{"id":"038a892c-6bc1-4520-b6e8-ef7917692324","arxiv_id":"2505.11529","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"DynamicDTA combines a drug molecular graph, a protein sequence encoder, and four per-protein dynamic descriptors with cross-attention and tensor fusion to predict binding affinity, outperforming seven baselines in most settings but with questionable train-test separation.","lead":"A new deep learning model for drug-target binding affinity prediction adds four measurements of protein flexibility from molecular dynamics simulations. The authors report better accuracy than seven existing models on three affinity datasets, though the evaluation may leak protein-level information between training and test sets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central performance claim is internally contradicted by Table 2: DynamicDTA is worse than ImageDTA on IC50* and no dataset reaches the claimed 'at least 3.4%' RMSE improvement.","rationale":"The reader's weakest assumption concerns the lack of target-disjoint splits, which is a legitimate methodological threat: dynamic descriptors are constant per protein, so if a protein appears in both training and test folds, the model can exploit target-level affinity biases. That concern is plausible and worth testing. However, it is not the most certain or most direct failure of the central claim. The paper's own Table 2 and Section 3.4 text openly disagree with the abstract: IC50* results are worse than ImageDTA in RMSE, and the 'at least 3.4%' improvement is not reproducible from any row of Table 2. This internal contradiction does not depend on assumptions about how the folds were constructed. It independently invalidates the headline claim. The split-leakage issue, if confirmed, would invalidate the comparison more broadly, but the abstract/table inconsistency is already sufficient to reject the paper in its current form. Therefore I agree with the reader's overall REJECT verdict but not with the exact prioritization: the most load-bearing concern is the self-contradictory performance claim, with the split protocol as a separate compounding risk. The verdict remains unchanged.","tokens_in":18287,"tokens_out":10716,"duration_ms":109958,"concrete_test":"Recompute the relative eRMSE improvement of DynamicDTA over the best baseline for each dataset from Table 2 using (best_baseline_eRMSE - DynamicDTA_eRMSE) / best_baseline_eRMSE. For IC50* this is (0.600 - 0.611) / 0.600 = -1.8%, for Kd* it is (1.029 - 1.007) / 1.029 = 2.1%, and for Ki* it is (0.792 - 0.727) / 0.792 = 8.2%. None equals or exceeds 3.4%, and IC50* is negative. If the authors instead intend '3.4%' as an average relative improvement over all seven baseline methods, recompute that average explicitly from Table 2; otherwise the stated central claim must be amended or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim, stated in the abstract and Section 3.4, is that DynamicDTA 'consistently outperforms all baseline methods' and improves RMSE 'by at least 3.4%'. The paper's own Table 2 contradicts this. On IC50*, ImageDTA has eRMSE 0.600 while DynamicDTA has 0.611, so DynamicDTA is worse by 1.8%. On Kd*, the best baseline is GraphDTA with eRMSE 1.029 and DynamicDTA achieves 1.007, a 2.1% improvement, not 3.4%. On Ki*, DynamicDTA's 0.727 versus GraphDTA's 0.792 is an 8.2% improvement, but no dataset supports the 'at least 3.4%' statement as a comparison against the second-best method, and the IC50* result directly violates 'consistently outperforms all baseline methods'. Because the central claim is a quantitative superiority claim, this internal inconsistency is directly load-bearing: the reported evidence does not support the stated conclusion. The reader's target-split concern is an additional and independent threat to the comparison, but the contradiction alone falsifies the abstract claim as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DynamicDTA, a deep learning framework for drug-target binding affinity (DTA) prediction that combines three input modalities: a molecular graph representation of the drug processed by a graph convolutional network, the protein sequence encoded with dilated convolutions, and a four-dimensional vector of MD-derived dynamic descriptors (Avg.RMSF, Avg.Gyr, Div.SE, Div.MM) processed by an MLP. The protein sequence and dynamic vectors are fused with a multi-head cross-attention mechanism, and a tensor fusion network integrates all modalities before regression. The authors evaluate on three BindingDB-derived datasets (Kd*, Ki*, IC50*) and an external Kiba* dataset, comparing against seven baselines. The central claim is that DynamicDTA 'consistently outperforms all baseline methods' and achieves 'at least 3.4% improvement in eRMSE' across the three datasets; a case study on HIV-1 and interpretability analyses are also presented.","tokens_in":18516,"tokens_out":4447,"duration_ms":41835,"significance":"If the reported results were valid, the idea of injecting MD-derived protein flexibility descriptors as an additional input modality would be a useful contribution to DTA prediction, and the public availability of code and data is a strength. However, the evaluation evidence as presented is internally inconsistent with the headline claims, and the cross-validation protocol is described too loosely to rule out data leakage through per-protein dynamic descriptors. These issues directly