Vaxrank is the neoantigen ranking component of the OpenVax pipeline for designing personalized cancer vaccines. Given either (a) a patient's somatic mutations + tumor RNA-seq + HLA type, or (b) a pre-computed neoepitope report from LENS or pVACseq, Vaxrank selects and ranks the mutant antigens most likely to elicit a T-cell response and emits them as the vaccine type(s) the user requests — peptide pools, mRNA constructs, or analysis reports for review.

INPUT (one of)
  ├── --vcf + --bam              full pipeline: variant calling → Isovar
  │                              transcript assembly → MHC prediction → ranking
  └── --input-lens / --input-pvacseq
                                 use a pre-computed neoepitope report;
                                 ranking happens against existing
                                 (peptide, allele) predictions

SHARED MIDDLE
  ranked_variants_with_vaccine_peptides   (the canonical intermediate;
                                           same shape from both inputs)

VACCINE-TYPE DISPATCH (multi-valued; --vaccine-type)
  ├── peptide   →  FASTA + JSON manifest + vendor order-form CSV
  │                (sub-modes via --peptide-mode: slp / minimal_epitope /
  │                multi_epitope)
  ├── mrna      →  three FASTAs (cds / no_polyA / full) + JSON manifest +
  │                long-format CSV
  └── (future)  →  dna, etc. plug in as one entry in the dispatch table

REPORTS (orthogonal to vaccine-type)
  CSV / XLSX / ASCII / HTML / PDF / JSON / neoepitope-report

Vaxrank always ranks; whether each vaccine-type writer fires depends on both --vaccine-type and the corresponding --output-<type> path being set. The reports are independent and stack with any vaccine type.

Personalized cancer vaccines (also called neoantigen vaccines) work by training the immune system to recognise peptides that arise from somatic mutations unique to a patient's tumor. Designing such a vaccine requires a computational pipeline that bridges raw sequencing data and the peptide synthesiser:

1. Variant calling — Whole-exome or whole-genome sequencing of the tumor and matched normal identifies somatic mutations. This is typically done with tools such as MuTect or Strelka, upstream of Vaxrank.
2. Mutant transcript assembly — Tumor RNA-seq reads overlapping each mutation are assembled by Isovar to determine the true mutant protein sequence. This step phases nearby germline variants and captures any mutation-associated splicing differences, producing a more accurate reading frame than DNA-only prediction.
3. MHC binding prediction — Candidate epitopes (short peptide subsequences spanning the mutation) are scored for predicted binding to the patient's HLA class I molecules using mhctools, which wraps predictors such as MHCflurry, NetMHCpan, and BigMHC.
4. Vaccine peptide selection — Vaxrank assembles longer synthetic long peptides (SLPs, typically 25-mers) around the mutation, scores them by the number and strength of their predicted MHC-binding epitopes, filters out peptides that appear in the reference proteome, annotates known cancer hotspot mutations, and ranks candidates by a combined immunogenicity and manufacturability score.
5. Vaccine-type dispatch — the ranked candidates are written out as one or more of the vaccine types selected via --vaccine-type: a peptide pool ready for synthesis, an mRNA construct ready for IVT, or both. Analysis reports are emitted independently. Steps 1-3 are skipped when an external neoepitope report is supplied via --input-lens or --input-pvacseq; the ranking and dispatch steps are identical.

Vaxrank's vaccine design space is two orthogonal axes (shared across vaccine types) plus the type itself:

Combined, the matrix yields 8 distinct designs — 4 per vaccine type:

A third knob, --epitopes-per-antigen, controls how many top MHC ligands to take per ranked vaccine peptide when content is minimal_epitope. The default 1 is the legacy "single top ligand" semantics; >1 packs multiple top ligands from the same variant as separate antigens.

