{"id":"f9d53fec-5e8b-4ecc-aca6-7dea66954cf0","arxiv_id":"2411.15208","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"M2oE fuses sequence and graph encodings of peptides with MoE routing and cross-attention, achieving 86.2% AMP classification accuracy and R2=0.951 on AP regression.","lead":"This paper proposes M2oE, a deep learning model that combines peptide sequences with molecular graph structures using sparse mixture-of-experts routing and cross-attention, and tests it on antimicrobial peptide classification and aggregation propensity regression. Against published baselines it reports a small accuracy gain on classification, a tie on regression, and no error bars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"M2oE's own Table III shows it does not surpass the best single-modality baseline on AP (R2 tie, worse MAE/MSE), and the sole AMP gain is a single-seed 1.5-point difference.","rationale":"I read the paper in good faith: the M2oE architecture is plausible, the ablation study is useful, and the authors provide code. However, the central empirical claim is undercut by the paper's own comparison table, not merely by the absence of statistical tests. On AP, SwitchTransformer achieves the same rounded R2 with better MAE and MSE, so the statement that M2oE surpasses baseline results is false for that dataset. On AMP, the only clear win is a 1.5-point accuracy gain from a single split and seed, while other M2oE fusion variants perform below GraphSAGE. This is exactly the kind of claim that needs repeated-seed evidence before it can be believed. I agree with the reader's emphasis on missing error bars, but I would sharpen it: the AP comparison is a direct numerical contradiction, not just an underpowered tie. The appropriate verdict remains conditional rather than reject, because the flaw is fixable by correcting the claim and providing multi-seed statistics; the model may still be competitive. The garbled equations and swapped dataset labels are secondary presentation issues, not the main load-bearing weakness.","tokens_in":6018,"tokens_out":6760,"duration_ms":69526,"concrete_test":"Re-run the complete benchmark (AP and AMP) for M2oE(Parallel), SwitchTransformer, and GraphSAGE using the same 8:1:1 split with at least 10 independent seeds. Report mean plus/minus standard deviation for MAE, MSE, R2, and accuracy, and perform paired bootstrap or Wilcoxon tests for M2oE versus SwitchTransformer on AP and M2oE versus GraphSAGE on AMP. If on AP M2oE's R2 remains tied and its MAE/MSE are not better than SwitchTransformer's, or if on AMP the 1.5-point accuracy advantage does not exceed seed-to-seed variability, then the central claim of surpassing baselines on both tasks is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is that the paper's own Table III contradicts the central claim of surpassing baselines on both datasets. On the AP regression task, M2oE(Parallel) reports R2 = 0.951, exactly equal to the single-modality SwitchTransformer's rounded R2 = 0.951, and M2oE's MAE (3.68E-2) and MSE (2.21E-3) are worse than SwitchTransformer's (3.65E-2 and 2.15E-3). So on AP the proposed multimodal model does not improve over the best sequence baseline; it matches it on R2 and is slightly worse on the other two metrics. On AMP classification, the reported 86.2% accuracy versus GraphSAGE 84.7% is a 1.5-point gain from a single 8:1:1 split with no repeated runs, error bars, or significance test (Section III). The same table shows M2oE(WS) and M2oE(Concat) at 82.0% and 82.4%, both below GraphSAGE, so the advantage is specific to one fusion variant. Thus the headline claim that M2oE surpasses baseline results across both tasks is not supported by the numbers as reported: it is false on AP and unverified on AMP.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes M2oE, a multimodal peptide prediction model that combines a Transformer-based sequence encoder and a GCN/GraphSAGE-style graph encoder with a sparse cross mixture-of-experts (SCMoE) fusion module, cross-attention between modalities, a learnable fusion weight alpha, and auxiliary MoE load-balancing losses. The model is evaluated on two tasks from the Liu et al. benchmark: AMP classification (accuracy) and AP regression (MAE, MSE, R2), using an 8:1:1 split. The authors claim that M2oE outperforms single-modality and mixed baselines on both datasets, and that ablation experiments confirm the contribution of each module.","tokens_in":6314,"tokens_out":4736,"duration_ms":47091,"significance":"The paper has a clear and explicit architecture, provides a public code repository, and evaluates on an external benchmark with held-out test data, which are strengths. If the reported gains are