{"id":"c5ffc51a-1cec-4c3f-b502-ec09ccdc494a","arxiv_id":"2506.05427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MTPNet conditions molecular graph representations on receptor protein and pocket features, improving activity cliff prediction across 30 MoleculeACE targets.","lead":"This paper introduces MTPNet, a deep learning method that predicts activity cliffs by combining molecule structure with the target protein's sequence and binding pocket. It reports lower error than molecule-only models across 30 drug-target datasets, which could help drug design teams prioritize compounds faster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains do not isolate receptor conditioning from dataset identity and joint training; without matched baselines or a non-protein target-embedding control, the central claim that MTPNet's improvement comes from protein/pocket guidance is not established.","rationale":"The paper's strongest claim is that receptor proteins, especially pocket-level interactions, drive a substantial RMSE improvement. For that claim to hold, target information must be the causal variable. The current evidence cannot establish causality because the evaluation protocol is ambiguous and the target representation is constant within each dataset. If MTPNet is trained on all datasets jointly, its F_target is essentially a per-target identifier; the model can learn target-specific calibration that molecule-only baselines trained per dataset cannot exploit. The scale-up experiment addresses model capacity but not this target-identity confound. I agree with the reader that pocket-feature construction is underspecified (Section 3.2c), but that is a reproducibility concern; the identity/joint-training confound directly threatens the attribution of the headline improvement. A one-hot target-ID control is cheap and decisive. If it performs as well as ESM2-conditioned MTPNet, the paper's claimed novelty ('first to employ receptor proteins as guiding information') is not supported by its experiments. The verdict remains conditional because the claim could still survive the test, but the condition is more binding than a request for pocket details: the authors should provide matched-protocol comparisons and a non-semantic target control.","tokens_in":12370,"tokens_out":8085,"duration_ms":103871,"concrete_test":"Run a controlled MoleculeACE experiment with the released code: train every baseline and MTPNet on exactly the same combined 30-dataset split with identical hyperparameters and early stopping, then train MTPNet variants in which F_target is replaced by (a) a per-dataset one-hot embedding and (b) randomly permuted ESM2 embeddings across targets, keeping all other components fixed. If the one-hot or permuted variant retains the RMSE gain over matched molecule-only baselines, or if matched joint training closes the reported gap, the receptor/pocket-guidance explanation is not supported; if the real ESM2 variant significantly beats both controls, the central claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that receptor/pocket information, not extra parameters or a different training protocol, explains the large RMSE improvements. The paper never specifies whether the baselines (GCN, GAT, GIN, GraphTrans, MolCLR, Mole-BERT) are trained per target, as in the standard MoleculeACE benchmark, while MTPNet is trained jointly on all 30 targets, as its 'unified' design suggests. If so, the comparison changes two variables at once: target conditioning and joint multi-target training. Within each MoleculeACE dataset the protein is fixed, so F_target is a constant vector. In a jointly trained model, that vector can act as a dataset/target embedding that calibrates each dataset's label distribution (pKi ranges differ across targets). A one-hot target-ID embedding, or even a per-target bias, could plausibly reproduce a large part of the reported RMSE drop. The 'w/o MTS & MPS' row in Table 3 (RMSE 0.917) is the molecule-only backbone; if it was also trained jointly, it is handicapped by not knowing which dataset each molecule comes from, which inflates the apparent value of the MTP module. The scale-up control in Table 1 increases parameters but does not add target identity, so it cannot rule out this confound. The reader's pocket-extraction concern is real, but the more basic issue is that the experimental protocol does not yet establish that protein-semantic or pocket guidance—rather than target identity and multi-task training—drives the claimed 18.95% improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MTPNet, a conditional framework that augments molecular GNNs with receptor protein information at two levels: Macro-level Target Semantic (MTS) guidance, which uses adaptive layer normalization conditioned on pooled target features, and Micro-level Pocket Semantic (MPS) guidance, which uses cross-attention between molecule and pocket features. The authors report experiments on 30 MoleculeACE datasets, plug-and-play gains over six GNN backbones, an ablation study, a CYP3A4 classification result, and an attention visualization. The central claim is that receptor/pocket guidance yields large RMSE improvements and that