# Default: SLP peptide pool
vaxrank --vcf v.vcf --bam r.bam --output-peptide pool.fasta

# Multi-epitope concatenated peptide
vaxrank --vcf v.vcf --bam r.bam \
        --output-peptide pool.fasta \
        --peptide-antigens-per-construct 5 --peptide-linker AAY

# Minimal-epitope peptide (single ligand per construct)
vaxrank --vcf v.vcf --bam r.bam \
        --output-peptide pool.fasta \
        --antigen-content minimal_epitope

# BioNTech FixVac canonical mRNA (default for --vaccine-type mrna)
vaxrank --vcf v.vcf --bam r.bam --vaccine-type mrna --output-mrna out/

# String-of-beads mRNA (concatenated minimal epitopes)
vaxrank --vcf v.vcf --bam r.bam --vaccine-type mrna --output-mrna out/ \
        --mrna-antigen-content minimal_epitope --mrna-antigens-per-construct 8 \
        --mrna-linker AAY

# Top-2 ligands per variant in a string-of-beads mRNA
vaxrank --vcf v.vcf --bam r.bam --vaccine-type mrna --output-mrna out/ \
        --mrna-antigen-content minimal_epitope \
        --mrna-epitopes-per-antigen 2 --mrna-antigens-per-construct 16

# Both modalities at once
vaxrank --vcf v.vcf --bam r.bam --vaccine-type peptide mrna \
        --output-peptide pool.fasta --output-mrna mrna_out/

The legacy --peptide-mode {slp, minimal_epitope, multi_epitope} flag still works as a shorthand (slp ≡ mutation_spanning + 1, etc.) but the orthogonal axes are preferred for new designs.

Vaccine-type selection is controlled by --vaccine-type (multi-valued, default peptide). Each type's writer fires only if its --output-<type> path is also set. Reports are orthogonal — they run regardless of vaccine type and can be combined with any of the construct outputs.

# Peptide pool (default vaccine type)
vaxrank --vcf v.vcf --bam r.bam --output-peptide pool.fasta

# mRNA construct
vaxrank --vcf v.vcf --bam r.bam --vaccine-type mrna --output-mrna mrna_out/

# Both at once
vaxrank --vcf v.vcf --bam r.bam --vaccine-type peptide mrna \
        --output-peptide pool.fasta --output-mrna mrna_out/

# Reports only (no vaccine constructs)
vaxrank --vcf v.vcf --bam r.bam --output-pdf-report report.pdf

# Drive vaccine design from a pre-computed LENS report
vaxrank --input-lens patient.lens.tsv --vaccine-type mrna \
        --output-mrna mrna_out/ --output-mrna-csv layers.csv

The peptide and mRNA construct JSON manifests share a back-compat schema (modality, name, length, length_unit, antigen_names, components, manufacturability). The mRNA manifest additionally exposes cds, no_polya_nt, full_nt, per-antigen antigens (each with AA + nt), and a structured elements dict with one entry per layer (5' UTR, signal peptide, antigens, linkers per junction, MITD, stop codon, 3' UTR, polyA) — every layer carrying both AA (where applicable) and nt forms for direct inspection.

Both vaccine types consume the same set of linker names so a single construct design can be ported between peptide and mRNA backbones.

Static entries:

Compositional grammar (parsed at lookup time):

Repeat counts are capped at 100. 2A entries (codon-frozen, positional) are rejected in repeat forms — use the base linker once.

Every name resolves through vaccine_library.get_linker(name) and returns a Linker with primary-source citations attached. The default mRNA inter-antigen linker is (G4S)2 (BioNTech FixVac canonical, Sahin Nature 2017); the default peptide linker is G4S3. Per-junction MHC-aware linker swap (--mrna-optimize-linkers, on by default) considers G3S, G4S, (G3S)2, (G4S)2, AAA per junction and substitutes whichever minimizes predicted presentation of chimeric k-mers spanning the junction.

All sequences carry primary-source citations in vaxrank/vaccine_library.py.

Vaxrank is the ranking engine behind the OpenVax neoantigen vaccine pipeline, which has been used in several clinical trials of personalized cancer vaccines at Mount Sinai:

• PGV001 (NCT02721043) — A phase I study of personalised neoantigen vaccines in patients with solid and haematologic malignancies. All 11 treated patients developed neoantigen-specific T-cell responses (Bortman et al., Cancer Discovery 2025).
• PGV001 + atezolizumab in urothelial cancer (NCT03359239) — A phase I trial combining PGV001 with checkpoint inhibition. The combination was safe and induced neoantigen-specific CD4+ and CD8+ T-cell responses in all evaluated patients (Galsky et al., Nature Cancer 2025).
• PGV001 + TTFields in newly diagnosed glioblastoma (NCT03223103) — A phase I trial combining PGV001 with tumor treating fields and standard-of-care temozolomide (paper in preparation).

