acceptodds
Under review as a conference paper at ICLR 2027

SPECIES AND TOPIC-AWARE REPRESENTATION LEARNING FOR ANTIMICROBIAL PEPTIDE ACTIVITY

Abstract

Antimicrobial peptides (AMPs) play a crucial role in addressing the growing threat of antimicrobial resistance. However, their activity varies markedly across microbial targets, indicating that primary peptide sequence data alone cannot fully explain antimicrobial function. To overcome this limitation, we present STAMP: a Species- and Topic-aware representation-learning framework that models AMP activity by integrating sequence, motif, and biological context. STAMP unifies contextual protein language model representations, latent motif composition, and microbial species information within a single model, addressing data sparsity by enabling cross-species training. We evaluated STAMP on four benchmark datasets curated from DBAASP, dbAMP, and DRAMP. As a single cross-species model, STAMP matches methods that rely on tens of independently trained species-specific models and outperforms the most recently published benchmark, reaching a maximum Pearson correlation coefficient (PCC) of 0.837 and an R2 of 0.698. To demonstrate its utility for candidate prioritization, we applied STAMP to score a large library of computationally generated peptides. Among the 18 candidates selected for experimental validation, 16 exhibited experimentally measured MIC values below 10 μM against E. coli, while 17 exhibited MIC values below 10 μM against S. epidermidis, providing experimental support for the predicted antimicrobial activity. Additionally, experiments with Tilapia Piscidin 4 (TP4)-derived peptides and residue-level attribution analysis revealed that STAMP captures target-dependent activity shifts and pinpoints sequence regions crucial for species-conditioned predictions. Collectively, these results demonstrate the value of integrating molecular and microbial context to model AMP activity, supporting computationally guided and efficient AMP discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antimicrobial peptides, Antimicrobial Resistance, Evolutionary Scale Modeling (ESM), Minimum Inhibitory Concentration (MIC), Large Language Models, Bio-informatics, Drug design, DBAASP

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_E31nOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript presents a antibacterial peptide activity prediction framework named STAMP, which takes peptide sequences and target bacterial species as inputs to predict the minimum inhibitory concentration (MIC). The model combines the micro-tuned ESM-2 sequence representation, the LDA topic representation constructed based on overlapping amino acid quadruplets, and species one-hot encoding. Through a shared regression network, it predicts the antibacterial activity under different species conditions. The authors conducted performance comparisons and ablation experiments on three public benchmarks and a self-built dataset and selected 18 candidate sequences from approximately 70000 computationally generated sequences for experimental verification. Additionally, the manuscript examines the model's response to changes in species conditions and the sequence regions relied upon for prediction through TP4-derived peptide analysis and residue gradient attribution, aiming to evaluate the feasibility of the unified model for predicting the activity of multi-species antibacterial peptides and candidate screening.

Critical strengths

This study addressed an interesting issue that the activity of antimicrobial peptides varies with the target species. It used a shared model to predict the MIC values for multiple species, which potentially has the value of reducing the repetitive modeling and maintenance work for each species. The method combines pre-trained sequence representations with local sequence statistical features, has a relatively simple structure, and is convenient for implementation and component analysis. The experiments covered datasets of different species scales and reported the species-specific results, providing a basis for evaluating the model's performance under different data conditions. Additionally, the authors clearly stated that LDA only fits on the training sequences and discussed the limitations such as prediction for unknown species, differences in experimental conditions, and handling of MIC annotations, which helped to define the scope of the method's applicability.