MTPNet is the first unified, target-aware activity-cliff predictor.","tokens_in":12753,"tokens_out":5122,"duration_ms":54628,"significance":"If the claimed gains are reproducible and can be attributed to protein/pocket conditioning rather than to target identity, joint training, or added parameters, the paper would make a useful contribution: it offers a pluggable module for existing GNNs and a unified framework for activity-cliff prediction. The code release is a strength, and the idea of combining global target semantics with local pocket semantics is reasonable. However, the current evidence does not yet establish that protein semantics, rather than target identity or training protocol, explains the reported improvements, so the significance of the central claim remains conditional on additional experiments.","major_comments":[{"comment":"The paper never states whether the baseline models (GCN, GAT, GIN, GraphTrans, MolCLR, Mole-BERT) are trained per target, as in the standard MoleculeACE benchmark, while MTPNet is trained jointly across the 30 datasets as a 'unified' model. Within each MoleculeACE dataset the protein is fixed, so F_target is a constant vector and can act as a dataset/target embedding. A jointly trained molecule-only backbone (Table 3, 'w/o MTS & MPS', RMSE 0.917) is not shown to be matched to this protocol, and the scale-up control in Table 1 varies parameters but not target identity or training protocol. To support the central claim that receptor/pocket information drives the 18.95% improvement, the authors should report per-target results and matched baselines, including a per-target-trained MTPNet, a jointly trained molecule-only backbone, and a control with a one-hot target-ID embedding or per-target bias in place of F_target.","section":"Section 4.2, Table 1, Table 3"},{"comment":"Only aggregate RMSE, PCC, and R² values are reported, with no per-dataset breakdown, error bars, or significance tests. The three headline improvements are mutually inconsistent: the abstract and Section 4.2 state an average RMSE improvement of 18.95% over mainstream GNN architectures, the introduction states a 7.2% improvement over SOTA, and the plug-and-play results in Table 1 are summarized as a 19.0% RMSE decrease. Additionally, Table 3's 'w/o MTS & MPS' baseline (RMSE 0.917) does not match the Mole-BERT baseline in Table 1 (RMSE 0.947), so the reader cannot tell whether the two tables use the same training protocol. Per-dataset tables, confidence intervals, and a precise definition of each reported improvement are needed before the claim of significant improvement can be evaluated.","section":"Table 1 and Table 3"},{"comment":"The pocket feature extraction is underspecified. The text states only that F_pocket ∈ R^{p×d} is extracted from F_target using a Pocket Prior Extractor (e.g., Cavity Plus), without saying whether 3D structures are available for all 30 MoleculeACE targets, how pocket residues are selected, how p is set, or how these residue embeddings are projected into the same space as ESM2 sequence features. If pocket extraction is noisy or unavailable for some targets, the MPS cross-attention operates on missing or approximate inputs, and the reported gains could come from added parameters or from MTS alone. The authors should specify the exact pocket extraction pipeline and include an ablation that removes MPS while keeping the same parameter count.","section":"Section 3.2c"},{"comment":"There are internal inconsistencies in the MPS formulation. Eq. (6) calls Φ_MPS with F_target, while Eq. (13) defines Φ_MPS(F_mol, F_target) = CrossAttention(F_mol, F_pocket); moreover, the text says K_pocket and V_pocket are generated by concatenating F_mol and F_pocket, but Eq. (12) defines K_pocket = W_k F_pocket and V_pocket = W_v F_pocket with no concatenation. These mismatches make the exact micro-level mechanism ambiguous and need to be corrected.","section":"Equations (6), (12), and (13)"}],"minor_comments":[{"comment":"The word 'summaried' in the contributions list should be 'summarized.'","section":"Section 1"},{"comment":"The Feature Abstractor is said to 'reduce the feature dimensions' via average pooling; since F_target is n×d, average pooling over the n dimension gives a 1×d vector, so the description should clarify whether the reduction is in sequence length or in d.","section":"Section 3.2b"},{"comment":"The text refers to results presented in 'Appendix,' but no appendix is included in the manuscript; either add the per-dataset tables or remove the reference.","section":"Section 4.2"},{"comment":"The cross-attention text says F_mol and F_pocket are concatenated to generate K_pocket and V_pocket, but the equation shows only F_pocket; the discrepancy should be resolved.","section":"Section 3.2c"},{"comment":"The caption says RMSE values are shown in reciprocal form, but the axis labels and legend are not explained; please state what the plotted quantity is and how the 7.2% improvement is computed.","section":"Figure 1 caption"},{"comment":"The interpretation analysis is qualitative and does not provide any quantitative comparison (e.g., attention-overlap scores or performance on molecules with and without