Research map / Probabilistic and scientific modeling / Molecules and proteins
Flow Autoencoders are Effective Protein Tokenizers
Rohit Dilip, Evan Zhang, Ayush Varshney, David Van Valen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add best-of-2 sampling by log-likelihood | Kanzi-AR (eta=0.66) · Unconditional generation | designability: 0.562 -> 0.617 | Improved |
| add score/noise scaling (eta=0.66) in sampling | Kanzi-AR · Unconditional generation | designability: 0.328 -> 0.562 | Improved |
| add score/noise scaling (eta=0.66) in sampling | Kanzi-AR · Unconditional generation | alpha helix %: 71.9 -> 88.7 | Hurt |
| add classifier-free guidance and score/noise scaling (starred setting) | Kanzi (30M) · CAMEO reconstruction | RMSD: 0.936 -> 0.817 | Improved |
| add pair bias (ablation) | Kanzi 11M · CAMEO | RMSD: 1.812 -> 1.741 | Improved |
| add pair bias (ablation) | Kanzi 11M · CATH | RMSD: 1.519 -> 1.551 | Hurt |
| add self-conditioning (ablation) | Kanzi 11M · CAMEO | RMSD: 1.812 -> 1.670 | Improved |
| add pair bias and self-conditioning together (ablation) | Kanzi 11M · AFDB | RMSD: 1.282 -> 1.233 | Mixed |
| add token mixing (windowed attention) in encoder instead of point-wise MLP (ablation) | Kanzi encoder · downstream generation | generation quality | Improved |
| use invariant encoder with non-invariant decoder instead of learning pose (ablation) | Kanzi · reconstruction and generation | quality | Hurt |
| add learned rotational invariance (ablation) | Kanzi · CATH | reconstruction RMSD: about 0.4 A worse | Hurt |
| add cross-attention conditioning in decoder (ablation) | Kanzi · reconstruction | performance | Hurt |
Closest papers by meaning
- Protein Structure Tokenization: Benchmarking and New Recipe (ICML 2025 · similarity 0.90)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.87)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.87)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.86)
- Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design (ICML 2024 · similarity 0.86)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
