Pith. sign in

REVIEW 2 major objections 4 minor 56 references

AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction

T0 review · 2 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read High AMP-recognition scores do not predict potency or safety readouts, a new benchmark shows.

desk verdict A genuinely useful homology-controlled AMP benchmark resource whose headline dissociation claim overreaches its tiny paired-model diagnostic. read the letter →

arxiv 2607.25518 v1 pith:4BV5RWQW submitted 2026-07-28 cs.LG q-bio.QM

classification cs.LGq-bio.QM
keywords antimicrobialpeptidesbenchmarkMICregressionhomology-controlledsplitactivityspectrumhemolysisselectivityproteinlanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper builds a benchmark that keeps antimicrobial peptide (AMP) recognition, species-specific potency, activity spectrum, and safety-proxy endpoints in one homology-controlled protocol, and uses it to argue that a model's AMP/non-AMP recognition score is not evidence of how the model will behave on assay-derived endpoints. Across 161 endpoint-specific model evaluations, binary performance and MIC/spectrum behavior were nearly uncorrelated, with a matched-model diagnostic giving Spearman -0.04 between binary MCC and MIC error. The paper claims evaluation should therefore report potency, spectrum, and safety-proxy readouts separately, not collapse them into a single recognition leaderboard. If right, this changes how AMP discovery models are selected and trusted.

What carries the argument

AMPBench-MT itself: a provenance-preserving pipeline that standardizes records from multiple AMP databases into canonical 20-amino-acid sequences and endpoint tables, assigns endpoint labels only from explicit assay evidence (missing evidence stays unassigned), splits data by MMseqs2 30% sequence-identity clusters so that no exact sequence or cluster appears in both train and test, and evaluates 161 model rows across seven model families with endpoint-matched metrics (MCC, balanced accuracy, MAE, Spearman, calibration). The load-bearing design is the combination of cluster-level splits plus endpoint-specific label construction, which lets the benchmark separate recognition from potency and s

What would settle it

A prospective evaluation with test species fully held out from training, or a controlled paired experiment with many models matched across binary and MIC endpoints, would settle the claim: if binary MCC and MIC error correlate strongly under those conditions, or if the PLM MIC lead disappears when species overlap is removed, the paper's central dissociation would not generalize.

Watch

Extended reading notes

Core claim

The central claim is that high binary AMP-recognition performance does not reliably indicate assay-endpoint behavior, so benchmark comparisons should be organized by endpoint rather than by a single recognition score. The evidence: PLM embeddings lead species-conditioned pMIC regression (ESM-C 300M at MAE 0.504) while binary leaders differ; the spectrum task shows AUPRC 0.999 with MCC 0.000 because the best model predicts every test row positive at the selected threshold; low-toxicity, HC50, and selectivity show smaller but measurable assay-facing signals; and a paired-model diagnostic finds essentially no correlation between binary MCC and MIC error. The paper frames this as a benchmark for

Load-bearing premise

The conclusion that binary performance does not transfer to endpoints rests on a tiny paired-model diagnostic (10 MIC rows, 4 spectrum rows) and on the assumption that residual species/publication overlap between train and test does not inflate the reported MIC and endpoint numbers.