The computational pipeline used in these trials is described in Kodysh & Rubinsteyn, Methods Mol. Biol. 2020.

vaxrank \
    --vcf tests/data/b16.f10/b16.vcf \
    --bam tests/data/b16.f10/b16.combined.bam \
    --vaccine-peptide-length 25 \
    --mhc-predictor netmhc \
    --mhc-alleles H2-Kb,H2-Db \
    --padding-around-mutation 5 \
    --output-ascii-report vaccine-peptides.txt \
    --output-pdf-report vaccine-peptides.pdf \
    --output-html-report vaccine-peptides.html

Inputs:

• --vcf — Somatic variants (VCF from any variant caller)
• --bam — Tumor RNA-seq alignments (used by Isovar to assemble mutant transcripts)
• --mhc-alleles — Patient HLA alleles (e.g. HLA-A*02:01,HLA-B*07:02)
• --mhc-predictor — Which MHC binding predictor to use (see table below)

pip install vaxrank

Requirements: Python 3.9+

Vaxrank uses PyEnsembl for reference genome annotation. Install an Ensembl release matching your reference genome:

# GRCh38
pyensembl install --release 113 --species human
# GRCh37 (legacy)
pyensembl install --release 75 --species human

PDF report generation uses wkhtmltopdf by default:

brew install --cask wkhtmltopdf

Alternatively, pass --pdf-backend=weasyprint to use WeasyPrint (experimental), which has no external binary dependency:

pip install weasyprint
# macOS also needs: brew install pango

On Apple Silicon, WeasyPrint loads Pango via dyld, which doesn't search Homebrew's /opt/homebrew/lib by default. Add this to your shell profile:

export DYLD_FALLBACK_LIBRARY_PATH="/opt/homebrew/lib:$DYLD_FALLBACK_LIBRARY_PATH"

(Intel macOS doesn't need this — Homebrew's /usr/local/lib is in dyld's default fallback path.)

Common parameters can be stored in a YAML file to avoid repeating them on every run:

vaxrank --config my_config.yaml --vcf variants.vcf --bam tumor.bam

Example my_config.yaml:

epitopes:
  min_score: 0.00001                        # drop epitopes below this score
  scoring_mode: affinity                    # "affinity" or "percentile_rank"
  logistic_midpoint: 350.0                  # IC50 (nM) at which score = 0.5
  logistic_width: 150.0                     # steepness of logistic curve
  affinity_cutoff: 5000.0                   # IC50 >= this → score 0
  percentile_rank_cutoff: 10.0              # rank >= this → score 0 (percentile mode)
  top_epitopes_per_candidate: 1000          # 0 = keep all

vaccine_peptides:
  preferred_length: 25                      # target amino acids per vaccine peptide
  min_length: 25                            # minimum vaccine peptide length
  max_length: 25                            # maximum vaccine peptide length
  padding_around_mutation: 5                # off-centre windows to consider
  per_mutation: 1                           # peptides to keep per variant
  max_epitopes_per_candidate: 1000          # 0 = keep all
  score_fraction_of_best: 0.99              # drop candidates scoring < 99% of best
  manufacturability:                        # GRAVY = mean hydropathy
    max_c_terminal_hydropathy: 1.5          # max GRAVY of C-terminal 7-mer
    min_kmer_hydropathy: 0.0                # min max-7mer GRAVY (floor)
    max_kmer_hydropathy_low_priority: 1.5   # low-priority max-7mer GRAVY cap
    max_kmer_hydropathy_high_priority: 2.5  # high-priority max-7mer GRAVY cap

For anything beyond the scalar logistic / percentile-rank defaults, set epitopes.filter_expr and/or epitopes.score_expr to a topiary DSL string. Both accept the full topiary 5.0 expression grammar (kind accessors like affinity / presentation, arithmetic, & / |, .logistic(...) / .clip(...) transforms, column(col_name) for raw DataFrame columns, etc.).

epitopes:
  # Drop rows wholesale before scoring
  filter_expr: "affinity <= 500 & affinity.rank <= 2.0"
  # Compute a per-(peptide, allele) score in [0, 1] (binder-quality score)
  score_expr:  "affinity.logistic_normalized(350, 150)"

When filter_expr is omitted, no rows are dropped up-front; the default score_expr is synthesized from the scalar fields above (binding_affinity_cutoff, logistic_midpoint, logistic_width, etc.) and masked so ic50 >= affinity_cutoff → 0, reproducing the pre-5.0 behavior byte-for-byte.

