sctherapy-artifacts

Trained checkpoints, derived data, and evaluation results for the sctherapy_pytorch project — predicting drug response (% inhibition) from gene expression, benchmarked on a 12-patient AML cohort from the scTherapy paper.

Code: Tino3141/sctherapy_pytorch (FT-Transformer, GRANDE, LightGBM baseline) — the training/eval code, scripts, and configs that produced these artifacts. This repo is self-contained: checkpoints, the full training set, both eval cohorts, and results are all here. (Training data is also mirrored at Tino3141/lincs-pharmaco-training / Tino3141/pharmaco_flat.)

The Python snippets below (from eval.model_registry import ..., from src.model.lgbm import ...) assume you're running inside a clone of the GitHub repo; this repo holds the artifacts, not the code.

Checkpoints

Neural (checkpoints/neural/) — the exact checkpoints used in the AML eval

Pulled from Weights & Biases (cpinkl/pmlr-sctherapy). These are the precise epoch checkpoints behind the per-seed predictions in results/aml_evals/ — not the final best-val (epoch-10) snapshots. The eval used an earlier epoch (6–9) per run; the filename encodes the epoch, and each .pt's stored epoch field matches it. Complete {42, 43, 44} seed sweep × 4 variants = 12 files.

File	Arch	Head	Seed	Epoch used	W&B run	val_rmse @ epoch
ft_transformer_hill__s42__epoch7__run-mwwp1nej.pt	FT-Transformer	Hill	42	7	mwwp1nej	10.87
ft_transformer_hill__s43__epoch7__run-k36em957.pt	FT-Transformer	Hill	43	7	k36em957	10.81
ft_transformer_hill__s44__epoch7__run-dnyzw2jk.pt	FT-Transformer	Hill	44	7	dnyzw2jk	10.72
ft_transformer_scalar__s42__epoch7__run-lv6eowft.pt	FT-Transformer	Scalar	42	7	lv6eowft	11.01
ft_transformer_scalar__s43__epoch9__run-1lx7r1vp.pt	FT-Transformer	Scalar	43	9	1lx7r1vp	10.91
ft_transformer_scalar__s44__epoch8__run-5mtffjfa.pt	FT-Transformer	Scalar	44	8	5mtffjfa	10.74
grande_hill__s42__epoch8__run-fhszu55u.pt	GRANDE	Hill	42	8	fhszu55u	11.47
grande_hill__s43__epoch8__run-0vsn6313.pt	GRANDE	Hill	43	8	0vsn6313	11.56
grande_hill__s44__epoch6__run-xutm14sz.pt	GRANDE	Hill	44	6	xutm14sz	11.72
grande_scalar__s42__epoch7__run-izgn7e8j.pt	GRANDE	Scalar	42	7	izgn7e8j	30.78
grande_scalar__s43__epoch6__run-iqngqwll.pt	GRANDE	Scalar	43	6	iqngqwll	31.00
grande_scalar__s44__epoch8__run-k2bepvf1.pt	GRANDE	Scalar	44	8	k2bepvf1	30.45

Each .pt is a dict with model_state_dict, epoch, val_rmse. Load with the project's eval/model_registry.py (state-dict introspection rebuilds the architecture — no separate config needed), or via build_model after reading the shapes.

Why these epochs (not best-val): the evaluated checkpoint was chosen on a combination of the mean and standard deviation of validation performance — favouring a stable epoch over the single lowest-mean snapshot. So the selected epoch (6–9) deliberately trades a marginally lower best-val for lower variance, rather than chasing the minimum mean val_rmse alone.

Recovering other snapshots: the absolute best-val epoch (usually epoch 10) scores marginally lower mean val_rmse (FT-hill ~10.5–10.6, GRANDE-hill ~11.3–11.6). To pull it instead, fetch cpinkl/pmlr-sctherapy/model-<runid>:best (alias best → epoch 10). Any other epoch: model-<runid>:vN where v0=epoch1 … v9=epoch10.

LightGBM (checkpoints/lgbm/)

Path	Notes
cell_seed42_ts0.8_nb1000_es50/model.txt	Cell-split baseline (held-out cell lines), 1000 rounds, ES@50. Headline LGBM used in the registry. Val RMSE ≈ 6.96, AUC ≈ 0.70 on the 12-patient benchmark.
lgbm_baseline_model.txt	Alternative local LGBM.