the highlighted groups); please add a quantitative measure or soften the claim.","section":"Section 4.4"},{"comment":"The paper does not report hyperparameters, training epochs, data splits, or number of random seeds for the CYP3A4 classification experiment; please add these details to allow reproducibility.","section":"Section 4.1 and Table 2"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between Table 1 (Mole-BERT RMSE 0.947) and Table 3 ('w/o MTS & MPS' RMSE 0.917) is particularly concerning because the latter is the only molecule-only baseline in the ablation; the authors should clarify whether these are different training protocols or whether one table contains an error. I did not independently run the code; my assessment is based solely on the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful novelty here is real but narrow: MTPNet is, to my knowledge, the first activity cliff model that conditions molecular representations on receptor sequence and pocket features, and the macro/micro split is sensible. The architecture is mostly assembled from ESM2, AdaLN, and cross-attention, so don't expect a new learning principle. What would make the paper valuable is a clean demonstration that receptor/pocket context, not extra capacity or target identity, explains the reported RMSE gains. That demonstration is not in the current text.\n\nThe plug-and-play framing is a good idea, and Table 1's scale-up control is the right instinct: adding parameters without the MTP module does not reproduce the gains. The ablation in Table 3 shows both guidance branches matter relative to the unguided backbone. The paper also ships code and uses the standard MoleculeACE benchmark.\n\nBut several things block a confident reading. The headline numbers are inconsistent: the abstract reports 18.95%, the introduction reports 7.2% over SOTA, and Section 4.2 reports a 19.0% plug-and-play RMSE decrease. No per-dataset results or error bars appear, and the classification result is one dataset. More importantly, the experimental protocol is underspecified. MoleculeACE normally evaluates each target separately; if MTPNet is trained jointly on all 30 targets while baselines are trained per target, the comparison changes two variables at once. Since the protein is constant inside each dataset, F_target can act as a target/dataset embedding that calibrates the different pKi scales, and a one-hot target ID or even a per-target bias might reproduce much of the drop. The 'w/o MTS & MPS' row in Table 3 would then be a jointly trained model denied target identity, which inflates the apparent value of the MTP module. The scale-up control adds parameters but not target identity, so it cannot rule this out. The pocket extraction is also underspecified: Section 3.2c names 'Pocket Prior Extractor (e.g., Cavity Plus)' but does not say which structures were used, how residues were selected, or how pocket residues enter the same embedding space as ESM2.\n\nThis is a paper for researchers working on structure-aware activity cliff prediction and multi-target molecular models. It deserves a serious referee, but the burden is on the authors to add matched baselines (per-target MTPNet, jointly trained backbones, target-ID and non-protein control embeddings), per-dataset results with error bars, and a full pocket extraction specification. As submitted, the central claim is not established; I would not cite the quantitative gains in my own work, but I would bring it to a reading group to discuss the confound.","headline":"Receptor-conditioned activity cliff prediction is a useful idea, but the reported RMSE gains are confounded by joint training and target identity until matched baselines are added.","tokens_in":13232,"tokens_out":2914,"would_cite":false,"duration_ms":34015,"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":"MTPNet conditions molecular representations on the receptor protein—at whole-target and binding-pocket granularity—and reports average activity-cliff error reduced by 18.95% across 30 benchmark datasets.","keywords":["activity cliff prediction","drug discovery","graph neural networks","protein language model","cross-attention","adaptive layer normalization","receptor-ligand interactions","MoleculeACE"],"falsifier":"Rerun the plug-and-play comparison that produces the headline 19% RMSE gain, but replace the pocket prior with the same number of residues sampled uniformly at random from each target's sequence; if RMSE stays near the reported MTPNet values instead of drifting back toward the molecule-only baselines, pocket semantics are not the operative ingredient. A complementary check is to verify that all 30 MoleculeACE targets actually have a 3D structure or cavity annotation available, since the micro-level branch has no defined input for a target without one.","tokens_in":12206,"feed_emoji":"💊","tokens_out":12072,"duration_ms":121420,"temperature":0.7,"pith_summary":"Activity cliffs are the drug designer's trap: two molecules that differ by a single substituent can differ wildly in potency, and standard graph-neural-network models that look only at the molecule routinely miss the jump. This paper claims that the missing information