# Flow Autoencoders are Effective Protein Tokenizers

Source: [Flow Autoencoders are Effective Protein Tokenizers](<https://www.sota2.com/research/paper/68fa5f871e209131ce9eab2e>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T19:20:46.140Z

Authors: Rohit Dilip, Evan Zhang, Ayush Varshney, David Van Valen

Published: 2025-09-30T23:29:39.000Z

## Abstract

Protein structure tokenizers enable the creation of multimodal models of protein structure, sequence, and function. Current approaches to protein structure tokenization rely on bespoke components that are invariant to spatial symmetries, but that are challenging to optimize and scale. We present Kanzi, a flow-based tokenizer for tokenization and generation of protein structures. Kanzi consists of a diffusion autoencoder trained with a flow matching loss. We show that this approach simplifies several aspects of protein structure tokenizers: frame-based representations can be replaced with global coordinates, complex losses are replaced with a single flow matching loss, and SE(3)-invariant attention operations can be replaced with standard attention. We find that these changes stabilize the training of parameter-efficient models that outperform existing tokenizers on reconstruction metrics at a fraction of the model size and training cost. An autoregressive model trained with Kanzi outperforms similar generative models that operate over tokens, although it does not yet match the performance of state-of-the-art continuous diffusion models. Code is available here: https://github.com/rdilip/kanzi/.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Protein Generation on Protein Structure Generation (test)](<https://www.sota2.com/research/sota/protein-generation-on-protein-structure-generation-test.md>): Protein Generation; Protein Structure Generation (test); Designability: 0.617.

- [Protein Reconstruction on AFDB 512 samples](<https://www.sota2.com/research/sota/protein-reconstruction-on-afdb-512-samples.md>): Protein Reconstruction; AFDB 512 samples; RMSD: 1.18.

- [Protein Reconstruction on CATH 512 samples](<https://www.sota2.com/research/sota/protein-reconstruction-on-cath-512-samples.md>): Protein Reconstruction; CATH 512 samples; RMSD: 1.09.

- [Protein Reconstruction on CAMEO (512 samples)](<https://www.sota2.com/research/sota/protein-reconstruction-on-cameo-512-samples.md>): Protein Reconstruction; CAMEO (512 samples); RMSD: 0.98.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2510.00351>)

[PDF](<https://arxiv.org/pdf/2510.00351.pdf>)
