{"id":"b16b289b-d83f-49c8-9ece-1d747c4ddfc7","arxiv_id":"2509.06465","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CAME-AB fuses five protein sequence, structural, and chemical encodings with adaptive fusion, mixture-of-experts, and contrastive learning to improve antibody epitope class prediction.","lead":"This paper introduces CAME-AB, a deep learning model that predicts which known virus region an antibody binds to using only the antibody's sequence. It merges five protein representations with several training enhancements and reports better F1 and MCC scores than four comparison models on a curated dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) conditions the fused representation on the ground-truth label, and no inference-time rule is given; the reported gains may reflect label leakage rather than multimodal learning.","rationale":"The reader's weakest assumption was baseline comparability in Table 3. That concern is real but downstream: even with perfectly matched baselines, the proposed method's own forward pass is not specified at inference. Eq. (4) is the only description of the fusion step, and it conditions on ground truth. Because the classification metrics are the entire empirical support for the abstract and conclusion, this label-conditioning ambiguity is the most load-bearing issue in the paper. I do not attribute intent; a missing inference description or an implementation that drops γ at test time would resolve it, and the proposed test would reveal which is the case. The ablation tables do not settle the issue, since every row removes a component other than the class-aware γ. The Transformer, MoE, contrastive loss, and SWA components are established techniques and could plausibly form a working predictor; the single point that must be settled before the central claim can be assessed is how γ is handled when labels are unavailable. I therefore recommend REJECT for the current version, with the understanding that a clear test-time protocol and re-run without label conditioning could lead to a revised conditional acceptance.","tokens_in":11883,"tokens_out":5326,"duration_ms":54732,"concrete_test":"Inspect the model forward pass in the anonymous repository: does the fusion module receive y_i as an input at test time? Then re-run the full pipeline with γ_m^(y_i) removed (set γ = 1) and with γ computed from the model's own predicted labels, keeping splits, seeds, and hyperparameters identical. If either variant's test F1 or MCC falls below the Full Model row of Table 2 by more than the reported standard deviation, the headline gains come from training-time label conditioning, not from multimodal fusion. A second check is to train the same model with shuffled labels used only for γ; if test metrics remain near the reported level, label leakage is confirmed.","verdict_should_be":"REJECT","load_bearing_attack":"The central empirical claim is that CAME-AB achieves the best F1, MCC, precision, and recall in Table 3. The architecture as written makes this claim unverifiable. Eq. (4) defines F_fused = Σ_m α_m β_m^(i) γ_m^(y_i) 𝐼̃F^(m), where γ_m^(y_i) = Embed(y_i)[m] is explicitly a function of the ground-truth epitope class y_i. F_fused is then passed through the Transformer encoder (Eqs. 6–8), the MoE module, and the classifier (Eq. 12). During training, the label is therefore an input feature: the classifier can learn to read the class-specific weighting and ignore the antibody-derived modalities, potentially inflating all reported metrics. The paper never states how γ is computed at inference. If γ is omitted at test time, the train and inference distributions differ, so the reported test numbers do not describe the deployed model. If γ is computed from the model's own predicted label, the evaluation is circular. If γ is intended only for an auxiliary loss, then Eq. (4) and Section 3.2.2 contradict the stated forward path. None of these readings supports the superiority claim as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAME-AB, a multimodal deep-learning framework for antibody binding site prediction. The model combines five feature modalities derived from antibody VH/CDR sequences: one-hot encoding, BLOSUM profiles, ESMC protein-language-model embeddings, ESMC structure-output features, and GCN-refined biochemical graph embeddings. These are projected into a shared space and fused by an adaptive modality fusion module that uses global, sample-specific, and class-aware weights. The fused representation is processed by a Transformer encoder and a Mixture-of-Experts module, with supervised contrastive learning, focal loss, expert diversity regularization, and stochastic weight averaging. The authors evaluate on a curated