{"id":"0ee737bf-0220-4b7e-8609-dc3bf4fe032b","arxiv_id":"2506.13196","paper_version":6,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"KEPLA jointly optimizes knowledge graph embeddings and cross attention to improve protein-ligand affinity prediction, reporting RMSE 1.202 on PDBbind core and 1.459 on CSAR-HiQ, beating all tested baselines.","lead":"KEPLA is a deep learning model that predicts how strongly a drug-like molecule binds to a protein by combining protein sequences and molecular graphs with biological knowledge from Gene Ontology and ligand properties. If the reported gains hold, it offers a fast, interaction-free way to rank drug candidates, with attention maps that point to the atoms and residues driving the binding.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"KG triplet sampling ambiguity could leak test-entity supervision in cross-domain and cold splits; the paper must clarify and verify before the generalization claims are accepted.","rationale":"The reader's weakest assumption identifies a genuine ambiguity in the KG training protocol, but it is not a confirmed error. The text says L_KGE is defined over all KG triplets (Eq. 8), which would imply test entities are scored, yet the mini-batch description says triplets are selected for sampled training pairs, which would exclude test entities from gradients. If the latter is implemented faithfully, no entity-specific test signal leaks through the KG objective; if the former is implemented, the cold-start claim is violated. Because the paper does not specify which behavior holds, and because cross-domain results are presented only as box plots without tabulated means and standard deviations, the central claim that KEPLA 'consistently outperforms' cannot be fully verified from the manuscript alone. Secondary concerns, such as the absence of knowledge-enhanced baselines like KG-MTL [32] and the 2025 Pattern Recognition work [6] from the comparison, and the use of the 3-billion-parameter ESM backbone that may account for a large share of the gains over older baselines, do not change the conditional verdict but reinforce the need for code-based evidence and tabulated cross-domain numbers. The reader's verdict of CONDITIONAL is therefore appropriate and no adjustment is required.","tokens_in":20729,"tokens_out":9942,"duration_ms":106957,"concrete_test":"Open the released code and inspect the KG-loss dataloader: determine whether L_KGE for a given batch is computed only from triplets whose head protein/ligand appears among the training pairs in that batch, or whether all KG triplets, including those for cold/cross-domain test entities, contribute. Then rerun the cold pair split (Section V-B) with a KG rebuilt exclusively from training-domain protein and ligand entities. If cold-test RMSE shifts by more than approximately 0.02 (the reported between-run standard deviation), the published cold-start results depend on test-entity supervision; if it does not shift materially, the ambiguity is resolved in the paper's favor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states the KG 'consists of triplets involving all proteins and ligands in the refined set of PDBbind dataset,' while Section V-B defines cold and clustering-based pair splits on the same PDBbind data. Section III-D defines the KG objective as an average over 'the set of all KG triplets' (Eq. 8), yet the training description in Section III-E says that for each mini-batch of sampled pairs, 'we select the associated KG triplets.' These statements are consistent only if triplet sampling is restricted to heads that appear in the training pairs. If, instead, all KG triplets are scored every epoch, then test-fold proteins and ligands that reside in the KG are directly supervised during training: their encoded global representations are optimized toward their GO/LP tail embeddings, injecting entity-specific signal that the cold-start protocol is specifically designed to forbid. The paper never states that the KG is rebuilt per fold or that test entities are excluded, and it provides no tabulated cross-domain numbers to allow independent checking. This ambiguity directly threatens the cross-domain generalization claims, which are central to the abstract's 'consistently outperforms state-of-the-art baselines' assertion. The resolution is straightforward: the authors should clarify the exact triplet sampling procedure and demonstrate, with code and a controlled experiment, that no test-entity triplets contribute to training.