AMPBench-MT

AMPBench-MT is a homology-controlled benchmark for antimicrobial peptide endpoint prediction. The release is dated 2026-07-08.

Repository: https://huggingface.co/datasets/ZihengZhou06/AMPBench-MT

The benchmark is organized around endpoint-aware prediction rather than binary AMP recognition alone. It contains processed task tables for AMP/non-AMP classification, species-conditioned MIC regression, activity spectrum positive-evidence audits, low-toxicity classification, HC50 hemolysis regression, selectivity regression, and a joint multitask endpoint table. Low toxicity, HC50, and selectivity are safety-proxy readouts rather than clinical safety labels. All released peptide sequences use uppercase symbols from the 20 standard amino acids.

Authors

• Ziheng Zhou, Shanghai Ocean University, Shanghai, China, zihengzhouac@outlook.com
• Huiyu Luo, Tongji University, Shanghai, China, 2211285@tongji.edu.cn
• Xiaohu Zhu, Center for Safe AGI, Shanghai, China, aleph@csagi.org
• Nan Wang, DP Technology, Shanghai, China, wangnan411570@gmail.com
• Xuebiao Qin, Shanghai Ocean University, Shanghai, China, xbqin@shou.edu.cn
• Chaoyan Zhang, Shanghai Ocean University, Shanghai, China, chyzhang@shou.edu.cn
• Jun Yan, Shanghai Ocean University, Shanghai, China, yanjun@ieee.org (corresponding author)

Dataset Tasks

Config	Rows	Target	Main input fields	Splits
binary_amp	60,946	binary_label	sequence	train / validation / test
mic_regression	73,656	pmic_median	sequence, species	train / validation / test
mic_unit_sensitivity	65,658	pmic_median	sequence, species	all
spectrum	37,022	label_active	sequence, group_label	train / validation / test
spectrum_consensus	30,048	label_active	sequence, group_label	train / validation / test
low_toxicity	7,143	low_toxicity_label	sequence, cell_category	train / validation / test
hc50_hemolysis	1,595	pHC50_median	sequence, cell_category	train / validation / test
selectivity	9,774	log10_selectivity_index	sequence, species, cell_category	train / validation / test
joint_multitask	119,416	target	sequence, condition, task_name	train / validation / test

The all.csv files are retained for audit and convenience. For model evaluation, use the split files listed in the dataset card metadata. The mic_unit_sensitivity configuration is an all-only direct-unit sensitivity diagnostic and was not used as a separate train/validation/test model-evaluation split.

File Layout

• data/<task>/all.csv: full released table for each task.
• data/<task>/train.csv, validation.csv, test.csv: homology-controlled splits for tasks with model-evaluation splits.
• metadata/task_overview.csv: task-level row counts, targets, inputs, and split sizes.
• metadata/schema.json: column-level schema and descriptions.
• metadata/split_summary.json: split and cluster summaries.
• metadata/source_databases.json: source database provenance and conservative source-term audit.
• metadata/validation_report.json: package-level validation flags for row counts, splits, sequence alphabet, and cluster disjointness.
• dataset_infos.json: Hugging Face style configuration metadata.
• MANIFEST.json and metadata/checksums.sha256: file inventory and SHA256 checksums.

Loading Examples

Load the public Hub release with its namespace-qualified repository id:

from datasets import load_dataset

REPO_ID = "ZihengZhou06/AMPBench-MT"
mic = load_dataset(REPO_ID, "mic_regression")
mic_train = mic["train"]

The same tables can be loaded directly from the local release package:

import pandas as pd

mic_train = pd.read_csv("data/mic_regression/train.csv")
sequences = mic_train["sequence"]
species = mic_train["species"]
targets = mic_train["pmic_median"]

Split and Evaluation Protocol

AMPBench-MT uses MMseqs2 30% sequence-identity cluster splits with coverage 0.8, cov-mode 0, cluster-mode 2, and seed 42. The binary task uses the historical 10-50 amino-acid length range. Assay-derived endpoint tasks use the 5-100 amino-acid range and retain endpoint-specific observation units rather than forcing every peptide into every endpoint. In the CSV split column, the validation partition is encoded as valid; Hugging Face metadata and filenames expose the same partition as validation and validation.csv.