Editorial extensions

If this is right

  • Binary AMP recognition scores should not be used as a proxy for potency, spectrum, or safety in model selection or discovery pipelines.
  • Models should be evaluated with thresholded and calibration-aware metrics (MCC, balanced accuracy, negative-class recall) alongside AUPRC, especially when observed negatives are scarce.
  • Frozen protein-language-model embeddings form a competitive baseline for species-conditioned pMIC regression, but the leading cluster is tight and R2 stays below 0.3, so quantitative potency prediction is far from solved.
  • Spectrum labels derived from MIC thresholds currently support only positive-evidence audits; observed-negative evidence is too sparse to rank models on spectrum discrimination.
  • The released benchmark tables and splits provide a standardized way for future AMP work to report endpoint-level results.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the binary-to-endpoint dissociation holds more broadly, many published AMP 'high-accuracy' classifiers may have little bearing on which peptide kills which pathogen, so candidate triage should be revalidated on potency endpoints.
  • A direct test of the provenance-overlap worry: retrain the same MIC models with test species excluded from training; if the ESM-C lead shrinks, the reported PLM cluster partly reflects species leakage rather than potency understanding.
  • The same endpoint-auditing protocol could be applied to other peptide properties (e.g., anti-inflammatory or cell-penetrating activity) where recognition-style benchmarks dominate but assay endpoints guide decisions.
  • Reporting a calibration curve and threshold-dependent MCC for spectrum-like tasks would give a cheaper diagnostic than full benchmark reruns.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper introduces AMPBench-MT, a multi-endpoint benchmark for antimicrobial peptides that integrates binary AMP/non-AMP recognition, species-conditioned pMIC regression, spectrum, low-toxicity, HC50 hemolysis, and selectivity tasks under MMseqs2 30% sequence-identity cluster splits with retained provenance. The authors curate labels from a heterogeneous set of AMP databases, deliberately leave missing assay evidence unassigned, and evaluate 161 endpoint-specific model configurations across seven model families. The headline empirical claim is that high binary recognition performance does not reliably transfer to assay-endpoint behavior, supported by endpoint-transfer diagnostics, a spectrum threshold diagnostic showing AUPRC≈0.999 with MCC=0, and a PLM-led pMIC error cluster. The paper concludes that AMP evaluation should move from recognition leaderboards toward endpoint-aware evidence auditing.

Significance. If the empirical claims hold, AMPBench-MT would be a useful community resource: it is one of the few AMP benchmarks that jointly exposes potency, spectrum, and safety-proxy endpoints under a single homology-controlled protocol, and it publishes complete result matrices, split files, provenance metadata, a manifest, and checksums. The candor about residual provenance overlap, single-run predictions, and unaudited pretraining overlap is a genuine strength. The paper also provides a convincing, reproducible diagnostic of how PR-oriented metrics can mislead under extreme class imbalance. However, the central empirical claim — that binary performance does not reliably indicate endpoint behavior — is currently supported by a very small matched-model diagnostic, so the manuscript's value is greater as a benchmark construction and audit contribution than as a fully powered study of cross-endpoint transfer.

major comments (2)
  1. [§4.5 / Appendix Table 4] The headline claim that high binary performance does not transfer to assay endpoints rests on the endpoint-transfer diagnostic: n=10 for binary MCC vs. MIC −MAE (Spearman ρ=−0.04) and n=4 for spectrum with endpoint MCC fixed at 0. With n=10, the 95% confidence interval for Spearman is very wide (roughly ±0.6), so ρ=−0.04 cannot distinguish no association from a substantial positive or negative association; n=4 is effectively uninformative. The table itself states these are not controlled paired experiments and all rows are single-run, so seed variance is unknown. This is load-bearing for the abstract's first claim. Please either add more matched model rows and multi-seed predictions so the cross-endpoint correlation is estimable, or rephrase the claim to the defensible statement that binary recognition and endpoint behavior are different evidence types and should be reported separately,
  2. [§3.2 (Eq. 2) and §4.4] The spectrum task is constructed by thresholding pMIC (pMIC≥5 positive, pMIC≤4 negative, middle decade excluded). The striking AUPRC≈0.999 / MCC=0.000 pattern is therefore an expected consequence of the extreme positive:negative test ratio (7,578:30) and validation-threshold selection, not an independent assay-endpoint discovery. The threshold diagnostic in Table 4 (all test rows predicted positive) is a useful metric audit, but the abstract's phrase 'spectrum labels further reveal that PR-oriented metrics can be misleading' should explicitly note that these labels are derived from MIC evidence. As written, a reader may over-interpret the spectrum result as evidence that models fail on a biologically independent endpoint. Please label spectrum as a derived MIC diagnostic in the task overview and in the abstract, or add at least one truly independent endpoint comparison.
minor comments (4)
  1. [§3.3 / Table 4] The provenance overlap (573/868 MIC test species in training; 551 shared PubMed IDs covering 7,632 test rows) is disclosed honestly, but it should also appear in the main-text conclusion near the MIC results. Since species identity is an explicit model input, the 'species-conditioned' MIC numbers should be read as within-benchmark performance, not as evidence about novel-species potency prediction. The current wording is mostly careful, but the phrase 'species-conditioned potency' could be qualified more consistently.
  2. [§4.3 / Figure 5] The bootstrap intervals in Figure 5 are single-run prediction resamples, not multi-seed uncertainty. The paper says this, but it is easy to miss. Consider adding a sentence in the main text that the overlap among ESM-C 300M, ESM-2 variants, SaProt-650M, and ProtT5-XL means the paper does not claim a decisive winner among these embeddings.
  3. [§4.4 / Figure 8] The selectivity regression result (best R² = 0.077) is reported as 'harder,' but the lack of any model with meaningful explained variance suggests the task may be dominated by label noise or by the small paired-evidence table (9,774 rows). A brief discussion of label-quality limits for selectivity would strengthen the interpretation.
  4. [Table 4] The 'direct-unit MIC sensitivity table' is described but no model evaluation is run on it. If the purpose is a unit-conversion diagnostic, state explicitly in Section 4.3 why the main split is used instead and what the sensitivity table adds beyond shared-row counts.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild self-definitional spectrum label; otherwise the benchmark evaluation is empirical and not circular.

