HypSeek: Hyperbolic Protein–Ligand Framework

• HypSeek is a hyperbolic embedding framework that models protein–ligand interactions using Lorentz-model representations for improved virtual screening and affinity ranking.
• It employs a three-tower architecture combining SE(3)-equivariant graph transformers and sequence encoders to map diverse modalities into hyperbolic space.
• Empirical results show notable gains over Euclidean models, with tiered cone supervision effectively separating activity cliffs and enhancing ranking performance.

Searching arXiv for the cited HypSeek-related papers and adjacent work to ground the article. HypSeek is, in its most specific usage, a retrieval-based protein–ligand binding framework that learns ligands, protein pockets, and protein sequences in Lorentz-model hyperbolic space and uses those embeddings for both virtual screening and affinity ranking (Wang et al., 21 Aug 2025). The available literature also uses the name as a systems label in several technical syntheses: a HypSeek-style hybrid search engine, a hypothesis-seeking interactive search companion, a Hypencoder-based retrieval system, and a HyperSearch-like hyperedge prediction engine (Wang et al., 2 Aug 2025, Bink et al., 16 Jan 2026, Killingback et al., 7 Feb 2025, Choo et al., 20 Oct 2025). This suggests a dual status for the term: a concrete model name in computational chemistry, and a reusable shorthand for search-oriented architectures in information retrieval and higher-order network analysis.

1. Nomenclature and scope

The sources support multiple, non-identical uses of the name. In one case, HypSeek is the formal name of a model. In several others, it is a hypothetical or synthesized system name used to explain how a paper’s methods could be assembled into a broader architecture.

Usage	Domain	Defining description
HypSeek	Protein–ligand modeling	Hyperbolic representation learning framework for virtual screening and affinity ranking (Wang et al., 21 Aug 2025)
HypSeek-style hybrid search system	Information retrieval	Decoupled FTS, SVS, DVS, and TenS paths with fusion and re-ranking (Wang et al., 2 Aug 2025)
HypSeek as hypothesis-seeking companion	Search interaction	Context-aware interactive companion integrated into SERPs (Bink et al., 16 Jan 2026)
HypSeek based on Hypencoders	Neural IR	Query-specific relevance function produced by a hypernetwork (Killingback et al., 7 Feb 2025)
HypSeek as HyperSearch-like engine	Hypergraph learning	Unconstrained hyperedge search with anti-monotonic pruning (Choo et al., 20 Oct 2025)

A further source uses the phrase “the ‘HypSeek’ profile of DeepSeek” only as a synthesized characterization of DeepSeek’s behavior on HPC code-generation tasks, rather than as the name of a model or deployed system (Nader et al., 15 Mar 2025). A common misconception is therefore to treat HypSeek as a single established architecture across all of these areas. The sources do not support that interpretation. They instead show a concrete chemistry model and several design extrapolations that reuse the label for systems concerned with seeking, ranking, or pruning hypotheses.

2. Hyperbolic protein–ligand HypSeek

In drug discovery, HypSeek targets two tasks: virtual screening, where a protein target is used to rank very large ligand libraries, and affinity ranking, where ligands within a single assay are ordered by relative binding strength (Wang et al., 21 Aug 2025). The framework is explicitly retrieval-based. It does not attempt global absolute affinity calibration across assays; the source states that affinity values are only comparable within an assay, so the model learns within-assay rankings.