Critical weaknesses
  1. The author performs species merging at the data level for joint training, which may result in the training set and test set not being mutually exclusive at the global sequence level, namely two records corresponding to one peptide may respectively enter the training set and test set. This can be used to examine the activity completion of known peptides on another target, but it is a different task from predicting new peptides that have never been seen before. The author should report the intersection of training-validation-test sequences after merging all species, as well as the maximum sequence similarity distribution of test peptides to the training set. After dividing according to global sequences and homologous clusters, does the advantage of STAMP still hold? Do the existing results measure the generalization of new peptides or the cross-target activity completion of known peptides?
  2. The model adopted by the author is very simple. Simple methods can indeed be valuable. However, when the algorithm itself is relatively conventional, rigorous experiments are even more necessary to prove new and generalizable findings. Currently, the explanatory power of the method exceeds the experimental evidence and the claimed contributions. Apart from conventional feature concatenation and conditional feedforward networks, what are the unique contributions of STAMP's method? Where is the necessity of LDA compared to simple local sequence features? Which independent evidence can support the explanations of "functional motifs" and "biological context", rather than just species identity conditioning and the utilization of sequence similarity?
  3. Table 1 clearly labels the baseline as "average across species", while STAMP is labeled as "across species". The manuscript does not clearly state whether the two use the same method for aggregating indicators. The total MAE index calculated based on the self-built dataset's data volume for each species provided in Table 5 and the MAE index for each species provided in Table 9 does not match the 0.285 reported in the manuscript. The authors need to clarify and explain the reason. Additionally, the authors did not provide confidence intervals and results from multiple runs. After these issues are clarified, will the main performance conclusions still be held?
  4. Among the 18 candidates, most showed strong activity, which was indeed a positive result. However, the authors did not provide random candidate controls from the same generation library or other method-based control groups. Therefore, it is impossible to determine whether the high hit rate mainly comes from the generation library itself, or whether STAMP has brought additional enrichment.
  5. The proposed is not compared with the recent published methods, so it is difficult to evaluated whether the results obtained by the proposed method are SOTA.
Questions for authors
  1. The author performs species merging at the data level for joint training, which may result in the training set and test set not being mutually exclusive at the global sequence level, namely two records corresponding to one peptide may respectively enter the training set and test set. This can be used to examine the activity completion of known peptides on another target, but it is a different task from predicting new peptides that have never been seen before. The author should report the intersection of training-validation-test sequences after merging all species, as well as the maximum sequence similarity distribution of test peptides to the training set. After dividing according to global sequences and homologous clusters, does the advantage of STAMP still hold? Do the existing results measure the generalization of new peptides or the cross-target activity completion of known peptides?
  2. The model adopted by the author is very simple. Simple methods can indeed be valuable. However, when the algorithm itself is relatively conventional, rigorous experiments are even more necessary to prove new and generalizable findings. Currently, the explanatory power of the method exceeds the experimental evidence and the claimed contributions. Apart from conventional feature concatenation and conditional feedforward networks, what are the unique contributions of STAMP's method? Where is the necessity of LDA compared to simple local sequence features? Which independent evidence can support the explanations of "functional motifs" and "biological context", rather than just species identity conditioning and the utilization of sequence similarity?
  3. Table 1 clearly labels the baseline as "average across species", while STAMP is labeled as "across species". The manuscript does not clearly state whether the two use the same method for aggregating indicators. The total MAE index calculated based on the self-built dataset's data volume for each species provided in Table 5 and the MAE index for each species provided in Table 9 does not match the 0.285 reported in the manuscript. The authors need to clarify and explain the reason. Additionally, the authors did not provide confidence intervals and results from multiple runs. After these issues are clarified, will the main performance conclusions still be held?
  4. Among the 18 candidates, most showed strong activity, which was indeed a positive result. However, the authors did not provide random candidate controls from the same generation library or other method-based control groups. Therefore, it is impossible to determine whether the high hit rate mainly comes from the generation library itself, or whether STAMP has brought additional enrichment.
  5. The proposed is not compared with the recent published methods, so it is difficult to evaluated whether the results obtained by the proposed method are SOTA.
5 comments
AuthorsOpenReview ↗

Response to Reviewer E31n Question 1

Reviewer E31n Q1. The author performs species merging at the data level for joint training, which may result in the training set and test set not being mutually exclusive at the global sequence level, namely two records corresponding to one peptide may respectively enter the training set and test set. This can be used to examine the activity completion of known peptides on another target, but it is a different task from predicting new peptides that have never been seen before. The author should report the intersection of training-validation-test sequences after merging all species, as well as the maximum sequence similarity distribution of test peptides to the training set. After dividing according to global sequences and homologous clusters, does the advantage of STAMP still hold? Do the existing results measure the generalization of new peptides or the cross-target activity completion of known peptides?