antibody-antigen dataset with seven epitope classes and report F1, MCC, precision, recall, and AUC-ROC, claiming state-of-the-art performance. Ablations assess the contribution of each modality and architectural component.","tokens_in":12153,"tokens_out":3910,"duration_ms":33061,"significance":"If the central claims were fully supported, the paper would contribute a useful multimodal architecture for antibody-antigen specificity prediction, and the emphasis on combining pretrained embeddings, structural proxies, and graph features is timely. The availability of code and the use of multiple metrics are also strengths. However, the evaluation as written does not establish the claimed superiority: the class-aware fusion term depends on the ground-truth label with no stated inference-time rule, the headline 'consistently outperforms' is contradicted by the reported AUC-ROC of ME-ACP, and several ablation descriptions do not match the tables. These are load-bearing issues for the paper's main empirical conclusion, not merely presentation problems.","major_comments":[{"comment":"The fused representation is defined as F_fused = sum_m alpha_m * beta_m^(i) * gamma_m^(y_i) * tilde F^(m), where gamma_m^(y_i) = Embed(y_i)[m] is explicitly a function of the ground-truth epitope class y_i. This makes the label an input feature in the training forward path, and the paper never specifies how gamma is computed at inference. If gamma is omitted at test time, the training and inference distributions differ and the reported test numbers do not describe the deployed model; if gamma is derived from the model's own predicted label, the evaluation is circular. The manuscript must state the inference-time computation of gamma, or remove the class-aware conditioning from the forward path. As written, the reported gains cannot be attributed to multimodal learning rather than label leakage.","section":"Section 3.2.2, Eq. (4)"},{"comment":"The abstract and Section 4.3 claim CAME-AB 'consistently outperforms' strong baselines, but Table 3 shows ME-ACP achieving AUC-ROC 0.9587 versus CAME-AB's 0.9351. The text dismisses this as 'potential overfitting or miscalibration' without any supporting evidence, such as calibration curves or analysis on validation subsets. The claim should either be weakened to a partial superiority statement or be backed by an explicit calibration comparison. As it stands, the central empirical claim is not supported by the reported numbers.","section":"Section 4.3, Table 3, and Abstract"},{"comment":"The ablation text and Table 2 contain internal inconsistencies. The text states that disabling AMF reduces F1 by 0.92% and MCC by 0.0144, but Table 2 shows a drop of 0.0058 in F1 (0.8185 to 0.8127) and 0.0101 in MCC (0.7134 to 0.7033). The text also states that removing SWA 'leads to the best AUC', yet Table 2 shows that removing the MoE gives AUC 0.9440, which is higher than the w/o SWA value of 0.9397. Additionally, 'removing the MoE block significantly reduces recall' is an overstatement for a drop of 0.0073. These discrepancies need to be corrected before the ablation conclusions can be assessed.","section":"Section 4.2.2, Table 2, and surrounding text"},{"comment":"The comparative evaluation is not fully specified. The paper does not report the dataset size, class distribution, or the exact protocol used for the baselines ABS, ME-ACP, xDeep-AcPEP, and PreAlgPro. In particular, it is not stated whether these baselines were retrained on the same 80/10/10 split of the same curated dataset or whether their reported numbers were imported from prior publications. Without this information, Table 3 cannot be interpreted as a fair comparison, and the claim of state-of-the-art performance is not verifiable.","section":"Section 4.1.1 and Section 4.3"},{"comment":"The total loss in Eq. (17) includes the term lambda_aux * L_modal, but L_modal is never defined anywhere in the manuscript or supplementary material. The text refers to 'auxiliary modality losses' but provides no formula or description. This missing definition prevents reproduction of the training objective and should be supplied or the term removed.","section":"Section 3.2.4, Eq. (17)"}],"minor_comments":[{"comment":"The title promises 'antibody binding site prediction', but the experiments predict one of seven epitope classes (S:NTD, S:RBD, S:S2, Others, HA:Head, HA:Stem, HIV) from antibody VH/CDR sequences. This is better described as antibody specificity or epitope class prediction; the terminology should be aligned to avoid overclaiming.","section":"Title and Section 3.1"},{"comment":"The cosine similarity function is written as 'com_sim' in the contrastive loss; it should be 'cos_sim' for consistency with Eq. (11) and the surrounding text.","section":"Section 3.2.4, Eq. (14)"},{"comment":"The grid search over embedding dimension, transformer depth, attention heads, number of experts, contrastive temperature, and loss weights is mentioned, but the final selected values are not reported. Reporting these hyperparameters is necessary for reproducibility.","section":"Section 4.1.2"},{"comment":"The t-SNE plot lacks axis labels that identify which colors correspond to which classes. The caption says the model exhibits class separability, but the reader cannot map the clusters to the seven epitope classes without a legend.