tsarina 1.4.4

Personalized cancer immunotherapy target selection from curated shared antigen data

tsarina

Personalized cancer immunotherapy target selection from curated shared antigen data.

Perseus weaves patient-specific tumor characteristics (mutations, CTA expression, viral infections, HLA type) together with curated public mass spectrometry evidence to produce prioritized lists of targetable peptide-MHC complexes. The name reflects the goal: using shared, public knowledge to personalize cancer immunotherapy -- like Perseus using borrowed divine weapons to slay Medusa.

Concept

The core insight is that many cancer-targetable peptides are shared across patients: cancer-testis antigens are recurrently activated in tumors, oncogenic viruses produce the same foreign proteins in every infected cell, and hotspot driver mutations generate identical mutant peptides across thousands of patients. Unlike private passenger-mutation neoantigens that require individual whole-exome sequencing, these shared targets can be curated once and reused.

Perseus combines curated shared targets with per-patient tumor data to produce a prioritized list of peptide-MHC complexes.

Shared targets (curated once, reused across patients):

• CTA genes — 358 curated from 5 databases, 258 after HPA tissue expression filtering
• Viral proteomes — 9 oncogenic viruses (HPV, EBV, HBV, HCV, HTLV-1, HIV, HHV-8, MCPyV, MCV)
• Hotspot mutations — 19 recurrent mutations across 8 driver genes

Public annotation data (used to score and filter targets):

• Mass spec evidence — IEDB/CEDAR immunopeptidomics observations
• Tissue expression — HPA RNA (50 tissues) + IHC protein (63 tissues)
• HLA allele panels — population-representative panels (27–53 alleles per region)

Patient data (per-individual):

• HLA type (Class I alleles)
• Tumor RNA-seq (CTA expression in TPM)
• Detected mutations (cross-referenced against hotspot list)
• Viral status (HPV, EBV, etc.)

Perseus filters shared targets through the patient's HLA type and tumor profile, then ranks the results into a prioritized target list annotated with:

• Public MS evidence — number of independent IEDB/CEDAR references, source context (cancer vs. healthy tissue)
• Source protein abundance — RNA expression in TPM, estimated protein abundance where HPA data permits
• Predicted presentation — MHCflurry presentation percentile, NetMHCpan binding affinity
• Target category — CTA, viral, or mutant, with full provenance

Install

pip install tsarina

# With full functionality (pyensembl for peptide generation + gene partition):
pip install tsarina[all]

Three target categories

CTA (cancer-testis antigens)

Proteins normally restricted to reproductive tissues (testis, ovary, placenta) that become aberrantly expressed in tumors. Their tissue restriction means immune responses against them should not damage normal somatic tissues. Thymus expression is expected (AIRE-mediated central tolerance) and excluded from restriction checks.

358 genes from 5 source databases, systematically filtered using Human Protein Atlas v23 tissue expression data to 258 expressed CTAs with predominantly reproductive-restricted expression (some pass the adaptive filter with minor somatic RNA signal).

from tsarina import CTA_gene_names, CTA_gene_ids, CTA_evidence

genes = CTA_gene_names()    # recommended default set (258 expressed CTAs)
df = CTA_evidence()          # full evidence table with HPA columns + 3-axis tiers

Per-modality restriction classifies each CTA independently by protein (IHC), RNA, and MS evidence, then synthesizes a unified restriction with confidence:

Modality	Column	Values
Protein IHC	protein_restriction	TESTIS / PLACENTAL / OVARIAN
RNA	rna_restriction	TESTIS / PLACENTAL / OVARIAN / REPRODUCTIVE
RNA quality	rna_restriction_level	STRICT / MODERATE / PERMISSIVE
MS (runtime)	ms_restriction	CANCER_ONLY / EXPECTED_TISSUE / SINGLETON_HEALTHY / RECURRENT_HEALTHY
Synthesized	restriction	TESTIS / PLACENTAL / OVARIAN / REPRODUCTIVE
Confidence	restriction_confidence	HIGH / MODERATE / LOW

from tsarina import CTA_testis_restricted_gene_names, CTA_by_axes

testis = CTA_testis_restricted_gene_names()  # 229 genes (synthesized TESTIS)
strict_testis = CTA_by_axes(restriction="TESTIS", rna_restriction_level="STRICT")
high_conf = CTA_by_axes(restriction="TESTIS", restriction_confidence="HIGH")

Source	Genes	Reference
CTpedia	167	Almeida et al. 2009, NAR
CTexploreR/CTdata	62 new	Loriot et al. 2025, PLOS Genetics
Protein-level CT genes	89 new	da Silva et al. 2017, Oncotarget
EWSR1-FLI1 CT gene binding sites	12	Gallegos et al. 2019, Mol Cell Biol
Meiosis, piRNA, spermatogenesis genes	28	Multiple sources (see docs)

CTA curation pipeline

Step 1: Collect. Take the union of protein-coding CT genes across all 5 source databases → 358 genes. Duplicates are merged; non-protein-coding genes (e.g., lncRNAs) are excluded.