undermine the central quantitative claim, so the current version does not establish the stated improvements.","major_comments":[{"comment":"The abstract and Section 3.4 claim that DynamicDTA 'achieves by at least 3.4% improvement in eRMSE' and 'consistently outperforms all baseline methods' on the three datasets. Table 2 directly contradicts this: on IC50*, DynamicDTA has eRMSE 0.611, worse than ImageDTA's 0.600, an 1.8% deficit; on Kd*, the improvement over GraphDTA is from 1.029 to 1.007, i.e., 2.1%, not 3.4%; only on Ki* does the improvement reach 8.2%. The Section 3.4 text even acknowledges the IC50* shortfall ('falls slightly behind ImageDTA in eRMSE'), which contradicts 'consistently outperforms all baseline methods.' The headline claim as stated is thus falsified by the paper's own results and must be corrected or substantially qualified.","section":"Abstract; Section 3.4; Table 2"},{"comment":"The five-fold cross-validation splits are described only as 'each dataset was split into five parts,' without specifying whether the splits are disjoint at the protein or ligand level. Because the dynamic descriptor vector is a constant per protein (Section 2.2.1), a protein appearing in both training and test folds would leak its descriptor and enable the model to memorize protein-level affinity biases, while the baselines, which lack such per-protein constants, would not benefit equally. This makes the comparison potentially unfair. Additionally, the min-max normalization in Eq. (2) uses x_min and x_max computed 'across the dataset'; if computed over the full dataset including test folds, that is another form of leakage. The authors must clarify the split granularity, and ideally rerun the evaluation with a target-disjoint split to demonstrate generalization to unseen proteins.","section":"Section 2.1; Section 2.2.1; Eq. (2)"},{"comment":"The ablation study claims that replacing dilated convolutions with standard convolutions 'resulted in a decline in both eRMSE and R metrics.' However, on IC50*, the 'w/o Dilated' row reports eRMSE 0.611 and R 0.923, identical to the full DynamicDTA model; on Kd* and Ki*, the eRMSE differences are 0.014 and 0.005, respectively. These differences are well within the reported standard deviations (e.g., Kd* std of 0.050-0.053) and do not support the stated conclusion that dilated convolution is important. The text should be revised to acknowledge that the effect is negligible on IC50* and not statistically significant, or the claim should be supported with significance testing.","section":"Section 3.5.1; Table 3"},{"comment":"The Kiba* external validation is presented as evidence of 'better generalization ability,' but the absolute performance is weak: DynamicDTA achieves eRMSE 5.090 and R 0.202, and the second-best model has R 0.176. The claimed '13% improvement in R' is relative to an almost-zero baseline, and the eRMSE values are several-fold larger than on the main datasets, suggesting the model does not transfer well to a fully disjoint dataset. The statement 'further highlight the better generalization ability' is an overstatement given these numbers; at minimum, the authors should discuss why the absolute performance is so much worse and temper the generalization claim.","section":"Section 3.7; Table 5"}],"minor_comments":[{"comment":"There are numerous typographical errors that should be corrected: 'crated' (Section 2.1), 'benefical' (Abstract), 'aromaitic' (Section 2.2.1), 'enhancs' (Section 3.2), 'outperformes' (Section 3.4), and 'Liner Regression' (Table 5).","section":"Throughout"},{"comment":"The HIV-1 case study lists five of ten top drugs as 'Unconfirmed' in PubMed evidence, and the docking scores reported in Fig. 6 (3.58-4.75) are not contextualized against any threshold or comparison, so the statement that these are 'relatively high' is not substantiated.","section":"Section 3.8.2; Table 6"},{"comment":"The hyperparameter settings are described, but no information is given about how they were chosen; the parameter sensitivity analysis in Section 3.6 is performed only on Ki*, and it is unclear whether the same optimal values were used for the other datasets.","section":"Section 3.1"},{"comment":"The scatter plots in Figure 2 compare only DynamicDTA against DEAttentionDTA, not against the strongest baselines (ImageDTA on IC50* or GraphDTA on Kd*/Ki*), so the visual claim of 'tighter clustering' is not a complete comparison.","section":"Section 3.4"}],"recommendation":"reject","confidential_remarks":"The manuscript has a central evaluation flaw: the reported results contradict the abstract's quantitative claim, and the cross-validation protocol does not rule out protein-level leakage via the per-protein dynamic descriptors. These are fixable in principle (by correcting claims and rerunning with target-disjoint splits), but the current evidence does not support the paper's main conclusions, and the presentation shows signs of haste. I would encourage the authors to redo the evaluation carefully before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe interesting part of this paper is the idea: take four per-protein MD descriptors (RMSF, Gyr, Div.SE, Div.MM) from ATLAS and treat them as a compact third input modality alongside drug graph and protein sequence. The architecture is clearly described, the code is public, and the ablation study is thorough. I think the motivation is sound and the external Kiba* test is a reasonable addition.