RESPONSE: We thank the reviewer for highlighting the important distinction between generalization to unseen peptide sequences and prediction of unseen peptide-species combinations.

Dataset A already employs sequence-disjoint partitions, with 9,181 unique peptide sequences distributed across training, validation, and test sets (7,344/918/919), ensuring zero sequence overlap after merging all species. Datasets B, C, and our curated dataset originally employ peptide-species pair-disjoint partitions, which evaluate prediction of previously unobserved peptide-species combinations but may contain identical peptide sequences-species pairs across partitions.

To directly address this concern, we conducted sequence-disjoint evaluations for Datasets B, C, and our curated dataset (Table 6), grouping all measurements of the same peptide sequence into a single partition and ensuring zero exact sequence overlap across training, validation, and test sets. Five distinct partitions were generated using different random seeds. Under sequence-disjoint evaluation, STAMP achieves mean PCC values of 0.723, 0.707, 0.706, and 0.718 across the four datasets, respectively, consistently maintaining PCC > 0.70 for previously unseen peptide sequences.

Our evaluations therefore address both tasks the reviewer distinguishes:

  • Cross-target activity completion of known peptides: evaluated under the original peptide-species pair-disjoint partitions (Table 1), where the model predicts activity of a known peptide against a new species.
  • Generalization to new peptides: evaluated under sequence-disjoint partitions (Table 6), where the model predicts activity of peptides whose sequences were entirely absent from training.

Both evaluation settings demonstrate competitive predictive performance, and the manuscript now explicitly distinguishes these objectives (Table 11).

Furthermore, our prospective experimental validation provides the strongest evidence of generalization: 15 computationally generated peptides with no overlap with any training data were synthesized and tested, with 14/15 exhibiting MIC < 10 uM against E. coli and 15/15 against S. epidermidis (Table 2). This prospective validation on entirely novel peptide sequences cannot be attributed to memorization or sequence overlap.

We acknowledge that homology-cluster-disjoint evaluation would provide additional insight into generalization to more distant regions of peptide sequence space, and we identify this as a direction for future work.

AuthorsOpenReview ↗

Response to Reviewer E31n Question 2

Reviewer E31n Q2. The model adopted by the author is very simple. Simple methods can indeed be valuable. However, when the algorithm itself is relatively conventional, rigorous experiments are even more necessary to prove new and generalizable findings. Currently, the explanatory power of the method exceeds the experimental evidence and the claimed contributions. Apart from conventional feature concatenation and conditional feedforward networks, what are the unique contributions of STAMP's method? Where is the necessity of LDA compared to simple local sequence features? Which independent evidence can support the explanations of "functional motifs" and "biological context", rather than just species identity conditioning and the utilization of sequence similarity?

RESPONSE: We thank the reviewer for this comment. We agree that simple methods require rigorous experimental support, and we believe our revised manuscript provides this.

Regarding the unique contributions of STAMP: the contribution is not feature concatenation itself, but the formulation and evaluation of antimicrobial activity as a context-dependent property of the peptide-target pair, operationalized through a unified cross-species framework. Existing methods (LLAMP, Ensemble, ANIA) train separate models for each species, requiring 3-35 independent models. STAMP replaces this with a single shared model that conditions predictions on species identity while sharing sequence-activity relationships across targets. This is not merely a modeling convenience -- the cross-species formulation enables data-sparse species to benefit from shared patterns (Section 2.4, Eq. 8-9), as demonstrated by the largest performance gap between per-species baselines and STAMP occurring on Dataset A (35 species, many with limited data).