Step 2: Annotate for tissue restriction. The goal is to answer: "is this gene's expression restricted to reproductive tissues?" We use two independent data modalities from Human Protein Atlas v23:

• RNA expression (50 tissues): What fraction of total expression comes from reproductive tissues (testis, ovary, placenta)? Raw fractions are misleading because many genes have low-level basal transcription (< 1 nTPM) across dozens of tissues, which inflates the denominator. The deflated reproductive fraction fixes this by zeroing out sub-1 nTPM values before computing the ratio, so only tissues with meaningful expression count. Example: CTCFL has testis nTPM = 10.8 but ~40 other tissues at 0.1–0.9 nTPM each. Raw reproductive fraction: 54%. Deflated fraction: 100%, because only testis exceeds 1 nTPM.

• Protein expression (63 tissues): Does IHC staining detect protein outside reproductive tissues? Each antibody carries a reliability tier — Enhanced (orthogonal validation), Supported, Approved, or Uncertain — which indicates how much to trust the staining result.

Thymus is excluded from all restriction checks because AIRE drives ectopic expression of tissue-restricted antigens in medullary thymic epithelial cells (mTECs) as part of central tolerance. CTA expression in thymus is expected and does not indicate somatic tissue leakage.

Step 3: Filter. Two rules determine whether a gene passes:

1. Protein exclusion (hard): If protein is detected in any non-reproductive somatic tissue (excluding thymus), the gene fails — regardless of RNA data.

2. RNA threshold (tiered): The required deflated reproductive fraction scales with protein data confidence. When high-quality protein data confirms reproductive restriction, we can tolerate more RNA noise in other tissues. When protein data is absent or unreliable, we demand near-perfect RNA restriction:

   Protein evidence	Min. deflated RNA reproductive fraction
   Enhanced + reproductive only	>= 80%
   Supported + reproductive only	>= 90%
   Approved + reproductive only	>= 95%
   Uncertain or no protein data	>= 98%

Result: 258 of 358 genes pass the filter as expressed CTAs; 279 total genes pass filters when the never-expressed low-evidence candidates are included.

See full curation documentation for the deflated fraction formula, never-expressed flag, and figures.

Viral (oncogenic virus proteins)

Foreign proteins from oncogenic viruses -- entirely absent from normal human tissue, making them ideal immunotherapy targets when the virus is present in the tumor.

from tsarina.viral import (
    human_exclusive_viral_peptides,
    viral_peptides,
)

peps = viral_peptides("hpv16")                    # all viral peptides
human_exclusive = human_exclusive_viral_peptides("hpv16")  # default clinical helper

personalize() and target_peptides() use the human-exclusive viral helper by default, dropping viral k-mers that also occur anywhere in the human proteome. cancer_specific_viral_peptides() is available for exploratory workflows that allow overlaps with CTA proteins while excluding non-CTA overlaps.

Virus	Cancers	Key oncoproteins
HPV-16, HPV-18	Cervical, oropharyngeal, anal	E6, E7
EBV/HHV-4	Burkitt lymphoma, NPC, Hodgkin lymphoma	LMP1, EBNA1, LMP2A
HTLV-1	Adult T-cell leukemia/lymphoma	Tax, HBZ
HBV	Hepatocellular carcinoma	HBx
HCV	HCC, B-cell lymphoma	Core, NS3, NS5A
KSHV/HHV-8	Kaposi sarcoma, primary effusion lymphoma	vFLIP, vCyclin, LANA
MCPyV	Merkel cell carcinoma	Large T, small T
HIV-1	Kaposi sarcoma, non-Hodgkin lymphoma	Tat, Nef

Mutant (recurrent somatic hotspot mutations)

Shared neoantigens from driver mutations that recur across thousands of patients. Unlike private passenger mutations, these produce the same mutant peptide in every patient carrying the same hotspot mutation.

