pVR: p-adic Bi-Filtered Simplicial Complexes for Genomic Classification

Code and data for the paper p-adic Bi-Filtrations for Topological Machine Learning on Genomic Sequences.

_{Overview of the pVR pipeline. The figure is adapted from Figure 3 of the paper.}

Notation

The code uses D_H for the compositional distance matrix (L1 distance on k-mer frequency vectors); the paper denotes this D_c. Same object.

Requirements

A full conda environment is given in pvr_env.yml, but it includes many extra packages. Minimally, the pipeline needs: numpy, scipy, scikit-learn, xgboost, gudhi, joblib, biopython, pandas.

The Nucleotide Transformer comparison (nt_compare.py) additionally needs torch, transformers, and umap-learn. Plotting needs matplotlib.

GUDHI is easiest via conda: conda install -c conda-forge gudhi.

Data

The benchmark sequences and labels are included under data/ (low-sample) and data_large/ (large-sample). To rebuild them from NCBI instead, see download_data.py and download_large.py (an NCBI Entrez email is required).

Running

Full pipeline (main results, ablations, sensitivity) on each regime:

python run_full_eff.py --datadir data/       --outdir results/
python run_full_eff.py --datadir data_large/ --outdir results_large/ --skip_sensitivity

Each dataset completes in seconds on a 12-core CPU; no GPU is needed.

Reproducing the tables

python make_tables_ci.py   --small results/all_results.json \
                           --large results_large/all_results.json --out tables_ci.tex
python make_aux_tables.py  --small_results results/all_results.json \
                           --large_results results_large/all_results.json \
                           --small_data data/ --large_data data_large/

Precomputed results are in results/all_results.json and results_large/all_results.json.

Files

• pvr_eff.py — core method (encoding, distances, bi-filtration, classification)
• baselines.py — FFP-JS, NVM, MinHash baselines
• run_full_eff.py — full experiment pipeline
• nt_compare.py — Nucleotide Transformer v2 comparison
• make_tables_ci.py, make_aux_tables.py — regenerate paper tables
• plot_results.py, make_figure.py — figures
• download_data.py, download_large.py — NCBI data acquisition

Citation

If you use this code or data, please cite our paper:

@article{dash2026pvr,
  title   = {$p$-adic Bi-Filtrations for Topological Machine Learning on Genomic Sequences},
  author  = {Dash, Tirtharaj and Sachdeva, Gunja},
  journal = {arXiv preprint arXiv:2606.06117},
  year    = {2026},
  doi     = {10.48550/arXiv.2606.06117}
}