Regarding LDA vs. simple local sequence features: LDA topic mixture probabilities are fundamentally different from frequency-based features such as k-mer counts or TF-IDF. TF-IDF weights individual k-mers independently based on occurrence frequency. LDA learns latent co-occurrence patterns among 4-mers and represents each peptide as a probability distribution over latent topics, where each topic captures a group of co-occurring motifs. The output is a compositional summary (probabilities summing to 1) that captures higher-level motif structure, not a frequency vector of individual patterns.

The biological relevance of LDA-derived motifs has been independently established in the literature. Padi et al. (Scientific Reports, 2025) demonstrated that:

  • LDA-derived topic motifs are associated with physicochemical properties relevant to antimicrobial activity, including isoelectric point, hydrophobic moment, hydrophobicity, and structural characteristics.
  • LDA-derived motifs differ from conventional frequency-based motifs in diversity, redundancy, and biochemical interpretation.
  • Topic-derived motif groups capture recurring co-occurrence patterns that correspond to biologically meaningful peptide characteristics.

Building on this independently established evidence, STAMP incorporates LDA topic-mixture representations into species-conditioned MIC prediction. Table 4 provides predictive evidence within the STAMP framework: Sequence+Motif consistently outperforms Sequence-only across all four datasets in both per-species and cross-species settings, and STAMP outperforms Sequence+Species across all four datasets. These results demonstrate that motif-topic information contributes predictive value beyond sequence embeddings and species identity alone.

The species-conditioned analysis (Section 3.4, Table 3) and residue-level attribution (Section 3.5, Figure 2) provide additional evidence that the model captures target-dependent activity differences consistent with known Gram-negative and Gram-positive membrane properties, rather than merely utilizing species identity as a categorical label.

  1. Padi, S., Mondal, K., Hoogerheide, D.P. et al. AI-driven antimicrobial peptide characterization unveils novel motifs for drug design. Sci Rep 16, 829 (2026). https://doi.org/10.1038/s41598-025-30419-1
AuthorsOpenReview ↗

Response to Reviewer E31n Question 3

Reviewer E31n Q3. Table 1 clearly labels the baseline as "average across species", while STAMP is labeled as "across species". The manuscript does not clearly state whether the two use the same method for aggregating indicators. The total MAE index calculated based on the self-built dataset's data volume for each species provided in Table 5 and the MAE index for each species provided in Table 9 does not match the 0.285 reported in the manuscript. The authors need to clarify and explain the reason. Additionally, the authors did not provide confidence intervals and results from multiple runs. After these issues are clarified, will the main performance conclusions still be held?

RESPONSE: We thank the reviewer for identifying these issues and apologize for the lack of clarity.

Metric aggregation: The distinction between "average across species" and "across species" in Table 1 refers to the modeling strategies, not the metric computation. LLAMP, Ensemble, and Attention-based methods employ species-specific models, whereas STAMP trains a single unified model. For all methods, we employ micro-averaging: MSE, MAE, R2, and PCC are computed directly from the pooled measured and predicted MIC values across all test samples. We have revised the Table 1 column headers and caption for clarity.

MAE discrepancy: Upon re-examination, we identified and corrected an error in the reported per-species metrics in the original Table 9. The corrected values (Table 15 in the revised manuscript) are now consistent with the overall STAMP results computed via micro-averaging.

Confidence intervals: We conducted five independent evaluations across all four datasets and report mean and standard deviation in Table 5. STAMP achieves mean PCC values of 0.723 +/- 0.010, 0.692 +/- 0.007, 0.788 +/- 0.004, and 0.824 +/- 0.013 across the four datasets, respectively. Although the mean performance on Dataset B (PCC: 0.692) is lower than the original single-run result (PCC: 0.758), results for Datasets A, C, and our curated dataset remain consistent with the original findings.

