Benchmarking single-cell dynamics: code, model checkpoints, and processed single-cell datasets

Code, trained model checkpoints, and processed single-cell .h5ad datasets for a benchmark of single-cell dynamics / trajectory-inference / optimal-transport flow methods. Code benchmarking_code.tar.gz — the full benchmarking pipeline: per-method prepare/train/evaluate scripts, the cord-blood benchmark driver and its SLURM job definitions, cross-method aggregation (fate accuracy, Wasserstein-2 global and per-clone), and the generated report tables and figures under reports/. Includes environment.yml, pip requirements, and a MANIFEST.txt recording the originating git commit and per-file SHA-256 checksums. Cord-blood dataset (rebuild, not redistributed) The cord-blood expression data is not deposited here. It is available from doi:10.6084/m9.figshare.27908142 (CordBlood_Refine.zip → adata_update.h5ad, CC BY 4.0). That release carries the same 24,885 cells and the same expression matrix, but not the annotations and embeddings this benchmark trains on. cordblood_sidecar.tar.gz supplies exactly those missing fields — 6 obs columns (Well, Annotation, Time_point, label_man, split, timepoint_tx_days), 8 obsm embeddings (diffusion-map eigenvectors, scaled PCA, and the delta-embeddings), 2 obsp graphs, and 9 uns entries (population priors and the PCA/DM scalers). All are keyed by cell barcode, so the join does not depend on row order. To reconstruct the exact h5ad used in the paper: python scripts/cordblood/data_prep/01_rebuild_from_figshare.py \ --sidecar cordblood_sidecar --out data/cordblood_addpop.h5ad python scripts/cordblood/data_prep/02_verify_rebuild.py \ --ref data/cordblood_addpop.h5ad --new data/cordblood_rebuilt.h5ad The rebuild has been verified to reproduce the original h5ad exactly (76/76 equality checks: X, raw, layers, every obs column, obsm, obsp and uns key). Checkpoints (one tarball per method) DeepRUOT_checkpoints.tar.gz MIOFlow_checkpoints.tar.gz otcfm.tar.gz (OT-CFM) pdp+_checkpoints.tar.gz (pseudodynamics+) PRESCIENT_checkpoint.tar.gz scDiffeq-checkpoints.tar.gz sf2m_checkpoints.tar.gz (SF2M) TIGON_checkpoints.tar.gz TNJ_checkpoints.tar.gz (TrajectoryNet) Datasets (AnnData / HDF5) klein_addpop.h5ad — the processed LARRY clonal dataset (126,861 cells), the main cross-method benchmarking dataset. meahr_monocle.h5ad tom_pos.h5ad synthetic_FP.h5ad — 2-D synthetic Fokker-Planck simulation. synthetic_FP_5D.h5ad — 5-D synthetic Fokker-Planck simulation. Used by the C1/C2 extrapolation analyses (scripts/rebuttal_C1C2/); it is not a cross-method benchmarking dataset. synthetic_FP_5D_divide_s4.h5ad — 5-D synthetic Fokker-Planck simulation with cell division (seed 4). This is the dataset behind the pseudodynamics+ growth-rate, lambda_g and RCG-sweep experiments. The scripts refer to it as synthetic_FP_5D_divide_formal.h5ad; symlink or rename it to that path before running them. Runtime: the code is intended to run inside the Singularity container archived separately at 10.5281/zenodo.18944745.