Use affinity.logistic_normalized(m, w) for a [0, 1] binder-quality score (the topiary 5.1+ primitive); the plain affinity.logistic(m, w) is the raw sigmoid and caps below 1 (≈0.912 at default m=350, w=150).

Invalid DSL strings are rejected at config load (not mid-pipeline), so typos in the YAML surface before any predictions run.

CLI arguments override YAML values. You can also use --config-value to override individual keys without editing the file:

vaxrank --config my_config.yaml \
  --config-value vaccine_peptides.score_fraction_of_best=0.95 \
  --config-value epitopes.percentile_rank_cutoff=5.0

Use --config-text when the right-hand side should be kept as a raw string instead of being YAML-parsed.

Config values are resolved in order (later wins):

1. Compiled-in defaults (see vaxrank/config/defaults.py)
2. YAML config file (--config)
3. --config-value / --config-text overrides
4. Dedicated CLI flags (e.g. --vaccine-peptide-length)

The four *_hydropathy* fields control the manufacturability tie-breaking in vaccine peptide ranking. See VaccinePeptide.peptide_synthesis_difficulty_score_tuple for details on how each threshold is applied.

Vaxrank integrates with MHC binding predictors via mhctools. Use --mhc-predictor <name> to select one:

Vaxrank accepts two distinct input shapes, both producing the same ranked-vaccine-peptides intermediate:

Full pipeline (VCF + BAM): Vaxrank does not perform variant calling or read alignment itself. Those steps happen upstream, typically as part of a larger bioinformatics pipeline (e.g. neoantigen-vaccine-pipeline):

1. Tumor and matched-normal DNA are sequenced and aligned; a variant caller (MuTect, Strelka, etc.) produces a VCF of somatic mutations.
2. Tumor RNA is sequenced and aligned to produce a BAM file.
3. The patient's HLA class I alleles are typed (from sequencing data or clinical records).

Vaxrank takes these three inputs — the VCF, the tumor RNA BAM, and the HLA alleles — runs Isovar transcript assembly + MHC binding prediction

• ranking, and produces vaccine peptide candidates.

External-input mode (--input-lens or --input-pvacseq): when an upstream tool (e.g. LENS or pVACseq) has already produced a per-(peptide, allele) neoepitope report, Vaxrank skips Isovar + MHC prediction and consumes the report directly. The per-row pep_context (LENS) or Best Peptide (pVACseq aggregate) is used as the SLP-style antigen window. Downstream dispatch — reports + peptide constructs + mRNA constructs — is identical to the full pipeline.

For each somatic variant, Isovar extracts RNA-seq reads overlapping the mutant locus and assembles them into a mutant protein fragment. This is more accurate than simply applying the DNA variant to the reference transcript because it:

• Phases adjacent germline and somatic variants that fall on the same read, producing the true amino acid sequence
• Captures splicing differences such as intron retention events that may alter the reading frame near the mutation
• Confirms expression — variants with no supporting RNA reads are filtered out

Each mutant protein fragment is sliced into overlapping subsequences of epitope length (typically 8–15 amino acids). These candidate epitopes are scored for predicted MHC binding affinity using the selected predictor. Binding predictions are converted to a score between 0 and 1 via a logistic function parameterised by the EpitopeConfig settings.

Candidate vaccine peptides (longer SLPs, typically 25-mers) are constructed around each mutation. Each candidate is scored by the combined immunogenicity of the epitopes it contains. Candidates are then filtered and ranked by:

1. Epitope content — total predicted immunogenicity score
2. Reference proteome filtering — peptides matching the human reference proteome are removed to ensure only truly novel sequences are selected
3. Cancer hotspot annotation — variants at known recurrently mutated positions (bundled data from cancerhotspots.org, ~2,700 mutations across cancer types) are flagged
4. Manufacturability — tie-breaking by hydropathy-based synthesis difficulty (C-terminal and 7-mer window GRAVY scores)

