{"id":"a119f8a5-ed44-49ea-8976-0354924d4120","arxiv_id":"2412.07778","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MIN, a multi-channel neural network with conservation-based residue selection, reports state-of-the-art drug-target interaction prediction on DUD-E and Human.","lead":"MIN is a machine-learning model that predicts whether a drug binds to a target protein by combining sequence, structural contact maps, and evolutionary conservation information. It reports the best accuracy on two standard drug-target interaction benchmarks, DUD-E and Human, and its residue-selection step can highlight regions overlapping known binding pockets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hard threshold in Eq. (2) blocks DTI-gradient flow to the C-Score Predictor, so the claimed joint-training distillation mechanism is not implemented as written.","rationale":"The reader's weakest_assumption identifies the same internal inconsistency I find most load-bearing: the central contribution described in the abstract and Section IV-A is a C-Score Predictor that distills residues using both conservation and the DTI objective, but Eq. (2) makes the DTI gradient path impossible. This is not a disagreement with community consensus; it is a checkable property of the model's computation graph. The reported AUC/RE numbers and the PDBBind case study provide some independent support, but they do not establish the claimed DTI-driven distillation unless the gradient path exists or is replaced by an explicit relaxation. The DUD-E target-fold split is a positive feature, and the ablations are internally consistent, but the Fig. 4 comparison between 'Distill by C-Score Predictor' and 'Distill by C-Score' depends exactly on the missing gradient. A gradient-norm check plus an STE retrain would settle whether the mechanism is real or whether the predictor is merely a conservation-score regressor. I therefore keep the reader's CONDITIONAL verdict: the paper needs code, the missing relaxation, or a corrected claim.","tokens_in":16059,"tokens_out":6009,"duration_ms":58776,"concrete_test":"Implement the architecture exactly as written in an autodiff framework with the full loss in Eq. (11). For a training batch where the threshold mask is active, compute the gradient norm ||∂L_CE/∂θ_CScore||: if it is zero (or at float-noise level) while ||∂L_MSE/∂θ_CScore|| is nonzero, the claimed joint-training path does not exist. Then retrain with a straight-through estimator that routes ∂L_CE through the threshold decision (or with a Gumbel-Sigmoid relaxation) and compare (i) DUD-E AUC and (ii) the Fig. 4 'Distill by C-Score Predictor' curve. If both are unchanged, the DTI objective was not the source of the predictor's behavior; if they change materially, the paper must specify which relaxation was used to obtain the reported numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A defines residue selection by a hard threshold: in Eq. (2), each residue embedding is either kept or replaced by e[MASK] depending on whether the predicted C-Score exceeds r0. This operation is piecewise constant in the predicted score, so the chain-rule factor ∂r̂_i/∂ŷ_s^i is zero almost everywhere. Consequently, ∂L_CE/∂θ_CScore is zero: no gradient from the cross-entropy loss in Eq. (10) can reach the C-Score Predictor through the mask. The only training signal reaching that predictor is the MSE loss in Eq. (1), which fits ground-truth conservation scores. Therefore the paper's repeated claim that the predictor is 'also trained with the DTI objective' (Section IV-A, and the 'Distill by C-Score Predictor' strategy in Figure 4) is not supported. Without a straight-through estimator, Gumbel-Sigmoid, or another relaxation—none of which is described—the predictor cannot be influenced by DTI labels. The reported binding-pocket overlap in the case study would then reflect conservation statistics alone, not a DTI-informed mechanism, and the stated advantage over directly thresholding C-Score in Figure 4 has no evident source.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MIN, a deep-learning framework for drug-target interaction (DTI) classification. MIN has three main components: a representation-learning module that uses a C-Score Predictor to distill protein sequences by masking residues predicted to have low conservation scores, a multi-channel interaction module with structure-agnostic, structure-aware, and extended-mixture channels, and a contrastive-learning objective that aligns sequence and structure representations. Experiments on the DUD-E and Human benchmarks report state-of-the-art AUC, ROC enrichment, precision, and F1 scores, and a case study claims that the selected residues overlap known binding pockets. The central claims are that MIN outperforms all compared baselines and that the C-Score Predictor is trained jointly with the DTI classification objective while also using biological conservation information.","tokens_in":16259,"tokens_out":5415,"duration_ms":49599,"significance":"If the results are correct, MIN would be a strong DTI classifier, particularly because the DUD-E evaluation uses target-fold cross-validation, which is a more realistic and stringent setting than random splits. The multi-channel interaction design and the use of contrastive learning for