The main performance conclusions hold: STAMP provides a unified cross-species MIC prediction framework with competitive performance, consistently achieving PCC > 0.69 across all datasets over multiple runs, without requiring independently trained models for each species. The additional sequence-disjoint evaluations (Table 6) further demonstrate that STAMP maintains PCC > 0.70 for previously unseen peptide sequences.

Reviewer_bJ6eOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STAMP predicts MIC of antimicrobial peptides from a concatenation of three components: a fine-tuned ESM-2 (8M) embedding, a fixed 100-dim LDA topic vector over amino-acid 4-mers, and a one-hot bacterial species code, fed to a one-hidden-layer MLP regressing log10(MIC). The claim is that one cross-species model replaces the per-species model collections of prior work. Evaluation covers three published benchmarks plus a curated DBAASP-v3 set, a four-way ablation, Gradient × Input attribution, species-conditioned predictions on TP4 variants, and 18 synthesized peptides assayed against E. coli and S. epidermidis.

Critical strengths

The paper addresses a problem that is genuinely framed better here than in most of the prior literature: MIC is a property of the peptide-target pair, not of the peptide alone, and conditioning a single shared encoder on species identity is the right structural response to that. The execution of the analysis supporting this idea is careful. The ablation isolates each representational component on every dataset, the interpretability work probes the species channel behaviorally rather than asserting it works, and the authors report per-species numbers that are unfavorable to their own narrative. My two main strengths are listed below.

1. The ablation and the interpretability analysis are the best part of the paper and they do support the representational claim.

Table 4 is clean, every configuration is reported on every dataset, and the conclusion that motif and species information contribute complementarily is earned. The attribution analysis and the TP4 variant study in Table 3, where only the species input changes and the model preserves the measured direction of the activity shift for TP4-noR5, give behavioral evidence that the species channel is more than a per-species intercept. This matters because it tests the paper’s actual hypothesis rather than aggregate fit.

2. Transparent per-species reporting.

Tables 6 to 9 report every species individually, including strongly negative R² (L. innocua -0.965, A. hydrophila -0.929, L. lactis -0.489) that cuts against the paper’s own narrative about sparse species benefiting from pooling. Reporting numbers that undercut your headline is good practice. The prospective synthesis and assay work is also real effort in a subfield dominated by retrospective benchmarks, and I would not want it discouraged.

Critical weaknesses

My concerns separate into two claims the paper makes and treats as one. The representational claim, that molecular and microbial context improves MIC prediction, is supported by the ablation. The performance claim, that a single cross-species model matches or outperforms multi-model state of the art, is not supported by the evidence as presented: the comparison is thin, the one favorable number carries a protocol risk, the one that is strongest faces no competitor at all, and the tables typeset losses as wins. Underlying both is an evaluation protocol that never controls for sequence homology, never tests transfer across datasets, and never characterizes the data it is built on, which leaves me unable to judge how much of the reported accuracy is real generalization. The specific points are listed below.

1. There is no rigorous comparison against the state of the art, and Table 1 presents losses as wins.

Only one competing method is run per dataset, and on the two datasets where the comparison is close STAMP is worse on every metric: Dataset A (MSE 0.275 vs 0.272, MAE 0.385 vs 0.379, R² 0.531 vs 0.536, PCC 0.732 vs 0.735) and Dataset B (MSE 0.260 vs 0.233, R² 0.566 vs 0.603, PCC 0.758 vs 0.779). STAMP’s row is bolded on both, which misrepresents the result. The abstract’s best numbers (PCC 0.837, R² 0.698) come from the authors’ own curated dataset, where no baseline is run at all. No sequence-only SOTA method is reimplemented and no motif-based method is included, so the ablation’s Sequence-only row is the only stand-in for a competitor. No result anywhere carries a seed, a standard deviation or a significance test, while the differences being argued over are 0.02 to 0.03 PCC at batch size 8. The one clear win is Dataset C, which Table 5 shows has no validation partition although Appendix A.3 states training uses early stopping; if selection was done on the test split, the single result supporting the strongest claim is contaminated.