  1. self definitional [Section 3.2, Eq. (2); Section 4.4 "Spectrum" results; Appendix Table 4 spectrum-threshold diagnostic]
    "Exact MIC evidence is mapped into spectrum labels by y_spectrum = 1, pMIC≥5.0, 0, pMIC≤4.0, excluded, 4.0<pMIC<5.0. ... Spectrum is the most pronounced metric-failure case: CatBoost [33] reaches AUROC/AUPRC 0.810/0.999 in the strict spectrum setting, but MCC and balanced accuracy remain 0.000 and 0.500."

    The spectrum endpoint is defined by thresholding the same pMIC values that constitute the MIC-regression target (Eq. 1), so the extreme imbalance (7,578 test positives vs. 30 negatives), the near-baseline AUPRC, and the triviality of the AUPRC/MCC mismatch are fixed by the label-construction rule rather than discovered from independent spectrum evidence. Presenting the spectrum AUPRC/MCC pattern as evidence that "PR-oriented metrics can be misleading" partly restates Eq. (2)'s threshold design. This touches one diagnostic (spectrum as a cross-endpoint illustration) and not the benchmark's central empirical value.

full rationale

AMPBench-MT's central contribution is an empirical benchmark evaluation: the 30% MMseqs2 splits, endpoint tables, and 161 model evaluations are test-set measurements, not a derivation in which an input is renamed as a prediction. There are no load-bearing self-citations, no uniqueness theorem imported from the authors' prior work, and no fitted parameter presented as an independent prediction. The main dissociation claim relies on a small matched-model diagnostic (MIC n=10, spectrum n=4) that the paper itself labels "diagnostic only" and "not controlled paired experiments"; underpower is a statistical limitation, not circularity. The one exhibitable circular step is the spectrum endpoint: Eq. (2) derives spectrum labels from the same pMIC target used for MIC regression, so the spectrum "metric-failure" pattern is largely a consequence of the threshold rule and class imbalance rather than an independent cross-endpoint finding. Since this affects only one diagnostic display and the paper openly discloses the construction, the appropriate circularity score is low (2).

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The benchmark's central claims rest mainly on data curation choices (thresholds, clustering, label aggregation) and domain assumptions about what missing evidence means. No new physical or biological entities are introduced.