The central claim is geometric. Euclidean retrieval models such as DrugCLIP and LigUnity embed ligands and pockets into a shared Euclidean space, but the source argues that Euclidean geometry compresses subtle functional differences and handles activity cliffs poorly. HypSeek instead uses an $n$-dimensional hyperbolic space of constant negative curvature $-\kappa$ in the Lorentz model, with Lorentzian inner product

$$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$

hyperboloid

$$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$

and geodesic distance

$$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$

The paper’s interpretation of this geometry is hierarchical and affinity-sensitive. Radial position can encode graded properties such as global activity or binding strength, while angular position can encode fine functional differences. Because hyperbolic volume grows exponentially with radius, small angular deviations at larger radial depth can produce large geodesic separations. The source uses this to motivate improved handling of activity cliffs: structurally similar ligands with large affinity gaps can remain locally similar while becoming more separable geodesically than in Euclidean space. The paper formalizes this with the proposition that, under constant radial norm and small angular deviation, hyperbolic embeddings can yield significantly larger distances between structurally similar but functionally divergent ligands than Euclidean embeddings (Wang et al., 21 Aug 2025).

3. Architecture, objectives, and inference

HypSeek adopts a protein-guided three-tower architecture consisting of a ligand tower, a pocket tower, and a sequence tower (Wang et al., 21 Aug 2025). The encoders are specified as follows: $f_\theta$, an SE(3)-equivariant 3D graph transformer for ligands; $g_\phi$, an SE(3)-equivariant 3D graph transformer for protein pockets; and $h_\psi$, an ESM-2-based transformer sequence encoder for proteins. Their Euclidean outputs are lifted into hyperbolic space by the exponential map at the origin

$$\mathbf 0 = \Bigl(\tfrac1{\sqrt\kappa},0,\dots,0\Bigr).$$

The resulting embeddings are

$$\mathbf{h}_{\mathrm{poc}} = \exp_{\mathbf 0}^{\kappa}\bigl(E_{\mathrm{poc}}\bigr),\quad \mathbf{h}_{\mathrm{mol}} = \exp_{\mathbf 0}^{\kappa}\bigl(E_{\mathrm{mol}}\bigr),\quad \mathbf{h}_{\mathrm{seq}} = \exp_{\mathbf 0}^{\kappa}\bigl(E_{\mathrm{seq}}\bigr).$$

Training combines three components. First, a contrastive InfoNCE-style retrieval loss aligns pocket–ligand and sequence–ligand pairs for virtual screening. Second, a listwise Plackett–Luce ranking loss models within-assay ligand ordering for affinity ranking. Third, a hyperbolic cone-hierarchy objective imposes pocket-centered radial and angular constraints on ligands by discretizing affinities into tiers and penalizing radial and angular violations. The source adds two regularizers: an angular margin regularizer to avoid angular collapse, and a heterogeneity regularizer to better model within-assay variation.

The cone formulation is the distinctive structural prior. Each pocket defines a cone with half-aperture

$-\kappa$0

and each ligand receives a tier-specific radial limit $-\kappa$1 and angular scaling $-\kappa$2. Stronger ligands are required to lie closer to the pocket and within narrower cones; weaker ligands may lie farther away and be more diffuse. This is an explicit affinity-aware neighborhood model rather than an implicit similarity prior.

Inference is deliberately simple. After encoding pocket and ligands, HypSeek extracts the spatial components $-\kappa$3 and $-\kappa$4, computes

$-\kappa$5

and ranks ligands by descending $-\kappa$6. This design preserves the efficiency properties of retrieval models while using hyperbolic training and cone supervision to structure the embedding space. The reported training setup uses 4×NVIDIA A100 GPUs for 50 epochs with Adam, learning rate $-\kappa$7, and curvature parameter $-\kappa$8 (Wang et al., 21 Aug 2025).

4. Empirical performance and ablation profile

The most detailed quantitative evaluation of HypSeek appears in protein–ligand prediction benchmarks (Wang et al., 21 Aug 2025). On DUD-E, the paper reports EF$-\kappa$9 = 51.44 for HypSeek versus 42.63 for LigUnity$$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$0, characterized as a $$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$1 relative improvement. It also reports AUROC = 0.9435 and BEDROC$$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$2 = 0.7892. On LIT-PCBA, HypSeek reports AUROC = 0.6210, BEDROC$$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$3 = 0.1196, and EF$$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$4 = 6.81.

