REDICAT

RNA Editing Cellular Assessment Toolkit

REDICAT is a high-performance Rust toolkit for RNA editing quantification in single-cell RNA-seq, with an emphasis on strand-aware mismatch logic, sparse-matrix efficiency, and reproducible end-to-end processing from indexed BAM/CRAM to analysis-ready AnnData (.h5ad).

Citation

If REDICAT supports your research, please cite:

Wei, T., Li, J., Lei, X., Lin, R., Wu, Q., Zhang, Z., Shuai, S., & Tian, R. (2025). Multimodal CRISPR screens uncover DDX39B as a global repressor of A-to-I RNA editing. Cell Reports, 44(7). https://doi.org/10.1016/j.celrep.2025.116009

Installation

# Clone and build
git clone https://github.com/aStudyingTurtle/redicat.git
cd redicat
cargo build --release

# Optional: install into Cargo's bin directory
cargo install --path .

1) Scientific overview

RNA editing introduces biologically meaningful RNA–DNA mismatches (for example A-to-I observed as A>G, and C-to-U observed as C>T). In single-cell data, these events are sparse and confounded by sequencing noise, alignment artifacts, and SNPs. REDICAT addresses this by:

• extracting per-cell, per-site strand-aware base counts from BAM/CRAM,
• applying biologically informed mismatch filters,
• generating reference/alternate/other sparse matrices,
• computing cell-level editing metrics (including CEI), and
• preserving metadata in standard AnnData format for downstream statistical analysis.

Innovation highlights

• Parallel genomic scheduling with bounded-memory streaming.
• UMI-level consensus deduplication to suppress conflicting molecules.
• Strand-aware editing inference (AG, CT, and related mismatch classes).
• Sparse linear-algebra-first implementation for scale.

2) Code architecture (from source-level review)

2.1 Top-level modules

• src/main.rs: CLI entry point with subcommands: bulk, bam2mtx, call.
• src/commands/*: user-facing command orchestration.
• src/lib/core/*: shared infrastructure (errors, IO, sparse ops, filtering).
• src/lib/engine/*: parallel genomic scheduler (par_granges) and position models.
• src/lib/pipeline/bam2mtx/*: BAM/CRAM → sparse base-count AnnData conversion.
• src/lib/pipeline/call/*: editing annotation/filtering/ref-alt matrix/CEI pipeline.

2.2 Runtime interaction logic

1. bulk (src/commands/bulk)

   • Uses ParGranges scheduler to iterate genomic chunks in parallel.
   • BaseProcessor performs pileup counting and emits per-position records.
   • Output is gzipped TSV (or bgzip-compatible path handling).

2. bam2mtx (src/commands/bam2mtx)

   • Reads position manifest TSV (or runs bulk first pass with --two-pass).
   • Splits positions into depth-aware chunks (special handling for near-max-depth loci).
   • OptimizedChunkProcessor performs per-chunk pileup traversal with:
     • mapping/base quality filtering,
     • barcode whitelist matching,
     • UMI consensus conflict suppression,
     • optional stranded counting.
   • AnnDataConverter assembles sparse layers (A1/T1/G1/C1 and optional A0/T0/G0/C0) and writes .h5ad.

3. call (src/commands/call + src/lib/pipeline/call)

   • Loads input AnnData and known editing-site whitelist.
   • Annotates candidate sites, computes coverage/base summaries, adds reference base from FASTA.
   • Applies mismatch classification thresholds.
   • Constructs strand-aware ref/alt/others matrices.
   • Computes cell-level CEI = alt / (ref + alt).
   • Computes site-level mismatch summaries and writes output .h5ad.

2.3 Biological mismatch logic implemented

EditingType supports: ag, ac, at, ca, cg, ct.

For each site, REDICAT uses strand-aware rules:

• expected reference bases on either strand,
• expected alternate base conditional on strand-consistent reference,
• rejection of sites with excess non-reference/non-alt support.

This operationalizes your stated biology:

• A-to-I: forward A>G with reverse-complement signal,
• C-to-U: forward C>T with reverse-complement signal, while retaining generalized mismatch classes for exploratory analyses.

3) Installation

3.1 Requirements

• Linux/macOS (validated workflow shown for Linux).
• Rust toolchain (recommended stable):
  • rustup, cargo, rustc.
• Typical native build dependencies for rust-htslib/compression stack:
  • pkg-config, clang, zlib, bzip2, xz, curl, OpenSSL dev headers.
• For CRAM input: reference FASTA should be available and indexed.

3.2 Build

git clone https://github.com/aStudyingTurtle/redicat.git
cd redicat
cargo build --release

Binary path:

./target/release/redicat

Quick check:

./target/release/redicat --help
./target/release/redicat bam2mtx --help
./target/release/redicat call --help
./target/release/redicat bulk --help

4) Input/output data contracts

4.1 BAM/CRAM prerequisites

• Coordinate-sorted and indexed alignment file.
• Cell barcode tag (default CB) and UMI tag (default UB) present if single-cell mode is expected.

4.2 bulk output schema (TSV.gz)

Serialized PileupPosition includes columns:

• CHR, POS, DEPTH, A, C, G, T, N, INS, DEL, REF_SKIP, FAIL, NEAR_MAX_DEPTH
• Optional REF_BASE when available.

POS is 1-based by default (0-based with --zero-base).

4.3 Position manifest required by bam2mtx

bam2mtx expects at least these columns in TSV:

• CHR, POS, DEPTH, INS, DEL, REF_SKIP, FAIL, NEAR_MAX_DEPTH

The easiest way to produce a valid manifest is --two-pass, which internally runs bulk.

