Modeling Non-Canonical Amino Acids and Post-Translational Modifications

This direction is split into other directions: non-canonical residue representation. Its last version stays readable and citable here.

Version 1 of 134 papers37 conceptsfrozen 2026-10-04

Modern protein language models are remarkably good at predicting the structure and function of natural proteins. However, they suffer from a fundamental "vocabulary bottleneck": they are almost exclusively trained on the 20 standard amino acids defined by the genetic code. In reality, the functional proteome is far more complex, featuring post-translational modifications (PTMs), non-canonical amino acids (NCAAs), and synthetic chemical linkages that standard models simply cannot "see."

Recent research is breaking this bottleneck by shifting from discrete biological tokens to atom-level chemical representations, allowing AI to design and simulate the full spectrum of modified biomolecules.

The Vocabulary Bottleneck

Most state-of-the-art models, such as ESM-2, treat proteins as strings of 20 characters. While efficient, this abstraction fails when a residue is modified—for example, by phosphorylation or the addition of a synthetic side chain. As noted by Jiahui Zhang and colleagues in their work on PepBenchmark, this limitation creates a significant gap in therapeutic development, as many modern peptide drugs rely on non-natural components for metabolic stability and potency.

When a model encounters a non-canonical residue, it typically treats it as an "unknown" token or defaults to its wild-type precursor. This ignores the critical biochemical diversity that dictates how a protein actually behaves in a cell.

From Tokens to Atoms

To overcome this, researchers are moving toward all-atom representations. Instead of viewing a residue as a single point or a categorical label, these models treat the protein as a collection of atoms in 3D space.

Daniel Flam-Shepherd and Kevin Zhu demonstrated that chemical language models can be scaled to generate entire proteins atom-by-atom. By representing molecules as atomic strings rather than amino acid sequences, these models can "break" the genetic code to design hybrid molecules like protein-drug conjugates.

Similarly, Han Tang and Wouter Boomsma introduced UNAAGI, which uses E(3)-equivariant diffusion to generate amino acid side chains atom-by-atom. This allows the model to predict the effects of both natural and synthetic substitutions by focusing on the underlying physics of the atoms rather than a fixed list of residues.

Bridging Biology and Chemistry

Another approach involves fusing the strengths of protein-centric models (which understand sequence context) with chemical-centric models (which understand molecular structure).

  • PepDoRA: Leyao Wang, Pranam Chatterjee, and their team developed PepDoRA, a unified peptide language model. By fine-tuning chemical models using Weight-Decomposed Low-Rank Adaptation, they created a representation capable of processing both natural and chemically modified peptides.
  • PTM-Mamba: To specifically address modifications, Peng Fei and Pranam Chatterjee introduced PTM-Mamba. This model explicitly incorporates PTMs into its architecture using bidirectional gated Mamba blocks, enabling high-accuracy predictions for how modifications influence disease and druggability.

Designing Beyond Nature

The ultimate goal of this research is not just to model existing modifications, but to invent new ones. Generative frameworks like PepINVENT, developed by Gökçe Geylan and others, use transformer-based text-infilling to design peptides that include entirely novel non-natural amino acids.

This expansion of the searchable chemical space is vital for overcoming the limitations of natural peptides, such as poor cell permeability and rapid degradation. By incorporating non-natural building blocks and complex topologies—such as the cyclic structures modeled by Xiangxin Zhou in CPSDE—researchers can now navigate a chemical landscape orders of magnitude larger than what nature provides.

Why It Works: The Shift to Generalization

The success of these methods stems from a shift in perspective: treating biological sequences as specific instances of general chemistry.

Whether it is Steven A. Benner expanding the genetic alphabet to create new DNA folding motifs, or Anthony Marchand and Bruno Correia designing proteins to bind to neosurfaces formed by small molecules, the core mechanism is the same. By moving away from rigid, predefined vocabularies and toward flexible, atomistic, or chemically-aware representations, these models gain the ability to generalize to any molecular structure, natural or synthetic.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_modeling_non_canonical_amino_acids_and_p_f27ae15f,
  title = {Modeling Non-Canonical Amino Acids and Post-Translational Modifications},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-non-canonical-amino-acids-and-post-translational-modifications/rdr_f27ae15f010c4be48c84bfd41c341e60},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_modeling_non_canonical_amino_acids_and_p_f27ae15f_v1,
  title = {Modeling Non-Canonical Amino Acids and Post-Translational Modifications},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-non-canonical-amino-acids-and-post-translational-modifications/rdr_f27ae15f010c4be48c84bfd41c341e60/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}