2. No generalization test, no homology-aware splitting, and the candidate panel is compromised.

The paper never trains on one dataset and tests on another, which is the natural test of a unified cross-species model and the cheapest experiment available. Within datasets, Appendix A.2.2 and A.2.3 control only for exact duplicates, with no clustering or identity threshold, so near-identical analogues will straddle the split; this plausibly explains why the authors’ own dataset yields the best numbers, which the paper instead credits to preprocessing. The same issue is concrete in Table 2: P1 (FIHHIIGGLFSAGKAIHRLIRRRRR, measured 4 and 1, predicted 4.54 and 1.24) is identical to TP4 in Table 3 including its predictions, and TP4 is a characterized natural peptide present in DBAASP, the source of the Dataset A training data used here. It is presented among 18 candidates selected from a generated library while almost certainly being a training example. P2 and P3 are one and two substitutions away, and eight further candidates (P5, P9, P10, P11, P13, P15, P17, P18) share one C-terminal scaffold, so the panel is roughly three or four templates, not eighteen trials. With no negative control, no random-library arm and no baseline predictor comparison, a 16/18 hit rate does not establish screening utility.

3. The data are never characterized, and two design choices are unexplained.

There is no peptide length distribution, no MIC distribution, no per-species sample counts placed next to the per-species results. The binarization threshold used for the AMP/non-AMP analysis (A.5) is stated without provenance, and it sits exactly at the censored values that A.2.2 creates by stripping inequality symbols and treating them as exact targets. The ablation omits the Motif + Species configuration, so the necessity of the sequence channel is never tested.

Questions for authors

The questions below ask mainly for clarification of the evaluation protocol and for experiments that appear feasible with the models and data the authors already have. Some concern points where the text and the tables disagree, others concern design choices that are stated but not justified. They are listed below.

1. Table 1 bolds STAMP on Datasets A and B although the baseline is better on every metric. Please correct the bolding and restate in the abstract what “matches” means.

2. Dataset C has no validation split in Table 5 but Appendix A.3 uses early stopping. What was used for model selection on Dataset C? Please rerun with a validation split carved from training.

3. Please report mean and standard deviation over at least five seeds for Tables 1 and 4, and state whether the Dataset C gain is significant.

4. Why is no competing method evaluated on your curated dataset, where your best results are reported?

5. Why only one baseline per dataset? Can you reimplement two or three sequence-only predictors and at least one motif-based method, and run them on all four datasets under your splits? This is the result I would weigh most heavily.

6. I suggest adding a cross-dataset generalization experiment: train on one benchmark and evaluate on another. The datasets share target species, so this is feasible with the models you already have, and it would be the clearest evidence that the species-conditioned representation transfers beyond a single split.

7. Why was the Motif + Species ablation not run? Without it the contribution of the sequence alone channel is untested.

8. How sensitive are results to the choice of protein language model? ESM-2 8M is used with no scaling ablation and no comparison against other encoders in current use (larger ESM-2 variants, ProtT5, ProtBERT, peptide-specific encoders). ESM2AMP (Haimeng et al. 2026), described in A.1 as the closest prior method, is absent from Table 1; why?

9. How does the LDA vector compare against a 4-mer count or TF-IDF vector of matched dimensionality, and how sensitive are results to K and to α, η (currently 100, 0.01, 0.01, with no sensitivity analysis)? Since ESM-2 is fine-tuned end to end, please show the topic vector carries information the encoder cannot recover.

10. Please add peptide length and MIC distributions, and per-species training counts.

11. Where does the AMP/non-AMP threshold in A.5 come from, and why binarize at all rather than evaluate the regression directly?

12. On Dataset A, Sequence + Species underperforms the per-species baselines (R² 0.334 vs 0.445) and several sparse species have negative R². Please plot per-species R² against training count and state plainly what the method does and does not deliver for sparse targets.

