Han Tang

Han Tang

PhD Candidate @ University of Copenhagen, interested in AI for Science (Chemistry and Life Science)

UNAAGI: Atom-Level Diffusion for Non-Canonical Amino Acid Substitutions

9 minute read

Published:

What amino acid could replace a residue while still fitting its local environment? With UNAAGI, we approach this question by generating the replacement side chain directly at the atom level. The model builds atom types, positions and bonds, so its proposals can include both canonical and non-canonical amino acids.

This post adapts our presentation at the Gubra meeting on 1 October 2026, with work by Han Tang, Alexander Tong and Wouter Boomsma. It walks through the method, examples of generated substitutions, and comparisons with published binding measurements. The original slide deck is also available.

Generating a side chain from noise

Diffusion models learn to generate structures through a pair of processes. During training, a forward process adds noise to examples. The model learns how to reverse that corruption, allowing it to start from noise and progressively construct a new sample.

Diffusion schematic showing data becoming noise in the forward process and noise becoming data in the reverse process.
The forward and reverse diffusion processes. The same general idea can be applied to molecular structures.

For UNAAGI, the object being generated is a single amino acid side chain in its structural context. We remove the side chain beyond Cα, hold the backbone and surrounding structure fixed, and ask the model to build a replacement using the surrounding atoms as context. Atom types, positions and bonds are generated together.

This representation lets the model propose side chains without selecting from a predefined amino acid list. In the sampling setup shown here, UNAAGI receives no peptide sequence, assay results or original residue identity. Its input is the local atomic environment that remains after the side chain is removed.

UNAAGI method schematic: a side chain is removed beyond C alpha and replaced by jointly generated atom types, coordinates and bonds in a fixed structural environment.
Local side chain generation with a fixed backbone and surrounding structure. Open the figure to inspect it at full size.

The animation below illustrates this process at position 7 of the CP2 macrocycle bound to KDM4A (PDB 5LY1), where the native residue is serine. In the first half, the side chain loses its structure and atom identities. In the second half, UNAAGI reconstructs serine from noise.

The forward half is an illustration made by playing the reconstruction backwards. The pause separates it from the reverse generation process; the backbone and pocket stay fixed throughout.

Testing substitutions in two peptides

To examine whether the proposed substitutions are useful, we compare them with the nonproteinogenic deep mutational scanning study by Rogers, Passioura and Suga (2018). The study tested a set of 40 amino acids, including 21 non-canonical amino acids, by replacing one residue at a time in two peptides:

  • PUMA, a helical peptide that binds MCL-1: 34 positions × 39 substitutions.
  • CP2, a cyclic peptide that binds KDM4A: 12 positions × 39 substitutions.

The non-canonical set includes longer side chains, N-methylated residues and D-alanine, extending the substitutions beyond the usual canonical amino acid vocabulary.

The PUMA assay uses a reference peptide containing the M144A substitution. Ala144 in the examples below therefore refers to this benchmark reference.

The amino acids and CP2 complex used in the benchmark
Chemical structures of the 21 nonproteinogenic amino acids in the experimental benchmark.
Non-canonical amino acids in the benchmark. Panel A from Rogers et al. (2018), CC BY-NC-ND 4.0.
Structure of the cyclic CP2 peptide bound to KDM4A.
CP2 bound to KDM4A. Panel F from Rogers et al. (2018), CC BY-NC-ND 4.0.

The experimental readout is the change in binding free energy relative to the original peptide, ΔΔG, in kcal/mol. Negative values mean stronger binding; positive values mean weaker binding. For this analysis, we call a substitution tolerated when ΔΔG < 0.5 kcal/mol, allowing a small loss in binding.

PUMA–MCL-1 mutational scan: red cells indicate stronger binding and blue cells weaker binding across positions and amino acid substitutions.
PUMA binding to MCL-1. Panel E from Rogers et al. (2018), CC BY-NC-ND 4.0.
CP2–KDM4A mutational scan showing stronger binding in red and weaker binding in blue.
CP2 binding to KDM4A. Panel G from Rogers et al. (2018), CC BY-NC-ND 4.0. In both heatmaps, red indicates stronger binding and blue indicates weaker binding.

How to read the examples: UNAAGI generates substitutions without access to these assay results. The ΔΔG values below come from the published experimental benchmark. They are measurements for the corresponding substitutions, rather than binding energies predicted by the diffusion model or new experiments on these generated structures.

What the model generates at individual sites

We can first look at the structural context of one site. The following animation zooms into Leu141 in PUMA, shown in green, within the MCL-1 complex. Removing the side chain beyond Cα leaves the local environment from which UNAAGI generates a replacement.

The native Leu141 site in PUMA–MCL-1. Use the video controls to play or replay each animation.

A hydrophobic site: PUMA Ala144

At Ala144, UNAAGI samples several residues whose substitutions improve measured binding. These include the non-canonical residue norleucine, alongside isoleucine, phenylalanine, leucine and tryptophan. In this example, direct atom generation reaches a non-canonical alternative with a binding effect similar to isoleucine.

Ala144 — benchmark referenceThe local environment before substitution.
IsoleucineMeasured ΔΔG: −1.87 kcal/mol.
Norleucine · non-canonicalMeasured ΔΔG: −1.86 kcal/mol.
PhenylalanineMeasured ΔΔG: −1.68 kcal/mol.
LeucineMeasured ΔΔG: −1.64 kcal/mol.
TryptophanMeasured ΔΔG: −1.08 kcal/mol.

