Published comparison policies

Inspect the bundled definitions without providing biological inputs:

vaxrank --list-policies
vaxrank --show-policy lens-v1
vaxrank --show-policy pvacseq-aggregate-v1
vaxrank --show-policy tesla-presentation-v1
vaxrank --show-policy tesla-recognition-v1

Select a bundle with --config builtin:NAME. --show-policy prints its resolved configuration, expanded Topiary expressions, required evidence, source version and citation. The frozen openvax-v1 bundle remains available. Overrides create a derived configuration; the saved effective definition and digest identify it.

Bundle	Evidence and behavior
lens-v1	LENS 1.9 documented score: max(0, 1 − IC50/1000) × log2(coding-support reads)/max(log2(coding-support reads)) × CCF. Missing optional CCF contributes 1.
pvacseq-aggregate-v1	Retains upstream best-peptide/Tier decisions and reproduces pVACtools 7.1.4 default aggregate ordering: Tier, summed dense ranks of expression/IC50/percentile, IC50 rank, Gene, AA Change.
tesla-presentation-v1	Experimental Wells et al. presentation gates: affinity <34 nM, tumor epitope abundance >33 TPM, and pMHC binding stability >1.4 hours. Passing epitopes order by affinity.
tesla-recognition-v1	Adds recognition gates: mutant/WT affinity <0.1 with positive WT affinity, or published foreignness >10⁻¹⁶. A known passing OR branch is sufficient; otherwise missing evidence excludes the occurrence.

These bundles compare epitope selection. Vaccine construction, multiple-epitope aggregation and manufacturability remain Vaxrank procedures. Their combined score uses target_epitope_score, preventing an additional RNA factor.

LENS evidence and normalization

Use a LENS report or normalized Topiary evidence with the reader's actual lens_rna_reads_covering_genomic_origin_with_peptide_cds column. This is the reported coding-sequence support count, not Isovar partial support, gene TPM or arbitrary coverage. Affinity is in nM and optional ccf is a fraction.

Normalization covers all occurrences in each input report/source partition. Evaluate patients/reports separately; report-wide operations reject mixed named samples. Repeated predictor rows do not add support. No pseudocount is added: zero support or a zero log-support maximum stays unscorable. Upstream discovery and filtering are retained in the input, not reconstructed by this score.

LENS 1.9 scoring specification

pVACseq aggregate scope

Use the aggregate report columns Tier, Allele Expr, IC50 MT, %ile MT, Gene, and AA Change. The Topiary reader exposes these as pvacseq_tier, rna_alt_expression, affinity.value, affinity.rank, gene, and aa_change. The preset requires the default median IC50/combined percentile metric choices. Nondefault source metrics need a derived policy.

The upstream Tier carries presentation, binding, expression, clonality, transcript, reference-match and mutation-anchor decisions. This bundle consumes those decisions; it does not assign tiers to raw all-epitopes rows. Missing numeric sort measurements rank at the bottom. Unknown tiers are excluded. Expression ranks descending, matching the pinned implementation despite the documentation's ascending wording. The score 1/ordinal_rank(...) represents the ordering and exact sort-key ties remain tied. Conflicting scores for multiple occurrences of one candidate require an explicit duplicate-resolution choice.

pVACseq aggregate criteria, pVACtools 7.1.4 sorting implementation

TESLA evidence

Provide normalized Topiary rows with affinity in nM, pMHC_stability values in hours, and tumor_abundance_tpm containing tumor epitope abundance. Gene TPM and read counts are not automatically substituted. Recognition also requires matched WT affinity and/or the published IEDB similarity-derived foreignness feature. The CLI does not calculate foreignness or obtain missing experimental measures.

Thresholds are strict; missing required evidence stays unknown. These experimental bundles apply the published gates, then order survivors by affinity.

Wells et al., Cell 2020

Save and replay

For a report, select its corresponding bundle and save native evidence:

vaxrank --input-lens report.tsv --no-processing-aware-annotation --config builtin:lens-v1 \
  --output-epitopes evidence.tsv --output-csv lens-comparison.csv
vaxrank --input-pvacseq aggregate.tsv --no-processing-aware-annotation --config builtin:pvacseq-aggregate-v1 \
  --output-epitopes evidence.tsv --output-csv pvacseq-comparison.csv
vaxrank --input-topiary measured.tsv --config builtin:tesla-recognition-v1 \
  --output-epitopes evidence.tsv --output-csv tesla-comparison.csv
vaxrank --input-epitopes evidence.tsv --output-csv replay.csv

Native evidence retains the effective policy, all source measurements and per-occurrence decisions. The adjacent .policy directory records configuration derivation, source citations, model choices and replayable Topiary audit tables. Evaluation and replay do not run predictors. Percentile-only rows remain in the evidence even when they cannot create a quantitative native prediction object.