Protpardelle

Sidechains do nearly all of a protein's chemistry, yet generative models of protein structure have largely ignored them. The reason is circular: to know which sidechain atoms exist at a position you need the amino acid, and to know the amino acid you need the sequence, at which point the design problem is already solved. The usual workaround — generate a backbone, thread a sequence onto it, pack sidechains afterwards — leaves the chemical groups that do the work with no influence over the fold built around them.

Protpardelle breaks the circularity with a representation its authors call a superposition. Instead of committing to one residue identity per position, it carries the atoms of all twenty possible sidechains at once and noises them jointly with the backbone. At each denoising step the superposition is collapsed against a current sequence estimate to yield a single chemically valid protein, that structure is denoised, and a fresh sequence estimate updates the superposition. The forward noising process only ever sees real proteins with one sidechain per residue, so the superposition never has to be modeled during training — the machinery lives entirely in the sampler.

The model was developed by Alexander Chu and colleagues in Po-Ssu Huang's group at Stanford University, released as a preprint in May 2023 and published in PNAS in June 2024. A backbone-only generator falls out as a special case of the same network, and the successor Protpardelle-1c replaces its guidance-based motif conditioning with direct channel-wise conditioning.

#Key Features

• Sidechain superposition: All twenty candidate sidechains at a position are denoised at once and collapsed onto the residue the sequence model picks, so structure and sequence constrain each other throughout generation rather than in separate stages.
• Per-atom integration steps: Backbone atoms denoise at every iteration while a sidechain atom advances only when its amino acid is selected, so different atoms follow different discretizations of the same ODE.
• Co-design with a fast sequence predictor: A noise-conditional "miniMPNN" — the ProteinMPNN architecture with its autoregressive mask removed, cutting sampling from O(N) to O(1) — supplies sequence estimates along the trajectory, and full ProteinMPNN designs the final sequence.
• Backbone- and rotamer-free motif conditioning: The model can be conditioned on bare chemical groups — three glutamate carboxylates and a histidine imidazole ring, say — and asked to invent a fold that presents them.
• Deliberately lightweight: Both models train in two to three days on a single NVIDIA A40, and the backbone model samples at roughly 0.002 seconds per residue, far faster than comparable structure generators.

#Technical Details

The score network is a U-ViT with a hidden dimension of 256 and six residual noise-conditional transformer layers, with convolutional up- and down-sampling removed and a patch size of one residue, so every residue is its own token; preconditioning follows the EDM formulation of Karras et al. The network is not SE(3)-equivariant, relying instead on random rotation and translation augmentation, which buys speed. Training used the CATH S40 domain set with the Ingraham splits, on crops up to length 512 for the backbone model and 256 for the all-atom model. The backbone model generates proteins from 50 to 400 residues, with at least 90% of samples reaching self-consistency RMSD below 2 Å up to length 300 and about 70% beyond it. The all-atom model covers 50 to 256 residues at roughly 60% success under the same threshold up to length 150, and exceeds 80% under the looser scTM > 0.5 criterion across its full range, using ESMFold as the structure oracle. Generated sidechains reproduce natural distributions of bond lengths, bond angles and chi angles, and score in a tolerable range under the Rosetta fa_dun rotamer energy without relaxation.

#Applications

The model suits protein engineers who want a design organized around a chemical requirement rather than a shape. The paper demonstrates two such cases: scaffolding the TGF-β1-binding loop of a de novo monobody with its sidechains intact, and building a fold around an iron-coordinating cluster from cytochrome b1 given only the coordinating functional groups. Both are the shape of problem in enzyme and metalloprotein design, where a site's geometry is the specification and the backbone around it is free. The backbone-only model serves as a fast unconditional generator.

#Impact

Protpardelle established sidechain superposition as a workable answer to the discrete-continuous problem in all-atom protein generation. Its limits are real: the models are single-chain only, conditioning on sparse input often needs reconstruction guidance because the network will otherwise ignore the motif, and sample quality degrades above roughly 300 residues for the backbone model and 150 for the all-atom one. All reported results are computational, resting on ESMFold self-consistency rather than experimental characterization. Code and model parameters are released under the MIT license, the training code was distributed separately on request, and the repository was retired in August 2025 in favor of Protpardelle-1c.