from tsarina.mutations import HOTSPOT_MUTATIONS, mutant_peptides

df = mutant_peptides()  # all mutation-spanning 8-11mer peptides

Gene	Mutations	Cancer types
KRAS	G12C, G12D, G12V, G12R, G13D	Pancreatic, colorectal, NSCLC
BRAF	V600E, V600K	Melanoma, colorectal, thyroid
TP53	R175H, R248W, R273H, G245S, R249S	Pan-cancer
PIK3CA	H1047R, E545K	Breast, endometrial
IDH1	R132H	Glioma, AML (peptidomics-validated vaccine target)
NRAS	Q61R, Q61K	Melanoma
EGFR	L858R, T790M	NSCLC

Positive and negative peptide sets

Perseus constructs both positive sets (cancer-specific peptides from the three target categories) and negative sets (peptides observed on normal non-reproductive, non-thymic tissues) using the same IEDB/CEDAR scanning infrastructure with consistent tissue classification.

Tissue source classification

Every IEDB/CEDAR mass spec observation is classified by biological context:

Category	IEDB criteria	Meaning
src_cancer	Process Type = "Occurrence of cancer"	Peptide detected on tumor MHC
src_healthy	Process Type = "No immunization", Disease = "healthy" or empty	Peptide detected on normal tissue
src_reproductive	Source Tissue in {testis, ovary, placenta, ...}	Expected for CTAs
src_thymus	Source Tissue = thymus	Expected for CTAs (AIRE-mediated)
src_cell_line	Culture Condition = "Cell Line / Clone"	In vitro, not direct tissue
src_ebv_lcl	Culture Condition contains "EBV transformed, B-LCL"	EBV-immortalized B cells (special case)
src_ex_vivo	Culture Condition = "Direct Ex Vivo"	Highest confidence tissue evidence

Positive set criteria: peptide has src_cancer evidence AND is exclusive to CTA/viral/mutant source proteins (not found in non-target human proteins).

Negative set criteria: peptide has src_healthy + src_ex_vivo evidence from non-reproductive, non-thymic tissues. These are peptides confirmed to be presented on normal somatic tissue -- targeting them would cause on-target, off-tumor toxicity.

Patient personalization

The main entry point for clinical use:

from tsarina.personalize import personalize

targets = personalize(
    # Patient HLA type
    hla_alleles=["HLA-A*02:01", "HLA-A*24:02", "HLA-B*07:02", "HLA-B*44:02",
                 "HLA-C*07:02", "HLA-C*05:01"],

    # CTA expression (gene symbol -> TPM from RNA-seq)
    cta_expression={"MAGEA4": 142.5, "PRAME": 87.3, "CTAG1B": 215.0},

    # Detected mutations (match against hotspot list)
    mutations=["KRAS G12D", "TP53 R175H"],

    # Viral status
    viruses=["hpv16"],

    # Data sources
    iedb_path="mhc_ligand_full.csv",
)

Returns a DataFrame with columns:

Column	Description
peptide	Peptide sequence
category	cta, viral, or mutant
source	Gene name, virus, or mutation label
source_abundance_tpm	RNA expression in tumor (CTAs only)
ms_hit_count	Number of IEDB/CEDAR MS observations
ms_alleles	MHC restrictions observed in public data
ms_in_cancer	Detected in cancer samples
ms_in_healthy_somatic	Detected in normal non-reproductive, non-thymic tissue (safety flag)
presentation_percentile	MHCflurry presentation percentile for best patient allele
best_allele	Patient HLA allele with best predicted presentation
binding_affinity_nm	Predicted binding affinity (nM)

Prioritization is by: (1) public MS evidence strength, (2) source protein abundance, (3) predicted presentation quality, (4) absence of healthy-tissue MS evidence.

CTA x HLA panel matrices

Build a CTA x HLA pMHC matrix for off-the-shelf panel design:

tsarina panel

Defaults:

• up to 25 downstream non-empty CTAs ranked by bundled HPA tumor RNA prevalence breadth/sample prevalence, with lower-ranked candidates scanned as needed to backfill empty downstream targets and clinical allowlist anchors pinned ahead of lower-ranked candidates
• automatic safety gates remove CTAs with vital-tissue RNA / unique healthy-MS evidence, while allowlisting PRAME, CTAG1A/CTAG1B, and MAGEA4
• automatic selection excludes MAGE-family CTAs other than MAGEA4 unless they are explicitly requested or allowlisted
• CTAG1A/CTAG1B is treated as one grouped CTA target, with NY-ESO-1 accepted as an input alias
• CTAs with identical enumerated peptide sets or final selected pMHC panels are grouped so paralogous targets do not consume multiple automatic panel slots
• global53_abc HLA-A/B/C panel
• 8-11mer CTA-exclusive peptides
• MHCflurry presentation scoring
• MS-evidence-first cell selection
• up to 3 peptides per CTA x HLA cell, ranked by MS source count, then prediction
• readable terminal table plus coverage summary

