pipnn

Fast, graph-based approximate-nearest-neighbor indexing for building the k-NN graph in single-cell / scanpy workflows — a Rust implementation (with PyO3 bindings) of the PiPNN method. It plugs into scanpy as a drop-in sc.pp.neighbors backend and also exposes a small standalone API. It builds the graph quickly at recall close to exact, and scales to tens of millions of cells.

Independent implementation of the PiPNN algorithm. Please cite the paper: Tobias Rubel, Richard Wen, Laxman Dhulipala, Lars Gottesbüren, Rajesh Jayaram, Jakub Łącki. PiPNN: Ultra-Scalable Graph-Based Nearest Neighbor Indexing. arXiv:2602.21247

Install

uv pip install pipnn      # or: pip install pipnn

Prebuilt wheels for Linux (x86_64, aarch64), macOS (arm64, x86_64), and Windows (x86_64); Python 3.9+. No Rust toolchain needed to install.

Quickstart

As a scanpy backend (drop-in for the default neighbor step):

import scanpy as sc
from pipnn import PiPNNTransformer

sc.pp.neighbors(adata, n_neighbors=15, use_rep="X_pca",
                transformer=PiPNNTransformer())
# then sc.tl.umap / sc.tl.leiden as usual

Standalone (just the k-NN graph):

import numpy as np
from pipnn import self_knn_graph

X = np.random.default_rng(0).normal(size=(100_000, 50)).astype("float32")
indices, distances = self_knn_graph(X, n_neighbors=15)
# indices, distances: shape (n, k+1), self edge first, sorted by distance

What it does

• Produces the exact-distance k-NN graph scanpy expects — a scipy CSR of shape (n, n) with k+1 entries per row, the self edge first.
• euclidean (default, matches scanpy on PCA space) and cosine metrics.
• Parallel Rust build; deterministic given a fixed random_state.
• Optional comparison backends under pipnn.contrib behind the same transformer API: a native HnswTransformer, FaissTransformer (pip install pipnn[faiss]), and GlassTransformer (pyglass).

Benchmarks & source

Full benchmarks — SIFT, and the Tahoe-100M single-cell atlas from 1M to 10M cells (build time, recall, and peak memory vs FAISS, pyglass, and pynndescent) — plus development notes and the comparison harness, are in the repository:

https://github.com/iandriver/pipnn-scanpy

MIT licensed.