HypSeek

Virtual screening and binding-affinity ranking are the two computational workhorses of early drug discovery, and they are usually served by different machinery: a fast retrieval model sifts a library of millions of compounds, then a separate scoring function ranks the survivors by potency. Retrieval-based screening embeds ligands and protein pockets into a shared Euclidean space, which turns a screen into a nearest-neighbour lookup. The geometry is the weak point. Chemical series branch hierarchically, and a flat space has to compress an exponentially branching tree into fixed volume, leaving little room for the fine-grained potency differences that matter most — above all activity cliffs, where two nearly identical ligands differ by orders of magnitude in affinity.

HypSeek replaces that flat geometry with Lorentz-model hyperbolic space, where volume grows exponentially with distance from the origin. Hierarchies embed there with much less distortion, so the same embedding budget can carry both the coarse question of whether a compound is active at all and the subtle affinity gradients within a congeneric series. The model places ligands, protein pockets, and protein sequences on a single manifold through a protein-guided three-tower encoder, and covers virtual screening and affinity ranking within one framework instead of two unrelated specialist systems.

HypSeek comes from the Institute for AI Industry Research (AIR) at Tsinghua University, with collaborators at the University of Electronic Science and Technology of China and the Beijing Academy of Artificial Intelligence, and was accepted at AAAI 2026. It sits alongside other protein-ligand potency models that reach the same question by different routes, including AQAffinity, which scores affinity directly from a sequence and a SMILES string, and BioLM-Score, which conditions geometric scoring potentials on language-model priors.

#Key Features

• Hyperbolic embedding space: Ligands, pockets, and sequences are lifted onto a Lorentz-model manifold whose negative curvature accommodates hierarchical chemical relationships that Euclidean embeddings flatten out.
• Protein-guided three-tower encoder: Separate ligand, pocket, and sequence towers share one embedding geometry, so sequence-level protein context shapes the structure-based representation rather than being modelled in isolation.
• Screening and ranking in one framework: The same architecture and training objective serve both library-scale retrieval and potency ordering within a series, with a task-specific checkpoint for each.
• Activity-cliff sensitivity: Exponential geometry gives the embedding room to separate structurally near-identical ligands with large affinity gaps, the failure mode that most limits similarity-based screening.
• Evaluation on unseen targets: Proteins appearing in the screening benchmarks are excluded from training, so reported results reflect frozen-checkpoint inference on targets the model has never seen.

#Technical Details

The ligand and pocket towers are SE(3)-equivariant 3D graph transformers built on the Uni-Mol and Uni-Core stack; the sequence tower is an ESM-2 protein language model encoder, the same family documented in the catalog's ESM-2 and ESMFold entry. Each tower produces a Euclidean vector that is mapped onto the Lorentz manifold by an exponential map, and scoring is then computed as a hyperbolic distance rather than a dot product. Training uses an assay-level dataset curated from ChEMBL, BindingDB, and PDBbind, with every UniProt target present in the DUD-E and LIT-PCBA test sets removed beforehand.

On DUD-E's 102 targets, HypSeek raises early enrichment from 42.63 to 51.44, a 20.7% relative improvement, and on the JACS affinity set it lifts ranking correlation from 0.5774 to 0.7239, a 25.4% gain. The paper additionally reports results on LIT-PCBA and the Merck free-energy-perturbation sets, benchmarked against docking programs such as Glide and Gnina, machine-learned scoring functions, and recent structure-based models including Boltz-2.

#Applications

The screening checkpoint is aimed at library triage: encode a pocket once, encode a compound collection once, and rank candidates by hyperbolic distance, which keeps the cost of adding compounds linear rather than requiring a docking run per pair. The ranking checkpoint targets the later stage, ordering analogues within a congeneric series where medicinal chemists need relative potency rather than a binary active call. Because the towers accept standard inputs — a pocket structure, a ligand conformer, a protein sequence — the model fits existing structure-based drug design pipelines without bespoke featurisation.

#Impact

HypSeek's contribution is less a new encoder than a demonstration that the geometry of the embedding space is itself a usable inductive bias for molecular recognition, and that the gain shows up on both retrieval and regression tasks that are normally optimised separately. The code, including training and evaluation scripts, is released on GitHub, where the README carries an MIT badge but the repository contains no license file, leaving the terms of reuse undefined; the evaluation scripts take a checkpoint path as an argument, but the trained screening and ranking checkpoints are not distributed with the repository, so reproducing the reported numbers requires retraining from the described data pipeline. As with any retrieval-style screening model, performance is bounded by the assay data it was trained on: the LIT-PCBA results, drawn from harder and more realistic experimental screens, remain far below the DUD-E figures.