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KEPLA, an interaction-free deep learning framework for protein-ligand binding affinity prediction. KEPLA encodes protein sequences with a pretrained ESM model and ligand molecular graphs with a GCN, and jointly optimizes a knowledge-graph embedding loss over Gene Ontology and ligand-property triplets together with a cross-attention-based affinity regression loss. Experiments are reported on PDBbind and CSAR-HiQ under random, clustering-based cross-domain, and cold splits, with ablations and interpretability case studies. The central claim is that KEPLA consistently outperforms state-of-the-art baselines, including interaction-based methods, with RMSE improvements of 5.28% on PDBbind core and 12.42% on CSAR-HiQ.","tokens_in":21042,"tokens_out":4713,"duration_ms":46642,"significance":"If the results hold, KEPLA would be a useful contribution to interaction-free affinity prediction: it combines an established protein language model with a structured knowledge graph in a multi-task objective, provides code and a constructed KG dataset, and reports five-run statistics with ablations. The interpretability analyses are also a strength, although they are post-hoc rather than independent validation. The main value of the paper depends on whether the cross-domain and cold-start evaluations are leak-free and whether the comparison to baselines is controlled for the large pretrained encoder; these points are currently not established.","major_comments":[{"comment":"The knowledge graph is built from 'all proteins and ligands in the refined set of PDBbind' (Section IV-A), while the cross-domain clustering split and cold split in Section V-B are defined on the same PDBbind data. Section III-D defines the KG loss as an average over 'the set of all KG triplets' S (Eq. 8), but Section III-E says that for each mini-batch 'we select the associated KG triplets.' The paper never states that triplets whose head is a test-fold protein or ligand are excluded, nor that the KG is rebuilt per fold. If test entities appear as heads in triplets scored during training, their global representations receive entity-specific supervisory signal from GO/LP tails, which would violate the claim that test proteins and ligands are unseen in the cold setting and would inflate cross-domain performance. Please specify the exact triplet sampling procedure and verify experimentally that no test-entity triplets contribute to training, for example by reporting the intersection between KG head entities and test entity sets for each split.","section":"Section IV-A, Section III-D, Section III-E, Section V-B"},{"comment":"KEPLA uses ESM2_t36_3B, a 3-billion-parameter pretrained protein language model, as its protein encoder, while the interaction-free baselines in Table II (DeepDTA, GraphDTA, DrugBAN) use much smaller CNN/GCN encoders trained from scratch. The reported improvements of 5.28% on PDBbind core and 12.42% on CSAR-HiQ over GIANT could therefore be due to the capacity of the pretrained backbone rather than the proposed KG and cross-attention modules. The paper should include a controlled comparison where a baseline or an ablated KEPLA variant uses the same ESM backbone, or at least report results with a smaller protein encoder, to isolate the contribution of the proposed components.","section":"Section IV-D and Table II"},{"comment":"Cross-domain and cold-split results are presented only as box plots without a numeric table, making the central cross-domain claims unverifiable. The text states improvements of 12.25% in RMSE and 8.37% in R over DrugBAN on the PDBbind dataset, but these numbers cannot be checked from the figure. Please provide a table analogous to Table II with means and standard deviations for all metrics under both the clustering-based and cold pair splits, along with the number of test pairs and, ideally, statistical significance tests across the five runs.","section":"Section V-B and Figure 2"},{"comment":"The cold pair split description is ambiguous about how pairs are formed after selecting 70% of proteins and ligands. If 70% of protein entities and 70% of ligand entities are selected independently, it is unclear which pairs constitute the training set and whether pairs with one entity in the selected set and the other outside are discarded. This ambiguity affects the validity of the statement that neither proteins nor ligands in the test set are observed during training. Please clarify the exact pair-generation procedure for the cold split.","section":"Section V-B"}],"minor_comments":[{"comment":"The KEPLA row in Table II appears to repeat each value three times (e.g., '1.202 (0.017)1.202 (0.017)1.202 (0.017)'), which makes the table difficult to read and the reported numbers ambiguous; please fix the formatting.","section":"Table II"},{"comment":"The knowledge-level interpretability is generated by applying the same KG score function used during training, so it is not an independent validation of the retrieved knowledge. This is acceptable as a post-hoc explanation, but the paper should describe it as such rather than as confirmatory evidence.","section":"Section V-E"},{"comment":"The claim of being the 'first general framework' for deep integration of biochemical knowledge into PLA prediction is not consistent with the cited works [6] and [32], which also integrate knowledge with sequence and structural features; please temper or qualify the novelty claim.","section":"Abstract and Section I"},{"comment":"The phrase 'nanoscale affinity' in Section V-D should be 'nanomolar affinity'; the current wording is imprecise.","section":"Section IV-A"},{"comment":"Reference [35] (RotatE) is missing its publication year; please add the complete citation information.