For affinity ranking, the paper gives both ensemble and seed-averaged results. On JACS, the five-seed ensemble reports Pearson $$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$5 and Spearman $$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$6, while the mean $$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$7 std over seeds is $$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$8 and $$\langle \mathbf p, \mathbf q\rangle_{\mathbb L} = -p_0 q_0 + \langle \tilde{\mathbf p}, \tilde{\mathbf q} \rangle_{\mathbb E},$$9. On Merck, the ensemble reports $$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$0 and $$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$1.

Setting	Benchmark	Reported result
Virtual screening	DUD-E	EF$$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$2 = 51.44; AUROC = 0.9435; BEDROC$$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$3 = 0.7892
Virtual screening	LIT-PCBA	AUROC = 0.6210; BEDROC$$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$4 = 0.1196; EF$$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$5 = 6.81
Affinity ranking	JACS	Ensemble $$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$6, $$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$7
Affinity ranking	Merck	Ensemble $$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$8, $$\mathbb{L}^n = \Big\{ \mathbf{p} \in \mathbb{R}^{n+1} : \langle \mathbf{p}, \mathbf{p} \rangle_{\mathbb{L}} = -\tfrac{1}{\kappa},\; p_0 > 0 \Big\},$$9

The ablation studies attribute the gains to both geometry and supervision. Removing hyperbolic supervision lowers DUD-E BEDROC from 0.7892 to 0.7671 and EF$$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$0 from 51.44 to 49.14, while the purely Euclidean version drops to BEDROC = 0.6565 and EF$$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$1 = 42.87. Removing the sequence tower lowers DUD-E BEDROC to 0.7351 and EF$$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$2 to 47.70. The paper also reports pairwise analyses on JACS showing that the Euclidean model’s accuracy and correlation drop sharply as ECFP4 similarity increases, whereas the hyperbolic model maintains high accuracy and correlation. This is presented as empirical support for the claim that hyperbolic geometry helps separate activity-cliff pairs (Wang et al., 21 Aug 2025).

A plausible implication is that HypSeek’s main contribution is not merely replacing Euclidean distance with hyperbolic distance, but coupling that geometry to a tiered pocket-centered supervision scheme and a multimodal three-tower encoder. The paper’s own ablations support that reading.

5. HypSeek as a search-systems shorthand

Outside protein–ligand modeling, the name is used to describe several search-oriented systems rather than a single standardized implementation.

In hybrid retrieval, a “HypSeek-style hybrid search system” is defined by deliberate choices over retrieval paradigms, fusion, and re-ranking under accuracy, latency, and cost constraints (Wang et al., 2 Aug 2025). The source evaluates four paradigms—Full-Text Search (FTS), Sparse Vector Search (SVS), Dense Vector Search (DVS), and Tensor Search (TenS)—and three late-fusion strategies: Reciprocal Rank Fusion, Weighted Sum, and Tensor-based Re-ranking Fusion (TRF). It emphasizes a “weakest link” phenomenon: on TOUC(en), single-path FTS achieves nDCG@10 = 0.650, single-path DVS 0.390, but FTS+DVS with RRF drops to 0.604. The same source identifies TRF as a high-efficacy alternative to mainstream fusion; for DBPE(en), FTS+DVS with RRF yields 0.668, while TRF yields 0.722, reported as $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$3. The architectural recommendation is decoupled storage, parallel retrieval, and configurable fusion, optionally using TenS-like MaxSim only for re-ranking rather than full retrieval.

