Summary

Add Rp-Bp as a Tier-1 opt-in ORF caller. Benchmarked alongside RiboCode as the recommended two-caller default combination: RiboCode is permissive and fast; Rp-Bp is Bayesian-strict and slow. Activated by --run_rpbp (default false).

Blocked by: #163 (ribotricer removal must complete first).

Benchmark results

From the nf-core/riboseq benchmark (Krueger + García Bediaga, May 2026; 6 biological replicates, genome-wide):

• Mean Spearman replicate concordance: 0.893 (Tier-1)
• Mean Jaccard set overlap: 0.673 (Tier-1)
• uORF/dORF ratio: 3.4–5.2× (biology-correct)
• Start-codon composition: 99.94% AUG
• ORF volume: ~18k–26k per replicate (comparable to RiboCode's ~22k)
• Runtime: 19h38m–24h11m per replicate (Bayesian MCMC fit dominates)

Software

• Bioconda: rpbp=4.0.1 (released 2024-11-21) — conda install -c bioconda rpbp
• PyPI: rpbp==4.0.1 (released 2024-11-20), Python 3.11–3.13
• GitHub: dieterich-lab/rp-bp, actively maintained (Etienne Boileau), not archived

Implementation

1. New nf-core module at modules/nf-core/rpbp/. Rp-Bp has two phases:

   • prepare-rpbp-genome (index build, run once per pipeline execution): takes genome FASTA + GTF, outputs a Rp-Bp config YAML and genome index
   • rpbp (per-sample run): takes BAM + config YAML, outputs BED/CSV of predicted ORFs with Bayes factor scores

2. Pipeline wiring in workflows/riboseq/main.nf:

   if (params.run_rpbp) {
       RPBP_PREPARE_GENOME(ch_fasta, ch_gtf)
       RPBP(ch_genome_bam, RPBP_PREPARE_GENOME.out.config)
   }

3. --rpbp_config parameter for the required YAML configuration file (genome paths, read-length range, minimum read count per length class). Provide a helper script or pipeline step to auto-generate this from pipeline inputs rather than requiring manual YAML authoring.

4. Multi-caller merge: update the ORF caller output aggregation (issue feat: wire hybrid GTF (canonical + novel intergenic) into ORF callers #165) to include Rp-Bp calls when --run_rpbp is active.

5. Documentation: note the ~24h/rep genome-wide runtime expectation prominently so users schedule accordingly.

Notes

• prepare-rpbp-genome check_programs_exist requires STAR on PATH even when not running alignment — account for this in the Nextflow process environment.
• Memory: comfortable with default Python/Stan settings at genome-wide scale (no -Xmx flags needed, unlike PRICE).
• The Bayesian fit is the bottleneck, not I/O — runtime scales with read depth, not genome complexity.

References

• Malone et al. (2017) PLOS Comput Biol. DOI: 10.1371/journal.pcbi.1005926
• Bioconda recipe
• dieterich-lab/rp-bp