Immunopeptidogenomics — cryptic peptide database construction from RNA-seq. An nf-core-style port of the 31-step bash pipeline used by the Li Lab and Purcell Lab at Monash University, implementing the cryptic peptide discovery method described in Scull et al. Mol Cell Proteomics 2021.
sanjaysgk/ipg turns paired-end RNA-seq reads into a cryptic peptide search database
suitable for downstream MS/MS immunopeptidomics searches. Starting from FASTQs, it:
- Aligns reads with STAR (two-pass) and infers strandedness with RSeQC.
- Assembles novel transcripts with StringTie and reconciles them with the reference
annotation via gffcompare (
-R -V -Cfor the consensus combined GTF). - Builds a variant-calling-ready BAM via the GATK4 RNA-seq Best Practices chain (FastqToSam → MergeBamAlignment → MarkDuplicates → SplitNCigarReads).
- Recalibrates base qualities with two passes of BaseRecalibrator + ApplyBQSR.
- Calls somatic variants with Mutect2 in tumour-only mode against a gnomAD-style germline allele-frequency database, then CalculateContamination / FilterMutectCalls / SelectVariants for clean PASS-only sites-only VCFs.
- Builds the cryptic peptide FASTA database via the IPG custom C tools
(
curate_vcf,revert_headers,alt_liftover,triple_translate,squish) plusgff3sortandgffreadfrom bioconda.
Optionally, a post-MS analysis step (--step post_ms) runs the two-phase
db_compare + origins workflow on PEAKS (or other search engine) results to
identify and annotate cryptic-only peptides.
An MS search step (--step ms_search) runs up to four open-source search
engines (MSFragger, Comet, Sage, PEAKS) in parallel against the
cryptic FASTA, applies mokapot FDR control per engine, rescores PSMs with
MS2Rescore, and merges results at 1% peptide-level FDR. Optional
immunoinformatics gates — --run_netmhcpan, --run_netmhciipan,
--run_gibbscluster, --run_flashlfq, --run_blastp_host — pick up the
integrated peptide table and emit a per-sample HTML report summarising HLA
binding, motif clusters, quantification, and host-background hits. See
docs/usage.md for the full invocation.
The 31 legacy steps span three pipeline modes (--step db_construct | ms_search | post_ms) selected per run; the metro map below colours each mode as a separate transit line.
Diagram source:
docs/images/ipg_overview.mmd— re-render withpixi run nf-metro render docs/images/ipg_overview.mmd -o docs/images/ipg_overview.svg.
The repository ships a pixi project that pins every tool
(nextflow 25.10.4, nf-core 3.5.2, nf-test 0.9.5, GATK4, STAR, samtools, bcftools,
stringtie, gffcompare, gffread, RSeQC, FastQC, MultiQC, OpenJDK 17, plus the
linting toolchain) to bit-for-bit reproducible versions via pixi.lock.
git clone https://github.com/sanjaysgk/ipg.git
cd ipg
pixi installIf you don't have pixi: curl -fsSL https://pixi.sh/install.sh | bash.
pixi run bash scripts/build_test_bundle.shBy default this writes to /fs04/scratch2/xy86/sanjay/ipg-test-data/. Override
on a non-Monash machine:
TEST_BUNDLE_DIR=/some/path \
REFERENCE_DIR=/path/to/GRCh38 \
VARIANT_DIR=/path/to/variant-resources \
FASTQ_R1=/path/to/sample_R1.fq.gz \
FASTQ_R2=/path/to/sample_R2.fq.gz \
pixi run bash scripts/build_test_bundle.shThe bundle is never committed to git — only the build script is. See
scripts/build_test_bundle.sh for the exact recipe.
| Profile combination | Container engine | Use when |
|---|---|---|
-profile test,pixi |
none — tools from local PATH | Fastest iteration on a workstation/HPC compute node where pixi can run all tools natively |
-profile test,singularity |
apptainer / singularity | Standard HPC; pulls biocontainers from community.wave.seqera.io |
-profile test,docker |
docker | Laptops, cloud, GitHub Actions CI |
-profile test,monash,singularity |
singularity | Monash M3 SLURM cluster (comp partition, xy86 account) |
# Fastest (no containers, all tools from pixi):
pixi run nextflow run . -profile test,pixi --outdir resultsThe pipeline takes ~2 minutes end-to-end against the chr22 test bundle on a
single Linux box and produces the cryptic peptide database at
results/squish/<sample>_cryptic.fasta.
