Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design

Daniel Kuhn; G\"unter Klambauer; Lisa Schneckenreiter; Lukas Friedrich; Sohvi Luukkonen

REVIEW 3 major objections 4 minor 2 references

Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read ConGLUDe unifies structure- and ligand-based drug design in one contrastive geometric model, removing the need for predefined binding pockets.

desk verdict The 'unified training' headline outruns the ablation evidence, but the architecture, the target-fishing result, and the new pocket-selection task are worth a serious look. read the letter →

arxiv 2601.09693 v3 pith:KGUOANBR submitted 2026-01-14 cs.LG stat.ML

Lisa Schneckenreiter , Sohvi Luukkonen , Lukas Friedrich , Daniel Kuhn , Günter Klambauer This is my paper · ORCID

classification cs.LGstat.ML

keywords contrastivelearningprotein–ligandinteractionbindingsitepredictionvirtualscreeningtargetfishingligand-conditionedpocketselectiondrugdesignequivariantgraphneuralnetwork

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that one contrastive geometric model, ConGLUDe, can be trained jointly on experimentally resolved protein–ligand complexes and on tens of millions of assay-based bioactivity measurements, and that this joint training is what lets a single pocket-agnostic encoder handle virtual screening, target fishing, binding-site prediction, and ligand-conditioned pocket selection. If true, drug discovery no longer has to choose between structure- and ligand-based paradigms, nor know a binding pocket in advance for large-scale screening: screening becomes embedding similarity, and pocket information is produced on the fly. The authors report competitive zero-shot virtual screening, substantially better target fishing than all compared methods including docking, and top ligand-conditioned pocket selection, at orders-of-magnitude lower inference cost. The paper itself flags that behavior on predicted structures is unverified, that ConGLUDe does not generate docked ligand conformations, and that it assumes bioassays with a uniquely associated protein target.

What carries the argument

The load-bearing object is the protein encoder: VN-EGNN, an equivariant graph neural network with virtual nodes, extended by a non-geometric global protein node. It predicts candidate binding-site centers, clusters them with DBSCAN, and emits pocket-level embeddings plus a whole-protein embedding, all linearly projected into a shared contrastive space. Three InfoNCE-style losses pull (protein + pocket) to the ligand, the ligand to the whole protein, and the ligand to the closest predicted pocket; a separate sigmoid contrastive loss on large-scale bioactivity data trains only the protein projection and the ligand MLP while the geometric encoder stays frozen. This lets a single architecture le

What would settle it

Train an ablation with a randomly initialized, frozen protein encoder plus only the ligand-based bioactivity data and compare target-fishing AUROC and LIT-PCBA enrichment with the jointly trained model; if the two are statistically indistinguishable, structure-based training contributes nothing to bioactivity semantics. Alternatively, unfreeze the geometric encoder during ligand-based training; if performance does not improve, the frozen-encoder assumption is the operative constraint.

Watch

Extended reading notes

Core claim

ConGLUDe is presented as the first end-to-end architecture integrating binding-site prediction with large-scale protein–ligand interaction modeling. A VN-EGNN protein encoder outputs a whole-protein embedding plus candidate binding-site embeddings—no predefined pockets needed—while a fast MLP ligand encoder yields split protein-/pocket-matching embeddings. Three contrastive axes align ligand, protein, and nearest predicted pocket; a sigmoid bioactivity loss tunes projections with the geometric encoder frozen. Reported results: competitive zero-shot virtual screening, markedly better target fishing than any compared baseline, and top ligand-conditioned pocket selection, at orders-of-magnitude

Load-bearing premise

The unified benefit rests on the assumption that a geometric protein encoder frozen on structural complexes already yields protein representations aligned with ligand-based bioactivity semantics, so that the tens of millions of activity labels only need to tune a linear layer and the ligand encoder.

