acceptodds
Under review as a conference paper at ICLR 2027

GOA-PGN: Global Orientation-Aligned Physics Graph Network for Quantitative Antimicrobial Peptide Discovery

Abstract

The rise of antibiotic-resistant bacteria drives urgent interest in antimicrobial peptides (AMPs). However, current computational discovery methods are hindered by single-modality representations that fail to capture the complex spatial mechanisms of membrane disruption. In this work, we propose a global orientation-aligned physics graph network (GOA-PGN) for predicting the minimum inhibitory concentration (MIC) of AMPs. GOA-PGN operates through two parallel paths to integrate localized sequence motifs with structural biophysical properties. One path leverages a peptide-tuned language model and partial convolutions to extract fine-grained sequential patterns. Simultaneously, the other path is used to construct coordinate-free topological graphs on the basis of the cosine similarity of side chain orientations to bypass static conformational noise. We integrate these dual representations via an attention mechanism to comprehensively capture global spatial amphipathicity, which is crucial for membrane disruption. Furthermore, the network incorporates a deep imbalanced regression (DIR) paradigm to mitigate the extreme label skewness prevalent in real-world bioactivity data. This paradigm uses dynamically weighted multitask loss to jointly predict the MIC and auxiliary physicochemical properties and employs a two-stage optimization strategy of adapter warm-up and subsequent joint fine-tuning to explicitly prioritize highly potent candidates. When evaluated on the DBAASP and CAMP datasets under the sequence homology constraints established by the quantitative mapping of antibacterial peptide (QMAP) benchmark, GOA-PGN demonstrates robust predictive accuracy across multiple bacterial strains. The code is publicly available at https://anonymous.4open.science/r/AMP-F316.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antimicrobial Peptides, Physics Graph Network, Deep Imbalanced Regression

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_iZoPOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1(−2)
Soundness
2
Read review
Paper summary

GOA-PGN predicts log-transformed minimum inhibitory concentration (MIC) for antimicrobial peptides. A peptide-tuned ESM-2 supplies residue embeddings, partial 1D convolutions refine local sequence patterns, and a graph aggregates these features using positive cosine similarities between ESMFold-derived side-chain orientations. Attention pooling and SE-equipped heads predict MIC, potency class, charge, and hydropathy. Training uses a frozen-backbone warm-up followed by joint fine-tuning and threshold-based upweighting of potent examples. Evaluation covers three DBAASP bacterial species and CAMP E. coli, with five homology-aware splits. The model improves several regression metrics, particularly on DBAASP E. coli, but ranking gains are not consistent across all high-potency subsets.

Critical strengths
  1. Quantitative MIC estimation and potent-peptide ranking are more useful for screening than AMP classification, and the QMAP split below 60% identity, five split evaluations and separate high-potency reporting guard against overly favorable random splits.
  2. GOA-PGN beats the closest baselines on DBAASP E. coli (Spearman 0.549 and RMSE 0.646 versus LLAMP’s 0.514 and 0.679) and on the high-potency subset (0.313 and 0.697 versus MBC-Attention’s 0.277 and 0.727).
Critical weaknesses
  1. The sequence and structure fusion idea is not new, since ANIA, AMPredictor and GGN-ABPMIC already combine local, global and graph information for MIC prediction, the last two are missing from the comparison tables, and without matched comparisons against other graph designs the orientation-only adjacency is not shown to be an advance.
  2. The central claim is under-validated, because cosine similarities still depend on a predicted conformation, the linked models.py uses exp(-dist / 3.0) * relu(cos_sim) rather than the distance-free Eqs. 4 to 7, and the generalization claims do not hold uniformly (LLAMP has the higher Spearman on high-potency CAMP, 0.303 versus 0.288), so I recommend 1: clear rejection.
Questions for authors
  1. What is the essential advance over AMPredictor and GGN-ABPMIC, can they be compared under identical cleaning, MIC transformation, homology splits and tuning budgets, and which existing controls isolate orientation-based adjacency from simply adding graph aggregation to the same backbone?
  2. Which code and configuration produced each main and appendix result, how do the released exp(-dist / 3.0), the distance-free Eqs. 4 to 7 and Appendix D’s constant 8.0 reconcile, and was model selection done without access to test results?
  3. What direct evidence supports robustness to conformational noise or membrane-relevant amphipathicity, such as perturbation, structure-confidence, alternative-graph or orientation-shuffling analyses, and should invariance to a common rotation be distinguished from robustness to internal conformation changes?
  4. Please report high-potency test counts and paired per-split differences, say whether CAMP was trained or tuned separately and how cross-database duplicates were handled, and clarify that the “dynamic” DIR weights in Eq. 11 are fixed thresholds.
1 comment

