ProteinZen

ProteinZen is a generative model for all-atom protein design that builds novel protein structures using SE(3) flow matching. Developed by Alex J. Li and Tanja Kortemme at UCSF and released as a preprint in October 2025, it addresses a persistent gap in structure-generation methods: most backbone generators produce only the coordinates of the protein main chain and defer side-chain placement to a separate step, which can compromise atomic-scale precision at functional sites.

ProteinZen instead generates protein structure at the all-atom level in a single unified framework. The method decomposes each residue into oriented rigid bodies and applies flow matching over the SE(3) manifold — the space of rigid-body rotations and translations — to iteratively transform noise into physically plausible structures. This lets the model reason jointly about backbone geometry and the atoms that determine chemistry and binding, rather than treating them as independent problems.

The model targets both unconditional generation, where new structures are sampled from scratch, and conditional motif scaffolding, where a fixed functional motif must be embedded within a newly designed protein. It sits alongside recent all-atom and frame-based generators such as Protpardelle and frame-flow methods, distinguishing itself through its rigid-body decomposition and strong sequence-structure consistency.

#Key Features

• All-atom generation: Produces complete atomic structures in one framework rather than generating a backbone and appending side chains separately, improving atomic-scale precision at functional sites.
• SE(3) flow matching: Decomposes residues into oriented rigid bodies and learns a continuous flow over rotations and translations, a stable and efficient alternative to score-based diffusion for structure generation.
• Motif scaffolding: Supports both indexed and unindexed conditioning modes to graft a specified functional motif into a designed scaffold, a core task for binder and enzyme design.
• Strong self-consistency: Achieves high sequence-structure consistency while retaining competitive structural diversity, meaning designed backbones can be refolded to the intended shape from their predicted sequences.

#Technical Details

ProteinZen frames protein generation as a flow-matching process on the SE(3) manifold, representing each residue as a set of oriented rigid bodies so that the full complement of atoms is denoised together with the backbone frame. Generation proceeds by integrating the learned flow over a configurable number of denoising timesteps (400 by default), with sampling available in both ODE and SDE modes. The released checkpoints are trained on single-chain monomers, with separate weights provided for unconditional generation and for motif scaffolding. On the Protpardelle motif-scaffolding benchmark the model reports low motif-grafting failure rates, performing best in indexed mode where the placement of the motif within the sequence is specified. Code is released under the MIT license, with pretrained weights distributed through a UCSF Box download.

#Applications

ProteinZen is aimed at protein engineers and computational biologists designing new proteins from scratch or scaffolding functional sites — for example, building a stable protein around a binding epitope, a catalytic motif, or a metal- coordinating cluster. Because it generates side-chain atoms alongside the backbone, it is particularly relevant when atomic detail at the functional site drives the design objective, such as enzyme active-site or binder-interface engineering. Designs can be filtered by refolding predicted sequences and checking agreement with the generated structure.

#Impact

ProteinZen contributes to the rapid shift from backbone-only generators toward unified all-atom design, where structure and chemistry are modeled jointly. By combining SE(3) flow matching with a rigid-body all-atom representation and reporting state-of-the-art motif-scaffolding results, it offers a concrete recipe for atomic-precision generation with an openly available implementation. As a preprint with released code and weights, its broader adoption and experimental validation will develop as the community applies it to real design campaigns.