Editorial extensions

If this is right

Virtual screening at genome-wide scale: with protein and pocket embeddings precomputed, millions of ligands are scored by dot-product similarity in seconds, and adding protein targets costs little because ligand embeddings are reused.
Pocket identification becomes a by-product: the same encoder that screens also outputs ranked candidate binding sites, either by confidence or conditioned on a query ligand.
Target fishing becomes a fast zero-shot task: on an out-of-domain chemoproteomics benchmark, ConGLUDe clearly outperforms docking while requiring orders of magnitude less compute.
Methods that demand a known pocket lose their edge when pockets are absent: on a realistic benchmark, pocket-agnostic training beats models that are given the correct pocket as input.
Ligand-conditioned pocket ranking matters most where orthosteric priors fail: on allosteric-site data, conditioning on the ligand clearly improves over the unconditioned geometric baseline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the largest untested leverage is ligand-side data scale — since the geometric encoder is frozen during bioactivity training, the 56M labels mostly reshape the ligand MLP and one linear layer, so scaling ligand capacity or data may yield most of the benefit at a fraction of the structure-based training cost.
Editorial inference: the paper's stated uncertainty about predicted structures sets up a concrete stress test — if target-fishing accuracy holds for proteins whose only structures come from prediction, the model's reach extends far beyond experimentally solved complexes.
Editorial inference: ConGLUDe could serve as a cheap pre-filter before pose generation — rank pockets by the ligand-conditioned dot product, then run docking or co-folding on the top candidate only, a pipeline the paper implies but does not benchmark.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this paper deserves a referee, but the headline claim is one notch above what the evidence supports.

ConGLUDe is a real step for contrastive protein–ligand models. It replaces predefined pockets with a VN-EGNN that outputs whole-protein and candidate-site embeddings, adds a third contrastive axis for molecule-to-pocket alignment, and defines a sensible ligand-conditioned pocket selection benchmark on ASD. The target-fishing result on Kinobeads is the strongest empirical contribution: 65.6 AUROC, with SPRINT and DrugCLIP near random and DiffDock at 58.9. The speed plot is also honest — with precomputed fingerprints, screening millions of ligands takes seconds, which is orders of magnitude faster than docking. Code is public. That is concrete, reproducible value. Self-citations to VN-EGNN are appropriate since it is the backbone and an external baseline.

Soft spots, in order. First, the "match SPRINT" line in Table 1 is wrong as written. On LIT-PCBA the AUROCs are 64.1 vs 73.4; the early-enrichment metrics (BEDROC, EF0.5%, EF1%) do match within noise. That is what the text should say. Second, the frozen VN-EGNN during ligand-based training is the load-bearing caveat. The paper says the module "cannot be meaningfully optimized," but gradients with respect to it are well-defined — SPRINT optimizes a whole-protein encoder on the same MERGED data. The real reason is presumably stability or forgetting, but as written it is under-justified. With the encoder frozen, the 56M bioactivity labels tune only a linear projection and the ligand MLP. The ablation in Table F1 is consistent: ConGLUDe gets 66.25 LIT AUROC vs 67.94 for ligand-only training, and 82.04 DUD-E vs 83.88 for structure-only training. No synergy from joint training is demonstrated. That weakens the "unified" framing, though not the practical usefulness of the model. Third, the pocket-selection "state of the art" is a statistical tie on PDBbind-time and PoseBusters, and on ASD ConGLUDe and DiffDock have the same point estimate (0.35). The clean win is ConGLUDe over the unconditional baselines on ASD. Say "competitive," not "SOTA." Fourth, using the PDBbind time split despite the acknowledged leakage is defensible for comparability, but it should not be sold as evidence; they do disclose it, which I credit.

Who this is for: computational drug design researchers who care about fast pocket-agnostic screening and site selection. The paper should go to peer review, not be desk-rejected, but the authors should be pushed to reword the overclaims and ideally run an unfreezing experiment or explicitly frame ligand-based training as head-tuning. I would cite the target-fishing and pocket-selection parts in my own work.