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim rests on a label-conditioned fusion term whose inference-time behavior is unspecified, and on a baseline comparison whose protocol is not documented. These issues are fixable in principle, but they require substantive corrections to the method description and evaluation rather than cosmetic changes. I do not see evidence of intentional misconduct, but the manuscript as submitted is not yet verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline is simple: the paper's forward pass, as written, uses the ground-truth epitope class as an input feature, and no inference-time rule is given. Eq. (4) defines the fused representation as a weighted sum with γ_m^(y_i) = Embed(y_i)[m], a learned embedding of the true label. That fused vector then goes through the transformer, MoE, and classifier. During training the classifier can read the label from the input; at test time there is no equivalent. If γ is dropped, the input distribution changes; if it is computed from a predicted label, the evaluation is circular. Neither reading supports the reported F1 and MCC. This is a load-bearing flaw, not a style issue.\n\nThere are real things to credit. Five modalities (one-hot, BLOSUM, ESM-C, structure-aware features, GCN graph) is a sensible set, and combining them with adaptive fusion, MoE, supervised contrastive loss, and SWA is a plausible engineering stack. The writing is mostly clear, and the authors share an anonymous code link. The ablation design is the right idea.\n\nBut the evaluation has more problems. The abstract says 'consistently outperforms,' yet ME-ACP beats CAME-AB on AUC-ROC (0.9587 vs 0.9351). The text calls this 'overfitting or miscalibration' without evidence. The ablation table and surrounding text disagree on numbers: F1 drop for removing AMF is about 0.6 points, not 0.92; MCC drop is about 0.01, not 0.0144; and removing contrastive learning or MoE gives higher AUC than the full model, contradicting 'slightly boosts all metrics.' The dataset is described only by pipeline (80/10/10, CD-HIT at 90%, 80% refinement), with no size or class distribution. And the title says 'binding site prediction' but the task is classifying an antibody into one of seven epitope classes—either a misnomer or a much narrower problem.\n\nIn short, the paper is not ready for publication in its current form. The central numbers cannot be interpreted, and the comparison is not honest as written. I'd desk-reject rather than send to review unless the authors can clarify gamma, rerun the experiments without label conditioning (or with a properly defined inference procedure), correct the ablation reporting, and release the dataset. If they fix those, there is a useful contribution here for someone doing antibody epitope classification. But as it stands, I would not cite it.","headline":"The forward pass as written leaks the ground-truth label into the fused representation, and the missing inference-time rule makes the reported numbers uninterpretable.","tokens_in":12686,"tokens_out":4382,"would_cite":false,"duration_ms":36611,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CAME-AB fuses five modalities of antibody sequence information and reports the best F1, precision, recall, and MCC among four baselines on a seven-class epitope dataset.","keywords":["antibody binding site prediction","antibody-antigen interaction","multimodal learning","adaptive modality fusion","mixture-of-experts","supervised contrastive learning","protein language model","epitope classification"],"falsifier":"Re-running the four baselines on the exact train/validation/test split used for CAME-AB, with the same preprocessing and metric definitions, and finding that any baseline reaches or exceeds the reported F1 of 0.8185 or MCC of 0.7134 would falsify the superiority claim.","tokens_in":11680,"feed_emoji":"🧬","tokens_out":11031,"duration_ms":82923,"temperature":0.7,"pith_summary":"The paper sets out to show that antibody binding-site prediction is improved by combining several complementary descriptions of the antibody heavy