sawcheet/parpi-resistance

PARP-inhibitor adaptation score for single cells (in vitro)

Scores each cell in a single-cell RNA-seq sample by how far it has adapted to the PARP inhibitor olaparib (0 = untreated, 1 = adapted to the highest training dose), with the pathways behind the score. It extends the hierarchical model of Satyal, Long, Gao. Attention-Based Hierarchical Graph Autoencoder for Dose-Specific Single-Cell Resistance Dynamics to cell-level learning with held-out-line validation.

Research use only. Not a diagnostic, and not a tumor or patient resistance score.

Scope

Validated for

• Ranking olaparib doses in a ovarian cancer cell line it was not trained on (leave-one-line-out over kuramochi, cov362, ovsaho): dose Spearman 0.70–0.99, AUROC untreated vs top dose 0.76–0.94.
• kuramochi_10x (ovarian, GSE247688): AUROC 1.000 (gated).

Not validated for

• pdx_resistance (ovarian, GSE247690): AUROC 0.614 (reported, not part of the claim).
• pdx_resistance_allcells (ovarian, GSE247690): AUROC 0.650 (reported, not part of the claim).
• Breast cancer: not evaluated.
• Tumors or patients: the score is not a tumor- or patient-level resistance measure.
• Absolute levels: scores shift between cell lines; compare samples within one experiment.
• Uncertainty: the ensemble spread is not yet calibrated and understates the real error.
• Pathway explanations: top_pathways names Hallmark sets only, and they explain the Hallmark part of the score; the model also uses Reactome sets. Reactome names mislead here: four immune/viral-named sets (e.g. Downstream TCR Signaling) share the same 43 proteasome genes. A Hallmark-only model is planned for v0.2.

Released model

• Model: ridge, trained on 31,762 cells from kuramochi, cov362, ovsaho
• Doses (µM): [0.0, 1.0, 2.5, 5.0, 10.0, 20.0, 40.0, 80.0, 160.0, 320.0]
• Config fingerprint da7e460af890 · git commit 6772349b0f41b0608afae76bed7713974e945952

Evaluation

Cross-validation

Each training cell line is held out in turn; the model is trained on the others.

held-out line	dose Spearman	AUROC untreated vs top	cell MAE	mean baseline MAE
cov362	0.700	0.756	0.201	0.226
kuramochi	0.988	0.942	0.247	0.307
ovsaho	0.900	0.778	0.203	0.191

External datasets

AUROC separates treated/resistant from untreated cells; 0.5 = chance. Only gated datasets are part of the claim above.

dataset	cancer	AUROC	cells	role
kuramochi_10x	ovarian	1.000	8474	gated
pdx_resistance	ovarian	0.614	8515	report only
pdx_resistance_allcells	ovarian	0.650	19890	report only
pdx_persistence	ovarian	n/a (no resistant labels)	—	report only

Usage

from parpi.bundle import ModelBundle
from parpi.data.readers import read_any

bundle = ModelBundle.load("hf://sawcheet/parpi-resistance")
predictions = bundle.predict(read_any("my_sample.h5ad"))  # one row per cell

Input: raw counts, HGNC gene symbols, human cells; at least 60% of the model's genes must be present. Compare samples within one experiment (for example treated vs untreated of the same line).

Data and citation

Training and test data: França et al., Cellular adaptation to cancer therapy along a resistance continuum, Nature 631, 876–883 (2024) (GEO SuperSeries GSE247691). Model architecture: Satyal, Long, Gao. Attention-Based Hierarchical Graph Autoencoder for Dose-Specific Single-Cell Resistance Dynamics.

Reproduce

git clone https://github.com/sachitsatyal/parpi-resistance && cd parpi-resistance
uv sync --all-extras
uv run parpi reproduce -c configs/release.yaml