AutoZyme

AutoZyme is an autonomous multi-agent framework that speeds up scientific software on CPUs while preserving the original results. Install AutoZyme-Library for drop-in accelerators of Seurat, Scanpy, and 30+ packages, or run the framework to optimize a function it doesn't ship yet.

Using pre-built patches

Three steps in either language: install, activate a package, then call its functions exactly as before.

R

# install
remotes::install_github("ElliotXie/autozyme", subdir = "autozyme_r")

# activate  (Seurat, InferCNV, cellchat, and more)
library(autozyme)
autozyme::activate("seurat")

# run your existing code, now faster
markers <- FindAllMarkers(obj)

The R package compiles C++ kernels at install. This only really trips up Windows (no built-in compiler): install Rtools matching your R version first, or the build fails. macOS (Xcode Command Line Tools) and Linux (gcc) normally already have a compiler.

Python

# install
pip install autozyme

# activate  (Scanpy, Squidpy, CellPhoneDB, and more)
import autozyme
autozyme.activate("scanpy")

# run your existing code, now faster
sc.tl.rank_genes_groups(adata, "leiden")

Activate as many packages as you need. Run autozyme::dashboard() (R) or python -m autozyme (Python) to list every available patch.

Optimizing a new function

AutoZyme can also optimize functions it doesn't ship yet, driven by any coding agent that can run shell commands (Claude Code, Codex, Cursor, and others). The examples below target Seurat's FindAllMarkers; swap in your own function and repo. Run it two ways.

Autonomous (one session)

The manager agent runs the whole pipeline for you: it launches workers, waits for them, checks progress every 15 minutes, and advances through scaffold, init, validate, iterate, scaling, and package. Paste this to your agent:

Clone https://github.com/ElliotXie/autozyme.git,
install the CLI (pip install -e autozyme/autozyme_cli/).

I want to optimize FindAllMarkers from Seurat.
Repo: https://github.com/satijalab/seurat

Read and follow autozyme/autozyme_cli/zyme/prompts/manager/0_pipeline.md

Typical wall time is 4 to 10 hours, depending on the target function.

Manual (phase by phase)

Scaffold the task, either by pasting the bootstrap prompt to your agent:

I want to speed up FindAllMarkers from Seurat using AutoZyme.
Clone https://github.com/ElliotXie/autozyme.git,
install the CLI (pip install -e autozyme/autozyme_cli/),
then read and follow autozyme/autozyme_cli/zyme/prompts/Bio/0_bootstrap.md

or by running the CLI yourself:

git clone https://github.com/ElliotXie/autozyme.git
cd autozyme && pip install -e autozyme_cli/

mkdir ../test_findallmarkers && cd ../test_findallmarkers
zyme init https://github.com/satijalab/seurat FindAllMarkers

Then paste each phase prompt in a fresh agent session:

1. Init: prompts/1_init.md fills the scaffold, picks datasets, and records baselines. Then run zyme validate init.
2. Iterate: prompts/2_iterate.md runs the 50-round optimization loop. Then run zyme validate iterate.
3. Scaling validation: prompts/3_validate_scaling.md checks OOD and threading robustness.
4. Package: prompts/4_package.md lifts the converged patch into autozyme_r or autozyme_py.

Contributing

AutoZyme-Library grows by community demand, and contributions of every kind are welcome:

• Request or vote on what we optimize next at autozyme.com: nominate a function or package, upvote the bottlenecks you hit most, and track requests through optimization, validation, and release.
• Contribute a patch: optimize a function yourself with the framework above, then open a pull request to add it to AutoZyme-Library.
• Share feedback or report bugs: open a GitHub issue with ideas, comments, or problems.

Datasets

Benchmark datasets are hosted on HuggingFace: elliotxie/autozyme-datasets

hf download elliotxie/autozyme-datasets \
  --repo-type dataset --include "single_cell/raw/pbmc68k.rds"

Citation

@article{xie2026autozyme,
  title   = {AutoZyme: An Autonomous Agentic Framework to Optimize Bioinformatics Software},
  author  = {Xie, Elliot and Cheng, Lingxin and Cai, Yujia and Shireman, Jack and Kendziorski, Christina},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.06.12.731250},
  url     = {https://www.biorxiv.org/content/10.64898/2026.06.12.731250v1}
}