LEN-Seek

ProBiS, the maximum-clique surface-graph aligner behind template-based binding-site search, needs about 0.44 seconds per pairwise comparison — roughly 20 CPU-hours to screen one query against a reference set of 163,252 known ligand binding sites. The databases keep growing, the PDB past 240,000 structures and the AlphaFold Database past 200 million predicted ones, while the fast alternatives (geometric hashing, 1D distance distributions) discard the spatial detail that makes a pocket a pocket.

LEN-Seek (Latent Encoding Network for template Seeking) removes the alignment step entirely. A binding site is treated as a graph whose nodes are amino acid residues and whose edges carry roto-translationally invariant geometry, and a graph transformer variational autoencoder compresses that graph into a set of per-residue latent vectors — a point cloud on a continuous manifold. Two sites are then compared by Chamfer distance between their point clouds, an arithmetic operation on 256-dimensional vectors rather than a combinatorial search over surface graphs.

Kyunghwan Yeo, Dongwoo Kim, Jaemin Sim and Juyong Lee released it from Seoul National University and Arontier in August 2026. It sits upstream of the docking and co-folding tools that consume binding-site information, including AlphaFold 3 and Boltz-1, whose accuracy improves when the search space is narrowed to a known pocket.

#Key Features

• Latent point-cloud representation: Each residue becomes its own probabilistic latent vector, so a variable-size site maps to a variable-size set rather than one pooled embedding, and similarity is a set-to-set Chamfer distance.
• SE(3)-invariant edge geometry: Every edge carries a 39-dimensional vector built from a 32-dimensional RBF distance kernel, a 3-dimensional local-frame direction and a 4-dimensional quaternion, removing the need for equivariant layers and rotation augmentation alike.
• Frozen protein language model features: Node features are per-residue embeddings from Ankh, projected from 768 to 512 dimensions by a learned adaptor block that reweights the language model signal toward binding-site structure.
• Encode once, query cheaply: Building the latent database is a frozen forward pass, after which each comparison costs 0.00013 seconds on one GPU against 0.443 seconds for ProBiS on 16 CPU cores — a roughly 3,400-fold gap the authors are explicit is not hardware-normalized.
• Alignment-free validation: Among retrieved pairs called similar, 88.0% bind ligands with a Morgan-fingerprint Dice similarity above 0.3, against 19.09% of pairs called dissimilar.

#Technical Details

The 18.5M-parameter network stacks four graph transformer layers in the encoder and four in the decoder, eight attention heads each, with edge features injected as a learned bias on the attention logits in the style of the AlphaFold Evoformer. Sites are built over the 15 nearest neighbors and padded or truncated to 65 residues; sites with fewer than 10 residues are dropped. The encoder emits a 256-dimensional mean and log-variance per residue, and the decoder reconstructs the adapted Ankh embedding, the alpha-carbon coordinate and the local frame, trained against a weighted sum of FAPE, RMSD and feature-recovery losses plus a Free Bits KL term. Training used BsitePDB, an in-house set of 163,252 sites from 66,912 proteins deposited up to 6 November 2023, of which 114,662 passed filtering; evaluation held out 482 sites from 161 proteins deposited after 21 July 2023, with MMseqs2 removing anything above 80% sequence identity to training data. Reconstruction reached 0.994 mean cosine similarity with 2.10 Å FAPE and 1.07 Å RMSD, and top-1 retrieval scored 0.9080 precision, 0.9245 recall and 0.7273 MCC against ProBiS, correlating at Pearson −0.853 with the G-LoSA GA-score.

#Applications

Binding-site retrieval turns a bare structure into a docking-ready target, so LEN-Seek fits drug discovery workflows that must localize a pocket before docking or co-folding. Its scaling profile suits cases alignment methods cannot reach: annotating pockets across predicted-structure collections, screening a target against the whole PDB for polypharmacology and off-target liabilities, or acting as a fast pre-filter that hands a short candidate list to ProBiS for exact verification — which is how the paper itself evaluates it.

#Impact

LEN-Seek shows that binding-site search can move into a learned latent space without the representational loss that made earlier fast methods unattractive, and that SE(3)-invariant edge features suffice — no equivariant module of the kind RoseTTAFold2 retains, no augmentation of the kind SimpleFold relies on. The evaluation is candid about its limits: true and false positives are defined relative to the ProBiS baseline rather than absolute ground truth, and the residue-level Chamfer distance overestimates dissimilarity when a shared pocket sits in different surrounding folds, which the authors flag as the main target for future work. The code and the pretrained checkpoint are public, but the repository carries no license file, the precomputed latent database is distributed only on request, and BsitePDB has no public release — so reproducing the search space means supplying your own structure database. The work is a preprint, not yet peer reviewed.