Shared upstream:

• core_logic.py: Main vaccine peptide selection algorithm
• epitope_logic.py: Epitope scoring and filtering
• epitope_io.py: LENS / pVACseq / vaxrank-native I/O for epitope predictions
• external_input.py: Synthesize the canonical ranked-vaccine-peptides shape from a LENS / pVACseq report so external-input runs reach the same dispatch as VCF + BAM
• reference_proteome.py: Set-based kmer index for reference proteome filtering (O(1) lookup, built once and cached)
• cancer_hotspots.py: Cancer mutation hotspot annotation
• vaccine_peptide.py: Vaccine peptide scoring and manufacturability
• vaccine_library.py: Shared linker vocabulary + compositional grammar ((BASE)N, GnSm, AnY, An, Gn) with primary-source citations

Vaccine-type-specific (downstream):

• peptide.py: Peptide construct assembly + FASTA / JSON manifest / vendor order-form CSV writers; sub-modes slp / minimal_epitope / multi_epitope
• mrna.py: mRNA construct assembly + three-FASTA / structured manifest / long-format CSV writers. DnaChisel codon optimization, 2A frozen-codon handling, configurable polyA tail (default A120, optional segmented BNT162b2 pattern), per-junction MHC-aware linker swap (issue #247)
• mrna_library.py: mRNA-specific elements (5'/3' UTRs incl. tandem 2× HBB FI; signal peptides HLA-A / HLA-B / tPA / IgK / CD8A / CD28; MITD HLA-A / HLA-B)
• junction_swap.py: Per-junction linker optimizer that minimizes predicted MHC presentation of chimeric k-mers spanning antigen junctions

Reports:

• report.py: Analysis-report generation (ASCII, HTML, PDF, XLSX, CSV, JSON)

Vaxrank algorithm:

Rubinsteyn, A., Hodes, I., Kodysh, J. & Hammerbacher, J. Vaxrank: A Computational Tool For Designing Personalized Cancer Vaccines. bioRxiv (2017).

OpenVax pipeline (methods):

Kodysh, J. & Rubinsteyn, A. OpenVax: An Open-Source Computational Pipeline for Cancer Neoantigen Prediction. Methods Mol. Biol. 2120, 147–160 (2020).

PGV001 clinical results:

Bortman et al. PGV001, a Multi-Peptide Personalized Neoantigen Vaccine Platform: Phase I Study in Patients with Solid and Hematologic Malignancies in the Adjuvant Setting. Cancer Discovery 15(5), 930–945 (2025).

Galsky et al. Atezolizumab plus personalized neoantigen vaccination in urothelial cancer: a phase 1 trial. Nature Cancer (2025).

BibTeX for the Vaxrank paper:

@article {Rubinsteyn142919,
    author = {Rubinsteyn, Alex and Hodes, Isaac and Kodysh, Julia and Hammerbacher, Jeffrey},
    title = {Vaxrank: A Computational Tool For Designing Personalized Cancer Vaccines},
    year = {2017},
    doi = {10.1101/142919},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2017/05/27/142919},
    journal = {bioRxiv}
}

Vaxrank is built on the OpenVax ecosystem:

• pyensembl: Reference genome annotation
• varcode: Variant effect prediction from DNA
• isovar: RNA-based mutant transcript assembly and variant phasing
• mhctools: Unified interface to MHC binding predictors

Other key dependencies:

• msgspec: Configuration serialization (YAML/JSON)
• pandas, numpy: Data processing
• jinja2, pdfkit/weasyprint: Report generation

To install Vaxrank for local development:

git clone git@github.com:openvax/vaxrank.git
cd vaxrank
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install -e .
# Examples; adjust release to match your reference
pyensembl install --release 113 --species human
pyensembl install --release 113 --species mouse

Run linting and tests:

./lint.sh && ./test.sh

The first run of the tests may take a while to build the reference proteome kmer index, but subsequent runs will use the cached index.

• develop.sh: installs the package in editable mode and sets PYTHONPATH to the repo root.
• lint.sh: runs ruff on vaxrank and tests.
• test.sh: runs pytest with coverage.
• deploy.sh: runs lint/tests, builds a distribution with build, uploads via twine, and tags the release (vX.Y.Z). Deploy is restricted to the main/master branch.