free parameters (4)
  • MMseqs2 30% identity threshold = 30% identity, coverage 0.8, cov-mode 0, cluster-mode 2, seed 42
    Homology split parameter chosen by hand; controls cluster-level leakage but not source/species overlap.
  • strict spectrum thresholds = pMIC >= 5.0 positive; pMIC <= 4.0 negative
    Hand-chosen decade boundaries that create extreme imbalance (30 test negatives out of 7,608).
  • pMIC aggregation gates = median, SD <= 1.0, range 0-12
    Quality gates for repeated exact sequence-species records; not fit to data but determine labels.
  • sequence length filters = binary 10-50; endpoint 5-100
    Chosen filters; cross-task comparisons are not length-controlled.
assumptions (6)
  • domain assumption 20-standard-amino-acid canonical sequence representation is sufficient for endpoint prediction tasks
    Excludes D-amino acids, cyclization, modifications; stated in §3.1 as a boundary.
  • domain assumption Missing assay evidence is not negative evidence and should remain unassigned
    Label-construction rule in §3.2; increases precision but reduces coverage.
  • domain assumption Swiss-Prot subset provides a valid non-AMP negative set for the binary task
    Binary negatives are from UniProtKB; negative-data bias is known from [37] and the paper's balanced design mitigates but does not remove it.
  • domain assumption Median pMIC over repeated records is a valid standardized regression target
    Explicitly called a benchmark label rule, not a meta-analysis of lab variability (§3.2).
  • domain assumption 30% identity MMseqs2 clustering prevents meaningful near-duplicate leakage
    Zero cluster overlap is achieved, but source/species/publication overlap remains; the paper reports residual risks.
  • domain assumption Frozen PLM/LLM pretraining data have no contamination with the benchmark test sequences
    Stated as 'unaudited' in §4.1 and §5; if violated, reported model rankings could be inflated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction." pith.science (2026). https://pith.science/paper/4BV5RWQW

@misc{pith2026260725518,
  author       = {Pith},
  title        = {Pith review of: AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BV5RWQW}},
  note         = {Machine review of arXiv:2607.25518}
}
read the original abstract

Computational AMP discovery is often evaluated through AMP/non-AMP recognition, yet follow-up decisions depend on assay-derived evidence such as target-species potency, hemolysis, toxicity, and selectivity. Existing AMP and peptide benchmarks cover binary recognition, multilabel annotation, assay regression, or broader peptide-model comparison, but they do not jointly place AMP recognition, species-conditioned potency, spectrum, safety-facing proxy endpoints, and cross-endpoint behavior within one sequence-homology-controlled protocol. To address this problem, we introduce AMPBench-MT, a provenance-preserving benchmark that standardizes canonical peptide records and organizes them into binary recognition, species-conditioned pMIC regression, and endpoint-specific potency and safety-facing readouts. Across 161 endpoint-specific model evaluations, high binary performance does not reliably indicate assay-endpoint behavior. Frozen protein-language-model embeddings form the leading pMIC error cluster, while graph and classical regressors remain close. Spectrum labels further reveal that PR-oriented metrics can be misleading under scarce observed negatives, whereas low-toxicity, HC50 hemolysis, and selectivity expose smaller but more assay-facing signals. AMPBench-MT shows that AMP evaluation should move beyond recognition leaderboards toward endpoint-aware evidence auditing. Our proposed benchmark is available at https://huggingface.co/datasets/ZihengZhou06/AMPBench-MT.

Figures

Figures reproduced from arXiv: 2607.25518 by the authors.

Figure 1
Figure 1. AMPBench-MT overview: multi-source evidence, homology-controlled splitting, and endpoint-specific [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Record-level data curation and endpoint routing. Source records are standardized into canonical sequence, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Source-linked endpoint coverage in the released benchmark tables. Cells report provenance-retaining endpoint [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Endpoint-table sizes and homology-controlled train/validation/test allocation. The joint multitask total pools [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: MIC regression comparison on the homology-controlled pMIC test split. Points show MAE with 95% [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Complete binary AMP/non-AMP classification results. All 38 reported model rows are retained and ordered [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Complete species-conditioned MIC, spectrum, and low-toxicity results. The upper matrix retains all 37 MIC [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Complete HC50, selectivity, and joint multitask results. Panel (a) gives HC50 hemolysis regression, panel (b) [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 10 linked inside Pith

  1. [1]

    Marah Abdin et al. 2024. Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone. arXiv:2404.14219

  2. [2]

    Abdelrahman Abouelenin et al. 2025. Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs. arXiv:2503.01743

  3. [3]