chain: raw residue identity, evolutionary substitution scores, protein-language-model embeddings, predicted structural context, and biochemical similarity graphs. It describes CAME-AB, a network that projects these five views into a shared space, weights them adaptively per sample and per epitope class, and then refines the fused representation with a Transformer and a mixture-of-experts layer. A supervised contrastive loss shapes the latent space so antibodies against the same epitope cluster together. Evaluated on a curated seven-class dataset of antibody sequences, CAME-AB reports top F1 (0.8185), precision (0.8227), recall (0.8250), and MCC (0.7134) against four baselines, with one baseline slightly higher on AUC-ROC. The paper's claim is that the multimodality and the adaptive weighting, not any single feature, are what make the difference.","feed_headline":"Five-modality antibody model tops epitope prediction on F1 and MCC","feed_subtitle":"Fusing five feature types lifts epitope prediction to F1 0.8185 and MCC 0.7134.","key_machinery":"The load-bearing mechanism is the adaptive modality fusion module, which computes the fused representation as a weighted sum over modalities, with three kinds of weights: a learnable global importance per modality, a sample-specific gating weight derived from the pooled embedding, and a class-aware weight looked up from the ground-truth epitope class. This fused tensor goes through a two-layer Pre-LN Transformer encoder, whose mean-pooled output is passed to a mixture-of-experts layer with a gating network and a diversity regularizer that penalizes cosine similarity between experts. The final embedding feeds a classifier trained with focal loss plus a supervised contrastive loss, and stochastic weight averaging smooths the last training phase. The design's intent is to let the network keep all five views, but reweight them according to what the input and the target class require.","core_discovery":"On the paper's own terms, the discovery is that a multimodal architecture with adaptive, class-aware fusion can outperform established sequence-based predictors of antibody specificity. The model treats the problem as classification into seven epitope classes (spike subdomains, hemagglutinin head/stem, HIV, and others) from the VH region and the three CDR loops. The authors report that every modality contributes: dropping the ESMC language-model embeddings costs the most F1, removing BLOSUM collapses recall by 17.5 points, and the GCN and one-hot features give smaller but consistent gains. Architecturally, removing the adaptive fusion, the MoE block, contrastive learning, or stochastic weight averaging each lowers F1 or MCC, supporting the claim that the components are complementary. The authors conclude that integrated multimodal representation is a promising direction for immunoinformatics.","pith_inferences":["Editorial inference: the paper's title says binding-site prediction, but the experiments classify whole antibody sequences into seven epitope classes; whether the same fusion machinery improves per-residue paratope/epitope annotation remains untested.","Editorial inference: Section 4.1.1 does not state whether the four baselines were retrained on the paper's 80/10/10 split or their reported numbers imported from earlier papers, so the reader cannot yet check the fairness of the comparison.","Editorial inference: because the class-aware fusion weight is a lookup table over the seven training classes, a model trained this way may not transfer to antibodies against epitopes unseen at training time; a leave-one-epitope-class-out experiment would test that.","Editorial inference: the same adaptive fusion and contrastive framework is label-agnostic in design and could be attached to other antibody endpoints, such as neutralization potency or polyreactivity, by only replacing the classification head."],"forward_implications":["If the results hold, practitioners should include protein-language-model embeddings first: removing them costs the largest F1 drop (about 1.9 percentage points) among the five modalities.","The 17.5-point recall drop when BLOSUM is removed implies that evolutionary substitution profiles are what keep the model sensitive across epitope classes, so they should not be treated as redundant with language-model embeddings.","Each architectural addition—adaptive fusion, MoE, contrastive learning, and stochastic weight averaging—improves at least one of F1 or MCC, so the gains are attributed to the combination rather than to any single trick.","CAME-AB's F1 of 0.8185 and MCC of 0.7134 suggest that class imbalance, which is common in epitope data, is handled well enough by focal loss and contrastive shaping