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The knowledge-graph leakage issue is the most serious concern. If it turns out that test entities are present in the KG triplets used for training in the cross-domain and cold splits, the central generalization claims would be invalid. However, the paper currently leaves the sampling procedure ambiguous rather than explicitly stating a leak, so a careful clarification and verification could resolve the issue. The uncontrolled ESM backbone comparison is also a substantive concern that should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the KEPLA paper. The in-domain story is credible and worth taking seriously. An interaction-free model that edges out GIANT on PDBbind core and CSAR-HiQ, with five-run statistics, is a real result. The joint KG embedding objective plus cross-attention decoder is a genuinely new combination, and the released KG is a useful artifact. The ablations are honest: the KG contributes, and the protein-side knowledge matters more than the ligand side. The interpretability case studies are qualitative but well-chosen and not oversold.\n\nThe soft spots are mostly in the generalization section. The cross-domain results appear only as box plots; for a paper claiming consistent SOTA, the lack of a numeric table is a real weakness. More importantly, the KG leakage concern raised by the stress test is legitimate. The KG is built from all refined-set proteins and ligands, and the paper never explicitly says that test entities are excluded from KG supervision in the cold and clustering splits. The training description in Section III-E (\"select the associated KG triplets\" for each mini-batch) suggests the intended reading is that only triplets attached to training pairs are used, which would be leak-free. But that needs to be stated and verified, ideally with code or a controlled experiment. If test entities' triplets are scored during training, the cold-start claim collapses.\n\nTwo smaller points: the paper cites KG-MTL and Li et al. but does not include them in the baselines, which makes the \"first\" claim feel overreaching. And the cross-domain baselines are restricted to interaction-free methods, which is reasonable but limits the comparison.\n\nThis paper is for anyone working on sequence-based affinity prediction or knowledge-enhanced molecular representations. It deserves a serious referee; the cross-domain questions are answerable with a revision, and the in-domain contribution stands. My recommendation: conditional acceptance, with the leakage clarification and numeric cross-domain results as conditions.","headline":"A credible in-domain win for knowledge-enhanced affinity prediction, but the cross-domain claims need a leakage check and numeric tables before they can be trusted.","tokens_in":21534,"tokens_out":2485,"would_cite":true,"duration_ms":26564,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"KEPLA claims that an interaction-free model jointly trained on a biochemical knowledge graph and affinity regression outperforms 3D-structure-based state-of-the-art methods on protein–ligand binding benchmarks, with RMSE gains of 5.28% on…","keywords":["protein-ligand binding affinity","knowledge graph embedding","Gene Ontology","cross attention","interaction-free prediction","PDBbind","CSAR-HiQ","drug discovery"],"falsifier":"Re-run the cross-domain and cold splits with the knowledge graph built only from source-domain (training) entities, or with every triplet removed that involves a test protein or ligand; if the RMSE advantage over DrugBAN collapses, knowledge-graph leakage rather than knowledge-enhanced representations explains the reported generalization.","tokens_in":20570,"feed_emoji":"🧬","tokens_out":12030,"duration_ms":118468,"temperature":0.7,"pith_summary":"The paper sets out to show that protein–ligand binding affinity can be predicted more accurately and more interpretably when an interaction-free deep model is trained not only on sequence and molecular-graph structure but also on a small biochemical knowledge graph. KEPLA builds that graph from Gene Ontology annotations for proteins and from computed ligand properties (molecular