{"id":"0246fc33-8dea-47ba-a3d3-14bc4bab2d1e","arxiv_id":"2601.09693","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A contrastive geometric model jointly trained on protein–ligand complexes and bioactivity data achieves competitive virtual screening, strong target fishing, and ligand-conditioned pocket prediction without predefined binding sites.","lead":"This paper introduces ConGLUDe, a single AI model that learns from both 3D protein–ligand complex structures and large-scale bioactivity data, and can screen drugs, find targets, and predict binding sites. It matters because it unifies two traditionally separate drug-design approaches into one fast, pocket-agnostic model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'unified training' claim is undercut by the frozen VN-EGNN during ligand-based batches; the stated reason for freezing is not self-evident and the 56M bioactivity labels may only be tuning a linear head.","rationale":"I focused on the reader's identified weakest assumption and found it to be the most load-bearing threat to the central claim. The paper's stated reason for freezing the VN-EGNN encoder — that ligand-based data lack pocket structural information and therefore cannot meaningfully optimize the encoder — is internally weak: the ligand-based objective directly supervises the global protein embedding p, which is produced by the VN-EGNN, so nonzero gradients exist. The fact that SPRINT trains a protein encoder end-to-end on the same MERGED data shows such optimization is practical. The ablation results further undercut the 'unified' narrative: no configuration that uses both data sources beats the best single-source configuration on either virtual screening benchmark, which is exactly what one would expect if ligand-based training is only adjusting a linear head and ligand MLP. I considered other concerns, such as possible ligand-level leakage in the target-fishing benchmark (only proteins were filtered at 90% sequence identity, not ligands), which could also threaten a headline result. That concern is real and worth checking, but it is a benchmark-integrity issue rather than a challenge to the conceptual architecture. The frozen-encoder issue strikes at the core claim that ConGLUDe is a unified structure- and ligand-based model. The proposed test directly settles whether unfreezing the encoder changes outcomes. Because the reader's verdict is already CONDITIONAL and this concern reinforces the need for conditional acceptance rather than changing it, I recommend UNCHANGED.","tokens_in":24054,"tokens_out":8671,"duration_ms":90120,"concrete_test":"Train a variant of ConGLUDe in which the VN-EGNN encoder is not frozen during ligand-based batches, using a small learning rate (e.g., 1e-4) for the encoder and the original learning rate for the linear layer and ligand encoder, with all other hyperparameters and data splits identical. Evaluate on LIT-PCBA, DUD-E, and the Kinobeads target-fishing set, and also re-measure pocket prediction on COACH420/HOLO4K to check for catastrophic forgetting. If the unfrozen variant improves ligand-based task AUROC by more than ~1 point without destroying pocket prediction, then the reported frozen-encoder model is not actually realizing the claimed unified training. If performance is unchanged, the frozen choice is harmless and the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ConGLUDe is jointly trained on structure-based complexes and millions of ligand-based bioactivity measurements, and that this unified training is what enables one pocket-agnostic model to handle virtual screening, target fishing, and pocket selection. However, Section 3.2.2 says that during ligand-based training 'the VN-EGNN module cannot be meaningfully optimized and is therefore kept frozen,' with only the linear layer on the protein node and the ligand encoder being updated. This justification is not compelling: the ligand-based loss in Eq. 14 is a function of the global protein representation p, which is an output of the VN-EGNN encoder. Gradients with respect to the VN-EGNN parameters are therefore well-defined, and training the whole encoder on activity labels is exactly what SPRINT (McNutt et al., 2024) does on the same MERGED data. The real constraint may be that unfreezing risks catastrophic forgetting of pocket prediction, but that is a trade-off, not an impossibility. Because the encoder is frozen, the 56M bioactivity labels cannot reshape the geometric protein representations; they only calibrate a linear projection and the ligand MLP. The paper's own ablation (Table F1) is consistent with this: ConGLUDe achieves 66.25 AUROC on LIT-PCBA vs 67.94 for ligand-based-only training, and 82.04 on DUD-E vs 83.88 for structure-based-only training. No synergistic gain from joint training is demonstrated. Thus the 'unified' contribution may reduce to a fine-tuned head on a structure-pretrained encoder, and the claim that ConGLUDe 'enables learning from both structure-based ... and ligand-based ... data' at the geometric level is overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":24424,"tokens_out":5447,"duration_ms":50281,"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":[{"comment":"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":"Section 3.2.2, Eq. (14), Table F1"},{"comment":"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":"Section 4.1, Table 1"},{"comment":"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.","section":"Section 4.4, Table 4"}],"minor_comments":[{"comment":"Equation (5) uses φ_h^{(BR)} in the P→R message-passing step; this appears to be a typo for φ_h^{(PR)}.","section":"Section 3.1.1, Eq. (5)"},{"comment":"The phrase 'a ligand-protein pairs' should be 'ligand–protein pairs' or 'a pair of ligand and protein.'","section":"Section 3.2.1"},{"comment":"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.'","section":"Section G.3"},{"comment":"The S2Drug row in Table G2 has numbers that run together (e.g., '92.4679.25'); the formatting should be fixed.","section":"Table G2"}],"recommendation":"major_revision","confidential_remarks":"The paper is methodologically interesting and the code release is a plus. The main issues are that the 'unified training' contribution is weakened by the frozen protein encoder and the lack of synergistic gains in the ablation, and that the virtual-screening and pocket-selection claims overstate the reported numbers. These are fixable by reframing and by adding experiments or significance tests. I recommend a major revision rather than rejection, because the target-fishing results and the ligand-conditioned pocket selection on ASD are genuinely novel and worth publishing if properly scoped."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper deserves a referee, but the headline claim is one notch above what the evidence supports.\n\nConGLUDe 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.\n\nSoft 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.\n\nWho 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.","headline":"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.","tokens_in":24991,"tokens_out":3020,"would_cite":true,"duration_ms":33719,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ConGLUDe unifies structure- and ligand-based drug design in one contrastive geometric model, removing the need for predefined binding pockets.","keywords":["contrastive learning","protein–ligand interaction","binding site prediction","virtual screening","target fishing","ligand-conditioned pocket selection","drug design","equivariant graph neural network"],"falsifier":"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.","tokens_in":23920,"feed_emoji":"💊","tokens_out":10384,"duration_ms":98433,"temperature":0.7,"pith_summary":"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.","feed_headline":"One model unifies structure- and ligand-based drug design","feed_subtitle":"Trained on 3D complexes and 56M bioactivity data, it removes the need for predefined binding pockets.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["One model: no predefined pockets, unified drug design","ConGLUDe unifies structure and ligand design without defined pockets","Unified drug design with implicit pocket prediction","Contrastive learning ties structure and ligand design together","Pocket-free: one model for screening, fishing, and pocket selection"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One model: no predefined pockets, unified drug design","ConGLUDe unifies structure and ligand design without defined pockets","Unified drug design with implicit pocket prediction","Contrastive learning ties structure and ligand design together","Pocket-free: one model for screening, fishing, and pocket selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1274,"prompt_tokens":712,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":481}},"tokens_in":456,"tokens_out":562,"duration_ms":5167,"temperature":1.0,"reasoning_tokens":481,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:30:37.798998+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}