Referee Report

3 major / 4 minor

Summary. The paper introduces ConGLUDe, a contrastive learning architecture that couples a VN-EGNN geometric protein/pocket encoder with a lightweight MLP ligand encoder. The method is trained alternately on structure-based protein–ligand complexes (PDBbind subsets) and on large-scale ligand-based bioactivity data (MERGED from PubChem, BindingDB, ChEMBL). The authors claim that this unified training enables a single pocket-agnostic model to perform virtual screening, target fishing, binding site prediction, and a newly introduced ligand-conditioned pocket selection task. Experiments report competitive virtual screening on LIT-PCBA, strong target fishing on a Kinobeads proteomics dataset, retained binding site prediction performance, and strong ligand-conditioned pocket selection on allosteric sites. The paper also emphasizes orders-of-magnitude speed advantages over docking-based baselines.

Significance. If the unified-training claim were fully established, ConGLUDe would be a valuable step toward a foundation model that bridges structure-based and ligand-based drug design. The architecture is simple and computationally efficient, the code is publicly released, and the target-fishing results on an external chemical-proteomics dataset are genuinely promising. The ablation study and extended benchmark tables are useful. However, the current evidence does not demonstrate a synergistic benefit from joint training, and several headline performance claims are stronger than the reported numbers justify. The core idea is worth pursuing, but the paper needs substantial revision to align its claims with its results.

major comments (3)

[Section 3.2.2, Eq. (14), Table F1] The paper's central claim—that ConGLUDe is trained jointly on structure- and ligand-based data such that both modalities improve a single representation—is not supported by the current training scheme. In ligand-based batches the VN-EGNN protein encoder is frozen; the stated justification that it 'cannot be meaningfully optimized' is not self-evident, since Eq. (14) is differentiable with respect to p and therefore with respect to the encoder parameters. The consequence is that the 56M bioactivity labels only train the linear protein projection and the ligand MLP. More importantly, the ablation in Table F1 shows no synergistic benefit: on LIT-PCBA, ConGLUDe (66.25 AUROC) is below ligand-based-only training (67.94), and on DUD-E it is below structure-based-only training (82.04 vs 83.88). The text's claim that 'all components together contribute' is therefore not supported by the data. The
[Section 4.1, Table 1] The text states that ConGLUDe 'match[es] SPRINT' on LIT-PCBA. This is contradicted by the AUROC column: SPRINT reports 73.40, ConGLUDe 64.06 ± 3.25, a gap of about 9.3 points. While ConGLUDe's BEDROC and EF1% are close to SPRINT's, the AUROC discrepancy is large and should be acknowledged. The claim of 'substantially outperform[ing] methods that use binding site information as input' also needs quantification: for example, HypSeek has AUROC 62.10, which is within one MAD of ConGLUDe's 64.06. The authors should soften these statements or provide statistical significance tests.
[Section 4.4, Table 4] The conclusion claims 'state-of-the-art ligand-conditioned pocket selection,' but Table 4 does not support an unqualified SOTA statement. On ASD, ConGLUDe (0.35) is tied with DiffDock (0.35); on PoseBusters it is tied with VN-EGNN (0.27); on PDBbind Time the margin over VN-EGNN is 0.54 vs 0.51 with overlapping 95% confidence intervals. The text itself acknowledges 'limited improvement over the inherent VN-EGNN ranking' on PDBbind and PoseBusters. The SOTA claim should be restricted to the allosteric setting (where the gain is meaningful) or supported by a significance test across all datasets.

minor comments (4)

[Section 3.1.1, Eq. (5)] Equation (5) uses φ_h^{(BR)} in the P→R message-passing step; this appears to be a typo for φ_h^{(PR)}.
[Section 3.2.1] The phrase 'a ligand-protein pairs' should be 'ligand–protein pairs' or 'a pair of ligand and protein.'
[Section G.3] The opening sentence of G.3 begins with '4 shows that...' which should be 'Table 4 shows that...' Also, 'This is task can also be performed' should read 'This task can also be performed.'
[Table G2] The S2Drug row in Table G2 has numbers that run together (e.g., '92.4679.25'); the formatting should be fixed.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; central results rest on external benchmarks and standard supervised objectives, with only a minor self-citation to the VN-EGNN backbone.

