Restore the plot-simulation profile and activate this checkout's paths before running the commands below. See the data guide and reproduction instructions. The default author-machine paths documented below are overridden by the supplied activation script or scripts/run_study.py.
Evaluate recovery of directional MI edges and molecular programs in simulated spatial transcriptomics, producing analysis figures and statistics associated with Fig. 2a-d, S3 and S4.
Run from this directory in the existing SpiderNet environment with numpy, pandas, scipy, matplotlib, scikit-learn, scanpy, torch, nbformat, nbclient and ipykernel. Full spatial comparisons also need COMMOT and Julia with ScCChain/CSV/DataFrames.
python run_benchmarks.py --check
python run_benchmarks.py --plot-only
python run_benchmarks.pyThe default order is summary, sample, insitu: aggregate six settings by 30 repeats, calculate both loading-rank scopes from saved loadings/activity, plot a representative sample, then recompute its NMF-LR/COMMOT/ScCChain comparisons. The 180 SpiderNet models and full-grid comparator results are upstream inputs. Default sample selection is --setting 0 --experiment 18 (Dropout1/Experiment_18); summary still uses all 180 repeats.
Plot-only reads saved summaries, rank tables/definitions/checksums, and spatial-score caches verified against input/source hashes. It neither fits nor merges upstream tables and fails without valid caches. First calculate rank tables with --stage summary. --check --plot-only checks plotting inputs. Notebooks use the calling interpreter and save executed copies under output/executed/.
Optional --stage train and --stage generate use SpiderNet_Simulation.py and SimulationData_generation.py. Without sample arguments they process all 180 samples and write upstream Results/Data; generation overwrites selected data. The generator uses hash() seeds, which are not guaranteed to match across processes.
Inputs are the five raw CSVs per sample in D:/SpiderNet/Data/Simulation/<setting>/Experiment_<number>/, plus adata_simulation.h5ad for spatial comparisons and cell_neigh_metaIprop.csv for training. D:/SpiderNet/Results/Simulation/ supplies SpiderNet/COMMOT/NMF_LR metrics and ScCChain_analysis/Spacia_analysis *_metrics.csv. Rank summaries need gene/edge metadata, ProcessedData/LR_list.pkl, loading_LR_use.npy, loading_sender_use.csv, loading_receiver_use.csv, Factor_envir_use.npy, and EdgeProgramScores.csv for all samples. The representative sample also uses loading arrays/CSVs and processed model inputs.
Override roots with --data-root, --result-root, --spacia-root, --output-root, or SIMULATION_DATA_ROOT, SIMULATION_RESULT_ROOT, SIMULATION_SPACIA_ROOT. No other Tutorial directory is required; the Julia runner is local.
Outputs are under output/: Merged_Benchmark/ for combined CSVs and rank definitions; Merged_Benchmark/Summary_Plots/ for figures/statistics; SpiderNet/<setting>/Experiment_<number>/SpiderNet_Result_Mode_cell_class/ for sample results; Simulation_Data/ for diagnostics; and ScCChain_analysis/ for Julia intermediates. Two diagnostics requiring unavailable latent expression are copied with provenance.
Sender/receiver ranks compare 10 versus 10 features within 20 genes and 10 versus 70 within all 80; LR uses 10 versus 10 in both. Ranks increase with loading and use average ties, column normalization and one-to-one activity-Spearman matching. Ratios are averaged across two MIs, then tested against 1 across 30 repeats using two-sided t tests. _all80 identifies the second scope. LoadingRankRatio_boxplot_by_setting_loading.pdf is the separate share/top-1 measure, not the mean-rank ratio. Spatial comparisons retain LOAD_COMPATIBLE_SPACIA=False. See the 20-feature definition (output/validation/rank_ratio_addition.md, supplied in plot-simulation) and 80-gene definition (output/validation/rank_ratio_all80_addition.md, supplied in plot-simulation) for limitations.