    Jianxiu Cai, Jielu Yan, Chonwai Un, Yapeng Wang, Francois-Xavier Campbell-Valois, and Shirley WI Siu. 2025. BERT-AmPEP60: A BERT-based transfer learning approach to predict the minimum inhibitory concentrations of antimicrobial peptides for Escherichia coli and Staphylococcus aureus.Journal of Chemical Information and Modeling65, 7 (2025), 3186–3202

  4. [4]

    Denise B Catacutan, Jeremie Alexander, Autumn Arnold, and Jonathan M Stokes. 2024. Machine learning in preclinical drug discovery.Nature Chemical Biology20, 8 (2024), 960–973

  5. [5]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 785–794

  6. [6]

    Davide Chicco and Giuseppe Jurman. 2020. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation.BMC genomics21, 1 (2020), 6

  7. [7]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient finetuning of quantized LLMs. InAdvances in Neural Information Processing Systems, V ol. 36. 10088–10115

  8. [8]

    Ahmed Elnaggar et al . 2023. Ankh: Optimized protein language model unlocks general-purpose modelling. arXiv:2301.06568

Show all 56 references
  1. [9]

    Ahmed Elnaggar, Michael Heinzinger, Christian Dallago, Ghalia Rehawi, Yu Wang, Llion Jones, Tom Gibbs, Tamas Feher, Christoph Angerer, Martin Steinegger, et al. 2022. ProtTrans: Toward understanding the language of life through self-supervised learning.IEEE transactions on pat...

  2. [10]

    ESM Team. 2024. ESM Cambrian: Revealing the mysteries of proteins with unsupervised learning. Evolution- aryScale Website. December 4, 2024.https://evolutionaryscale.ai/blog/esm-cambrian

  3. [11]

    Noelia Ferruz, Steffen Schmidt, and Birte Höcker. 2022. ProtGPT2 is a deep unsupervised language model for protein design.Nature communications13, 1 (2022), 4348

  4. [12]

    Legana CHW Fingerhut, David J Miller, Jan M Strugnell, Norelle L Daly, and Ira R Cooke. 2020. ampir: an R package for fast genome-wide prediction of antimicrobial peptides.Bioinformatics36, 21 (2020), 5262–5263

  5. [13]

    Ulka Gawde, Shuvechha Chakraborty, Faiza Hanif Waghu, Ram Shankar Barai, Ashlesha Khanderkar, Rishikesh Indraguru, Tanmay Shirsat, and Susan Idicula-Thomas. 2023. CAMPR4: a database of natural and synthetic antimicrobial peptides.Nucleic Acids Research51, D1 (2023), D377–D383

  6. [14]

    Gemma Team. 2025. Gemma 3 Technical Report. arXiv:2503.19786

  7. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The Llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

  8. [16]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature645, 8081 (2025), 633–638

  9. [17]

    Thomas Hayes et al. 2025. Simulating 500 million years of evolution with a language model.Science387, 6736 (2025), 850–858

  10. [18]

    Junjie Huang, Yanchao Xu, Yunfan Xue, Yue Huang, Xu Li, Xiaohui Chen, Yao Xu, Dongxiang Zhang, Peng Zhang, Junbo Zhao, et al. 2023. Identification of potent antimicrobial peptides via a machine-learning pipeline that mines the entire space of peptide sequences.Nature Biomedica...

  11. [19]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. 2024. Qwen2.5-Coder technical report.arXiv preprint arXiv:2409.12186(2024)

  12. [20]

    Jiang et al

    Albert Q. Jiang et al. 2023. Mistral 7B. arXiv:2310.06825

  13. [21]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: A highly efficient gradient boosting decision tree. InAdvances in Neural Information Processing Systems, V ol. 30

  14. [22]

    Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations

  15. [23]

    Anthony Lavertu, Jacques Corbeil, and Pascal Germain. 2026. QMAP: A benchmark for standardized evaluation of antimicrobial peptide MIC and hemolytic activity regression.Scientific Reports(2026). Published online June 3, 2026

  16. [24]

    Yipin Lei, Xu Wang, Meng Fang, Han Li, Xiang Li, and Jianyang Zeng. 2024. PepGB: facilitating peptide drug discovery via graph neural networks.arXiv preprint arXiv:2401.14665(2024)