Create a samplesheet samplesheet.csv:
sample,fastq_1,fastq_2,strandedness
D100_liver,/path/to/D100-liver_R1.fastq.gz,/path/to/D100-liver_R2.fastq.gz,reverse
D101_pancreas,/path/to/D101-pancreas_R1.fastq.gz,/path/to/D101-pancreas_R2.fastq.gz,reverseThen run with explicit reference paths:
pixi run nextflow run . \
-profile monash,singularity \
--input samplesheet.csv \
--outdir /fs04/scratch2/xy86/sanjay/ipg-results \
--fasta /path/to/GRCh38.primary_assembly.genome.fa \
--fasta_fai /path/to/GRCh38.primary_assembly.genome.fa.fai \
--fasta_dict /path/to/GRCh38.primary_assembly.genome.dict \
--gtf /path/to/gencode.v44.primary_assembly.annotation.gtf \
--star_index /path/to/human_genome_index_GRCh38 \
--rseqc_bed /path/to/gencode_assembly.bed \
--dbsnp /path/to/dbsnp138.vcf \
--dbsnp_tbi /path/to/dbsnp138.vcf.idx \
--known_indels /path/to/Homo_sapiens_assembly38.known_indels.vcf.gz \
--known_indels_tbi /path/to/Homo_sapiens_assembly38.known_indels.vcf.gz.tbi \
--mills /path/to/Mills_and_1000G_gold_standard.indels.hg38.vcf.gz \
--mills_tbi /path/to/Mills_and_1000G_gold_standard.indels.hg38.vcf.gz.tbi \
--germline_resource /path/to/small_exac_common_3.hg38.vcf.gz \
--germline_resource_tbi /path/to/small_exac_common_3.hg38.vcf.gz.tbiBy default the cryptic peptide DB is built from the reference assembly only,
matching the legacy Scull et al. 2021 / D122_Lung run (verified empirically by
inspecting the legacy squish.log). Pass --include_variant_peptides true to
also fold the alt-reference variant-derived peptide branches (unmasked + indel)
into the final database:
pixi run nextflow run . -profile monash,singularity \
--input samplesheet.csv \
--include_variant_peptides true \
[other reference args][!IMPORTANT] >
--include_variant_peptidesdoes not switch the variant caller to matched tumour-normal mode. Variant calling is always performed in tumour-only Mutect2 mode against a gnomAD-style germline allele-frequency database; this pipeline does not support matched tumour-normal calling. The flag controls only whether the discovered variants get folded into the final cryptic peptide DB. Settrueonly when the sample is expected to harbour biologically meaningful somatic variants (e.g. tumour tissue, hypermutated cell lines, MMR-deficient samples). Leave atfalsefor normal tissue, cell lines, or any sample where variant peptides would mostly add noise. This is a pipeline-level flag — to mix modes for a heterogeneous cohort, run the pipeline twice.
After running the DB construction pipeline, the cryptic peptide FASTA is searched
against MS/MS data using PEAKS Online (or another search engine such as
MSFragger, Comet, or Sage). The resulting PSM CSVs are then analysed with the
two-phase db_compare + origins workflow
(Scull et al. 2021):
Phase 1: db_compare_v2.R → cryptic_only.txt
origins -s → origins_discard.txt + origins_unconventional.txt
Phase 2: db_compare_v2.R (with -j discard -u unconventional)
→ unambiguous_unconventional.txt
origins (full Ensembl mode) → deep origin annotation
Create a CSV with one row per sample:
sample,cryptic_psm_csv,uniprot_psm_csv,cryptic_decoy_score,uniprot_decoy_score
D122_liver,/path/to/D122_Liver_Cryptic_DB.db.psms.csv,/path/to/D122_Liver_Uniprot_DB.db.psms.csv,44.43,36.48
D101_heart,/path/to/D101_Heart_Cryptic_DB.db.psms.csv,/path/to/D101_Heart_Uniprot_DB.db.psms.csv,3.64,3.25The cryptic_decoy_score and uniprot_decoy_score columns are the -10lgP
decoy thresholds from the respective PEAKS searches.
pixi run nextflow run . -profile pixi,monash \
--step post_ms \
--post_ms_input post_ms_samplesheet.csv \
--uniprot_fasta /path/to/uniprotkb_human_canonical_isoform.fasta \
--transcriptome_fasta /path/to/D122_liver_transcriptome.fa \
--prefix_tracking /path/to/prefix.tracking \
--outdir results_post_msThe --transcriptome_fasta and --prefix_tracking files are outputs from the
DB construction step (gffread and gffcompare respectively). You can find them
in the pipeline output directory from the previous run.