The authors have partially addressed my concerns, but the presentation of the work remains unclear. For example, Figure 4 does not clearly demonstrate which components contribute most to the overall performance. Furthermore, the paper uses multiple definitions of α, referring to both attention scores and hyperparameters, which creates confusion. The sensitivity analysis of the hyperparameters is also missing, as is an ablation study of the deep imbalanced regression component. These omissions make it difficult to assess the key contributions and effectiveness of the proposed method. Therefore, I maintain my original score of 1.

Reviewer_PDowOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GOA-PGN predicts peptide minimum inhibitory concentration (MIC) by combining a peptide-tuned ESM-2 encoder, partial convolutions, and a graph constructed from the cosine similarity of predicted side-chain directions. The model uses attention pooling and multiple prediction heads for MIC regression, high-potency classification, and auxiliary physicochemical properties. Training combines threshold-weighted regression with a two-stage optimization schedule. The evaluation covers three species-specific DBAASP datasets and CAMP E. coli, with five reported homology-controlled splits. GOA-PGN reports the best mean metrics on the full datasets among the listed baselines. On the high-potency subsets, its prediction errors remain competitive, while ranking improvements vary across datasets. The paper also includes component ablations and screening-enrichment curves.

Critical strengths
  1. A useful quantitative task evaluated across several settings.MIC regression and high-potency ranking address peptide prioritization beyond binary activity classification. The evaluation includes four dataset–species settings, several relevant baselines, and both correlation and error metrics (Section 4; Appendices A–C). The five reported splits use homology-aware partitioning; Appendix A and the released splitter specify a 60% identity threshold with post-filtering. On DBAASP E. coli, GOA-PGN improves mean Spearman correlation from LLAMP's 0.514 to 0.549 and reduces RMSE from 0.679 to 0.646 (Table 1). These results support the method's potential utility.

  2. A plausible structural representation with supporting ablations. Relative side-chain orientations offer a plausible way to complement contextual sequence features without directly using pairwise distances. Table 5 examines six architectural or training components. Removing the graph lowers full-set Spearman correlation from 0.549 to 0.526 and increases RMSE from 0.646 to 0.660, supporting an incremental predictive contribution. The ablations support predictive value, but do not establish the proposed biological mechanism or immunity to conformational noise.

Critical weaknesses
  1. The released structural extractor has an indexing defect, but its effect on the original experiments needs clarification.

Section 3.1 defines one side-chain direction per residue, giving an L × 3 tensor. The released extractor selects outputs.positions[0] at line 160, then indexes the result as though it had shape L × atoms × 3 (lines 24–26 and 54). Under the standard Hugging Face ESMFold interface, this selection leaves a batch × L × 14 × 3 tensor. The subsequent indexing therefore selects residue positions instead of atom types, producing a direction tensor of shape 1 × 14 × 3 for a single input. The loader's flattening and padding operations can conceal this mismatch (GOA-PGN/dataset.py, lines 52–67).

The effect on the submitted results is less certain. Training loads precomputed tensors, and extraction skips existing caches, so the experimental features may have been generated by a different script. The original caches and their generating version were unavailable. The authors should identify that version and verify residue-to-vector correspondence in the original tensors. Correct caches would make this primarily a release correction; incorrect caches would undermine the interpretation of the graph's contribution.

