The Unreasonable Redundancy of Nature's Protein Folds

Natural Protein Folds are Highly Redundant

Natural protein sequences are vast, but their corresponding 3D folds are remarkably redundant. While metagenomic databases like MGnify provide billions of sequences, Ligo Biosciences' analysis reveals that these do not translate into a proportional increase in structural diversity. Instead, the majority of natural proteins reuse a relatively small set of stable, expressible, and adaptable fold solutions.

This redundancy creates a significant challenge for training generative biomolecular models. Simply scaling training data by folding more natural sequences often results in adding many sequence variants of the same fold families rather than introducing genuinely new structural examples. This mismatch between sequence diversity and fold diversity means that the "scale the data" recipe for improving deep learning models may yield diminishing returns if the goal is to expand the model's knowledge of possible protein architectures.

The Gap Between Sequence and Fold Diversity

There is a profound disconnect between how similar proteins appear in sequence space versus how similar they are in fold space. Proteins can exhibit very low sequence identity while maintaining nearly identical 3D structures.

For example, in structural cluster A0A242HMU2_f1, three proteins share the same fold (TM-score > 0.75) despite having pairwise global identities as low as 23.9% to 28.3%. This demonstrates that evolution frequently arrives at the same structural solution using widely divergent amino acid sequences.

Engineering a Scalable Clustering Pipeline

To quantify structural redundancy, Ligo developed a pipeline to convert raw predicted structures into clean, clusterable fragments. This process addresses the "predicted-structure problem," where AlphaFold predictions often include disordered tails, floppy linkers, and multi-domain proteins that obscure the core structural signal.

Noise Removal and Fragmentation

The pipeline begins by removing residues with a pLDDT below 65 to eliminate obvious disorder. Remaining residues are split into contiguous fragments, and spatially touching fragments are rejoined using a union-find approach. However, because high-confidence linkers can falsely merge independent domains, a more sophisticated graph-theoretic approach is required.

Graph-Theoretic Spectral Bisection

Ligo treats the protein as a graph where residues are nodes and edges connect spatial nearest neighbors. To identify the true boundaries between domains, the pipeline uses spectral bisection based on the Fiedler vector (the eigenvector corresponding to the second smallest eigenvalue of the normalized graph Laplacian).

This method identifies the "bottleneck" in the spatial graph—the narrow bridge connecting dense regions. By recursively applying this bisection, the system can surgically split multi-domain proteins into independent interacting units, preventing the model from wasting capacity on disordered regions or bizarre, elongated chains.

Quantifying the Redundancy of MGnify

After fragmenting and filtering approximately 2 million MGnify fragments, Ligo performed structural clustering using Foldseek and a subsequent TM-align audit to ensure accuracy.

The Distribution of Folds

The results indicate that the structural universe is dominated by a small "head" of common clusters:

Total Multi-member Clusters: Approximately 25,300 for MGnify.
Concentration of Mass: 71.5% of all MGnify multi-member fragments reside in the top 1,000 clusters.

This finding suggests that the "true" number of reusable structural neighborhoods is closer to tens of thousands than the millions of clusters often reported by faster, less rigorous clustering passes.

Implications for Data Sampling

Because the data is so skewed, uniform sampling from the dataset would over-represent common folds, while uniform sampling from clusters would over-represent rare folds. Ligo employs a balancing exponent $\gamma$ (cluster size exponent) to weight sampling probability as $N^\gamma$ (where $N$ is cluster size). A value of $\gamma \approx 0.5$ is used to preserve the dominance of common folds while flattening the long tail of rare ones.

Impact on Enzyme Design

The discovery of extreme structural redundancy suggests two distinct paths for the future of enzyme design:

The Nature-Like Route: Focus on refining the active-site neighborhood of familiar scaffolds. In this regime, the precise geometry of first-shell residues and ligand poses is more critical than global backbone novelty.
The Speculative Route: Explore the "unnatural" backbone space that evolution never visited. This assumes that natural fold space is constrained by historical accidents (duplication and divergence) rather than absolute physical limitations.

Community Perspectives

Technical discussion surrounding these findings highlights a tension between the novelty of the data and the established knowledge of biochemistry:

"Thirty years ago the professor in my Proteins class made a few relevant important points... Across species the 'same' protein can diverge in sequence by up to 40%, while keeping the same structure. Sometimes this goes as far as 80%."

Other contributors noted that evolution's tendency toward "liberal protein plagiarism" is a known phenomenon, but the scale of the redundancy revealed by these automated pipelines provides a concrete metric for how much "new" information is actually contained in metagenomic sequence databases.