4.4 Site whitelist required by call

• Tab-separated file, plain or gzipped.
• First two columns must be chromosome and position (CHR, POS convention).
• Key format used internally: CHR:POS matching AnnData.var_names.

4.5 call output

• .h5ad with layers (priority written):
  • ref, alt, others, coverage (when present)
• obs additions:
  • ref, alt, others (row sums over filtered editing sites)
  • CEI
• var additions:
  • is_editing_site, ref, Mismatch, filter_pass
  • site-level columns: <XY>_ref, <XY>_alt, <XY>_others for selected editing type X>Y

bam2mtx also writes a sidecar file for capped-depth loci:

• <output_stem>_skipped_sites.txt.

5) Command-line usage

5.1 bulk (first-pass site discovery / pileup profiling)

redicat bulk input.bam \
  --output first_pass.tsv.gz \
  --threads 16 \
  --chunksize 100000 \
  --mapquality 255 \
  --min-depth 5

Use --all to emit all covered positions (not only editing-enriched candidates).

5.2 bam2mtx (BAM to sparse single-cell base matrices)

Two-pass (recommended)

redicat bam2mtx \
  --bam input.bam \
  --barcodes barcodes.tsv.gz \
  --output base_counts.h5ad \
  --two-pass \
  --threads 16 \
  --stranded

Single-pass with precomputed manifest

redicat bam2mtx \
  --bam input.bam \
  --tsv first_pass.tsv.gz \
  --barcodes barcodes.tsv.gz \
  --output base_counts.h5ad \
  --threads 16

5.3 call (editing annotation and quantification)

redicat call \
  --input base_counts.h5ad \
  --output editing_calls.h5ad \
  --fa reference.fa \
  --site-white-list rediportal_sites.tsv.gz \
  --editingtype ag \
  --max-other-threshold 0.1 \
  --min-edited-threshold 0.001 \
  --min-ref-threshold 0.001 \
  --min-coverage 2 \
  --threads 16

Dry-run validation:

redicat call --input base_counts.h5ad --output out.h5ad --fa reference.fa --site-white-list sites.tsv.gz --dry-run

6) Parameter reference and tuning guidance

6.1 bulk parameters

6.2 bam2mtx parameters

6.3 call parameters

6.4 Mismatch decision rule used in call

For each site with coverage $C$:

• $other_max = \max(\lceil C \cdot max_other_threshold \rceil, 2)$
• $edited_min = \max(\lceil C \cdot min_edited_threshold \rceil, 1)$
• $ref_min = \max(\lceil C \cdot min_ref_threshold \rceil, 1)$

A site passes if:

1. reference base is valid for selected editing type,
2. reference count $\ge ref_min$,
3. edited count $\ge edited_min$,
4. sum of other bases $\le other_max$,
5. site coverage $\ge min_coverage$.

7) Reproducible workflow (recommended)

1. Generate candidate loci
   • redicat bulk with stringent quality filters.
2. Build base-count matrix
   • redicat bam2mtx --two-pass (or provide validated TSV).
3. Call editing signal
   • redicat call with matched reference build and curated whitelist.
4. Post-analysis
   • Use obs['CEI'], var['filter_pass'], and <XY>_alt/<XY>_ref for downstream statistics.

For publication-grade reproducibility, record:

• full command lines,
• software version (redicat --version),
• reference genome build and checksum,
• whitelist provenance and version,
• all threshold parameters.

8) Result interpretation

8.1 Cell-level metrics

• ref, alt, others in obs: aggregated counts across retained sites.
• CEI: cell editing index, interpreted as editing burden proxy:

$$ CEI = \frac{alt}{ref + alt} $$

Practical guidance:

• Compare CEI across cell states/conditions with coverage-aware covariates.
• Exclude cells with very low total informative counts (ref + alt) before differential analyses.

8.2 Site-level metrics

In var, fields such as AG_ref, AG_alt, AG_others (for editingtype=ag) summarize cohort-level support at each site.

Recommended biological interpretation:

• prioritize sites with high *_alt, low *_others, and filter_pass=true;
• remove known SNP loci (dbSNP/gnomAD) in downstream curation;
• inspect strand consistency for expected editing class behavior.

9) Performance and complexity

Let:

• $R$ = number of aligned reads inspected,
• $P$ = number of queried positions,
• $U$ = number of unique (cell, UMI, site) molecules after filtering,
• $nnz$ = non-zeros in sparse matrices,
• $T$ = worker threads.

9.1 bulk

• Pileup traversal is approximately $O(R_{region})$ over processed chunks.
• Parallel scheduling reduces wall-time to roughly $\sim O(R/T)$ in balanced workloads.

9.2 bam2mtx

• Site processing: $O(R_{sites})$ for pileup reads intersecting selected loci.
• UMI consensus hash aggregation: expected $O(U)$.
• Sparse assembly: near $O(nnz)$ plus sorting/merge costs for triplet consolidation.

9.3 call

• Coverage/base summaries and sparse reductions: near $O(nnz)$.
• Site annotation/filtering/reference lookup: $O(P)$ (with caching effects for FASTA access).
• Overall dominated by sparse matrix operations and site count.

In practice, REDICAT is throughput-optimized for sparse single-cell matrices with multi-core scaling and bounded-memory spill strategies in matrix assembly.

10) Notes and current implementation caveats

• Canonical contig filtering defaults to chr1..chr22, chrX, chrY, chrM; use --allcontigs for non-canonical references.
• call writes priority layers (ref, alt, others, coverage) and annotations; preserve intermediate matrices if additional layers are required for custom methods.