About AutoZyme

AutoZyme is building the optimization layer for scientific software: a way to make the tools researchers already trust run faster, without changing how they are used or the results they produce. The rest of this page is why.

Speed is the neglected axis

Bioinformatics software is built for correctness first, and that is the right call: a fast answer that is wrong is worthless. The tools the field depends on are good tools. But correctness is where the effort goes, and performance is left behind. Going back to make mature code run faster takes bandwidth, specialized optimization knowledge, and an incentive to maintain a hot loop someone else wrote, and the field rarely has all three at once. So as datasets keep growing, bottlenecks accumulate in code that no one has time to revisit.

Why now

AI coding agents are getting better and cheaper at a steady clip. We built a way to point them at this problem automatically and at scale, and to do it verifiably: every candidate speedup is checked against frozen reference outputs before it is accepted, so faster never means different results. What we keep finding, across 45 finalized tasks in 9 scientific domains, is that the headroom is large. The slowness is usually not the algorithm but the implementation, and for many functions a real speedup is simply sitting there unclaimed. See Benchmarks for the measured numbers.

Every optimization compounds

Optimizing one function is not a one-off win. Every patch we ship benefits everyone who runs that tool, and summed across the field the saved time and compute is enormous: this is leverage on the pace of science, not one lab's convenience. The framework is also self-evolving: each accepted patch feeds back into it, so the system gets better at finding the next one as the catalog grows.

Team

AutoZyme is developed in the Kendziorski Lab at the University of Wisconsin–Madison. Contact: [email protected] · github.com/ElliotXie.

Also check out CASSIA, our cell annotation tool published in Nature Communications: CASSIA: a multi-agent large language model for automated and interpretable cell annotation.

How to participate

• Use AutoZyme or build a patch: start with the documentation.
• Nominate the next slow function: Suggest & Vote.
• Contribute a verified optimization: follow the contribution workflow.

@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}
}