is the protein the molecule binds. It proposes MTPNet, a unified network that conditions the molecular representation on the receptor at two granularities—whole-target semantics via adaptive layer normalization, and binding-pocket details via cross-attention—and reports that on the 30 MoleculeACE datasets this cuts average RMSE by 18.95% on top of mainstream GNN backbones, with the same model reaching AUC 0.924 on CYP3A4 cliff classification. If the claim holds, a single target-aware model could replace the per-target models that dominate activity cliff prediction today, and the attention maps would double as a chemical rationale for each prediction.","feed_headline":"Activity-cliff error falls 19% with receptor context","feed_subtitle":"One model conditions on the target and its binding pocket to beat molecule-only rivals across 30 drug datasets.","key_machinery":"The load-bearing object is the Multi-Grained Target Perception (MTP) module, a stacked attention block that iteratively refines the ligand feature matrix $F_{\\mathrm{mol}}$ under two alternating protein conditions. Macro-level Target Semantic (MTS) guidance compresses the target features by average pooling, sends the compressed vector through a linear Weight Regressor to produce dynamic weights $(\\gamma_i, \\beta_i)$, and uses them as Adaptive Layer Normalization (AdaLN) parameters before a self-attention layer, aligning the molecule with the whole receptor's semantic context. Micro-level Pocket Semantic (MPS) guidance extracts a binding-pocket feature matrix $F_{\\mathrm{pocket}}$ from the target using a Pocket Prior Extractor such as Cavity Plus, then runs scaled dot-product cross-attention with the molecule features as the query and the pocket features as key and value, capturing local receptor-ligand contacts. Alternating the two mechanisms across $L$ stacked layers lets the network internalize both global interaction patterns and fine-grained structural sensitivity, which the paper argues is exactly what an activity cliff—a small molecular change producing a large activity change—requires.","core_discovery":"The paper's central claim is that receptor proteins can be used as conditional guidance, not just as side information, to improve activity cliff prediction; it states this is the first time receptor proteins are employed as guiding information to capture critical interaction details. MTPNet embeds each ligand with Mole-BERT and each target with ESM2, then runs a Multi-Grained Target Perception (MTP) module that alternates Macro-level Target Semantic guidance (average-pooled target features turned by a weight regressor into adaptive layer-normalization parameters gamma and beta) and Micro-level Pocket Semantic guidance (binding-pocket features from a Pocket Prior Extractor such as Cavity Plus, fused with the molecule by cross-attention). Through stacked layers the ligand representation is progressively aligned to the receptor, and the same network serves single-target and multi-target datasets. The paper reports that MTPNet beats all baselines on all 30 MoleculeACE datasets with an average RMSE improvement of 18.95% over mainstream GNN architectures, that the improvement is not explained by parameter scaling (scaling GCN's parameters from 1.11M to 3.17M only reduces RMSE from 0.950 to 0.915, while adding MTP reaches 0.744), and that the learned attention concentrates on functional groups and double or triple bonds, aligning with known chemistry.","pith_inferences":["A decisive control the paper does not run would be to feed the MTP module randomly chosen pocket residues of the same count; if the RMSE gain over molecule-only baselines survives, pocket semantics are contributing little and the gain should be credited to the macro guidance or to added capacity.","If receptor conditioning is genuinely the driver, the same conditioning idea should transfer to other target-dependent endpoints such as matched molecular pairs, solubility, or toxicity panels, and the cheapest probe would be swapping the ESM2 and Mole-BERT embeddings for plain fingerprints to see how much of the 18.95% depends on pretrained representations.","The paper's 'first time' framing implies molecule-only pretraining has plateaued on MoleculeACE; if that is so, the reported numbers set an upper bound on what receptor conditioning alone can buy on this benchmark, and combining it with larger protein language models is the natural next step.","Whether pocket annotations exist for all 30 targets is left open; a testable prediction is that MTPNet's performance degrades smoothly as pocket-residue noise is injected, which would locate the mechanism in pocket fidelity rather than in the attention machinery."],"forward_implications":["One unified network, MTPNet, predicts activity cliffs across all 30 MoleculeACE targets instead of requiring a separate model per binding target, eliminating the per-target redundancy of current approaches.","Adding the MTP module to six mainstream GNN backbones (GCN, GAT, GIN, GraphTrans, MolCLR, Mole-BERT) improves PCC by 11.6%, R-squared by 17.8%, and RMSE by 19.0% on average, and the scale-up control shows that simply enlarging the baselines recovers only a