All-Atom Protein Sequence Diffusion

Generative models of protein sequence almost universally operate over the 20 canonical amino acids. That alphabet excludes a large and useful part of protein chemistry: non-canonical amino acids, D-stereoisomers, and residues carrying post-translational modifications, all of which matter for peptide therapeutics, enzyme engineering and biomaterials. This work asks what happens if a sequence generator instead writes proteins at the level of atoms, using SELFIES — a molecular string representation designed so that every syntactically valid string decodes to a chemically valid molecule.

The model was developed by Amelia Villegas-Morcillo, Gijs J. Admiraal, Marcel J.T. Reinders and Jana M. Weber in the Department of Intelligent Systems at Delft University of Technology. A preprint appeared on bioRxiv in June 2025 and the peer-reviewed version in the Journal of Cheminformatics. It is a research-stage study rather than a production design tool, and the authors are explicit that the all-atom variant does not yet generate fully valid proteins consistently.

Architecturally it stays close to established sequence diffusion work: a modified ByteNet backbone inside the D3PM discrete denoising diffusion framework, adapted from EvoDiff. The novelty lies in the representation and in a matched evaluation pipeline that checks whether a generated SELFIES string decodes to a peptide with a continuous backbone, correct constitution and correct stereochemistry.

#Key Features

• All-atom SELFIES representation: Each residue is written as its atomic composition rather than a single letter, so the generative alphabet is not restricted to the canonical twenty and can express non-canonical side chains and modified residues.
• Head-to-head noise schedule comparison: Uniform noising (random token replacement) and absorbing noising (progressive masking) are trained under otherwise identical conditions, and the absorbing schedule wins on validity by a wide margin.
• Chemical validity pipeline: Generated strings are decoded and checked for a continuous peptide backbone, constitutionally correct residues, and L-stereochemistry — a validity metric letter-alphabet models never need and therefore never report.
• Novelty and diversity gains: Successfully generated all-atom sequences are more novel and more diverse by BLAST-based measures than their amino-acid counterparts, indicating the representation explores a wider region of sequence space.
• Open release: Data and code are published under the MIT licence, including the evaluation pipeline, reusable for any model generating peptides as molecular strings.

#Technical Details

The backbone is a 38M-parameter ByteNet — a dilated convolutional architecture that scales to long sequences — adapted from EvoDiff, trained within D3PM using transition matrices for the two noise schedules and the standard βt = (T − t + 1)⁻¹ timestep parameterisation. Training used UniRef50, roughly 14 million sequences, filtered to 30 to 100 residues containing only canonical amino acids, in a length-balanced 90/10 train/validation split; the SELFIES vocabulary comprises 21 tokens against the amino acid model's 20.

From 1,000 generated sequences, the absorbing-schedule all-atom model produced 239 with a continuous backbone (against 52 for uniform), yielding 77 fully canonical proteins and 150 containing non-canonical residues, versus 4 and 44 for the uniform schedule. Among continuous-backbone sequences, 98.0% of residues were constitutionally correct and 97.9% were in the L configuration. Novelty and diversity, measured by BLAST at e-value < 0.05, reached 99.9% and 99.7% for non-canonical outputs and 99.8% and 97.8% for canonical ones, against 95.7% and 94.6% for the amino-acid baseline and 83.6% and 94.1% for ProtGPT2. OmegaFold pLDDT averaged 64.2 ± 11.9 for canonical all-atom outputs under the uniform schedule against 58.5 ± 13.1 for the amino-acid absorbing model, with 32.1% of all-atom canonical proteins exceeding a pLDDT of 70 versus 11 to 19.4% for the amino-acid models. Generated proteins are short: 37.3 ± 15.3 residues for canonical outputs and 60.9 ± 30.7 for non-canonical ones.

#Applications

The design space this opens is peptides and small proteins containing residues a 20-letter model cannot write — non-canonical amino acids used to tune protease resistance or add bioorthogonal handles, and PTM-bearing residues that carry biological signal. In its current form the practical use is as a research platform rather than a design service: the low yield of valid sequences means a run produces few usable candidates, and there is no conditioning mechanism for a specified function, fold or binding target. The validity pipeline is independently useful for anyone evaluating molecular-string generators on peptides.

#Impact

The study's value is largely negative-result value, and it is honest about that. Moving proteins into an atom-level representation costs a great deal of generation reliability — most sampled sequences do not decode to a valid peptide — while buying measurable gains in novelty, diversity and, for canonical outputs, predicted foldability. The comparison of uniform against absorbing noise schedules is a transferable finding for anyone building discrete diffusion models over biological sequences. Adoption is limited by the short sequence lengths, the absence of conditional generation, and the lack of wet-lab validation; the released code and evaluation pipeline are the most reusable outputs.