  17. [25]

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, et al. 2023. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science379, 6637 (2023), 1123–1130

  18. [26]

    Rajat Kumar Mondal, Debarup Sen, Ankish Arya, and Sintu Kumar Samanta. 2023. Developing anti-microbial peptide database version 1 to provide comprehensive and exhaustive resource of manually curated AMPs.Scientific Reports13, 1 (2023), 17843

  19. [27]

    Mohsen Naghavi, Stein Emil V ollset, Kevin S Ikuta, Lucien R Swetschinski, Authia P Gray, Eve E Wool, Gisela Robles Aguilar, Tomislav Mestrovic, Georgia Smith, Chieh Han, et al. 2024. Global burden of bacterial antimicrobial resistance 1990–2021: a systematic analysis with for...

  20. [28]

    Mario Novkovi´c, Juraj Simuni´c, Viktor Bojovi´c, Alessandro Tossi, and Davor Jureti´c. 2012. DADP: the database of anuran defense peptides.Bioinformatics28, 10 (2012), 1406–1407

  21. [29]

    Sebastian Ojeda, Rafael Velasquez, Nicolás Aparicio, Juanita Puentes, Paula Cárdenas, Nicolás Andrade, Gabriel González, Sergio Rincón, Carolina Muñoz-Camargo, and Pablo Arbeláez. 2025. A standardized benchmark for multilabel antimicrobial peptide classification. InAdvances in...

  22. [30]

    Stefano P Piotto, Lucia Sessa, Simona Concilio, and Pio Iannelli. 2012. YADAMP: yet another database of antimicrobial peptides.International journal of antimicrobial agents39, 4 (2012), 346–351. 12 AMPBench-MTARXIVPREPRINT

  23. [31]

    Malak Pirtskhalava, Anthony A Amstrong, Maia Grigolava, Mindia Chubinidze, Evgenia Alimbarashvili, Boris Vishnepolsky, Andrei Gabrielian, Alex Rosenthal, Darrell E Hurt, and Michael Tartakovsky. 2021. DBAASP v3: database of antimicrobial/cytotoxic activity and structure of pep...

  24. [32]

    Francois Pognan, Mario Beilmann, Harrie CM Boonen, Andreas Czich, Gordon Dear, Philip Hewitt, Tomas Mow, Teija Oinonen, Adrian Roth, Thomas Steger-Hartmann, et al. 2023. The evolving role of investigative toxicology in the pharmaceutical industry.Nature reviews drug discovery2...

  25. [33]

    Liudmila Prokhorenkova, Gleb Gusev, Aleksandr V orobev, Anna Veronika Dorogush, and Andrey Gulin. 2018. CatBoost: Unbiased boosting with categorical features. InAdvances in Neural Information Processing Systems, V ol. 31

  26. [34]

    Qwen Team. 2024. Qwen2.5 Technical Report. arXiv:2412.15115

  27. [35]

    Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C Lawrence Zitnick, Jerry Ma, et al. 2021. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences.Proceedings of the nati...

  28. [36]

    Guobang Shi, Xinyue Kang, Fanyi Dong, Yanchao Liu, Ning Zhu, Yuxuan Hu, Hanmei Xu, Xingzhen Lao, and Heng Zheng. 2022. DRAMP 3.0: an enhanced comprehensive data repository of antimicrobial peptides.Nucleic acids research50, D1 (2022), D488–D496

  29. [37]

    Katarzyna Sidorczuk, Przemysław Gagat, Filip Pietluch, Jakub Kała, Dominik Rafacz, Laura B ˛ akała, Jadwiga Słowik, Rafał Kolenda, Stefan Rödiger, Legana CHW Fingerhut, et al. 2022. Benchmarks in antimicrobial peptide prediction are biased due to the selection of negative data...

  30. [38]

    Sandeep Singh, Kumardeep Chaudhary, Sandeep Kumar Dhanda, Sherry Bhalla, Salman Sadullah Usmani, Ankur Gautam, Abhishek Tuknait, Piyush Agrawal, Deepika Mathur, and Gajendra PS Raghava. 2016. SATPdb: a database of structurally annotated therapeutic peptides.Nucleic acids resea...