reproducible, the work offers a modest but useful contribution by demonstrating that combining sequence and structure via an MoE router with cross-attention can improve peptide prediction. However, the current empirical support is thin: the key claim of supremacy on AP is contradicted by the paper's own results, and the AMP advantage rests on a single split with no error bars. The architecture is largely a composition of known components (MoE, cross-attention, learnable fusion), so the significance hinges on the evaluation being reliable.","major_comments":[{"comment":"The headline claim that M2oE \"surpasses baseline results\" is not supported by Table III on the AP regression task. M2oE(Parallel) reports R2 = 0.951, which exactly matches SwitchTransformer's rounded R2 = 0.951, and M2oE's MAE (3.68E-2) and MSE (2.21E-3) are worse than SwitchTransformer's (3.65E-2 and 2.15E-3). Thus, the proposed model does not improve over the best sequence baseline on AP; it ties on R2 and is inferior on the other two metrics. The central claim as written is false on this task and must be revised or substantiated with further results.","section":"Section III, Table III"},{"comment":"On AMP, the only reported advantage of M2oE(Parallel) over GraphSAGE is 86.2% vs 84.7%, a 1.5 percentage-point gain from a single 8:1:1 split with no repeated runs, standard deviations, or significance tests (Section II-A). The fragility of this result is underscored by the fact that the other two fusion variants, M2oE(WS) and M2oE(Concat), perform worse than GraphSAGE on the same task. The authors should report multiple random seeds and appropriate significance tests before claiming superiority on AMP.","section":"Section III, Table III"},{"comment":"The ablation study in Table II is also based on single runs. The incremental R2 improvements from adding Cross-Attention and MoE are small (0.942 to 0.949 to 0.951, with the without-CRA variant at 0.949) and are likely within run-to-run variation. Without repeated experiments or confidence intervals, the conclusion that both components are effective is not statistically supported.","section":"Section III, Table II and Table III"},{"comment":"The comparison set in Table III is limited to generic sequence and graph models. No peptide-specific state-of-the-art predictors are included, notably the co-modeling method of Liu et al. [21] that provided the benchmark, or other peptide prediction methods mentioned in the introduction such as AMPEP [22]. Since the paper's contribution is specifically to peptide prediction, the absence of these baselines makes the significance claim difficult to assess.","section":"Section I and Section III"}],"minor_comments":[{"comment":"The sentence \"sequence and graph models are single-modality frameworks evaluated on classification (AP) and regression (AMP) datasets\" is reversed: per Table I, AMP is the classification task and AP is the regression task. This should be corrected.","section":"Section III, first paragraph"},{"comment":"Equation (2) is self-referential: the router formula uses alpha_j before it is defined on the following line, and the notation contains artifacts such as \"T opk\" and \"Sof tplus\". The formula needs to be rewritten so that all quantities are defined before use.","section":"Section II-C, Equation (2)"},{"comment":"The text says \"it can be represented as follows:\" and is immediately followed by \"where Fseq, Fgra denote features ... and dk is the scaling factor respectively,\" but no displayed equation appears before that sentence; the actual cross-attention expression is only implied in Equation (3). This formatting issue should be fixed.","section":"Section II-C, Cross-Attention paragraph"},{"comment":"The abbreviation \"M2oE(WS)\" is not defined; the authors should spell out that it refers to weighted-sum fusion.","section":"Section III, Table III"},{"comment":"The manuscript does not report training hyperparameters (number of experts C, top-k value, learning rate, batch size, epochs, hidden dimensions, random seed, or alpha initialization). The code link is appreciated, but the core experimental setup should be stated in the paper to allow independent reproduction.","section":"Section II and Section III"}],"recommendation":"major_revision","confidential_remarks":"The paper is borderline for this venue. The empirical evaluation as reported does not back the headline claim: the AP result is a tie at best, and the AMP result is a single-split 1.5-point gain. The authors should either provide multi-seed results with significance tests and update the claims to match (including narrowing the claim to AMP if AP does not improve), or the paper will likely not meet the bar. The architectural novelty is incremental; the contribution is more about careful integration of existing techniques in a peptide setting, so the evaluation rigor is the deciding factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read of M2oE. The architecture is a genuine but small novelty: sparse MoE routing combined with cross-attention between sequence and graph encoders for peptide tasks. Code is public, which is good. The ablation shows each added component nudges AP R2 up from 0.942 to 0.951, so the modules do something on that metric.