\n\nBut the central performance claim does not survive contact with the paper's own tables. The abstract says DynamicDTA improves eRMSE by at least 3.4% and 'consistently outperforms all baseline methods.' Table 2 shows ImageDTA has a lower eRMSE on IC50* (0.600 vs 0.611), and on Kd* the gain over GraphDTA is about 2%, not 3.4%. Only on Ki* is the advantage that large. The body of Section 3.4 admits the IC50* shortfall, so the abstract is simply not consistent with the data. That is a load-bearing error because the entire contribution is framed as empirical superiority.\n\nThe second issue is the evaluation split. The paper says 'each dataset was split into five parts' and does not say the splits are disjoint at the protein level. Since the dynamic descriptors are constant per protein, a train/test split at the sample level lets the model see the same protein's descriptors on both sides. Baselines also see the protein sequence, so they might also exploit protein identity, but the comparison is uncontrolled and the four scalar descriptors are a suspiciously convenient shortcut. A target-disjoint split would resolve this, and the authors should report it.\n\nI would not reject this outright because the underlying method is plausible and reproducible. But as written, the abstract overclaims and the evaluation is at best incomplete. A revision that fixes the numbers and adds a target-disjoint comparison could make this a useful contribution to the DTA literature. If I were the editor, I would send it to review with a clear request to re-evaluate under a proper split and to align the abstract with the tables. It deserves serious referee time, but the verdict would likely be major revision.","headline":"A sound architecture let down by an overstated abstract and an uncontrolled train/test split; the reported 'consistent' improvement is contradicted by the paper's own Table 2.","tokens_in":19073,"tokens_out":4456,"would_cite":false,"duration_ms":42053,"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":"Protein motion data sharpens drug-target affinity prediction.","keywords":["drug-target binding affinity","protein dynamics","molecular dynamics descriptors","graph neural network","cross-attention","tensor fusion network","binding affinity prediction"],"falsifier":"Re-run the same experiments with splits that never put the same protein in both training and test; if the eRMSE advantage over the best baselines on Kd* and Ki* falls below the reported 2-6% or disappears, the central claim of a dynamics-driven gain is falsified. A simpler check is to compare DynamicDTA against a variant that replaces the four descriptors with a learned per-protein embedding of the same size.","tokens_in":18053,"feed_emoji":"🧬","tokens_out":7158,"duration_ms":65888,"temperature":0.7,"pith_summary":"DynamicDTA proposes that drug-target binding affinity prediction is improved when a protein's dynamical behavior, not just its sequence or static structure, is given to the model. The paper's central claim is that adding four molecular-dynamics descriptors for each target—averaged residue fluctuation, radius of gyration, and two conformational-divergence measures—as an extra input modality yields consistently lower eRMSE and higher Pearson correlation than seven baselines on three affinity datasets. It reports gains of about 2.2% and 6.5% in eRMSE over the second-best model on the Kd and Ki benchmarks, and the best correlation on IC50 even though its eRMSE trails one image-based baseline slightly. If true, the result matters because it would mean cheap per-protein dynamics descriptors can supply information that sequence- and graph-only models miss, improving virtual screening without requiring target-ligand complex structures.","feed_headline":"Protein motion data sharpens drug-target affinity prediction","feed_subtitle":"Four molecular-dynamics descriptors cut error on two benchmarks and hold the best correlation on a third.","key_machinery":"The mechanism that carries the argument is a per-protein dynamic descriptor vector: four normalized numbers (average root mean square fluctuation, average radius of gyration, and two TM-score-derived divergence scores between simulation conformations) computed from molecular dynamics and attached to every target. The model pairs this vector with the drug's molecular graph embedding and the protein's dilated-convolution sequence embedding, lets the sequence and dynamics vectors attend to each other, and fuses all three through a tensor fusion network that computes their outer product. The descriptor vector is what distinguishes DynamicDTA from sequence- and graph-only baselines; the cross-attention and tensor fusion determine how that extra signal enters the prediction.","core_discovery":"On its own terms, the paper establishes that a deep network taking a drug molecular graph, a protein sequence, and a four-dimensional vector of per-protein dynamics descriptors can beat seven existing predictors. The descriptors are treated as a constant property of each protein, normalized and passed through an MLP; a multi-head cross-attention layer lets the sequence embedding and the dynamics vector refine each other, and a tensor fusion network forms the outer product of the ligand, target, and dynamics embeddings so that unimodal, bimodal, and trimodal interactions are all represented before regression. Across