Code for accumap, imaging-free reconstruction of bead positions in diffusion-based, optics-free spatial transcriptomics (Oakes & Pachter, Computational array reconstruction with accumap for spatial transcriptomics).
accumap models each receiver-by-sender count as Poisson, with the expected count set by bead capture, sender abundance and a transfer kernel of the distance between the two beads. Starting from a UMAP embedding, it fits the kernel and then warps and refines the layout to maximize the likelihood. Two physical priors supply what the counts leave undetermined: beads are deposited at uniform density, and contacts between neighboring beads have no preferred direction. Imaging is never used by the method; in the paper it is used only to score reconstructions.
src/accumap/: the installable package.count_model: the method (reconstruct,umap_layout; diagnosticscontact_anisotropy,fit_kernel_at).metrics: the paper's evaluation (evaluate,neighbors_kept,similarity_align).realistic_sim: the simulator calibrated to experimental data (simulate_realistic).real_data: loads the Hu et al. ground-truth replicates (load_hu).
analysis_scripts/: notebooks that walk through the paper and regenerate its figures and tables (below).paper_results/holds the cached results they read;baselines/holds the original method, similarity-weighted multidimensional scaling (similarity MDS), which the paper uses as a baseline and which is not part of the package.final_figures/: the paper's figures (Fig#_final.png,FigS#_final.png), as written by the notebooks.data/: input data (gitignored; see below).
git clone https://github.com/pachterlab/OP_2025_2.git
cd OP_2025_2
pip install -e . # accumap (installs PyTorch)
pip install -r requirements.txt # notebook dependencies
accumap runs on CPU, but a CUDA GPU is strongly recommended.
from accumap.count_model import reconstruct, umap_layout
# C: (n_receivers, n_senders) UMI counts, dense or scipy sparse
start = umap_layout(C, seed=0)
result = reconstruct(C, start, array_diameter=3000.0) # the array's physical diameter sets the scale (here um)
receivers, senders = result["receivers"], result["senders"]For arrays in which every bead both sends and receives, pass the square bead-by-bead matrix with shared=True (and umap_layout(C, shared=True)). n_starts=3 runs several UMAP initializations and keeps the most likely layout. accumap warns if the counts indicate that transfer favors one direction; for such arrays use iso_center=True (see SCOPE_Simulator.ipynb). The reconstruct docstring documents every option.
Each notebook draws its figures and tables from the cached results in analysis_scripts/paper_results/ and runs a small live example. Setting RECOMPUTE = True in a notebook reruns its analyses from scratch (CUDA GPU required) and writes to a separate folder, so the cache is never overwritten. Each notebook states its runtime and the data it needs.
| Notebook | Produces |
|---|---|
Method_Overview.ipynb |
Fig. 1 (the method, step by step on a toy array) |
Experimental_Ground_Truth.ipynb |
Fig. 2, Supp Figs S2–S3, Supp Tables S3–S4 |
Simulation_Benchmark.ipynb |
Supp Fig S1, Supp Tables S1–S2 |
Design_Choices.ipynb |
Supp Fig S4, Supp Table S5, the number of passes |
SCOPE_Simulator.ipynb |
the SCOPE results (Supplementary Note 1; Supp Table S2) |
Original_Method.ipynb |
Supp Fig S5 and Supp Table S6 (the original method, similarity MDS; Supplementary Note 3), and the reviewer-response Figs R1–R3 and Table R1 |
Original_Method.ipynb draws its figures and tables from cached results only; the runs behind them were development scripts.
The experimental ground truth (four replicates of Hu et al., 2025) is available from the Single Cell Portal study SCP2577. Place the files in data/ or point $ACCUMAP_DATA at their folder. Cached results let every notebook draw its figures without the data. Concordex domain analyses also need the Slide-seq cerebellum section (data/cerebellum_processed.h5ad).
See LICENSE.