Protpardelle-1c

Protpardelle-1c is a family of all-atom protein structure diffusion models from the Protein Design Lab at Stanford University, built for conditional generation: scaffolding a functional motif, refining sidechains, and generating multi-chain complexes against a specified target surface. It updates Protpardelle (Chu et al., PNAS 2024), which introduced Euclidean all-atom diffusion for proteins but handled motif conditioning only through reconstruction guidance — a brittle scheme whose guidance scale had to be retuned per problem, and which could not see the motif's sidechains at all.

The central change is architectural rather than scalar. Motif coordinates are concatenated channel-wise into the model input, so the network is trained to attend to both backbone and sidechain atoms of the motif and can be sampled with classifier-free guidance instead of gradient-based reconstruction. That single change moved the MotifBench score from 4.97 to 28.16, ahead of RFdiffusion's 21.27, and it arrives alongside a lighter noise-embedding scheme borrowed from DiT that cut trainable parameters from 33M to 22M.

The result is a small, fast model that is competitive with far larger ones. Sampling all 3,000 MotifBench backbones takes 40 minutes on a single NVIDIA A100-80GB, against 31 hours for RFdiffusion, and on the all-atom motif scaffolding benchmark the model matches La-Proteina with roughly ten times fewer parameters.

#Key Features

• Sidechain-conditioned motif scaffolding: The model conditions on motif sidechain atoms, not just the backbone, so generated scaffolds place polar and hydrophobic motif residues in chemically appropriate environments rather than burying or exposing them arbitrarily.
• Classifier-free guidance: Conditional and unconditional models are amortized into one network by dropping the conditioning input 5% of the time during training, removing the per-problem guidance-scale tuning that made reconstruction guidance unreliable.
• Multi-chain hotspot conditioning: The cc83 checkpoint treats binder design as motif scaffolding against a target chain, conditioned on 3-8 hotspot residues, and generalizes to two- and three-chain targets despite training only on chain pairs.
• A family of checkpoints, not one model: Separate weights cover unconditional generation, backbone-only and all-atom conditional generation, sidechain tip-atom conditioning, sequence- provided refinement, and molecular-dynamics-trained conformational sampling.
• Small and fast: At 22M parameters the model samples quickly enough to make large-scale screening of scaffold hypotheses routine on a single GPU.

#Technical Details

Protpardelle-1c follows the EDM diffusion framework and is deliberately non-equivariant: inputs are randomly rotated by a uniform SO(3) sample and translated during training, and the model picks a frame during denoising. Training uses two curated datasets derived from CATH. AI-CATH contains ESMFold-predicted structures of 32 ProteinMPNN redesigns per CATH domain, filtered to self-consistent structures with scRMSD below 2.0 Å and pLDDT above 80, yielding 337,936 structures; MD-CATH subsamples 32 conformations per domain from 320 K molecular dynamics trajectories of 5,398 CATH domains, followed by Rosetta cartesian minimization. The backbone-only motif scaffolding model cc58 was trained for 4.38M steps on one A100-80GB. On MotifBench it solves 22 of 30 problems with 164 unique solutions (RFdiffusion: 16 problems, 192 solutions). On the RFdiffusion/La-Proteina all-atom benchmark the all-atom cc91 model solves 22 of 26 tasks with 208 unique solutions, against 4 tasks and 4 solutions for the original Protpardelle. All-atom sampling interleaves a retrained noise-conditional MiniMPNN with the structure denoiser, with full ProteinMPNN designing the final sequence. On the BindCraft target set, cc83 is outperformed by BindCraft on every target except DerF21 — expected for an unfiltered base model that is not optimized against AlphaFold2 metrics. Motif indices must be pre-specified, unlike the unindexed conditioning offered by RFdiffusion2 and La-Proteina.

#Applications

The model suits protein engineers who need scaffolds that hold a functional site — an enzyme active site, a metal-binding cluster, a neutralizing epitope — in a precise geometry, and who want to sample broadly enough to find several structurally distinct solutions per problem. The multi-chain checkpoints extend this to binder design against a defined epitope, and the all-atom refinement checkpoint supports partial diffusion of loops such as antibody CDR3 while holding the framework fixed. Its modest compute footprint makes it practical for groups without large GPU allocations.

#Impact

Protpardelle-1c is a demonstration that careful conditioning design, self-consistent training data, and correct handling of unresolved residues matter more than parameter count for motif scaffolding. Code is released under the MIT license and weights are archived on Zenodo, making the whole family reproducible. The work is a preprint and has not completed peer review, and all results reported are in-silico: the benchmarks rely on ESMFold self-consistency and, as the authors note, do not penalize designs whose functional site is occluded, so success rates should be read as a filter for experimental testing rather than a substitute for it.