Use CSV formats for scripts:

tsarina panel --format long -o panel-long.csv
tsarina panel --format wide -o panel-wide.csv

The CLI prints progress for peptide enumeration, public-MS evidence loading, scoring, evidence-tier construction, and final selection. Interactive terminals also get a tqdm scoring progress bar; MHCflurry still scores all alleles in one batch by default because chunking repeats its allele-independent processing model work. Use --score-chunk-size only when you explicitly want chunking, or --no-progress / --no-progress-bars to suppress progress output.

Use --selection-allowlist, --no-vital-tissue-filter, --vital-tissue-max-ntpm, and --allow-non-magea4-mage-family to tune automatic CTA safety filtering. The default vital RNA cutoff is 2.0 nTPM; public healthy-MS observations in vital tissues remain exclusionary only when the peptide evidence maps uniquely to that CTA, unless allowlisted. Explicit --ctas accepts aliases such as NY-ESO-1 and MAGE-A4 and bypasses automatic CTA-family safety gates.

Default automatic ranking uses tumor_prevalence_panel_score, computed from bundled HPA cancer RNA prevalence at pTPM >= 2.0 and cancer-type breadth at a 5% sample-prevalence floor, with HPA cancer IHC as a weak tie-breaker. Public MS support and safety gates are recomputed from the current hitlist observations index for each candidate batch before pMHC scoring; packaged CTA evidence does not carry MS count columns. Use --cancer-rna-threshold, --cancer-type-prevalence-floor, or --cta-rank-by <column> to change the initial non-MS candidate ranking.

Automatic panel output scans lower-ranked CTA candidates to backfill CTAs with zero selected pMHCs after peptide enumeration, CTA-exclusivity, public-MS, and presentation-score gates; pass --show-empty-ctas to audit the top ranked candidates including those failures. Explicit --ctas requests are preserved even if a requested CTA has zero selected pMHCs. The "Expected Population Coverage Per CTA" rows are sorted by selected peptide count, then HLA-hit count, then estimated population coverage, and split monoallelic MS pMHC support from sample-genotype/deconvolved MS support.

Peptide enumeration may expand one target label to multiple Ensembl genes (CTAG1A/CTAG1B and the NY-ESO-1 alias expand to CTAG1A and CTAG1B), but output target names stay grouped. Pass --no-group-identical-cta-peptide-sets to keep peptide-identical paralog targets separate, or --no-group-identical-cta-pmhcs to keep duplicate pMHC panels separate. Pass --netmhcpan-affinity to annotate every selected pMHC row with NetMHCpan BA affinity nM and affinity percentile rank. This is opt-in because it requires the external NetMHCpan backend and adds a second scoring pass when the main selector is using MHCflurry.

Evidence tiers use configurable presentation percentile cutoffs:

Evidence tier	Default cutoff	Meaning
monoallelic_ms	< 2.0	Peptide observed in mono-allelic MS for that HLA
sample_allele_ms	< 1.0	Peptide observed in multi-allelic MS with a usable exact restriction set or donor allele set; the selected HLA must be the best predicted allele in that set, including when the row's reported restriction is only HLA class I
unrestricted_ms	< 0.5	Peptide observed by class-I MS with no usable exact or donor-set allele assignment; the selected panel HLA is assigned by prediction under this stricter cutoff
predicted_only	< 0.1	No MS support; excluded unless --include-predicted-only is passed

Available HLA panels:

Panel	Alleles	Coverage
iedb27_ab	27	Global baseline (HLA-A/B)
iedb36_abc	36	+ HLA-C
global44_abc	44	+ East Asia, South Asia, Sub-Saharan Africa
global48_abc	48	+ Latin America, MENA
global51_abc_ssa	51	Legacy Global-48 + additional Sub-Saharan Africa
global51_abc	51	Global reference panel: IEDB A/B backbone, frequent HLA-C allotypes, and IEDB/Paul common-A/B complements
global53_abc	53	Default global panel: Global-51 plus CTA-MS supported A*29:02, B*15:02, and B*27:05, keeping only C*14:02 from the MHCflurry-identical C*14 pair

