Experimental mixed-antigen design

What information is needed?

Stage	Minimum information
Consider a candidate	Patient/sample, sequence or unresolved sequence hypothesis, source identifiers and evidence provenance. Incomplete candidates remain report-only.
Admit to a construct	Exact amino-acid sequence, source coordinates, targetable intervals and a documented tumor-specificity decision. Held-out candidates cannot enter constructs.
Rank sources together	Same patient and declared host genotype/reference; compatible predictor versions, endpoints and units; explicit scoring, missing-data and duplicate policies.
Allocate by source	Eligibility rules and slot limits for mutation, CTA and viral candidates, plus product size and allele-coverage requirements.

• Mutation: somatic-call/reference provenance, mutant sequence and available RNA evidence. Missing RNA is not zero expression.
• CTA: gene/transcript identity, versioned OncoRef reference and patient tumor expression with assay, units and threshold. Gene RNA does not establish protein abundance. See the CTA input guide.
• Viral: virus/strain, accession/version, ORF sequence and evidence linking its expression to this tumor. Native VaccineAntigen evidence must carry the targetable mask and admission decision. A FASTA or infection history alone is insufficient; automatic viral admission is not available.

Rank together or select separate panels

For resolved sequences without predictions, use the unified antigen input. It accepts all supported kinds in one dataset and predicts them through the shared pipeline.

Save candidates with --output-epitopes, keep their .references directories, and combine them using an input manifest. Preserve viral accessions separately from the manifest's host reference.

For inputs with compatible saved MHCflurry affinity measurements, this explicitly uses binding alone for ranking:

vaxrank --input-manifest inputs.yaml \
  --config-text 'epitopes.score_expr=affinity["mhcflurry"].value.logistic_normalized(350, 150)' \
  --config-text 'vaccine_peptides.combined_score_expr=target_epitope_score' \
  --vaccine-type peptide --output-dir mixed \
  --output-epitopes mixed/native.tsv --no-processing-aware-annotation

Select a recorded version with affinity["mhcflurry", "VERSION"] when needed. For incompatible scores, use common-model prediction.

For separate source allocations, filter the saved pool and cap each panel:

vaxrank --input-epitopes mixed/native.tsv \
  --config-text 'epitopes.filter_expr=vaxrank_antigen_kind == "CTA"' \
  --vaccine-type peptide --peptide-antigens-per-construct 1 \
  --peptide-max-constructs 2 --output-dir cta-stratum \
  --output-epitopes cta-stratum/native.tsv --no-processing-aware-annotation

Use "mutation" or "viral" for the other panels. These source fields come from native antigen evidence. Compose additional eligibility rules into the same filter; it replaces the inherited scalar filter. Named policies use filter_by. Caps are maxima, not guaranteed slots. Review duplicates, allele coverage and total capacity across panels. Joint source quotas inside one mRNA construct are not implemented.

Saved CTA/viral reports need no Ensembl release; mutation reports still need annotation when it was not retained. Optional window exposure is also experimental and does not change selection.