Pallatom

De novo protein design has mostly been a relay. A backbone generator such as RFdiffusion proposes a trace of residue frames, a fixed-backbone design model such as ProteinMPNN writes a sequence onto it, and a folding model such as ESMFold checks that the sequence recovers the backbone. The atoms that actually make a protein work — the side chains forming the hydrophobic core, the hydrogen bonds, the salt bridges — never enter the generative step at all.

Pallatom, from LEVINTHAL Biotechnology in Hangzhou, collapses that relay into one model by making all-atom coordinates the object being generated. Learning P(all-atom) implicitly learns the joint distribution over structure and sequence, because a cloud of atoms in the right places already encodes hydrophobicity, polarity, and hydrogen-bonding capacity. The network diffuses that cloud directly; an auxiliary head then reads amino acid identity out of the denoised geometry, and redundant virtual atoms are discarded. The sequence is a consequence of the atoms, not a condition imposed on them.

The work was first posted in August 2024 and revised through 2025 under the title "P(all-atom) Is Unlocking New Path For Protein Design," with code and a trained checkpoint released alongside it.

#Key Features

• Joint structure-sequence generation: One sampling run yields all-atom coordinates and a matching sequence, so no separate inverse-folding pass is needed to make a design synthesizable.
• atom14 representation: Every residue is given 14 uniform atom slots, sidestepping the fact that amino acids differ in atom count without committing to a residue identity up front.
• Dual-track decoder: Local attention over atoms and global attention over residues are fused inside repeated decoder units, with a differentiable recycling path that carries atomic detail back into the residue track.
• Out-of-distribution length scaling: Trained on crops of at most 128 residues, the model still achieves its best designability at lengths of 150 to 400.
• Efficient sampling: 200 diffusion steps take 10.2 seconds for a 100-residue protein on a single RTX 4090, and 57.5 seconds at length 400 — roughly 5 times faster than RFdiffusion and 16 times faster than ProteinGenerator at that size.

#Technical Details

The denoiser is a JAX implementation built around an AtomFeature encoder feeding eight AtomDecoder units, using 256-dimensional single representations, 128-dimensional pair representations, 8 attention heads, triangle updates, and AlphaFold 3-style feature broadcasting between residue and atom levels. Diffusion follows the EDM formulation with a lognormal noise schedule and 100% self-conditioning. Training used Adam at learning rate 1e-3, batch size 32, and 3 x 10^5 steps over roughly 10 days on four A6000 GPUs. The corpus is deliberately small and heavily filtered: 7,459 PDB entries below 3 A resolution, plus 27,697 AFDB-cluster structures surviving a pLDDT floor of 80, a 128-residue cap, and packing-density, core-residue, secondary-structure, and Foldseek redundancy filters.

On the paper's co-design benchmark — 250 samples at each of seven lengths from 60 to 120 residues — Pallatom reaches 85.03% all-atom designability (mean pLDDT above 80 and all-atom RMSD under 2 A), against 62.74% for Multiflow, 43.14% for ProteinGenerator, and 30.00% for Protpardelle, while also producing the most structure and sequence clusters (291 and 1,466) and the lowest maximum TM-score to the PDB (0.719). Backbone-only designability is competitive rather than dominant: 89.89% under a pLDDT constraint, between ProteinGenerator's 93.14% and Multiflow's 84.69%. Ablations show both innovations matter — swapping atom14 for a Protpardelle-style hybrid representation drops all-atom designability from 87% to 5%, and removing recycling drops it to 21% while leaving backbone quality untouched.

#Applications

The model suits de novo design of small monomeric proteins where side-chain packing decides whether a design folds — mini-binder scaffolds, compact structural modules, and starting points for directed evolution. Its training distribution was chosen for proteins that commercial oligo-pool synthesis can produce cheaply, so generated candidates are ordering-ready. Because a single run returns both the structure and a sequence that folds to it, the usual generate-then-redesign-then-refold loop shortens to generate-then-verify, which matters most when a campaign needs thousands of candidates rather than a handful.

#Impact

Pallatom's contribution is a demonstration that sequence need not be modelled as a separate discrete variable at all: made explicit at the atomic level, geometry carries enough chemistry for identity to be recovered afterwards. That is a cleaner formulation than one-hot sequence diffusion or post-hoc side-chain packing, and the ablation against a sequence-guided variant shows why — noisy intermediate structures produce unreliable sequence predictions whose errors compound through reverse diffusion. Scope is honestly narrow. The released model handles single unconditioned monomers, with no support for binding-target conditioning, motif scaffolding, or complexes; validation is entirely in silico, with no expressed designs reported; and both code and weights are CC BY-NC-SA 4.0, restricting commercial use. The preprint has not been peer reviewed.