From Atoms to Fragments: A Coarse Representation for Efficient and Functional Protein Design

Leonardo V. Castorina, Christopher W. Wood, Kartic Subr

Abstract

Although deep learning has accelerated protein design, current protein representations such as sequences or full-atom structures scale non-linearly with protein length. We propose a sparse and interpretable representation for proteins, based on evolutionarily conserved fragments. Specifically, we use a curated set of 40 functional and evolutionarily conserved fragments as an alphabet to build Fragment Graphs and Fragment Sets. These fragment-based representations are both lightweight and functionally informative, capturing up to 55% more variance using fewer dimensions than traditional methods. On a dataset of 215 functionally diverse proteins, our approach creates more coherent functional clusters than traditional sequence- and structure-based methods, even among proteins with ≤ 30% sequence identity. Fragment-based searches execute ∼68.7× faster than RMSD-based structural methods and ∼1.64× faster than sequence-based methods. Additionally, our representation effectively guides RFDiffusion for protein backbone generation with functional recovery rates higher than 40%.

Introduction

Protein design and analysis are currently caught in a computational "pincer movement." Large Language Models (LLMs) treat proteins as sequences of amino acids, which is fast but often misses the 3D structural context essential for function. Conversely, diffusion models and structural aligners treat proteins as clouds of thousands of atoms, which is accurate but computationally expensive and scales poorly as proteins grow. This paper introduces a "coarse-grained" representation—an intermediate level of abstraction that is lightweight enough for massive database searches but detailed enough to capture the "functional signatures" governing biological activity.

Methodology

The authors propose representing proteins as Fragment Graphs or Fragment Sets based on a curated library of 40 ancient, evolutionarily conserved structural fragments (e.g., motifs that bind DNA, ATP, or metal ions).

  1. Detection: A sliding-window algorithm scans protein structures, using sequence identity and backbone torsion angles (ϕ and ψ) to match segments to the 40-fragment library.
  2. Representation:
    • Fragment Sets: A "bag-of-words" approach listing present fragments.
  3. Fragment Graphs: A map where nodes are fragments and edges represent physical peptide bonds or spatial proximity (<10 Å).
  4. Design Integration: These fragments serve as "blueprints" for generative models like RFDiffusion. By masking non-functional regions, the model is tasked with "filling in the gaps" to create functional protein backbones.
Overview of the fragment-based representation process, converting a full-atom structure into a Fragment Graph and a simplified Fragment Set.

Key Results and Insights

Efficiency and Scalability

The fragment-based approach offers dramatic improvements in computational performance. The "BagOfNodes" method is approximately 68.7× faster than RMSD-based structural methods and 1.64× faster than sequence-based BLOSUM searches. Furthermore, it reduces memory requirements by 90–99%, enabling massive database searches on standard consumer hardware.

Comparison of computational efficiency, showing that fragment-based methods (BagOfNodes) significantly outperform traditional methods in query time and memory requirements.

Functional Information Density

Fragment-based representations capture significantly more functional variance (up to 95%) in just 20 dimensions compared to traditional sequence or structural metrics. This allows the model to identify functional similarities between proteins that share little to no sequence identity (≤ 30%). In generative tasks, using these fragments as templates for RFDiffusion achieved functional recovery rates between 40% and 100%.

Evolutionary Stability

Fragment regions were found to have ∼15% more internal hydrogen bonds than non-fragment regions. This suggests these fragments are self-contained, stable structural units that evolution has preserved as modular components for biological function.

Conclusion and Future Work

While the current library is limited to 40 curated fragments, it demonstrates that a coarse, motif-based alphabet can outperform full-atom representations in both speed and functional clustering. Future work will focus on unsupervised machine learning to discover new fragments directly from the Protein Data Bank (PDB) and developing probabilistic models to handle subtle backbone variations more gracefully. Additionally, the authors emphasize the importance of biosecurity screening as these tools become more capable of designing specific functional motifs.

Summary updated

Abstract

Abstract Motivation Although deep learning has accelerated protein design, current protein representations such as sequences or full-atom structures scale non-linearly with protein length. We propose a sparse and interpretable representation for proteins, based on evolutionarily conserved fragments. Specifically, we use a curated set of 40 functional and evolutionarily conserved fragments as an alphabet to build Fragment Graphs and Fragment Sets. These fragment-based representations are both lightweight and functionally informative, capturing up to 55% more variance using fewer than of the dimensions required by traditional methods. Results On a dataset of 215 functionally diverse proteins, our approach creates more coherent functional clusters than traditional sequence- and structure-based methods, even among proteins with ≤ 30% sequence identity. Fragment-based searches of protein databases achieve accuracies comparable to traditional methods, while using 90% fewer tokens per protein. These searches execute ∼68.7× faster than RMSD-based structural methods and ∼1.64× faster than sequence-based methods, even including fragment pre-processing overhead. Additionally, we show that our representation effectively guides RFDiffusion for protein backbone generation with functional recovery rates higher than 40%. In summary, our fragment-based representation offers a scalable and interpretable alternative for the next generation of protein design tools for backbone design, sequence design, and functional similarity searches within protein structure databases. Availability https://github.com/wells-wood-research/tessera (Documentation to be made available upon acceptance)
BibTeX
@article{castorina2026from,
  author = {Leonardo V. Castorina and Christopher W. Wood and Kartic Subr},
  title = {{From Atoms to Fragments: A Coarse Representation for Efficient and Functional Protein Design}},
  journal = {Bioinformatics},
  year = {2026},
  doi = {10.1093/bioinformatics/btag172},
}
Figure