small fraction of that gain.","The same receptor-conditioned model transfers to classification, reaching AUC 0.924 on the CYP3A4 cliff dataset, above Mole-BERT (0.902) and MolCLR (0.896).","MTPNet's attention maps concentrate on functional groups (amino, carbonyl, sulfonyl, carboxyl, halogen) and on double and triple bonds, so the predictions come with a chemical rationale that matches known binding chemistry."],"supporting_citations":[{"why":"Supplies the MoleculeACE benchmark: the 30 activity-cliff datasets on which every central comparison in the paper is run.","marker":"[Van Tilborget al., 2022]"},{"why":"ESM2, the protein language model that produces the target sequence features F_target feeding both MTS and MPS guidance.","marker":"[Linet al., 2023]"},{"why":"Mole-BERT, the molecular graph encoder that embeds the ligands inside MTPNet and serves as the strongest molecule-only baseline to beat.","marker":"[Xiaet al., 2023b]"},{"why":"Cavity Plus, the Pocket Prior Extractor that produces F_pocket for the micro-level guidance; the load-bearing input of the MPS branch.","marker":"[Xuet al., 2018]"},{"why":"Supplies the CYP3A4 dataset on which the classification result (AUC 0.924) is obtained.","marker":"[Raoet al., 2022]"},{"why":"Source of the experimentally measured CYP3A4 actives and inactives underlying the classification evaluation.","marker":"[Veithet al., 2009]"},{"why":"GCN, one of the six mainstream backbones in the plug-and-play ablation that establishes the module's average RMSE gain.","marker":"[Kipf and Welling, 2017]"},{"why":"MolCLR, a pretrained baseline and one of the backbones augmented by the MTP module in the plug-and-play experiments.","marker":"[Wanget al., 2022]"}],"fun_headline_variants":["Receptor-guided model cuts activity-cliff error 19%","Pocket-aware network beats molecule-only cliff predictors by 19%","MTPNet: first to use receptor guidance, 19% lower cliff error","Multi-grained protein semantics slash activity-cliff RMSE 19%","Unified cliff prediction: receptor context delivers 19% gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a trustworthy binding pocket can be found for every target protein and turned into features compatible with the protein encoder; the paper never says whether 3D structures exist for all 30 datasets, how pocket residues are chosen, or how they are aligned with the sequence features—if pocket extraction is noisy or unavailable, the cross-attention is conditioning on missing or approximate inputs, and the reported gains could come from extra parameters or from the whole-target guidance alone.","fun_headline_variants_meta":{"raw":{"variants":["Receptor-guided model cuts activity-cliff error 19%","Pocket-aware network beats molecule-only cliff predictors by 19%","MTPNet: first to use receptor guidance, 19% lower cliff error","Multi-grained protein semantics slash activity-cliff RMSE 19%","Unified cliff prediction: receptor context delivers 19% gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00065,"raw_usage":{"total_tokens":3023,"prompt_tokens":1025,"completion_tokens":1998,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":1906}},"tokens_in":641,"tokens_out":1998,"duration_ms":16633,"temperature":1.0,"reasoning_tokens":1906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:36:03.514000+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the plug-and-play comparison that produces the headline 19% RMSE gain, but replace the pocket prior with the same number of residues sampled uniformly at random from each target's sequence; if RMSE stays near the reported MTPNet values instead of drifting back toward the molecule-only baselines, pocket semantics are not the operative ingredient. A complementary check is to verify that all 30 MoleculeACE targets actually have a 3D structure or cavity annotation available, since the micro-level branch has no defined input for a target without one.","supporting_citations":[{"cited_title":"Evolutionary-scale pre- diction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130,","cited_arxiv_id":null,"evidence_quote":"ESM2, the protein language model that produces the target sequence features F_target feeding both MTS and MPS guidance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cavity Plus, the Pocket Prior Extractor that produces F_pocket for the micro-level guidance; the load-bearing input of the MPS branch."},{"cited_title":"Quantitative evaluation of explainable graph neural networks for molecular property prediction","cited_arxiv_id":null,"evidence_quote":"Supplies the CYP3A4 dataset on which the classification result (AUC 0.924) is obtained."},{"cited_title":"Austin, David G","cited_arxiv_id":null,"evidence_quote":"Source of the experimentally measured CYP3A4 actives and inactives underlying the classification evaluation."},{"cited_title":"Kipf and Max Welling","cited_arxiv_id":null,"evidence_quote":"GCN, one of the six mainstream backbones in the plug-and-play ablation that establishes the module's average RMSE gain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MolCLR, a pretrained baseline and one of the backbones augmented by the MTP module in the plug-and-play experiments."}],"review_version":1}