discard positional and topological patterns. Furthermore, circulating RNA pools contain degraded fragments, heavily modified non-coding species, and non-templated additions that fail standard alignment against linear reference genomes.
To address these challenges, I built two complementary pipeline architectures:
- Parallel Pipeline (Parallel v2.0): A reference-anchored positional fragmentomics framework.
- PRISM Pipeline (PRISM v1.0): An alignment-free, poly-modal sequence discovery framework using k-mers and compacted de Bruijn graphs.
1. Parallel Pipeline (Parallel v2.0): Reference-Anchored Positional Fragmentomics
Parallel v2.0 shifts the analytical focus from whole-transcript counts to single-nucleotide cleavage coordinates across biological RNA domains. It uses the ShardMap hardware-accelerated mapping framework integrated with dual Bowtie and STAR alignment engines.
┌───────────────────────────────────────────────┐
│ Filtered FASTQ Reads (Reamp R1, 15–100 nt) │
└───────────────────────┬───────────────────────┘
│
┌───────────────────────▼───────────────────────┐
│ Exclusive Domain Cascade Alignment │
│ 1. Small RNA (miRNA via miRge3, tRNA, sno) │
│ 2. Long RNA (GENCODE v50 exons & lncRNAs) │
│ 3. Microbial QC (SILVA 138.2, post-human) │
└───────────────────────┬───────────────────────┘
│
┌───────────────────────▼───────────────────────┐
│ Immutable Positional Event Ledgers │
│ Unit biological mass (w <= 1) per read │
└───────────────────────┬───────────────────────┘
│
┌──────────────────────┴──────────────────────┐
▼ ▼
┌─────────────────────────────────────────┐ ┌─────────────────────────────────────────┐
│ Single-Base Cleavage Profiling │ │ Distribution Shape Metrics │
│ Exact 5' and 3' cut endpoints │ │ Total Variation (TV) distance (primary) │
│ BedGraph coordinate ledgers (w <= 1) │ │ Wasserstein 1D sensitivity check │
└────────────────────┬────────────────────┘ └────────────────────┬────────────────────┘
│ │
└──────────────────────┬──────────────────────┘
│
┌───────────────────────▼───────────────────────┐
│ Statistical Testing & Dispersion Guardrails │
│ PERMANOVA with PERMDISP checks │
│ Group-blind peak detection │
│ Within-domain Benjamini-Hochberg FDR control │
└───────────────────────────────────────────────┘
Key Architecture Principles:
- Exclusive Domain Cascade Alignment: Reads (specifically short inserts of 15–100 nt) are routed sequentially through an exclusive hierarchy:
- Small RNAs: Mature miRNA (with authoritative miRge3 and miRBase v22.1 assignment), tRNA, snRNA, snoRNA, and piRNA.
- Long RNAs: GENCODE v50 protein-coding exons and lncRNAs, mapped along native transcript axes (MANE Select or longest transcript models).
- Quality Control: Microbial rRNA (SILVA 138.2) kept strictly post-human to evaluate technical contamination without read stealing.
- Immutable Positional Event Ledgers: Every mapped read is assigned a unit biological mass (weight ≤ 1) and recorded in coordinate-resolved ledgers, guaranteeing mass conservation across multi-mapping regions.
- Cleavage & Shape Profiling:
- Evaluates exact 5' and 3' cleavage cut sites at single-base resolution (BedGraph and ledger formats).
- Normalizes positional coverage distributions per transcript.
- Measures shape divergence between clinical cohorts using non-parametric distance metrics: Total Variation distance as primary, with Wasserstein 1D as sensitivity verification.
- Statistical Testing & Dispersion Guardrails:
- Employs PERMANOVA on shape distance matrices coupled with PERMDISP checks to confirm that significant differences reflect true location shifts rather than unequal within-group variance.
- Identifies group-blind cleavage endpoint peaks with strict within-domain Benjamini–Hochberg false discovery rate control.
Primary Use Cases:
- Positional Biomarker Discovery: Pinpointing disease-associated cleavage motifs or fragment length shifts in known transcripts that are undetectable by conventional gene expression counting.
- RNase Degradation & Footprinting: Mapping which transcript segments remain protected in circulation through ribonucleoprotein complexes or extracellular vesicle encapsulation.
2. PRISM Pipeline (PRISM v1.0): Reference-Free Poly-Modal Discovery
PRISM (Poly-modal Reference-free Identification of Signatures & Markers) is an alignment-free pipeline. Rather than mapping against a reference genome, PRISM operates directly on raw sequence space using k-mers and compacted de Bruijn graphs (cDBG).