Regional allele frequency data from 7 geographic regions supports population-weighted coverage calculations. The frequency audit keeps those sub-population proxy rows separate from published global average allele frequencies on the same 0-1 allele-frequency scale. Panel coverage uses the regional weighted value when a numeric regional proxy exists and falls back to the published global average only when no regional proxy is available. For each CTA, covered allele frequencies are summed within each HLA locus, converted to locus carrier probability as 1 - (1 - locus_frequency)^2, and then combined across loci. All default global53_abc alleles have a published global average, source/proxy/resolution provenance, and a nonzero coverage frequency, preventing known-frequency HLA hits from reporting artificial 0.0% CTA coverage. The reference global51_abc panel keeps all 27 IEDB/TepiTool class-I A/B reference alleles, adds all 21 frequent HLA-C allotypes from the Sarkizova HLA-C peptidome coverage set, and fills the remaining 51-panel slots with the highest-frequency calibrated alleles missing from the IEDB/Paul 38 common HLA-A/B threshold set (B*18:01, B*40:02, B*46:01). The default global53_abc panel adds A*29:02, B*15:02, and B*27:05 because these were the top missing alleles in a public CTA-MS evidence audit while retaining MHCflurry percentile rank support. It keeps C*14:02 but excludes C*14:03 because MHCflurry uses the same pseudosequence and percentile-rank calibration for both, and C*14:02 had the CTA-MS support in the local audit. References: IEDB reference set https://help.iedb.org/hc/en-us/articles/114094151851-HLA-allele-frequencies-and-reference-sets-with-maximal-population-coverage, TepiTool allele-selection description https://pmc.ncbi.nlm.nih.gov/articles/PMC4981331/, IEDB/Paul 38 common A/B thresholds https://help.iedb.org/hc/en-us/articles/114094151811-Selecting-thresholds-cut-offs-for-MHC-class-I-and-II-binding-predictions, and Sarkizova et al. https://doi.org/10.1038/s41587-019-0322-9. Audit notes: all 53 default alleles resolve through MHCflurry's percent_rank_calibrated_allele lookup and produce numeric affinity percentile ranks. HLA-C*15:05 remains excluded because MHCflurry supports raw affinity and presentation predictions for it but does not have an affinity percentile-rank calibration.

Data management

Tsarina uses the shared hitlist data registry for external datasets:

# See what data is available
tsarina data available

# Auto-download viral proteomes from UniProt
tsarina data fetch hpv16
tsarina data fetch ebv

# Register manually downloaded IEDB/CEDAR exports
tsarina data register iedb /data/mhc_ligand_full.csv
tsarina data register cedar /data/cedar-mhc-ligand-full.csv

# Inspect what's installed
tsarina data list

# Resolve paths for use in scripts
tsarina data path iedb

Data sources

Dataset	Source	Size	How to get
IEDB MHC ligand	iedb.org	~2 GB	Manual download (terms of use)
CEDAR MHC ligand	cedar.iedb.org	~1 GB	Manual download
HPV-16 proteome	UniProt UP000006729	~3 KB	tsarina data fetch hpv16
EBV proteome	UniProt UP000153037	~50 KB	tsarina data fetch ebv
(9 viral proteomes total)	UniProt	varies	tsarina data fetch <name>

Storage location: ~/.hitlist/ (override with HITLIST_DATA_DIR env var). tsarina data delegates registry and cache management to hitlist.

IEDB column indices are resolved dynamically from CSV headers, with fallback to known defaults -- robust to IEDB schema changes.

Tissue definitions

Three tiers of reproductive tissue sets for CTA restriction analysis:

from tsarina.tissues import (
    CORE_REPRODUCTIVE_TISSUES,       # {testis, ovary, placenta}
    EXTENDED_REPRODUCTIVE_TISSUES,   # + cervix, endometrium, prostate, ...
    PERMISSIVE_REPRODUCTIVE_TISSUES, # + breast
    is_tissue_restricted,
    adaptive_rna_threshold,
)

MHCflurry scoring

from tsarina.scoring import score_presentation
from tsarina.alleles import get_panel

scores = score_presentation(
    peptides=["SLYNTVATL", "GILGFVFTL"],
    alleles=get_panel("iedb27_ab"),
)

Target naming convention

Perseus uses a unified naming scheme across all target categories:

Category	source column	source_detail column	Example
CTA	Gene symbol	Ensembl gene ID	MAGEA4 / ENSG00000147381
Viral	Virus short name	UniProt protein accession	HPV-16 / P03126
Mutant	Mutation label	Mutation string	KRAS G12D / G12D

Development

./develop.sh    # install in dev mode
./format.sh     # ruff format
./lint.sh       # ruff check + format check
./test.sh       # pytest with coverage