representation alignment are reasonable and potentially useful ideas. The paper also reports an ablation study that quantifies the contribution of each component, which is valuable. However, the central novelty regarding the C-Score Predictor--that it is trained with the DTI objective to select residues for the DTI task--is not supported by the equations as written, and this directly affects the interpretability claim and the interpretation of the distillation mechanism.","major_comments":[{"comment":"The paper claims that the C-Score Predictor is 'also trained with the DTI objective' and 'influenced by DTI prediction objective.' However, Eq. (2) applies a hard threshold to the predicted C-Score: residues with predicted score above r0 are replaced by a fixed e[MASK] embedding. This operation is piecewise constant in the predicted score, so the derivative of the masked embedding with respect to the C-Score Predictor parameters is zero almost everywhere. Consequently, the cross-entropy loss LCE in Eq. (10) cannot provide any gradient to the C-Score Predictor through Eq. (2). The only training signal reaching the predictor is the MSE loss in Eq. (1), which fits ground-truth conservation scores. Unless a straight-through estimator, a Gumbel-Sigmoid relaxation, or another differentiable approximation is used--none of which is described--the 'Distill by C-Score Predictor' strategy in Figure 4 is simply a learned regressor of conservation scores, and the stated advantage over 'Distill by C-Score' in Section V-D is not explained. Please either describe a concrete differentiable mechanism that allows DTI gradients to reach the C-Score Predictor or revise the claims and the interpretation of the ablation and case study accordingly.","section":"Section IV-A, Eq. (2), Eq. (10), Eq. (11), and Figure 4"},{"comment":"The contact map P_e is built from distances between C-alpha atoms, but the paper does not state where the three-dimensional structures come from. For the DUD-E and Human datasets, are these experimental structures (e.g., from the PDB) or predicted structures (e.g., AlphaFold)? This matters for reproducibility, for the method's applicability to proteins without known structures, and for whether the comparison to baselines that do not use structures is fair. Please specify the structure source and, if predicted structures are used, the prediction pipeline and confidence criteria.","section":"Section III-A(e) and Section V-A"},{"comment":"The case study numbers are internally inconsistent: the text says 'Among 37 residues in the binding pocket, 16 overlap residues' and then 'There are 37 residues in the real binding pocket and among them 9 residues are real binding positions.' If the predicted pocket contains 37 residues and the true pocket contains 37 residues, then '16 overlap' is the intersection size; but the next sentence says the real binding pocket has 9 residues. Please clarify the sizes of the real pocket, the predicted pocket, and the overlap, and also specify how the 'prediction binding residues' in Figure 6 are selected. Without this clarification, the explainability claim is hard to evaluate.","section":"Section V-D, Case Study Q2"}],"minor_comments":[{"comment":"The text says the model masks 'residues with low C-Score,' but Eq. (2) replaces residues with predicted C-Score greater than r0 with e[MASK] and keeps residues with score below or equal to r0. Since a lower C-Score is described as more conserved, the equation keeps conserved residues and masks highly variable ones. Please reword the sentence to avoid confusion.","section":"Section IV-A, Eq. (2)"},{"comment":"The contact threshold s0 in the definition of the contact map is never given a value. Only the conservation threshold r0 = 0.01 is reported. Please report s0 or state that it is a hyperparameter with its tuned value.","section":"Section V-A, Implementation Details"},{"comment":"The contrastive loss is defined with the symbol L used both for the denominator sum and for the loss function itself, making the expression difficult to parse. Please rename the denominator (e.g., S_neg) and make the positive/negative terms unambiguous.","section":"Section IV-C, Eq. (5)"},{"comment":"The dagger (†) indicates statistical significance with p < 0.01, but the statistical test is not described. Please specify the test (e.g., paired t-test across folds or seeds) and whether it is one-sided or two-sided.","section":"Section V-B, Tables I and II"},{"comment":"The Human dataset is evaluated with a random split, which can overestimate performance because the same target may appear in both training and test sets. The authors follow previous work, so this is not a flaw by itself, but it should be acknowledged as a limitation with respect to generalization to unseen targets.","section":"Section V-A, Human dataset"},{"comment":"There is a typo: 'tree-channel interaction network' should be 'three-channel interaction network.'","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The central numerical claims are likely to be reproducible if the implementation matches the described architecture, but the paper's core methodological novelty--the DTI-guided training of the C-Score Predictor--is not supported by the provided equations. This is fixable by adding a differentiable relaxation or