In interactive search support, the paper on a context-aware search companion does not name its system HypSeek, but one source explicitly maps it onto a hypothesis-seeking companion architecture (Bink et al., 16 Jan 2026). The described design is a right-hand sidebar integrated into a standard SERP, with four tips triggered by search context: clarifying information needs, improving query reformulation, encouraging result exploration, and mitigating bias through source comparison. In a user study with $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$4, the companion led users to issue about $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$5 more queries and view roughly twice as many results: baseline query count $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$6 versus companion $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$7, and baseline result views $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$8 versus companion $$d_{\mathbb{L}}(\mathbf p,\mathbf q) = \frac{1}{\sqrt{\kappa}}\,\cosh^{-1}\bigl(-\,\kappa\,\langle \mathbf p,\mathbf q\rangle_{\mathbb L}\bigr).$$9. Overall accuracy did not change globally, but difficult tasks showed slight gains. The intervention is framed as boost-style micro-learning rather than answer substitution.

In neural information retrieval, another source uses “HypSeek” to denote a production system built on Hypencoders rather than a model named HypSeek (Killingback et al., 7 Feb 2025). The underlying method replaces vector similarity with a query-specific neural relevance function:

$f_\theta$0

where a hypernetwork encodes the query into the weights of a small q-net applied to document vectors. The reported model reaches nDCG@10 = 0.736, RR = 0.885, and R@1000 = 0.871 on TREC DL ’19/’20, and the paper’s approximate graph search retrieves from a corpus of 8.8M documents in under 60 milliseconds. In this usage, HypSeek denotes an IR system whose core operation is $f_\theta$1, then $f_\theta$2, rather than a fixed query vector.

These uses share a family resemblance: modular retrieval paths, query- or context-conditioned scoring, and explicit concern with how hypotheses are generated, fused, or re-ranked. The sources do not establish a common codebase or benchmark under the name, but they do establish a recurring design vocabulary.

6. Hypergraph prediction, HPC synthesis, and broader significance

A HyperSearch-based system-level explanation also names a hypothetical hyperedge predictor “HypSeek” (Choo et al., 20 Oct 2025). Here the object is not document retrieval but higher-order interaction prediction in hypergraphs. The task is to search

$f_\theta$3

for new hyperedges not already observed. The defining contributions are an empirically grounded scoring function based on relaxed structural overlap, optional temporal weighting and node-feature similarity, and an efficient search mechanism that uses anti-monotonic upper bounds to prune the combinatorial candidate space. The paper proves that its upper bound $f_\theta$4 is anti-monotonic and introduces a practical upper bound $f_\theta$5 for pruning. On Enron, for example, Recall@1× is reported as 16.1 for HyperSearch versus 13.3 for NHP-C, 10.3 for CNS, and 1.7 for HPRA. The runtime scaling with $f_\theta$6 is described as almost linear, with regression slope approximately 1.05. In this usage, HypSeek denotes a branch-and-bound search engine over subsets rather than a neural encoder.

A final source applies the label only indirectly, by describing a “HypSeek” synthesis of DeepSeek’s behavior on HPC programming tasks (Nader et al., 15 Mar 2025). The underlying paper benchmarks DeepSeek-R1 on five canonical workloads—conjugate gradient, parallel heat equation, parallel matrix multiplication, DGEMM, and STREAM triad—across C++, Fortran, Julia, and Python. Its conclusion is restrained: DeepSeek generates functional code for HPC tasks but lags behind GPT-4 in scalability and execution efficiency. The failure modes are concrete and language-specific, including invalid OpenMP SIMD clauses, Fortran type mismatches, incorrect numba.prange usage, and undefined Julia symbols. This is not a HypSeek model in the same sense as the protein–ligand system, but it shows how the label can be extended to a “seeking” profile centered on search, generation, and evaluation under computational constraints.

Taken together, these sources position HypSeek as an unusually polysemous research term. Its most precise referent is the hyperbolic protein–ligand framework (Wang et al., 21 Aug 2025). Its broader usage marks a style of system design: modular, search-centric, and explicitly structured around the generation, ranking, pruning, or critical evaluation of candidate hypotheses. The literature therefore supports both a narrow definition and a wider conceptual lineage, but not the existence of a single unified architecture spanning chemistry, IR, hypergraphs, and HPC.