descriptors and chemical features), then ties the protein and ligand encoders to it through a joint knowledge-graph embedding objective while a cross-attention module forms the final affinity prediction. On PDBbind core and CSAR-HiQ, the authors report consistent improvements over state-of-the-art baselines, including 3D-structure-based methods, with RMSE reductions of 5.28% and 12.42% relative to the strongest baseline (GIANT), and further gains in cross-domain and cold-start splits. If this holds, sequence-plus-graph models could compete with structure-based scoring while remaining applicable to targets without known complex structures, and the knowledge-graph relations plus attention maps would give chemists grounded explanations instead of a black-box score.","feed_headline":"Sequence-only model beats 3D affinity models with knowledge-graph boost","feed_subtitle":"KEPLA trains encoders on GO and ligand-property triples, cutting RMSE by up to 12.4% over GIANT.","key_machinery":"Two coupled mechanisms carry the argument. First, a joint objective $L = L_{\\mathrm{PLA}} + \\beta L_{\\mathrm{KGE}} + \\lambda \\|\\Theta\\|_2^2$ makes the same protein and ligand encoders serve both affinity regression and knowledge-graph embedding: global average-pooled encoder outputs are linearly projected and scored as KG entities, with RotatE $\\|\\mathbf{h} \\circ \\mathbf{r} - \\mathbf{t}\\|$ for the protein–GO subgraph and TransE $\\|\\mathbf{h} + \\mathbf{r} - \\mathbf{t}\\|$ for the ligand–property subgraph. Second, a cross-attention interaction map $\\mathbf{V} = H_p^\\top H_d$ between protein fragments and ligand atoms produces softmax-normalized attention weights over both sides, and the weighted fragment and atom sums are concatenated and decoded by an MLP. This design lets the KG objective shape global representations while the attention objective preserves fine-grained local contact information; the ablation study attributes the reported gains to exactly this combination.","core_discovery":"KEPLA's central claim is that explicit biochemical knowledge, injected during encoding rather than bolted on afterwards, is what lets an interaction-free model surpass interaction-based state of the art; the paper presents this as the first general interaction-free framework to integrate such knowledge deeply. The paper constructs two knowledge graphs—protein–GO triplets and ligand–property triplets—from PDBbind refined-set entities, and uses the encoders' pooled global representations as entity embeddings optimized with RotatE for proteins and TransE for ligands. Simultaneously, local fragment-level representations are combined by cross attention into a joint vector decoded by an MLP into the affinity, and the two losses are minimized together. The authors report that KEPLA improves RMSE by 5.28% on the PDBbind core set and 12.42% on CSAR-HiQ against the best baseline (GIANT), outperforms interaction-free baselines in clustering-based and cold splits, and produces attention- and knowledge-based explanations that align with crystallographic binding-site residues for the three examined complexes.","pith_inferences":["If the leak-free condition holds, the cold-start gains imply that functional annotations, rather than protein identity, carry much of the predictive signal; a direct test is to compare affinity accuracy for test proteins that do versus do not have GO terms in the training knowledge graph.","The mechanism does not depend on Gene Ontology specifically; replacing it with pathway or disease annotations, or adding them, would test whether the benefit comes from functional semantics or from any relational side information.","The attention-map interpretability could become a quantitative benchmark by measuring agreement between top-attention protein fragments and crystallographic contact residues across the whole core set, not only the three showcased complexes.","Because the knowledge graph contains roughly 60,000 triplets, a scaled-up version with more GO evidence codes or ligand property types would test whether the reported gains grow or saturate with knowledge coverage."],"forward_implications":["Affinity prediction can be done from a protein sequence and a ligand's 2D graph with accuracy competitive with 3D-structure-based models, so targets without resolved complex structures remain addressable.","Gene Ontology terms and simple ligand descriptors such as hydrogen-bond donor and acceptor counts, aromatic rings, and rotatable bonds carry predictive signal beyond ESM and GCN features, making external annotations a cheap and reusable accuracy source.","The joint training recipe—shared encoders, a knowledge-graph embedding objective, and a cross-attention decoder—can be transferred to other molecular interaction tasks, such as drug–drug or protein–protein interaction