Skip to content

Signature reproducibility map — what RNA can and can't do

The curated gene-signature set for clone selection, with an honest map of what each is good for (#145) — so you don't reach for RNA where the TCR sequence is the right tool. Derived from the B1-2/B1-3 MART-1 in-vitro expansion pilot (paired scRNA + VDJ, two donors).

This map is queryable in code:

from tcrsift.signature_methods import (
    SIGNATURE_GUIDANCE, recommended_signatures, RNA_REPRODUCIBILITY_NOTE,
)
recommended_signatures()        # ['Differentiated', 'AcuteActivation']
SIGNATURE_GUIDANCE["Differentiated"].use_for
Signature Use for Why it's reproducible
Differentiated (effector − naïve/stem, signed; #141) clonal expansion — which clones expanded in culture Best RNA axis: separates expanded vs rare clones at AUROC up to 0.85, same direction in both donors.
AcuteActivation (TNFRSF9/4-1BB, MKI67; #142) the one reproducible RNA correlate of publicness Consistently lower in public (TRAV12-2) clones across all three normalizations and both donors (modest but stable). Private/specific clones carry more recent cognate activation + proliferation; the public pool is more bystander-like. Runs inverse to expansion.

Situational — document, don't default

  • Cytolytic / AntigenExperienced — effector program; correlated with an AIM sort by construction (#142). Prefer Differentiated for the expansion question.
  • AIM / AIMBroad — the transcriptional analogue of an AIMpos sort; use to check sort/transcriptome concordance, not as an independent axis.

Wrong biology for short PBMC culture

  • TumorReactive, exhaustion (CD39 / CXCL13 / TOX / CD103) — chronic in-situ exhaustion biology. Reads as noise in short PBMC culture; meaningful only in fresh tissue TILs.

Documented limitation

RNA does not reproducibly recover the precursor-frequency / cross-reactivity axis. Most signatures flip direction between donors; a cross-donor-validated transcriptome classifier (TCR genes removed, per

144) reaches only AUROC ~0.62–0.69 for publicness, on a diffuse,

non-interpretable gene set (ZBTB16, ISG15, IL7R, CCL20, stress genes — not a naïve program).

For high-specificity / high-avidity / low-cross-reactivity selection, use Pgen/Ppost + sequence features (tcrsift.seqprob,

143), not RNA. RNA is for differentiation / expansion state, not

specificity.

Methodology note

Do not fit/trim signatures on the selection target. With n=2 donors, per-gene direction "consistency" is ~chance (≈3.5/7 agree by coin-flip), so trimming a signature to the genes that separate a label here is overfitting. Refine signatures from curated external references and validate cross-donor / cross-cohort.