  1. Label construction and the consistency of evaluated samples need fuller documentation.

Section 4.1.1 and Appendix A, lines 660–668, describe averaging repeated MIC measurements for each peptide–species pair. The released DBAASP preprocessing script averages within each sequence, terminal-modification group, and species (lines 301–325). The E. coli splitter subsequently retains the first row per sequence (line 45). Thus, repeated measurements within one terminal form are averaged, but only one form is retained when multiple forms share a sequence. Distinguishing these chemical forms is reasonable. The definition of peptide identity and the rule for retaining one form need clarification.

The preprocessing script also removes inequality markers and replaces MIC ranges with midpoints (lines 107–163), choices not described in Appendix A. These conventions can affect both the regression target and high-potency membership. For example, converting a single “<10 μM” record to 10 would exclude it from the strictly <10 μM subset. The prevalence and impact of such cases in the actual datasets remain unknown.

Whether these choices affect comparison fairness is a separate question. All eight methods' visible evaluation paths use log10(MIC) and the same raw-MIC <10 threshold. MBC also reverses its training-target standardization before evaluation. However, each method reads files from its own data directory, and GOA-PGN and MBC filter samples according to feature availability (GOA-PGN/dataset.py, lines 15–18; MBC-Attention/dataset.py, lines 18–26). The splitting script can reuse existing CSVs, so the current source alone does not establish how the experimental labels were produced or whether the final evaluated samples match.

These findings leave a documentation and reproducibility gap. They do not establish an unfair advantage. The authors should first report the original label-processing rules, affected-case counts, and final test-set membership for each method. A focused sensitivity analysis would be appropriate if ambiguous labels or differing test subsets could materially affect the reported gains.

I narrowly favor weak acceptance because the quantitative task, results across four settings, and component ablations support predictive value. This recommendation depends on clarification of the original artifacts. I would reconsider if the structural tensors are incorrect, differing test samples explain the gains, or reasonable handling of ambiguous MIC labels changes the central comparison. If the original artifacts are consistent, the remaining work is to correct the release, document preprocessing, and moderate the stronger ranking and mechanistic claims.

Questions for authors
  1. Which script and ESMFold/Transformers version generated the tensors used for Tables 1–5? Please verify their shapes and residue-to-vector correspondence using the original experiment caches and corresponding coordinate records. Shape checks alone are insufficient, particularly for sequences of length 14. This would determine whether the indexing defect is confined to the release or also affects the reported experiments.

  2. What preprocessing version produced the original split files, and how were terminal forms and inequality/interval MIC records handled? Please report affected-case counts, particularly around the <10 μM threshold. Were the final sequence–label pairs identical across methods after feature generation and cache filtering? Per-split counts and consistently ordered sequence–label lists would help verify this. If test subsets differ, comparisons on their common examples using existing predictions would clarify the impact.

  3. Does Table 2 use the distance-free model or the distance-relaxed variant described in Section 4.2.2 and Appendix D? Please identify the configuration used and provide any existing paired per-split comparison between the variants. This would clarify which structural formulation the CAMP results support.

The first steps are to clarify the existing configurations, caches, split files, and predictions. No new wet-lab study or expanded benchmark is needed to answer these questions.

Reviewer_6k4COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GOA-PGN predicts log10 minimum inhibotry concentration (MIC) from peptide sequences using a peptide-tuned ESM-2 encoder, partial convolutions, and a graph whose edges depend on the positive cosine similarity of ESMFold-derived C-alpha-to-C-beta directions. Attention pooling feeds MIC regression/classification and physicochemical auxiliary heads. Training combines potency-weighted regression with backbone warm-up and joint fine-tuning. Evaluation uses five homology-controlled splits of DBAASP data for three bacterial species and CAMP E. coli, with component ablations and screening enrichment analysis.