to outperform the four compared baselines on the balanced metrics."],"supporting_citations":[{"why":"Supplies the curated dataset construction pipeline that the paper's 80/10/10 split follows, and provides one of the four baseline result sets.","marker":"[31]"},{"why":"CD-HIT clustering at 90% identity is used to de-duplicate sequences and prevent leakage between train, validation, and test splits.","marker":"[32]"},{"why":"One of the four compared baselines; its reported F1, AUC-ROC, and MCC are the numbers CAME-AB must beat.","marker":"[34]"},{"why":"One of the four compared baselines; provides the baseline precision/recall/F1 values in Table 3.","marker":"[35]"},{"why":"One of the four compared baselines; supplies the fourth row of comparison metrics.","marker":"[36]"},{"why":"Motivates and supplies the ESM-family pretrained protein language model embeddings used as the ESMC modality, which the ablation shows is the most important single feature.","marker":"[17]"},{"why":"Defines the graph convolutional network layer used to refine the biochemical similarity graph into the fifth modality.","marker":"[23]"},{"why":"Defines the sparsely-gated mixture-of-experts layer and gating formulation adopted in the model's MoE module.","marker":"[26]"},{"why":"Defines the supervised contrastive loss used to shape the latent space and enforce intra-class compactness.","marker":"[28]"},{"why":"Defines the focal loss used as the primary classification objective to handle class imbalance.","marker":"[30]"}],"fun_headline_variants":["Five-modality fusion tops antibody epitope prediction benchmarks","Multimodal model with MoE and contrastive loss wins on F1 and MCC","Adaptive feature weighting sharpens antibody binding site prediction","CAME-AB: cross-modal attention elevates epitope prediction accuracy","Dropping any modality hurts, so fusion lifts epitope scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes the four baseline scores in Table 3 were measured under the same CD-HIT-clustered 80/10/10 split of the same seven-class curated dataset as CAME-AB, with the same evaluation protocol.","fun_headline_variants_meta":{"raw":{"variants":["Five-modality fusion tops antibody epitope prediction benchmarks","Multimodal model with MoE and contrastive loss wins on F1 and MCC","Adaptive feature weighting sharpens antibody binding site prediction","CAME-AB: cross-modal attention elevates epitope prediction accuracy","Dropping any modality hurts, so fusion lifts epitope scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1413,"prompt_tokens":1009,"completion_tokens":404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":314}},"tokens_in":625,"tokens_out":404,"duration_ms":4128,"temperature":1.0,"reasoning_tokens":314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:15:37.559301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the four baselines on the exact train/validation/test split used for CAME-AB, with the same preprocessing and metric definitions, and finding that any baseline reaches or exceeds the reported F1 of 0.8185 or MCC of 0.7134 would falsify the superiority claim.","supporting_citations":[{"cited_title":"An explainable language model for antibody specificity prediction using curated influenza hemagglutinin antibodies.Immunity, 57(10):2453–2465, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the curated dataset construction pipeline that the paper's 80/10/10 split follows, and provides one of the four baseline result sets."},{"cited_title":"Cd-hit: accelerated for clustering the next-generation sequencing data.Bioinformatics, 28(23):3150–3152, 2012","cited_arxiv_id":null,"evidence_quote":"CD-HIT clustering at 90% identity is used to de-duplicate sequences and prevent leakage between train, validation, and test splits."},{"cited_title":"Me-acp: multi-view neural networks with ensemble model for identification of anticancer peptides","cited_arxiv_id":null,"evidence_quote":"One of the four compared baselines; its reported F1, AUC-ROC, and MCC are the numbers CAME-AB must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the four compared baselines; provides the baseline precision/recall/F1 values in Table 3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the four compared baselines; supplies the fourth row of comparison metrics."},{"cited_title":"Single-residue linear and conformational b cell epitopes prediction using random and esm-2 based projections.Briefings in Bioinformatics, 25(2):bbae084, 2024","cited_arxiv_id":null,"evidence_quote":"Motivates and supplies the ESM-family pretrained protein language model embeddings used as the ESMC modality, which the ablation shows is the most important single feature."}],"review_version":1}