┌───────────────────────────────────────────────┐
│ Input Reads (FASTQ / Residual Unmapped) │
└───────────────────────┬───────────────────────┘
│
┌───────────────────────▼───────────────────────┐
│ k-mer Decomposition (KMC3: k=21, k=31) │
└───────────────────────┬───────────────────────┘
│
┌──────────────────────┴──────────────────────┐
▼ ▼
┌─────────────────────────────────────────┐ ┌─────────────────────────────────────────┐
│ Branch A: Count Reduction & Clustering │ │ Branch B: Topological Graph Association │
│ - Abundance matrix filtering │ │ - Compacted de Bruijn graph (cDBG) │
│ - KaMRaT association ranking & contigs │ │ - Unitig phenotype testing (edgeR QL) │
│ - Linclust/MMseqs2 sequence clustering │ │ - Bubble extraction (splicing & loops) │
└────────────────────┬────────────────────┘ └────────────────────┬────────────────────┘
│ │
(Consensus Centroids) (Unitigs & Bubbles)
│ │
└──────────────────────┬──────────────────────┘
│
┌───────────────────────▼───────────────────────┐
│ Dual-Branch Convergence & DP Path Stitching │
│ Consolidated Golden Biomarker Signatures │
└───────────────────────┬───────────────────────┘
│
┌──────────────────────┴──────────────────────┐
▼ ▼
┌─────────────────────────────────────────┐ ┌─────────────────────────────────────────┐
│ Retrospective Biological Annotation │ │ Cross-Validated Machine Learning │
│ - GENCODE v50 protein-coding & lncRNAs │ │ - Nested CV generalizability framework │
│ - tRF databases, RepeatMasker, SILVA QC │ │ - ElasticNet & Random Forest ranking │
└─────────────────────────────────────────┘ └─────────────────────────────────────────┘
Dual-Branch Convergence Architecture:
- Branch A (Count Reduction & Clustering):
- High-throughput k-mer extraction using KMC3 at defined strata (k=21 and k=31).
- Abundance filtering and statistical association testing using KaMRaT to merge co-occurring k-mers into contiguous sequence fragments.
- Linear-time sequence clustering via Linclust and MMseqs2 with dynamic programming verification to generate non-redundant consensus centroids.
- Branch B (Topological Graph Association):
- Constructs compacted de Bruijn graphs directly from sample read pools.
- Tests unitigs for phenotypic association (GLM / edgeR quasi-likelihood) and extracts neighborhood subgraphs to isolate alternative splicing loops and sequence variation bubbles.
- Dual-Branch Convergence & Retrospective Annotation:
- Stitches contigs and unitigs via dynamic programming path alignment into consolidated biomarker signatures.
- Retrospectively annotates identified signatures against reference databases (GENCODE, tRF databases, RepeatMasker, and SILVA).
- Evaluates predictive performance via nested cross-validation (ElasticNet, Random Forest).
Primary Use Cases:
- Circulating "Dark Matter" Discovery: Capturing biological signals from reads that fail linear alignment due to dense post-transcriptional modifications (such as 1-methyladenosine [m1A] or 1-methylguanosine [m1G] in tRNA loops), non-templated nucleotide additions, or structural fusions.
- Unbiased Biomarker Identification: Identifying unannotated non-coding transcripts or microbial cfRNA fragments without reference bias.
Side-by-Side Comparison
| Feature | Parallel Pipeline (v2.0) | PRISM Pipeline (v1.0) |
|---|---|---|
| Analytical Paradigm | Reference-anchored alignment | Reference-free sequence discovery |
| Input Material | Filtered FASTQ (Reamp R1, 15–100 nt) | Filtered FASTQ or residual unmapped reads |
| Core Data Structures | Coordinate ledgers, BAMs, BedGraph | k-mer count matrices, compacted de Bruijn graphs |
| Target Scope | Known transcripts, defined RNA domains, cut sites | Canonical RNAs, unannotated RNAs, modified loops, dark matter |
| Positional Resolution | Single-nucleotide 5' and 3' coordinates | Assembled k-mer contigs, unitigs, bubble subgraphs |
| Statistical Methods | Total Variation, PERMANOVA, PERMDISP | edgeR QL on unitigs, graph association, ML ranking |
| Primary Goal | Mechanistic analysis of RNA cleavage and fragment stability | Unbiased biomarker discovery from altered sequence space |
Implementation & Reproducibility
Both pipelines are implemented with reproducible workflow managers (Nextflow/Python), containerized via Docker and Apptainer, and optimized for high-performance computing clusters with hardware-accelerated mapping routines.