FoldMatch

The explosion of predicted protein structures from tools like AlphaFold has made structure-based search a routine need, but classical pairwise alignment methods such as DALI and TM-align do not scale to hundreds of millions of models. Embedding-based approaches instead compress each structure into a vector that can be compared with fast nearest-neighbor search, trading exact alignment for tractable, proteome-wide retrieval.

FoldMatch, developed by the RCSB Protein Data Bank team at UC San Diego and released as a preprint in early 2025, is one such embedding model. It encodes each 3D protein structure into a single fixed-length vector, enabling similarity search and clustering across entire proteomes. Rather than learning a structural encoder from scratch, it repurposes residue-level representations from the ESM3 generative protein language model and aggregates them into a compact whole-structure embedding.

#Key Features

• Structure-derived residue embeddings: FoldMatch uses the ESM3 protein language model to compute per-residue embeddings directly from 3D structural input.
• Transformer aggregator: A dedicated aggregator network pools residue embeddings into a single 1,536-dimensional vector representing the whole structure.
• Scalable similarity search: FAISS-based nearest-neighbor search with multi-GPU support enables fast retrieval and clustering across proteome-scale collections.
• Two-stage sequence search: A complementary mode combines embedding similarity with exact Smith-Waterman alignment for higher-precision matching.
• Packaged and installable: Distributed as a pip-installable Python package with both command-line tools and a programmatic API.

#Technical Details

FoldMatch's aggregator is a transformer-based network with six encoder layers, each using 3,072-unit feedforward blocks with ReLU activations. It applies summation pooling over residue embeddings followed by twelve fully connected residual layers, producing a single 1,536-dimensional structure embedding. Residue-level features are provided by ESM3 computed from 3D coordinates. Retrieval is implemented with FAISS and supports multiple GPUs, and the toolkit can emit per-residue, sequence-based, and aggregated structure-level embeddings. The package requires Python 3.12 or later with PyTorch and Lightning. Because it incorporates ESM3, it is distributed under the EvolutionaryScale Cambrian Non-Commercial License, which restricts commercial use.

#Applications

FoldMatch is aimed at structural bioinformatics workflows that need to search or organize very large structure collections: finding structural homologs of a query protein across an entire proteome, detecting remote homology beyond sequence similarity, transferring functional annotations, and clustering structure space to reveal fold families. Its origin at the RCSB PDB positions it for integration with large public structure archives.

#Impact

FoldMatch reflects the move toward embedding-based structure search as predicted structures outpace the capacity of pairwise alignment, and its development within the RCSB Protein Data Bank connects it to a primary structural data resource. Its main practical constraints are the Cambrian Non-Commercial License inherited from ESM3, which precludes commercial deployment, and its status as a preprint.