by honestly reframing the predictor as a conservation-only module, but as written it undermines the interpretability story. I recommend major revision rather than rejection because the experimental results and the multi-channel design have independent value and the issue is local to one mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my private read. MIN is a real architecture paper, not a throwaway. The new bit is the C-Score Predictor residue-distillation scheme combined with three interaction channels and contrastive alignment. The evaluation is decent: target-fold splits on DUD-E, an external PDBBind case study, and ablations that show all components matter. The gains over HyperPCM are small in AUC (0.983 vs 0.982) but larger in early enrichment, and HyperPCM is the right comparator. I would not ding the citation pattern.\n\nThe dominant soft spot is the one the stress-test note flags, and it lands. Eq. (2) uses a hard threshold: keep residue embedding if predicted C-Score <= r0, else replace with e[MASK]. The kept embeddings do not depend on the C-Score Predictor's parameters, and the mask decision is piecewise constant in the predicted score. So the cross-entropy loss in Eq. (10) cannot backprop through the mask to the C-Score Predictor. The paper's repeated claim that the predictor is \"also trained with the DTI objective\" is not implemented as written. A straight-through estimator, Gumbel-sigmoid, or soft mask would fix it; none is described. As it stands, the case-study residue overlap is better explained by conservation statistics alone, and the claimed advantage over direct C-Score thresholding in Figure 4 has no evident mechanism. This is load-bearing for the distillation contribution and needs to be resolved.\n\nOther issues are smaller. The source of the 3D contact maps is unspecified—experimental structures or predicted? Baseline numbers appear quoted rather than re-run, and no code or data is released. The r0=0.01 threshold retaining 588/1608 residues in the case study is plausible but not analyzed. These are fixable in a revision.\n\nBottom line: the architecture is coherent and the evaluation is mostly sound, but the paper overclaims one mechanism. It deserves a serious referee, and I would accept it for review conditional on the authors fixing or clarifying the gradient-flow issue and releasing code/data. For my own work, I would not cite it as evidence for DTI-informed residue selection until that is fixed; the architecture may still be a useful baseline.","headline":"Genuinely new DTI architecture with a real bug: the C-Score distillation's claimed DTI training signal is blocked by a hard threshold, so the mechanism doesn't do what the paper says.","tokens_in":16847,"tokens_out":3902,"would_cite":false,"duration_ms":37460,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that screening proteins by predicted conservation before matching them to drugs yields drug–target interaction predictions that surpass all compared baselines, with selected residues overlapping known binding pockets.","keywords":["drug-target interaction prediction","protein distillation","C-Score Predictor","protein conservation score","multiple sequence alignment","multi-channel interaction","contrastive learning","binding pocket prediction"],"falsifier":"Check in the released training code whether the interaction-prediction loss can change the parameters of the residue-selection module through the masking step in Eq. (2); if it cannot, the claim that the module is trained with the interaction objective is not realised. Also run a direct comparison between the learned selector and a plain conservation-score cutoff on the same downstream model and the same threshold $r_0$; if AUC and enrichment are statistically indistinguishable, the learned selector adds nothing.","tokens_in":15789,"feed_emoji":"🧬","tokens_out":18420,"duration_ms":128576,"temperature":0.7,"pith_summary":"This paper claims that drug–target interaction (DTI) prediction gets better and more interpretable when the target protein is first stripped down to its most evolutionarily conserved residues, and when the remaining representation is matched to the drug through three complementary channels that use sequence-only, structure-only, and mixed information. The proposed model, MIN, contains a C-Score Predictor that learns per-residue conservation from multiple-sequence alignments and masks non-conserved residues before encoding; a contrastive loss then aligns the sequence and structure views of each drug and target. On the DUD-E benchmark the authors report an AUC of 0.983 with a 0.5% ROC enrichment of 197.741, and on the Human benchmark an AUC of 0.984 with precision 0.983 and F1 0.965, both described as surpassing all compared baselines. In a case study, 16 of the 37 residues MIN selects overlap the true binding pocket of a PDB structure. If these claims hold, conservation-based distillation is a practical prior for locating binding sites and for improving virtual screening at low false-positive rates.","feed_headline":"New model tops drug–target benchmarks by distilling conserved residues","feed_subtitle":"MIN reports AUC 0.983 on DUD-E and 0.984 on Human; kept residues overlap known binding pockets.","key_machinery":"The load-bearing object is the C-Score Predictor: a Transformer encoder $\\varphi_c$ plus a fully connected layer $f_c$ that maps MSA row-average embeddings, residue embeddings, and positional embeddings to a per-residue conservation score $\\hat{y}_s$, trained by mean-squared error against computed conservation scores. Its output is applied through Eq. (2), a hard threshold that replaces every residue whose predicted score exceeds $r_0=0.01$ with a mask embedding, so only conserved residues survive. The second mechanism is the three-channel interaction module, where three MLPs score sequence-level, structure-level, and mixed representations and a feed-forward layer fuses their outputs; contrastive InfoNCE losses align the sequence and structure representations of each drug and target. Together these components are intended to let the model focus on the conserved, likely pocket-forming residues while still exploiting both sequence and 3D contact information.","core_discovery":"MIN's central discovery claim is that treating protein conservation as a learnable screening signal, rather than as a fixed filter, improves DTI classification. The C-Score Predictor takes a protein sequence, its MSA, and positional embeddings, predicts a conservation score for every residue, and keeps only residues whose predicted score falls below the threshold $r_0=0.01$, because a lower score means higher conservation; the distilled sequence is then encoded by a Transformer while the MSA-guided contact graph is encoded by a GCN. Three interaction channels — structure-agnostic (sequence–sequence), structure-aware (graph–graph), and extended-mixture (concatenated sequence and structure) — are fused by a feed-forward layer, and InfoNCE contrastive losses pull the sequence and structure representations of the same entity together. The paper reports that MIN achieves the best AUC and enrichment numbers on DUD-E and the best AUC, precision, and F1 on Human among all baselines tested, and that ablating the C-Score Predictor produces the largest single-component drop in AUC (from 0.983 to 0.968). The case study on PDB entry 2vjx is offered as evidence that the residues MIN keeps are biologically meaningful, with 16 of 37 predicted pocket residues falling inside the annotated pocket.","pith_inferences":["The paper leaves implicit that Eq. (2) is a hard threshold, so the interaction-prediction loss cannot change the residue selector through it; a consequence is that, as stated, the predictor is trained only by the conservation regression, and the claimed joint optimization would require an explicit gradient bypass.","The paper does not say whether the contact maps use experimental or predicted protein structures; a testable extension is to compare the structure-aware channel's contribution under both structure sources.","The binding-pocket overlap is reported for a single structure; an extension is to run the same logit-difference procedure across many PDBbind complexes and report mean overlap, which would separate a generalisable behaviour from a favourable case.","The DUD-E evaluation splits targets into folds by protein family; an inference is that a split on unseen target families would test whether the reported AUC and enrichment transfer to genuinely novel proteins."],"forward_implications":["If MIN's reported numbers hold, conservation-based residue screening gives a practical way to shorten long protein inputs and reduce noise, with the largest single-component AUC gain in the ablations.","A 0.5% ROC enrichment of 197.741 on DUD-E, if reproducible, means true actives would be concentrated near the top of a large decoy set, the operating regime that matters for virtual screening.","The multi-channel fusion claim implies sequence-only and structure-only representations carry complementary signals, so future DTI models can gain from combining them rather than choosing one input modality.","The case-study overlap suggests the same learned representations can double as a binding-pocket localiser without a separate pocket-detection stage.","The ablation ordering, where removing the C-Score Predictor drops AUC from 0.983 to 0.968, implies that the conservation-distillation module is the most influential component in the architecture."],"supporting_citations":[{"why":"Supplies the DUD-E evaluation protocol, the 2D distance-map protein representation, and the DrugVQA baseline that MIN is compared against.","marker":"[15]"},{"why":"Provides the TransformerCPI sequence-based baseline and the label-reversal experiment idea for testing whether a model learns true interaction features.","marker":"[23]"},{"why":"AttentionSiteDTI is the binding-site-aware interpretable baseline whose structure-informed performance MIN compares with.","marker":"[26]"},{"why":"DrugCLIP is the contrastive virtual-screening baseline that MIN must beat on ROC enrichment.","marker":"[33]"},{"why":"HyperPCM is the strongest prior baseline on DUD-E (AUC 0.982) that MIN claims to surpass with statistical significance.","marker":"[34]"},{"why":"Uniclust30 is the protein database used to generate MSAs, which supply both the conservation signal and the MSA-guided contact graph.","marker":"[35]"},{"why":"Defines the conservation score and supports the premise that conserved residues are functionally important.","marker":"[36]"},{"why":"Supports the