prediction, by swapping in the appropriate knowledge graph.","Knowledge-graph triplet scores and cross-attention maps give a route to explain predictions at the level of GO functions and binding-site residues or atoms, which is useful for prioritizing which interactions to verify experimentally."],"supporting_citations":[{"why":"Supplies Gene Ontology annotations that form the protein-side KG triplets.","marker":"[16]"},{"why":"ESM is the pretrained protein language model used as the protein encoder backbone.","marker":"[19]"},{"why":"DrugBAN is the interaction-free baseline that KEPLA's cross-domain and cold-start gains are measured against.","marker":"[24]"},{"why":"GIANT is the state-of-the-art baseline for the headline RMSE comparisons on PDBbind core and CSAR-HiQ.","marker":"[30]"},{"why":"TransE is the KG embedding score function used for the ligand-property subgraph.","marker":"[34]"},{"why":"RotatE is the KG embedding score function used for the protein-GO subgraph.","marker":"[35]"},{"why":"PDBbind v2016 provides the refined/general/core data and the knowledge-graph entities.","marker":"[37]"},{"why":"CSAR-HiQ is the independent external benchmark used to test cross-dataset generalization.","marker":"[38]"}],"fun_headline_variants":["KEPLA folds knowledge graphs into sequence-only affinity prediction","Knowledge injection beats 3D models for binding affinity","Sequence and graph model with GO and ligand priors wins affinity","No 3D needed: KEPLA's knowledge boost tops affinity benchmarks","From sequences to affinity: KEPLA harnesses biological knowledge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that no test protein or ligand appears in the knowledge-graph triplets used during training; if a test entity is in the graph, the reported cross-domain and cold-start gains could come from reading entity identity off the knowledge graph rather than from generalizing.","fun_headline_variants_meta":{"raw":{"variants":["KEPLA folds knowledge graphs into sequence-only affinity prediction","Knowledge injection beats 3D models for binding affinity","Sequence and graph model with GO and ligand priors wins affinity","No 3D needed: KEPLA's knowledge boost tops affinity benchmarks","From sequences to affinity: KEPLA harnesses biological knowledge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1495,"prompt_tokens":933,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":476}},"tokens_in":549,"tokens_out":562,"duration_ms":7092,"temperature":1.0,"reasoning_tokens":476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:36:25.569995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the cross-domain and cold splits with the knowledge graph built only from source-domain (training) entities, or with every triplet removed that involves a test protein or ligand; if the RMSE advantage over DrugBAN collapses, knowledge-graph leakage rather than knowledge-enhanced representations explains the reported generalization.","supporting_citations":[{"cited_title":"Gene ontology: tool for the unification of biology,","cited_arxiv_id":null,"evidence_quote":"Supplies Gene Ontology annotations that form the protein-side KG triplets."},{"cited_title":"Evolutionary-scale prediction of atomic- level protein structure with a language model,","cited_arxiv_id":null,"evidence_quote":"ESM is the pretrained protein language model used as the protein encoder backbone."},{"cited_title":"Interpretable bilinear atten- tion network with domain adaptation improves drug–target prediction,","cited_arxiv_id":null,"evidence_quote":"DrugBAN is the interaction-free baseline that KEPLA's cross-domain and cold-start gains are measured against."},{"cited_title":"Giant: Protein-ligand binding affinity prediction via geometry- aware interactive graph neural network,","cited_arxiv_id":null,"evidence_quote":"GIANT is the state-of-the-art baseline for the headline RMSE comparisons on PDBbind core and CSAR-HiQ."},{"cited_title":"Translating embeddings for modeling multi-relational data,","cited_arxiv_id":null,"evidence_quote":"TransE is the KG embedding score function used for the ligand-property subgraph."},{"cited_title":"Rotate: Knowledge graph embedding by relational rotation in complex space,","cited_arxiv_id":null,"evidence_quote":"RotatE is the KG embedding score function used for the protein-GO subgraph."},{"cited_title":"Pdb-wide collection of binding data: current status of the pdbbind database,","cited_arxiv_id":null,"evidence_quote":"PDBbind v2016 provides the refined/general/core data and the knowledge-graph entities."},{"cited_title":"Csar benchmark exercise of 2010: combined evaluation across all submitted scoring functions,","cited_arxiv_id":null,"evidence_quote":"CSAR-HiQ is the independent external benchmark used to test cross-dataset generalization."}],"review_version":1}