Which immune cells predict whether melanoma responds to checkpoint immunotherapy?

Single-cell RNA-seq analysis of 16,291 immune cells from 32 melanoma patients (Sade-Feldman et al., Cell 2018, GEO GSE120575) treated with checkpoint immunotherapy (anti-PD-1 / anti-CTLA-4).

Headline result: the balance between memory-like and cytotoxic/exhausted CD8 T cells is the strongest predictor of response. A minimal 2-gene ratio (TCF7 − GZMB) separates responders from non-responders with AUC 0.83, and a data-driven CD8 signature reaches AUC 0.80 under leave-one-patient-out cross-validation — with zero label leakage in either case.

Responding tumors are also B-cell-rich and myeloid-poor (all FDR < 0.05 at the sample level) — reproducing the B-cell/tertiary-lymphoid-structure association reported in the original paper.

Pipeline

GEO download → QC → Harmony-integrated clustering (32-patient batch correction) → UMAP → canonical-marker cell-type annotation → sample-level composition testing → response-signature discovery with cross-validated ROC/AUC.

25 Leiden clusters were annotated to 7 immune cell types / 9 cell states, dominated by a CD8 T cell continuum from memory → cytotoxic → exhausted.

Read the full analysis

📄 Full report with methods, all figures, and caveats →

Repo contents

report/    full write-up (methods, results, honest caveats)
figures/   9 figures: QC, UMAP, marker annotation, composition, ROC
data/      signature AUC table + per-sample signature scores

The ~890 MB annotated AnnData (.h5ad) object is not included in this repo due to size; the CSVs above contain the derived results needed to reproduce the headline findings.

Methodological notes

• Sample-level statistical testing throughout (not per-cell) — respects the true unit of biological replication.
• Signature evaluation separates biology-driven (label-independent) from data-driven (fit on data) signatures, and reports the leave-one-patient-out cross-validated AUC as the credible estimate rather than the optimistic in-sample number.
• Smart-seq2 full-length TPM data (not UMI counts) — QC and normalization adapted accordingly.
• All findings are exploratory on a single public cohort; not a validated biomarker.

Tools

Analysis performed with Claude Science (scanpy, harmonypy, scikit-learn) — Anthropic's AI workbench for scientific research.