Critical strengths
  • The evaluation addresses a consequential failure mode through sequence identity below 60% across train/test sets and five splits.
  • GOA-PGN achieves the best reported full-set correlations and errors across all four settings, per Table 1. These results support the practical usefulness of the combined predictor.
Critical weaknesses
  • The central claim that the orientation graph avoids conformational noise is not established. Equation 4 is invariant to a common rigid rotation, but backbone conformational changes can alter relative C$\alpha$-to-C$\beta$ directions and therefore the adjacency. Thus the immunity claim (216-218) does not follow from omitting distances.
  • Table 5 supports a modest benefit from the graph: full-set Spearman increases from 0.526 to 0.549, and RMSE falls from 0.660 to 0.646. However, high-potency RMSE slightly worsens from 0.695 to 0.697. These gains do not establish robustness to alternative conformations or the asserted amphipathicity mechanism, which is the principal rationale for this distinctive graph construction.
  • The claimed advance in rare-potency learning is also insufficiently isolated. Equation 11 assigns fixed threshold-based weights, and Equation 13 uses fixed task coefficients, despite the repeated description of dynamic weighting. The code confirms a log-MIC threshold of 1.0, high-potency weight of 5.0, and auxiliary coefficients of 0.5 and 0.1. Table 5 does not isolate regression weighting from the rest of the pipeline.
  • Moreover, the MIC < 10 micromolar evaluation subset comprises around 40% of E. coli samples, whereas the ultra potent subset comprises 6.1%. Figure 3 shows promising enrichment for the latter but does not compare against the same model with uniform regression weights. The results do not identify what the proposed imbalance treatment contributes to the rare-tail gains, beyond the combined architecture and auxiliary supervision.
  • Novelty and informativeness is relatively low. The method incorporates well-known existing methods with a litany of loss functions to regularize the objective. While nothing is particularly damning about the experiments or the approach, the fundamental contribution of the paper is limited in insight.
Questions for authors

Clarification: Does Table 2 report the orientation-only model or the distance-relaxed CAMP variant, and which released configuration reproduces it?

Evidence needed for the current claims: Can the authors provide a matched uniform-weight regression ablation, including ultrapotent screening enrichment, and evaluate the graph across alternative or perturbed conformations? These would distinguish the claimed benefits of imbalance treatment and conformational robustness from the overall predictor's demonstrated gains.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Answers & Major Revisions to All Reviewers

We sincerely thank all reviewers for their highly constructive feedback and inspection of our codebase. We have uploaded a revised PDF and updated our anonymous code repository. Here we address the most critical shared concerns:

1.Clarification on the Codebase & "Distance-Free" Mechanism

Reviewers correctly pointed out that our initially uploaded models.py contained a residual distance decay term torch.exp(-dist/3.0). We apologize for uploading an untidy codebase snapshot used during early exploratory tuning. However, our reported results across all datasets are Distance-Agnostic, exactly as claimed. In our main experiments, the dataloader explicitly forces all 3D spatial coordinates to zero. Mathematically, this forces the pairwise distance dist = 0 for all residues. Consequently, the decay term continuously evaluates to exp(0) = 1.0, and the adjacency matrix mathematically collapses purely to orientation-based edges: adj = F.relu(cos_sim). Our final model evaluates peptides entirely without absolute distance information. We have updated the anonymous repository to explicitly reflect this mathematically clean version.

2.Matched Uniform-Weight Regression Ablation

To systematically isolate the contribution of our deep imbalanced regression (DIR), we conducted a matched uniform-weight ablation (all weights = 1.0) on the E. coli dataset. As newly detailed in Appendix E (Table 6) of the revised PDF, removing the potency weights causes the High-Potency RMSE to spike dramatically from 0.697 to 0.719, and Spearman to drop from 0.313 to 0.292. This explicitly confirms that the upweighting mechanism is required to stabilize absolute error gradients in the rare, ultra-potent tail.

3.Correction of "Dynamic Weighting" Terminology

Reviewers correctly noted that our DIR loss uses fixed threshold-based sample weights rather than dynamically adaptive weights. We fully agree. We have removed all instances of "dynamic weighting" and corrected the terminology to "threshold-based static weighting" throughout the revised PDF.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.