\n\nThe problem is the headline. Table III shows the model does not beat the best single-modality baseline on AP: R2 ties SwitchTransformer at 0.951, and MAE/MSE are slightly worse. On AMP, the gain over GraphSAGE is 1.5 accuracy points from a single 8:1:1 split with no error bars or repeats, and the other two fusion variants score lower than GraphSAGE. So the claim of 'surpassing baseline results' holds only for one variant on one metric, and even that is statistically unverified.\n\nThere are also presentation issues: Section III swaps the task labels (calls AMP regression and AP classification), and some equations are garbled. The learnable fusion weight alpha is standard, not a problem.\n\nNet: decent idea, weak evidence, overstated claims. A revision with repeated runs, error bars, and corrected text could make a useful benchmark contribution. As it stands, I would not cite it for the performance claims, but the architecture might interest people working on peptide multi-modal fusion. I would only send it to review if the venue accepts major revisions; otherwise desk reject with an invitation to resubmit after statistical rigor is added.","headline":"A plausible architecture whose own results table undercuts its main claim: no AP gain over the best baseline, and a single-seed 1.5-point AMP edge.","tokens_in":6808,"tokens_out":2204,"would_cite":false,"duration_ms":22761,"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":"M2oE fuses peptide sequence and graph structure with sparse mixture-of-experts and cross-attention, reporting 86.2% accuracy on antimicrobial peptide classification and an R-squared of 0.951 on aggregation propensity.","keywords":["peptide property prediction","multimodal learning","mixture of experts","cross-attention","antimicrobial peptides","aggregation propensity","graph neural networks"],"falsifier":"Re-run M2oE and the best baselines on the same datasets with many random seeds and multiple 8:1:1 splits and compare the distributions; if the AMP accuracy advantage over GraphSAGE is smaller than the run-to-run spread, the central claim of superiority is not supported.","tokens_in":5826,"feed_emoji":"🧬","tokens_out":10944,"duration_ms":91862,"temperature":0.7,"pith_summary":"The paper proposes M2oE, a model that predicts peptide properties by using both the amino-acid sequence and the molecular graph structure at once. The authors argue that single-modality models fail when the dataset carries little information in that modality, and that a sparse mixture-of-experts layer with cross-attention balances the two views. They report 86.2% accuracy on antimicrobial peptide classification and $R^2=0.951$ with low error on aggregation propensity regression, beating or matching the best unimodal baselines. If those numbers are stable, the work provides a practical recipe for multimodal peptide property prediction.","feed_headline":"Sequence-plus-graph peptide model hits 86.2% accuracy","feed_subtitle":"It also ties the best regression score of 0.951 on aggregation propensity, beating unimodal baselines on both tasks.","key_machinery":"The key object is the Sparse Cross Mixture of Experts (SCMoE) fusion module, which routes tokens from the sequence and graph encoders to shared experts using a learnable gating matrix with noisy top-$k$ selection, and then aligns the two modalities through cross-attention that swaps query vectors between the sequence and graph feature spaces. The gating noise gives low-ranked tokens a chance to be assigned to experts, and two auxiliary losses (load and importance) prevent expert collapse. A learnable scalar $\\alpha$ in the final MLP weights the sequence and graph contributions, so the fusion adapts to the data distribution rather than using fixed weights.","core_discovery":"M2oE jointly encodes a peptide's primary sequence with a Transformer and its bond structure with a graph encoder, then fuses both through a sparse cross mixture-of-experts (SCMoE) module that includes cross-attention between the two modalities and a learnable weight $\\alpha$ that decides how much each modality contributes. The central claim is that this collaboration