Protein structure search model fusing a graph attention encoder with ESM-2 sequence features into one embedding per protein for fast homolog lookup.
No providers recorded yet. Browse all providers
Searching a query structure against the 214 million entries of the AlphaFold Protein Structure Database is a problem of scale before accuracy. Pairwise aligners such as TM-align and Dali compare two structures precisely, but a single query against a million structures costs days of CPU time, and even Foldseek, which turns structures into a discrete 3Di alphabet so they can be aligned like sequences, still performs an alignment for every candidate that survives its prefilter. FoldExplorer removes the alignment step from search: each protein is encoded once into a fixed 512-dimensional descriptor, a database is embedded in advance, and a query becomes a cosine-similarity lookup.
FoldExplorer was developed by Yuan Liu and Hong-Bin Shen at the Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University, first as an arXiv preprint in November 2023 and then in the Journal of Molecular Biology in 2025, with Ying Zhang and Zhen Zhou joining as authors. It extends the lab's earlier GraSR, a contrastively trained graph neural network over protein structures, by adding a sequence channel: protein-level embeddings from the 650-million-parameter ESM-2 language model are added to the graph embedding. The sequence signal helps the descriptor rank proteins sharing function and overall shape above ones that only overlap locally.
The trained encoder is applied unchanged to new structures, experimental or predicted, and its descriptors are reused elsewhere: RNA&Protein SubCell Embedding, from the same lab, uses FoldExplorer vectors as frozen structure-aware protein tokens.
Each residue is a node whose initial features are GraSR's rotation- and translation-invariant distance and angle descriptors computed from Cα coordinates, and an edge joins any two residues closer than 10 Å. Stacked multi-head GAT layers are interleaved with self-attention graph pooling (SAGPooling); the output of every layer is also globally average-pooled and passed through an MLP, and these per-layer readouts are summed into the graph embedding. The encoder is trained with the MoCo momentum-contrast framework and an InfoNCE loss, using the DeepFold and GraSR definition of positives: protein B is a structural neighbour of A when their TM-score is at least 90% of A's best TM-score in the dataset. Training data are 13,265 SCOPe 2.07 domains filtered to below 40% sequence identity, split for 5-fold cross-validation.
On SCOPe cross-validation the preprint reports an AUPRC of 0.7101 against 0.6595 for GraSR and a top-1 hit ratio of 0.7830 against 0.7282. On indDomain, an independent set of 1,076 PDB domains released between October 2021 and May 2023 and filtered below 40% identity to SCOPe, the top-1 hit ratio is 0.7928 against 0.7537 for Foldseek. Logistic regression on the descriptors reaches 0.9434 accuracy for SCOPe class, and descriptor similarity correlates with TM-score at Spearman 0.79. An all-versus-all search of 542,378 AlphaFold SwissProt structures took 10.8 hours on one RTX 3090 GPU with 16 CPU threads, against 21 hours for Foldseek with its prefilter and 400 hours without it. The journal version adds that the descriptors remain effective on low-confidence predicted structures.
Structural biologists can retrieve candidate homologs for a newly solved or predicted structure from the full AlphaFold database in one query, without a local cluster. Groups annotating proteomes can pre-embed their structures with the released checkpoints and run their own similarity searches, and the fixed descriptors serve as input features for fold classification or as structure-aware protein representations in downstream models. A retrained residue-level variant, released later in a separate repository, fuses ESM-IF1 structure and ESM-2 sequence features into per-residue embeddings for tasks that need site resolution.
FoldExplorer belongs to the embedding-based structural search family that also includes TM-Vec 2, and its sequence-plus-structure fusion contrasts with sequence-only retrieval in PLMSearch. It shows that a small contrastive encoder trained on about thirteen thousand domains, enriched with language-model features, can match alignment-based ranking on out-of-distribution domains at database sizes where alignment becomes the bottleneck. Its limits are equally concrete: training covers only SCOPe domains, the descriptor summarizes a whole chain and so gives no residue alignment, and code and weights are released under CC BY-NC 4.0, with the web server free to academic users only.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.