Full-Atom Generative Modeling for Protein Interactions

This direction is merged into another direction: backbone-only structural modeling. Its last version stays readable and citable here.

Version 1 of 142 papers53 conceptsfrozen 2026-10-04

For years, the "gold standard" in computational protein design was the alpha-carbon backbone. By simplifying a protein to a ribbon of $C\alpha$ atoms, researchers could model global folds while ignoring the "noise" of thousands of side-chain atoms. However, as the field moves from predicting static shapes to designing functional binders and enzymes, this simplification has become a bottleneck.

The emerging research direction focuses on full-atom modeling, where every heavy atom in a protein—including the intricate side-chains—is modeled and designed simultaneously. This shift is driven by a simple realization: while the backbone determines the fold, the side-chains determine the function.

The Backbone Bottleneck

Traditional models often treat proteins as geometric paths in 3D space. While effective for folding, this "coarse-grained" view fails in the "lock and key" world of molecular recognition. As noted in Sable by Qiwei Ye and colleagues, side-chains are the primary drivers of chemical reactivity and binding specificity. Omitting them limits a model's ability to predict how an antibody actually "grabs" its target.

When models like KarmaLoop (Tianyue Wang et al.) or dyMEAN (Xiangzhe Kong and Wenbing Huang) incorporate full-atom detail, they capture the steric clashes and fine-grained hydrogen bonding that backbone-only models miss. This is particularly critical in loop regions, where Deniz Akpinaroglu showed in DeepSCAb that side-chain packing is actually a primary determinant of the backbone's own shape.

Sub-Angstrom Precision and Physical Feedback

The move to all-atom representations allows AI models to "speak the language" of physics. In Energy-based Alignment (EBA), Jiarui Lu and Vijil Chenthamarakshan demonstrate that all-atom models can receive direct feedback from classical force fields. Because the model "sees" every atom, it can calculate real-world thermodynamic properties, ensuring generated proteins are not just pretty shapes, but physically plausible structures that respect the Boltzmann distribution.

This precision is essential for targeting small, highly charged functional groups. Magnus S. Bauer and Jason Z. Zhang used RFD2-MI to design binders for post-translational modifications with sub-angstrom accuracy. By coordinating specific atoms within a phosphate group, they achieved a level of specificity that backbone-focused design simply cannot reach.

Co-Designing Sequence and Structure

A recurring theme across this research is the "co-design" of sequence and structure. Rather than designing a backbone and then "filling in" the amino acids (side-chain packing), new frameworks like FAMPNN (Talal Widatalla) and APM (Ruizhe Chen and Xiangxiang Zeng) generate both simultaneously.

This works because the chemical environment of the side-chains provides "geometric cues" that inform the sequence. In PepFlow, Jiahan Li and Shitong Luo use multi-modal flow matching to jointly model backbone geometry and side-chain torsions on their respective mathematical manifolds. They found that including side-chain information actually improves the accuracy of the backbone design itself, suggesting that the two are fundamentally inseparable.

From Scaffolding to Active Sites

The ultimate application of this all-atom shift is enzyme and drug design. Traditional "motif scaffolding" required researchers to guess where a functional site should sit within a sequence. RFdiffusion2 by Jason Yim and Woody Ahern flips this: it starts with the 3D coordinates of the active-site atoms (the "theozyme") and builds the entire protein scaffold around them.

Similarly, FAIR (Zaixi Zhang et al.) and DTMol (Haotian Teng et al.) use all-atom pocket representations to achieve docking success rates that rival or exceed traditional physics-based scoring. By treating the protein surface as a detailed atomic landscape rather than a smooth tube, these models can identify the precise "handshake" required for a drug to bind to its target.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_full_atom_generative_modeling_for_protei_151e1a62,
  title = {Full-Atom Generative Modeling for Protein Interactions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/full-atom-generative-modeling-for-protein-interactions/rdr_151e1a6286fb42e2adc257a9d1963d6a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_full_atom_generative_modeling_for_protei_151e1a62_v1,
  title = {Full-Atom Generative Modeling for Protein Interactions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/full-atom-generative-modeling-for-protein-interactions/rdr_151e1a6286fb42e2adc257a9d1963d6a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}