motivating claim that binding-pocket residues are more conserved than surrounding residues.","marker":"[38]"},{"why":"PDBbind refined set is used to verify that pocket residues have lower (more conserved) C-Scores than non-pocket residues.","marker":"[39]"},{"why":"ConSurf is the tool used to compute the conservation scores in the pocket-versus-non-pocket analysis.","marker":"[40]"}],"fun_headline_variants":["MIN distills conserved residues to top drug–target benchmarks","Distilling conserved residues: MIN tops DTI benchmarks","New network distills key residues for better drug–target prediction","Multi-channel network distills conserved residues for top DTI results","Conserved residue screening lifts drug–target prediction accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the residue-selection module actually learns from the interaction-prediction error, because the paper's masking step is a hard cutoff that blocks learning signals, and the paper does not describe a workaround; if no learning signal gets through, the module only copies conservation scores and the claimed benefit over a simple conservation filter disappears.","fun_headline_variants_meta":{"raw":{"variants":["MIN distills conserved residues to top drug–target benchmarks","Distilling conserved residues: MIN tops DTI benchmarks","New network distills key residues for better drug–target prediction","Multi-channel network distills conserved residues for top DTI results","Conserved residue screening lifts drug–target prediction accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000889,"raw_usage":{"total_tokens":3881,"prompt_tokens":1037,"completion_tokens":2844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":2763}},"tokens_in":653,"tokens_out":2844,"duration_ms":18891,"temperature":1.0,"reasoning_tokens":2763,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:17:16.733535+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check in the released training code whether the interaction-prediction loss can change the parameters of the residue-selection module through the masking step in Eq. (2); if it cannot, the claim that the module is trained with the interaction objective is not realised. Also run a direct comparison between the learned selector and a plain conservation-score cutoff on the same downstream model and the same threshold $r_0$; if AUC and enrichment are statistically indistinguishable, the learned selector adds nothing.","supporting_citations":[{"cited_title":"Predict- ing drug–protein interaction using quasi-visual question answering system,","cited_arxiv_id":null,"evidence_quote":"Supplies the DUD-E evaluation protocol, the 2D distance-map protein representation, and the DrugVQA baseline that MIN is compared against."},{"cited_title":"Transformer- cpi: improving compound–protein interaction predic- tion by sequence-based deep learning with self-attention mechanism and label reversal experiments,","cited_arxiv_id":null,"evidence_quote":"Provides the TransformerCPI sequence-based baseline and the label-reversal experiment idea for testing whether a model learns true interaction features."},{"cited_title":"Attentionsit- edti: an interpretable graph-based model for drug-target interaction prediction using nlp sentence-level relation classification,","cited_arxiv_id":null,"evidence_quote":"AttentionSiteDTI is the binding-site-aware interpretable baseline whose structure-informed performance MIN compares with."},{"cited_title":"Drugclip: Contrasive protein- molecule representation learning for virtual screening,","cited_arxiv_id":null,"evidence_quote":"DrugCLIP is the contrastive virtual-screening baseline that MIN must beat on ROC enrichment."},{"cited_title":"Hyperpcm: Robust task-conditioned modeling of drug–target interactions,","cited_arxiv_id":null,"evidence_quote":"HyperPCM is the strongest prior baseline on DUD-E (AUC 0.982) that MIN claims to surpass with statistical significance."},{"cited_title":"Uniclust databases of clustered and deeply annotated protein sequences and alignments,","cited_arxiv_id":null,"evidence_quote":"Uniclust30 is the protein database used to generate MSAs, which supply both the conservation signal and the MSA-guided contact graph."},{"cited_title":"Predicting functionally important residues from sequence conservation,","cited_arxiv_id":null,"evidence_quote":"Defines the conservation score and supports the premise that conserved residues are functionally important."},{"cited_title":"Conservation and rela- tive importance of residues across protein-protein inter- faces,","cited_arxiv_id":null,"evidence_quote":"Supports the motivating claim that binding-pocket residues are more conserved than surrounding residues."},{"cited_title":"The pdbbind database: Collection of binding affinities for protein- ligand complexes with known three-dimensional struc- tures,","cited_arxiv_id":null,"evidence_quote":"PDBbind refined set is used to verify that pocket residues have lower (more conserved) C-Scores than non-pocket residues."},{"cited_title":"Consurf-db: An accessible repository for the evolution- ary conservation patterns of the majority of pdb proteins,","cited_arxiv_id":null,"evidence_quote":"ConSurf is the tool used to compute the conservation scores in the pocket-versus-non-pocket analysis."}],"review_version":1}