outperforms any single modality: on aggregation propensity it reports $R^2 = 0.951$ with MAE $3.68\\times10^{-2}$ and MSE $2.21\\times10^{-3}$, and on antimicrobial peptide classification it reports 86.2% accuracy, compared with 84.7% for the best graph-only baseline and 0.951 for the best sequence-only baseline. The paper also claims, based on ablations, that both the cross-attention and the mixture-of-experts routing are necessary for the gain.","pith_inferences":["If the single-split results are stable, we would expect M2oE's AMP margin to persist across random seeds; otherwise the 1.5-point gain may not be reproducible.","The cross-attention weights could be inspected to see which amino acids and graph beads are aligned, offering a possible interpretability tool for peptide design.","The same SCMoE design could be applied to other molecular systems where sequence and 3D structure both matter, such as protein-protein interfaces or RNA property prediction."],"forward_implications":["Single-modality models underperform on peptide datasets where their modality is weak; M2oE's routed fusion gives more consistent results across classification and regression.","Removing the cross-attention or the mixture-of-experts routing degrades performance on the AP task, so both components are credited with the improvement.","The learnable $\\alpha$ lets the model adapt the sequence/graph balance to the dataset, which should transfer to other molecular property tasks.","The reported 86.2% AMP accuracy is 1.5 percentage points above the best graph baseline, while AP $R^2$ matches the best sequence baseline at 0.951.","The paper suggests the multimodal expert design can be extended to peptide generation tasks."],"supporting_citations":[{"why":"Provides the benchmark datasets for AMP and AP and the fixed 8:1:1 train/validation/test split used in every experiment.","marker":"[21]"},{"why":"Supplies the graph convolutional encoder that forms the structural modality for peptide graphs.","marker":"[12]"},{"why":"The SwitchTransformer sequence baseline that M2oE must match or exceed on the AP regression task.","marker":"[16]"},{"why":"The GMoE graph baseline that M2oE must beat on AMP classification.","marker":"[15]"},{"why":"The Git-MOL multimodal model whose cross-attention design motivates the SCMoE interaction mechanism.","marker":"[14]"},{"why":"Introduces noisy top-k expert routing, which the paper adapts so tokens ranked after the top-k still get assigned.","marker":"[23]"}],"fun_headline_variants":["Multimodal peptide model beats single-modality baselines","Sequence plus structure hits 86.2% peptide accuracy","Expert fusion with cross-attention lifts peptide scores","Two modalities outperform one in peptide prediction","M2oE: collaborative expert model boosts peptide tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage relies on the assumption that a single fixed 8:1:1 data split and one random seed produce stable evaluation numbers, so the 1.5-point accuracy gain is real rather than a fluctuation.","fun_headline_variants_meta":{"raw":{"variants":["Multimodal peptide model beats single-modality baselines","Sequence plus structure hits 86.2% peptide accuracy","Expert fusion with cross-attention lifts peptide scores","Two modalities outperform one in peptide prediction","M2oE: collaborative expert model boosts peptide tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1250,"prompt_tokens":871,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":304}},"tokens_in":487,"tokens_out":379,"duration_ms":4060,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:45:07.400596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run M2oE and the best baselines on the same datasets with many random seeds and multiple 8:1:1 splits and compare the distributions; if the AMP accuracy advantage over GraphSAGE is smaller than the run-to-run spread, the central claim of superiority is not supported.","supporting_citations":[{"cited_title":"Co-modeling the sequential and graphical routes for peptide representation learning,","cited_arxiv_id":null,"evidence_quote":"Provides the benchmark datasets for AMP and AP and the fixed 8:1:1 train/validation/test split used in every experiment."},{"cited_title":"Graph mixture of experts: Learning on large-scale graphs with explicit diversity modeling,","cited_arxiv_id":null,"evidence_quote":"The GMoE graph baseline that M2oE must beat on AMP classification."},{"cited_title":"Git-mol: A multi-modal large language model for molecular science with graph, image, and text,","cited_arxiv_id":null,"evidence_quote":"The Git-MOL multimodal model whose cross-attention design motivates the SCMoE interaction mechanism."}],"review_version":1}