More alternatives at CP2 Thr13 and PUMA Ala139

Other sites also admit several generated alternatives with favourable measured effects. At CP2 Thr13, examples include phenylalanine (−0.96), histidine (−0.47) and non-canonical norvaline (−0.20). At PUMA Ala139, the model produces leucine (−0.80), tyrosine (−0.20) and non-canonical norleucine (−0.04). All values are ΔΔG in kcal/mol.

Watch the CP2 Thr13 and PUMA Ala139 examples
CP2 Thr13Native site.
CP2 → phenylalanineMeasured ΔΔG: −0.96 kcal/mol.
CP2 → histidineMeasured ΔΔG: −0.47 kcal/mol.
CP2 → norvaline · non-canonicalMeasured ΔΔG: −0.20 kcal/mol.
PUMA Ala139Native site.
PUMA → leucineMeasured ΔΔG: −0.80 kcal/mol.
PUMA → tyrosineMeasured ΔΔG: −0.20 kcal/mol.
PUMA → norleucine · non-canonicalMeasured ΔΔG: −0.04 kcal/mol.

Tolerated substitutions at restrictive sites

The examples also include sites where fewer substitutions are tolerated. UNAAGI generates glycine at PUMA Ala145, and tert-butylalanine at Leu148 and Leu141. The Leu141 substitution has a slightly positive ΔΔG, so it is tolerated under our threshold but does not improve measured binding.

PUMA siteGenerated substitutionMeasured ΔΔG (kcal/mol)
Ala145Glycine−1.71
Leu148tert-Butylalanine−0.04
Leu141tert-Butylalanine+0.02
Ala145 → glycineMeasured ΔΔG: −1.71 kcal/mol. View the native site.
Leu148 → tert-butylalanineMeasured ΔΔG: −0.04 kcal/mol. View the native site.
Leu141 → tert-butylalanineMeasured ΔΔG: +0.02 kcal/mol.

Agreement with experimental rankings

Individual examples show what generation looks like, but the broader comparison asks whether model rankings agree with the measured effects on binding. In the comparison presented here, UNAAGI has the highest correlation for non-canonical substitutions in both peptides among the displayed baselines. Across all substitutions, it leads on PUMA and performs similarly to the AlphaFold 3 iPTM score on CP2.

Comparison of model rank correlations with measured binding effects for PUMA and CP2, with separate results for all substitutions and non-canonical substitutions.
Correlation with experimental binding effects for PUMA and CP2. Higher bars indicate closer agreement between model rankings and measured substitution effects.

We also evaluate canonical substitutions across 25 assays from ProteinGym as a complementary check. This comparison addresses conventional amino acid substitutions across a broader set of assays; the 25 assays are a subset of ProteinGym.

Spearman correlations across 25 ProteinGym deep mutational scanning assays comparing UNAAGI with ten baselines.
Canonical substitution evaluation on 25 ProteinGym DMS assays, comparing UNAAGI with ten baselines. Open the plot at full size to inspect individual assays.

These comparisons address agreement with experimental substitution effects in the evaluated settings. The examples remain local side chain substitutions with a fixed backbone and environment; extending the method to broader design tasks requires further evaluation.

Scoring chosen residues and learning from assays

Two directions are in progress.

Conditional diffusion for scoring a chosen amino acid. We are developing a model that takes both the local atomic environment and a target residue identity, and estimates how well that chosen amino acid fits. This would support scoring specific canonical and non-canonical substitutions, alongside generating candidate side chains.

Preference optimisation using experimental results. We plan to use assay measurements to favour substitutions with better binding: generate candidates, measure them, update the model, and use the new data to guide the next round of generation. This is a planned extension of the current work.

The results so far suggest that local atomic context can guide proposals beyond the canonical amino acid vocabulary. The next step is to make those proposals easier to score for a chosen residue and to improve them using experimental feedback.

References and figure credits

  1. Rogers, J. M., Passioura, T., & Suga, H. (2018). Nonproteinogenic deep mutational scanning of linear and cyclic peptides. Proceedings of the National Academy of Sciences, 115(43), 10959–10964. Benchmark panels A, E, F and G are reproduced as supplied in the deck, without modification. © 2018 the authors; published by PNAS under CC BY-NC-ND 4.0.
  2. Notin, P., et al. (2023). ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. Advances in Neural Information Processing Systems, 36, 64331–64379.

The UNAAGI schematics, generated structure animations and result plots accompany the original presentation by Han Tang, Alexander Tong and Wouter Boomsma, presented on 1 October 2026.

Leave a Comment

You May Also Enjoy

An Overview of Machine Learning Models for NCAA and PTM modeling

7 minute read

Published:

AlphaFold 2 [1] has largely solved the problem of predicting the 3D structure of proteins from their amino acid sequences, though there is still debate over the extent to which this problem is truly solved — a view I share. AlphaFold 3 [2] and a number of subsequent models [3,4,5] have extended this success to protein complexes and proteins with small-molecule binders. However, all of these are primarily designed for canonical amino acids (CAAs) and do not account for non-canonical amino acids (NCAAs). Given that NCAAs play an increasingly important role in drug discovery and protein engineering, it is crucial to review the current state of machine learning models for predicting the structure of proteins containing NCAAs and post-translational modifications (PTMs). In this post, I review the recent progress in this field, with a critical view on the extent to which these models offer fruitful solutions.