8 comments
AuthorsOpenReview ↗

Response to Reviewer bJ6e Questions 1 and 2

Reviewer bJ6e Q1. Table 1 bolds STAMP on Datasets A and B although the baseline is better on every metric. Please correct the bolding and restate in the abstract what "matches" means.

RESPONSE: We have revised the Table 1 caption to clarify that bold indicates our proposed method (STAMP), not the best-performing method. The abstract has been revised to use "competitive" and "comparable" rather than "matches," and performance differences are explicitly acknowledged in the main text.

Reviewer bJ6e Q2. Dataset C has no validation split in Table 5 but Appendix A.3 uses early stopping. What was used for model selection on Dataset C? Please rerun with a validation split carved from training.

RESPONSE: Since Dataset C does not provide a predefined validation split, we reserved 10% of the original training data for validation and used the remaining 90% for model training. The validation subset was used to monitor model performance and determine the early-stopping criterion. The predefined test dataset was kept separate and was not used during model training or model selection. These details have been added to the revised manuscript (Appendix A.2.1).

AuthorsOpenReview ↗

Response to Reviewer bJ6e Question 3

Reviewer bJ6e Q3. Please report mean and standard deviation over at least five seeds for Tables 1 and 4, and state whether the Dataset C gain is significant.

RESPONSE: We conducted five independent training and evaluation runs of the complete STAMP model (ESM2-8M backbone) across all four datasets. The mean and standard deviation of MSE, MAE, R2, and PCC are reported in Table 5 of the revised manuscript. For Datasets A, B, and C, we retained the original predefined training, validation, and test partitions while varying the model training seed. For our curated dataset, we generated five distinct partitions using different random seeds.

As shown in Table 5, STAMP demonstrates consistent predictive performance across all four datasets, with mean PCC values of 0.723 +/- 0.010 (Dataset A), 0.692 +/- 0.007 (Dataset B), 0.788 +/- 0.004 (Dataset C), and 0.824 +/- 0.013 (our curated dataset). The low standard deviations confirm that STAMP's performance is stable across independent runs and not sensitive to random initialization.

Due to the limited time, we conducted the five-run evaluation for the complete STAMP model only, rather than for all ablation configurations in Table 4. However, since the full STAMP model integrates all components (Sequence + Motif + Species), its consistent performance across five runs supports the overall reliability of the framework. The ablation results in Table 4 demonstrate the relative contributions of individual components, and the stability of the complete model provides confidence in these findings.

Furthermore, we conducted five-run sequence-disjoint evaluations across all four datasets (Table 6), where STAMP consistently maintains PCC > 0.70 for previously unseen peptide sequences, further demonstrating the robustness of the model.

Table 1 follows the same single-run reporting convention used by all baseline methods, whose published results are also based on single runs. The 5-run mean on Dataset B (PCC: 0.692 +/- 0.007) is lower than the single-run result reported in Table 1 (PCC: 0.758); however, it still substantially outperforms both RF (0.647) and XGBoost (0.637) baselines on the same dataset, confirming that STAMP's cross-species advantage is maintained across independent runs.

Regarding Dataset C, STAMP achieves higher predictive performance than the published Attention-based baseline (PCC: 0.798 vs. 0.775; R2: 0.634 vs. 0.585). However, because repeated-run results for the published baseline are unavailable, we cannot establish the statistical significance of this improvement and therefore describe it as an observed performance improvement.

AuthorsOpenReview ↗

Response to Reviewer bJ6e Questions 4 and 5

Reviewer bJ6e Q4. Why is no competing method evaluated on your curated dataset, where your best results are reported?

RESPONSE: To strengthen the evaluation, we have added Random Forest (RF) and XGBoost baselines to Table 1 for the curated dataset. STAMP (MSE: 0.173, PCC: 0.833) substantially outperforms both RF (MSE: 0.300, PCC: 0.698) and XGBoost (MSE: 0.285, PCC: 0.695). Table 4 additionally provides ablation results (Sequence-only, Sequence+Motif, Sequence+Species, and full STAMP) on the curated dataset.