the Kd*, Ki*, and IC50* datasets the model reports the best Pearson correlation in all three cases, the best eRMSE on Kd* and Ki*, and an eRMSE slightly above the best baseline on IC50*. On the held-out external set it reports roughly a 4.3% eRMSE improvement over the second-best method without fine-tuning, and ablations attribute part of the gain to the dynamics descriptors, especially the two TM-score-based divergence measures.","pith_inferences":["Because the descriptor vector is constant per protein, an important control would be to replace it with a one-hot protein identity: if the gain persists, the model is learning protein-level affinity biases rather than a dynamics-based interaction signal.","If protein-disjoint splits erase the reported gains on Kd* and Ki*, the improvement would be attributable to descriptor leakage across training and test folds rather than to generalizable dynamics information.","The same architecture could be tested on per-residue or per-complex dynamic features, such as ligand-induced flexibility changes, instead of whole-protein constants; that would reveal whether the signal comes from binding-site plasticity.","The paper's own proposed extension of generating MD descriptors with diffusion models would turn an expensive per-protein preprocessing step into a fast prediction step, making the method practical for large-scale screening."],"forward_implications":["Adding MD-derived descriptors as a fourth modality reduces eRMSE by about 2.2% on Kd* and 6.5% on Ki* relative to the second-best baseline, with the largest gains concentrated in the more compact affinity distribution of Ki*.","Removing the dynamics descriptors degrades performance, and removing gyration radius plus the two divergence measures together produces the largest drop, so conformational-divergence descriptors carry most of the signal.","Tensor fusion outperforms concat, sum, average, and Hadamard fusion, indicating that higher-order interactions across ligand, sequence, and dynamics matter for the prediction.","A model trained on the three affinity datasets transfers to an unseen external set without fine-tuning, reporting roughly 4.3% lower eRMSE than the next-best method.","The cross-attention weights highlight residues in experimentally characterized binding pockets of the 2FOS complex, suggesting the dynamic signal is not only predictive but also interpretable."],"supporting_citations":[{"why":"Supplies the four per-protein molecular-dynamics descriptors that form the new input modality.","marker":"[32]"},{"why":"Provides the experimental affinity values used to build the three main evaluation sets.","marker":"[41]"},{"why":"A sequence-based CNN baseline against which DynamicDTA's performance is compared.","marker":"[15]"},{"why":"A graph-based baseline, the second-best performer on Kd* and Ki*, that anchors the reported improvements.","marker":"[20]"},{"why":"Provides the attention-based baseline and the five-fold cross-validation protocol used in evaluation.","marker":"[17]"},{"why":"Supplies the tensor fusion network used to integrate ligand, sequence, and dynamics modalities.","marker":"[31]"},{"why":"Source of the affinity data used to build the external validation set.","marker":"[11]"}],"fun_headline_variants":["Protein motion data sharpens drug-target affinity scores","Dynamic descriptors cut error in binding affinity prediction","Motion-aware model beats seven baselines in drug-target affinity","Adding dynamics features boosts drug-target binding prediction","DynamicDTA: Dynamic protein features improve affinity forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the five-fold cross-validation splits are protein-disjoint; if the same protein appears in both training and test, its per-protein dynamics vector leaks and can inflate the reported gains over baselines that lack such a feature.","fun_headline_variants_meta":{"raw":{"variants":["Protein motion data sharpens drug-target affinity scores","Dynamic descriptors cut error in binding affinity prediction","Motion-aware model beats seven baselines in drug-target affinity","Adding dynamics features boosts drug-target binding prediction","DynamicDTA: Dynamic protein features improve affinity forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000154,"raw_usage":{"total_tokens":1221,"prompt_tokens":968,"completion_tokens":253,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":181}},"tokens_in":584,"tokens_out":253,"duration_ms":3359,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:53:04.377060+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same experiments with splits that never put the same protein in both training and test; if the eRMSE advantage over the best baselines on Kd* and Ki* falls below the reported 2-6% or disappears, the central claim of a dynamics-driven gain is falsified. A simpler check is to compare DynamicDTA against a variant that replaces the four descriptors with a learned per-protein embedding of the same size.","supporting_citations":[{"cited_title":"Bioinformat- ics 37(8):1140–1147","cited_arxiv_id":null,"evidence_quote":"A graph-based baseline, the second-best performer on Kd* and Ki*, that anchors the reported improvements."},{"cited_title":"Journal of Chemical Information and Modeling 54(3):735–743","cited_arxiv_id":null,"evidence_quote":"Source of the affinity data used to build the external validation set."}],"review_version":1}