  31. [39]

    Martin Steinegger and Johannes Söding. 2017. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets.Nature biotechnology35, 11 (2017), 1026–1028

  32. [40]

    Jin Su, Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou, and Fajie Yuan. 2024. SaProt: Protein language modeling with structure-aware vocabulary. InInternational Conference on Learning Representations

  33. [41]

    Zhelyazko Terziyski, Margarita Terziyska, Ivelina Deseva, Stanka Hadzhikoleva, Albert Krastanov, Dasha Mihaylova, and Emil Hadzhikolev. 2023. PepLab platform: Database and software tools for analysis of food- derived bioactive peptides.Applied Sciences13, 2 (2023), 961

  34. [42]

    The UniProt Consortium. 2025. UniProt: the universal protein knowledgebase in 2025.Nucleic acids research53, D1 (2025), D609–D617

  35. [43]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InAdvances in Neural Information Processing Systems, V ol. 30

  36. [44]

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph attention networks. InInternational Conference on Learning Representations

  37. [45]

    Daniel Veltri, Uday Kamath, and Amarda Shehu. 2018. Deep learning improves antimicrobial peptide recognition. Bioinformatics34, 16 (2018), 2740–2747

  38. [46]

    Fangping Wan, Felix Wong, James J Collins, and Cesar de la Fuente-Nunez. 2024. Machine learning for antimicrobial peptide identification and design.Nature Reviews Bioengineering2, 5 (2024), 392–407

  39. [47]

    Guangshun Wang, Cindy Schmidt, Xia Li, and Zhe Wang. 2026. APD6: the antimicrobial peptide database is expanded to promote research and development by deploying an unprecedented information pipeline.Nucleic Acids Research54, D1 (2026), D363–D374

  40. [48]

    Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, et al. 2023. Scientific discovery in the age of artificial intelligence.Nature620, 7972 (2023), 47–60

  41. [49]

    Jacob Witten and Zack Witten. 2019. Deep learning regression model for antimicrobial peptide design.BioRxiv (2019), 692681

  42. [50]

    Felix Wong, Erica J Zheng, Jacqueline A Valeri, Nina M Donghia, Melis N Anahtar, Satotaka Omori, Alicia Li, Andres Cubillos-Ruiz, Aarti Krishnan, Wengong Jin, et al. 2024. Discovery of a structural class of antibiotics with explainable deep learning.Nature626, 7997 (2024), 177...

  43. [51]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2019. How powerful are graph neural networks?. In International Conference on Learning Representations

  44. [52]

    Ke Yan, Meijing Li, Shutao Chen, Tianyi Liu, Jing Hao, Bin Liu, and Zhen Li. 2026. PepLM-GNN: A graph neural network framework leveraging pre-trained language models for peptide-protein binding prediction.PLOS Computational Biology22, 3 (2026), e1014084

  45. [53]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

  46. [54]

    Lantian Yao, Jiahui Guan, Peilin Xie, Chia-Ru Chung, Zhihao Zhao, Danhong Dong, Yilin Guo, Wenyang Zhang, Junyang Deng, Yuxuan Pang, et al. 2025. dbAMP 3.0: updated resource of antimicrobial activity and structural annotation of peptides in the post-pandemic era.Nucleic acids ...

  47. [55]

    Jiahui Zhang, Rouyi Wang, Kuangqi Zhou, Tianshu Xiao, Lingyan Zhu, Yaosen Min, and Yang Wang. 2026. PepBenchmark: A standardized benchmark for peptide machine learning. InInternational Conference on Learning Representations

  48. [56]

    Kang Zhang, Xin Yang, Yifei Wang, Yunfang Yu, Niu Huang, Gen Li, Xiaokun Li, Joseph C Wu, and Shengyong Yang. 2025. Artificial intelligence in drug development.Nature medicine31, 1 (2025), 45–59. 14 AMPBench-MTARXIVPREPRINT Appendix.The appendix begins with the reproducibility...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.