Reviewer bJ6e Q5. Why only one baseline per dataset? Can you reimplement two or three sequence-only predictors and at least one motif-based method, and run them on all four datasets under your splits? This is the result I would weigh most heavily.

RESPONSE: For Datasets A, B, and C, the original benchmark studies already evaluated multiple competing approaches, including RF, XGBoost, and various deep learning models, and identified their best-performing methods. We compare STAMP directly against these published best-performing methods rather than reimplementing all previously evaluated approaches, as we use the same predefined data partitions to ensure a fair comparison.

To further strengthen our evaluation, we additionally implemented RF and XGBoost baselines across all four datasets (Table 1). Table 4 presents systematic ablations across all four datasets evaluating the contributions of motif representations and species conditioning. Together, these experiments provide: (i) comparisons against published state-of-the-art methods on their respective benchmarks, (ii) consistent RF and XGBoost baselines across all four datasets, and (iii) component ablations evaluating reduced STAMP configurations.

Reviewer_Nyn1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes STAMP, which concatenates three features of a peptide–species pair — a mean-pooled ESM-2 (8M) embedding, a frozen 100-d LDA topic mixture over amino-acid 4-mers, and a one-hot target species — and regresses log10 MIC through a single-hidden-layer MLP. ESM-2 is fine-tuned while the other two components stay fixed. The design point is that one cross-species model replaces the per-species collections used by prior work (LLAMP: 35 models; an ensemble method: 8 per species; ANIA: 3). The paper compares against those baselines on three public benchmarks plus a dataset curated from DBAASP-v3, runs a representation ablation, synthesises 18 model-selected peptides for prospective validation (16/18 against E. coli and 17/18 against S. epidermidis measuring MIC < 10 µM), and performs residue attribution via Gradient × Input.

Critical strengths
  1. Treating MIC as a property of the peptide–target pair and conditioning one shared model on species identity is the correct formalisation for this task, and collapsing 35 separately trained and tuned models into one is worthwhile even at a small accuracy cost.
  2. Synthesising model-selected candidates and reporting measured MICs goes further than most work in this area. I discount it somewhat because the screen has no control arm and several candidates are not independent of the training data.
Critical weaknesses
  • The data unit is the peptide–species measurement, not the peptide. For Dataset A, Table 5's three partitions sum to 36,818 — exactly the measurement count — against only 9,181 unique sequences, i.e. about four measurements per peptide. Unless splits were grouped by sequence, a test peptide's remaining measurements almost certainly fall in training, making the test condition "seen peptide, new species".
  • Datasets A–C defer to the original authors' splits with no grouping policy stated, and the curated set's handling of duplicates is described only as following "the dataset curation protocol". No sequence-identity-grouped splitting appears anywhere, though it is standard in this field.
  • Two signs the effect is material: the frozen 100-d LDA vector adds +0.145 to +0.197 R² on top of an already fine-tuned ESM-2 (Table 4: A 0.334→0.531, B 0.421→0.566, C 0.476→0.634, curated 0.532→0.698) — implausibly large for an unsupervised 4-mer topic mixture, but exactly what a near-sequence-fingerprint buys when labels can be retrieved across the split. Also, Table 2's P1 is character-identical to TP4 in Table 3 with the same 4.54 prediction, and P2/P3 differ from it by single substitutions.
  • Behind the baselines on two benchmarks, yet described as matching. On Dataset A, STAMP is worse than LLAMP on all four metrics (0.275/0.385/0.531/0.732 vs 0.272/0.379/0.536/0.735); on Dataset B, worse than the ensemble on all three comparable metrics (MSE 0.260 vs 0.233, R² 0.566 vs 0.603, PCC 0.758 vs 0.779). STAMP leads on one of three external benchmarks.
  • The ablation cannot support its causal reading. Sequence-only and +motif are per-species models while the other two rows are cross-species, so the comparison used to isolate species conditioning changes the training regime at the same time.
Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.