full rationale

ConGLUDe is a contrastive-learning architecture; its training objectives (Eq. 8-14) are standard InfoNCE and sigmoid losses over protein, pocket, and ligand embeddings. There is no fitted constant that is later renamed a prediction: the virtual-screening score s(p, m_p) is literally the quantity optimized by Eq. 14 on activity labels, which is supervised learning rather than circularity. Pocket prediction inherits the VN-EGNN geometric objective (Eq. B.17); the Table 3 result is a sanity check that the modifications do not degrade the backbone, not a derivation of that backbone from its own outputs. The main self-citation, Sestak et al. (2025), is used as an architectural source and as a published baseline; it is not invoked as an external 'uniqueness theorem' and is independently benchmarked. The frozen VN-EGNN during ligand-based batches (Section 3.2.2) weakens the 'unified training' claim, and the paper's own ablation (Table F1) shows no synergistic virtual-screening gain, but that is an evidence/claim-strength problem, not an equation-level reduction. The Limitations section discloses scope limits (e.g., no docked conformations, target-agnostic assays unsupported). The admitted leakage in PDBbind Time (Section G.3) is a benchmark-validity concern, not circularity. No circular step can be exhibited; score 2 reflects only the minor, non-load-bearing self-citation of the VN-EGNN backbone.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The method relies on standard contrastive learning assumptions and a set of structural/bioactivity representations borrowed from prior work. No new physical entities are postulated. Most of the modeling choices (4 Å binding-site definition, graph construction, fingerprint representation) are explicit and traceable. The main unexamined assumptions are the adequacy of AlphaFold structures and the transferability of a frozen geometric encoder to ligand-based alignment.

free parameters (4)

Ligand-based loss scaling factor = 6
Hand-picked to match the magnitude of the structure-based loss; affects the balance between the two data sources (Appendix D).
DBSCAN clustering radius and min_samples = radius=4, min_samples=1
Used to cluster candidate pocket nodes; determines the number of final pockets K (Section 3.1.1).
Active:inactive sampling ratio and cap = 1:3, cap 10,000 actives per protein
Sampling strategy for ligand-based batches; chosen by the authors and affects the training distribution (Appendix D).
Batch sizes = 64 (structure-based), 16 proteins (ligand-based)
Affects the number of in-batch negatives used in the contrastive losses (Appendix D).

assumptions (7)

domain assumption Binding sites are defined as the geometric center of residues within 4 Å of any ligand atom.
Used to define ground-truth binding sites for structure-based training and evaluation (Section 2.1).
domain assumption A protein geometric graph with Cα nodes, ESM-2 features, and 10 Å/10-neighbor edges captures the structural information needed for binding.
Adopted from VN-EGNN; this representation is the sole structural input to the protein encoder (Section 2.1).
domain assumption Morgan fingerprints plus RDKit descriptors are sufficient ligand representations for contrastive binding prediction.
The ligand encoder architecture relies on this fixed representation, which has been shown effective in prior work but is still a modeling choice (Section 3.1.2).
domain assumption Large-scale bioactivity data (MERGED: PubChem, BindingDB, ChEMBL) provide reliable binary active/inactive labels that are comparable to structural binding data.
Used to train on ligand-based batches; label noise and assay heterogeneity are not modeled (Section 3.2.1).
domain assumption AlphaFold2-predicted structures are accurate enough for proteins without experimental PDB structures.
Used to create 3D structures for all MERGED proteins without PDB entries (Appendix C.2).
domain assumption In-batch negatives in the contrastive loss are a good approximation of the non-binding distribution.
Standard contrastive learning assumption; the quality of learned embeddings depends on it (Section 3.2.2).
domain assumption The pocket representation selected during training (closest predicted pocket to the ground-truth center) is representative of test-time predicted pockets.
Creates a train/test distribution gap: at test time, no ground-truth centers are available and the model must rank its own predicted pockets (Section 3.2.2, Eq. 12).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design." pith.science (2026). https://pith.science/paper/KGUOANBR