Usage

The neural .pt files are PyTorch state-dicts ({model_state_dict, epoch, val_rmse}) — weights only, no architecture. The repo's eval/model_registry.py re-infers the architecture from the tensor shapes, so loading needs nothing but the checkpoint and a registry key (ft_scalar, ft_hill, grande_scalar, grande_hill, or lgbm — the filename tells you which).

import torch, numpy as np
from huggingface_hub import hf_hub_download
from eval.model_registry import REGISTRY, load_predictor

# Download one of the archived checkpoints (best FT-hill, seed 43, epoch 7)
ckpt = hf_hub_download(
    "Tino3141/sctherapy-artifacts",
    "checkpoints/neural/ft_transformer_hill__s43__epoch7__run-k36em957.pt",
)

pred = load_predictor(
    REGISTRY["ft_hill"],                 # key must match the checkpoint's arch/head
    device=torch.device("cpu"),
    checkpoint_override=ckpt,
)

# Inputs: gene z-scores (N, 978), ECFP4 bits (N, 1024), dose in µM (N,)
gene  = np.zeros((2, 978), dtype=np.float32)
ecfp4 = np.zeros((2, 1024), dtype=np.float32)
dose  = np.array([1.0, 10.0], dtype=np.float32)

y = pred.predict(gene, ecfp4, dose)      # → predicted % inhibition, shape (N,)
print(y)

LightGBM (model.txt) loads directly:

from src.model.lgbm import LGBMDrugPredictor
lgbm = LGBMDrugPredictor.load("cell_seed42_ts0.8_nb1000_es50/model.txt")
# feature layout: np.hstack([gene(978), ecfp4(1024), dose(1)])

The AML patient inputs are in eval/aml_eval/ (model_inputs, with the ground-truth DSS to score against); the separate multi-cancer held-out set is in eval/zenodo_eval/; derived/gene_names.json gives the gene column order and derived/lincs_gene_stats.npz the (mean, std) used to z-score new expression data. To download everything at once: hf download Tino3141/sctherapy-artifacts --local-dir ./sctherapy-artifacts.

Derived data (derived/)

Small generated files needed to reproduce inference (not the raw sources):

File	What
gene_names.json	978 LINCS landmark gene symbols, in feature order
feature_names.json	Full flat-feature names `[genes
lincs_gene_stats.npz	Per-gene LINCS (mean, std) for z-scoring patient inputs
matched_samples.parquet	LINCS×PharmacoDB matched (cell, drug) rows
training_data.parquet	Full training set — 3.24M rows [sig_id, cell_iname, smiles, dose, inhibition, gene_expression(978), ecfp4(1024)] (1.5 GB)

The full training parquet is included here so the repo is self-contained; it is also mirrored at Tino3141/lincs-pharmaco-training / Tino3141/pharmaco_flat.

Eval datasets (eval/) — two separate cohorts

eval/aml_eval/ — the 12-patient AML benchmark

The primary patient benchmark (scTherapy paper; consolidated from the public Tino3141/aml12-drug-response-eval dataset). Patients patient1 … patient12.

Path	What
ground_truth/dss.csv / .parquet	Ground-truth Drug Sensitivity Scores per (patient, drug) — the eval target
ground_truth/supplementary_scTherapy.xlsx	scTherapy supplementary Excel the DSS were extracted from
model_inputs/full_inputs.parquet	Assembled (patient × drug) model inputs
model_inputs/patient_deg_vectors.parquet	978-gene log2FC vector per patient
model_inputs/drug_fingerprints.parquet	1024-bit ECFP4 per drug
model_inputs/example_full_input.parquet	Small worked example of the input layout
degs/	Differential-expression gene sets per patient (raw + filtered + meta)

eval/zenodo_eval/ — multi-cancer held-out set (NOT AML)

A separate single-cell drug-response set from public GEO/Zenodo studies — HNSCC (SCC47, JHU006, HN120/137), lung (PC9, H1975), CLL, HCC, breast (FCIBC02), etc., across 10 drugs. Each pseudo-bulk carries a binary sensitive/resistant label. Built from the ../tino/*.pt source (named after the author).

Path	What
expression_zscore.parquet	Dose-expanded expression, LINCS z-score (built by scripts/build_zenodo_zscore.py)
expression_logfc.parquet	Dose-expanded expression, LINCS log2FC (built by scripts/build_zenodo_logfc.py)

eval/reference/ — shared lookups

Path	What
lincs_landmark_genes.csv	The 978 LINCS landmark genes
drug_name_to_smiles.json	Drug-name → canonical SMILES map

Each folder has its own README.md. Raw AML patient .RDS files (~5 GB) are not here — re-downloadable from Zenodo 13340927.

Results (results/)

Per-drug / per-patient prediction CSVs, metric tables, and plots from the headline runs (exploratory/superseded variants were pruned):

Dir	Cohort	What
aml_evals/	AML	per-arch/head/seed AML predictions — the main neural-vs-LGBM comparison (filenames encode the epoch used)
aml_lgbm_seed42/	AML	LGBM baseline AML eval (summary, per-drug, per-patient AUC)
zenodo_cluster/	Zenodo	run with all 5 models in one aggregated.csv + ROC plots
zenodo_logfc_all_final_final/	Zenodo	final full Zenodo-cohort run
zenodo_logfc_within_sweep/	Zenodo	within-file hyperparameter sweep grid
split_comparison/	—	random vs cell vs drug vs cell_drug split analysis (the data-leakage finding)