salad

Backbone generation is the first stage of the machine-learning protein design pipeline: produce a plausible fold, thread a sequence onto it, then filter with a structure predictor. The generators that dominated this stage inherited AlphaFold's pair representation — triangle multiplication, dense pair attention — which scales quadratically or worse with chain length. In practice that meant slow sampling and a designability cliff somewhere past 300 residues.

salad, from Michael Jendrusch and Jan Korbel at EMBL Heidelberg, removes the pair stack. Each residue attends only to a small set of neighbours drawn from sequence index, Euclidean proximity, and a random sample, reducing attention cost from quadratic to linear in chain length. The models denoise all-atom structures — backbone, sequence, and side-chain positions together — hence sparse all-atom denoising.

The second contribution is structure-editing: a way to apply a frozen salad checkpoint to design tasks it was never trained on, by editing the partially denoised structure at each step rather than conditioning the network. Aligning a motif into the intermediate coordinates scaffolds that motif; symmetrising them produces repeat proteins; sharing coordinates across parallel trajectories produces multi-state designs. This puts salad in the same family as RFdiffusion, Chroma, and Proteina, but with the task-adaptation happening entirely outside the weights.

#Key Features

• Sparse neighbour attention: A sparse variant of invariant point attention computes pair features and attention weights only over each residue's neighbour set, with no persistent pair representation and no triangle multiplication.
• Speed at length: Generating a 1,000-residue backbone takes about 19 seconds on a single RTX 3090 — up to an order of magnitude faster than Chroma and up to two orders of magnitude faster than RFdiffusion on large proteins.
• Noise schedule variants: Variance-preserving, length-scaled variance-preserving, and variance-expanding models trade designability against protein size; VP-scaled holds median scRMSD below 2 Å to 600 residues and VE to 800.
• Domain-shaped initial noise: Seeding the reverse process with noise clustered into domain-sized blobs rather than a single Gaussian raises designability at 1,000 residues to 36.7%, matching or improving on hallucination-based large-protein design.
• Zero-shot task transfer: Structure-editing handles motif scaffolding, point- and screw-symmetric repeat proteins, shape-conditioned design, and multi-state design without retraining or task-specific conditioning.

#Technical Details

The full model — sparse encoder, six-layer diffusion module, and amino acid decoder — has 11.9M parameters, against 18.5M for Chroma, 15.7M for Genie 2, and 59.8M for RFdiffusion. Training used a PDB snapshot from October 2023 restricted to entries deposited before 31 December 2020, filtered at 4 Å resolution and clustered with MMseqs2 at 30% sequence identity; each run took 200,000 iterations, roughly 3.5 days on eight RTX 3090-class GPUs.

Designability is measured by generating sequences with ProteinMPNN and refolding with ESMFold, counting a backbone designable at scRMSD below 2 Å. Between 50 and 400 residues, the VP model is comparable to Genie 2 and ahead of RFdiffusion despite a much smaller training set. On the Genie 2 motif-scaffolding benchmark, both the editing and the explicitly conditioned variants solve 23 of 24 single-motif and 5 of 6 multi-motif tasks, adding one multi-motif case Genie 2 does not solve. Shape conditioning generated backbones spelling out the model's name, with 55% designable under the strict scRMSD and pLDDT criteria and 92.5% refoldable under the looser TM-score criterion used for Chroma. The multi-state benchmark is, to the authors' knowledge, the first solved by a structure generative model.

#Applications

The obvious beneficiaries are design groups working beyond the few-hundred-residue regime — multi-domain scaffolds, large binding surfaces, extended repeat architectures — where earlier diffusion generators lose designability. The speed makes broad in silico screening practical on a single consumer GPU rather than a cluster. Structure-editing is the more general contribution: because the recipe operates on coordinates rather than weights, a design task with no training data, such as a protein required to adopt two distinct folds, can still be attacked with a pretrained model. Ready-made pipelines for unconditional design, motif scaffolding, and binder design ship in the companion flexcraft package, and a Colab notebook runs unconditional generation without local installation.

#Impact

salad is a demonstration that the AlphaFold-derived pair stack is not required for high-quality protein structure generation, and that dropping it buys both speed and length in one move. Its structure-editing framing has broader reach than the model itself, since the same editing operations apply to any denoising generator. The work appeared in Nature Machine Intelligence after a February 2025 preprint. Code is released under Apache 2.0 and model parameters under CC BY 4.0, with a Zenodo deposit carrying the checkpoints, generated structures, and a code snapshot; a known incompatibility with JAX 0.5.1–0.5.3 affects symmetric design. All reported evaluation is computational — designability means self-consistency against ProteinMPNN and ESMFold, with no wet-lab characterisation of the designs.