@misc{pith2026260109693,
  author       = {Pith},
  title        = {Pith review of: Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGUOANBR}},
  note         = {Machine review of arXiv:2601.09693}
}

read the original abstract

Structure-based and ligand-based computational drug design have traditionally relied on disjoint data sources and modeling assumptions, limiting their joint use at scale. In this work, we introduce Contrastive Geometric Learning for Unified Computational Drug Design (ConGLUDe), a single contrastive geometric model that unifies structure- and ligand-based training. ConGLUDe couples a geometric protein encoder that produces whole-protein representations and implicit embeddings of predicted binding sites with a fast ligand encoder, removing the need for predefined pockets. By aligning ligands with both global protein representations and multiple candidate binding sites through contrastive learning, ConGLUDe supports ligand-conditioned pocket prediction in addition to virtual screening and target fishing, while being trained jointly on protein-ligand complexes and large-scale bioactivity data. Across diverse benchmarks, ConGLUDe achieves competitive zero-shot virtual screening performance, substantially outperforms existing methods on a challenging target fishing task, and demonstrates state-of-the-art ligand-conditioned pocket selection. These results highlight the advantages of unified structure-ligand training and position ConGLUDe as a step toward general-purpose foundation models for drug discovery.

Figures

Figures reproduced from arXiv: 2601.09693 by the authors.

**Figure 1.** Joint training of ConGLUDe on structure- and ligandbased data enables (ligand-conditioned) pocket prediction, virtual screening, and target fishing. to modulate their function (Kinch et al., 2024). Understanding protein-ligand interactions (PLIs) through modeling atomic interactions at binding sites or more general ligand bioactivity measures enables rational drug design (Gohlke et al., 2000; Du et al., 2016). Co… view at source ↗

**Figure 2.** ConGLUDe architecture and training procedure. A: Message-passing scheme of ConGLUDe’s protein encoder based on VN-EGNN: 1. message exchange between residue nodes, 2. residue nodes to virtual pocket nodes, 3. pocket nodes to residue nodes, 4. residue nodes to virtual protein node, 5. virtual protein node to residue nodes. B: The protein encoder supplies a representation of the whole protein p, and of each detected po… view at source ↗

**Figure 3.** Average virtual screening time per protein target versus the number of screened molecules, averaged over 100 targets (error bars: standard deviation). For ConGLUDe, the thin line uses precomputed ligand fingerprints and descriptors, while the thick line includes pre-processing. SPRINT and DiffDock report end-toend time including pre-processing; DrugCLIP uses pre-computed pocket and ligand coordinates. For ConGLUDe … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

[2]

early recognition

Enrichment factors in database screening.Journal of Medicinal Chemistry, 47(7):1750–1759, 2004. Han, J., Hong, Y ., and Li, W.-J. DrugHash: Hashing Based Contrastive Learning for Virtual Screening.Proceedings of the AAAI Conference on Artificial Intelligence, 39(16): 17041–17049, 2025. Hansch, C., Maloney, P. P., Fujita, T., and Muir, R. M. Cor- relation ...

arXiv 2004
[2011]

Ester, M., Kriegel, H.-P., Sander, J., and Xu, X

doi: 10.1021/ci2003889. Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. A density- based algorithm for discovering clusters in large spatial databases with noise. InInternational Conference on Knowledge Discovery and Data Mining, pp. 226–231, 1996. Fan, J., Fu, A., and Zhang, L. Progress in molecular docking. Quantitative Biology, 7:83–89, 2019. Feng, B...

work page doi:10.1021/ci2003889 1996

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [2]

early recognition

Enrichment factors in database screening.Journal of Medicinal Chemistry, 47(7):1750–1759, 2004. Han, J., Hong, Y ., and Li, W.-J. DrugHash: Hashing Based Contrastive Learning for Virtual Screening.Proceedings of the AAAI Conference on Artificial Intelligence, 39(16): 17041–17049, 2025. Hansch, C., Maloney, P. P., Fujita, T., and Muir, R. M. Cor- relation ...

arXiv 2004

[2] [2011]

Ester, M., Kriegel, H.-P., Sander, J., and Xu, X

doi: 10.1021/ci2003889. Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. A density- based algorithm for discovering clusters in large spatial databases with noise. InInternational Conference on Knowledge Discovery and Data Mining, pp. 226–231, 1996. Fan, J., Fu, A., and Zhang, L. Progress in molecular docking. Quantitative Biology, 7:83–89, 2019. Feng, B...

work page doi:10.1021/ci2003889 1996