results_post_ms/
├── post_ms/
│ ├── phase1/
│ │ ├── db_compare/ Phase 1 cryptic-only peptide lists + plots
│ │ └── origins/ Phase 1 origins (simple mode) — discard + unconventional lists
│ └── phase2/
│ ├── db_compare/ Phase 2 refined unambiguous unconventional peptides
│ └── origins/ Phase 2 origins (full Ensembl annotation)
└── pipeline_info/
results/
├── star/ STAR alignment BAMs + .Log.final.out
├── samtools/ sorted/indexed BAMs
├── rseqc/ strandedness inference reports
├── stringtie/ assembled transcript GTFs
├── gffcompare/ prefix.combined.gtf, prefix.tracking, .stats
├── gatk4/ Mutect2 / BQSR / contamination tables
├── revert/ alt-reference FASTAs (only when --include_variant_peptides=true)
├── gff3sort/ sorted assembly GTF
├── tt/ triple_translate per-branch peptide FASTAs
├── squish/
│ └── <sample>_cryptic.fasta ← THE DELIVERABLE
├── multiqc/
│ └── multiqc_report.html aggregated QC report
└── pipeline_info/
├── execution_report_<timestamp>.html
├── execution_timeline_<timestamp>.html
└── pipeline_dag_<timestamp>.html
| Profile | Purpose |
|---|---|
test |
Use the chr22 test bundle (built by scripts/build_test_bundle.sh) |
pixi |
Run every process from the local pixi env, no containers |
singularity / apptainer |
Pull biocontainers via singularity/apptainer (HPC default) |
docker |
Pull biocontainers via docker (laptop / cloud / CI default) |
monash |
SLURM executor on the Monash M3 comp partition under the xy86 project, with shared singularity cache |
sanjaysgk/ipg/
├── main.nf entry point
├── workflows/ipg.nf main workflow — chains the 6 subworkflows
├── subworkflows/local/
│ ├── align_qc/ steps 1-3
│ ├── transcript_assembly/ steps 4-5
│ ├── bam_prep/ steps 6-12
│ ├── bqsr/ steps 13-16
│ ├── mutect_calling/ steps 17-23
│ ├── db_construct/ steps 24-31 (branches on --include_variant_peptides)
│ └── post_ms_analysis/ --step post_ms: 2-phase db_compare + origins
├── modules/
│ ├── nf-core/ 23 upstream nf-core modules (STAR, samtools, GATK4, etc.)
│ └── local/ 10 local modules:
│ ├── curate_vcf/ IPG custom C tool (kescull)
│ ├── revert_headers/ IPG custom C tool (kescull)
│ ├── alt_liftover/ IPG custom C tool (kescull)
│ ├── triple_translate/ IPG custom C tool (kescull)
│ ├── squish/ IPG custom C tool (kescull)
│ ├── origins/ IPG custom C tool — peptide origin annotation (kescull)
│ ├── db_compare/ R script — cryptic vs UniProt PSM comparison (kescull)
│ ├── gff3sort/ bioconda gff3sort wrapper
│ ├── gatk4_validatesamfile/ missing-from-upstream wrapper
│ └── gatk4_fastaalternatereferencemaker/ missing-from-upstream wrapper
├── containers/
│ └── ipg-tools/ Reproducible Docker build of the kescull C tools
│ from a pinned commit SHA → ghcr.io/sanjaysgk/ipg-tools
├── conf/
│ ├── base.config
│ ├── modules.config per-process ext.args / ext.prefix / errorStrategy
│ ├── test.config chr22 test bundle paths
│ └── monash.config Monash M3 SLURM
├── bin/ runtime scripts staged onto PATH by modules
├── scripts/
│ └── build_test_bundle.sh reproducible chr22 test bundle builder (dev/CI helper)
├── tests/ nf-test pipeline-level tests
├── pixi.toml dev env definition (committed)
└── pixi.lock dev env lockfile (committed, bit-reproducible)
- Sanjay SG Krishna (@sanjaysgk) — pipeline port, Li Lab, Monash University
- Kate Scull — original IPG method + custom C tools, Purcell Lab, Monash University
(see
kescull/immunopeptidogenomics) - Chen Li — supervision, Li Lab, Monash University
- Anthony W. Purcell — supervision, Purcell Lab, Monash University
If you use sanjaysgk/ipg in your research, please cite the original method paper:
Scull KE, Pandey K, Ramarathinam SH, Purcell AW. > Immunopeptidogenomics: harnessing RNA-seq to illuminate the dark immunopeptidome. Mol Cell Proteomics. 2021;20:100143. doi: 10.1016/j.mcpro.2021.100143
A full reference list for every tool in the pipeline lives in CITATIONS.md.
This pipeline is built on top of Nextflow and the nf-core framework:
The nf-core framework for community-curated bioinformatics pipelines. Ewels PA, Peltzer A, Fillinger S, Patel H, Alneberg J, Wilm A, Garcia MU, Di Tommaso P, Nahnsen S. Nat Biotechnol. 2020 Mar;38(3):276-278. doi: 10.1038/s41587-020-0439-x.
MIT — see LICENSE.