The benchmark should be evaluated with endpoint-appropriate metrics. Binary classification uses threshold-aware and ranking-aware metrics. The spectrum configurations are positive-evidence audits rather than ordinary balanced classification tasks: missing target-group evidence is not treated as a confirmed negative, so AUPRC and F1 should be interpreted alongside MCC, balanced accuracy, and negative-class recall (specificity). MIC, HC50, and selectivity regression use error and rank/correlation metrics. The joint multitask table records the corresponding single-task split in single_task_split and the multitask split in split.

For concentration endpoints, pMIC = -log10(MIC in mol/L) and pHC50 = -log10(HC50 in mol/L); equivalently, values measured in micromolar units use 6 - log10(value in uM). The selectivity target is log10(HC50/MIC) = pMIC - pHC50. The conversion_methods field distinguishes direct micromolar records from mass-concentration records converted with an estimated unmodified peptide molecular weight.

Data Sources and Provenance

The released provenance fields contain 13 named source resources and dataset collections. The table below lists every source name found in source_databases_public across the released task tables, the task groups in which it appears, and its recommended source citation. MIC includes the main regression table and the direct-unit sensitivity diagnostic; spectrum includes the strict and consensus configurations. A merged row may retain more than one source name.

Source resource	Released task coverage	Recommended citation
UniProtKB/Swiss-Prot	Binary AMP classification	The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research 53(D1), D609-D617 (2025).
SATPdb	Binary AMP classification; spectrum	Singh et al. SATPdb: a database of structurally annotated therapeutic peptides. Nucleic Acids Research 44(D1), D1119-D1126 (2016).
DRAMP 3.0	Binary AMP classification; MIC; spectrum; selectivity	Shi et al. DRAMP 3.0: an enhanced comprehensive data repository of antimicrobial peptides. Nucleic Acids Research 50(D1), D488-D496 (2022).
CAMPR4	Binary AMP classification; MIC; spectrum; selectivity	Gawde et al. CAMPR4: a database of natural and synthetic antimicrobial peptides. Nucleic Acids Research 51(D1), D377-D383 (2023).
DBAASP	Binary AMP classification; MIC; spectrum; low toxicity; HC50; selectivity	Pirtskhalava et al. DBAASP v3: database of antimicrobial/cytotoxic activity and structure of peptides as a resource for development of new therapeutics. Nucleic Acids Research 49(D1), D288-D297 (2021).
dbAMP 3.0	Binary AMP classification; spectrum	Yao et al. dbAMP 3.0: updated resource of antimicrobial activity and structural annotation of peptides in the post-pandemic era. Nucleic Acids Research 53(D1), D364-D376 (2025).
AMPDB	Binary AMP classification	Mondal et al. Developing anti-microbial peptide database version 1 to provide comprehensive and exhaustive resource of manually curated AMPs. Scientific Reports 13, 17843 (2023).
GRAMPA	Binary AMP classification; MIC; spectrum; selectivity	Witten and Witten. Deep learning regression model for antimicrobial peptide design. bioRxiv 692681 (2019).
APD6	Binary AMP classification; MIC; spectrum; selectivity	Wang et al. APD6: the antimicrobial peptide database is expanded to promote research and development by deploying an unprecedented information pipeline. Nucleic Acids Research 54(D1), D363-D374 (2026).
PEP-Lab	Binary AMP classification; spectrum	Terziyski et al. PepLab Platform: Database and Software Tools for Analysis of Food-Derived Bioactive Peptides. Applied Sciences 13(2), 961 (2023).
EC-SA 2025 AMP regression collection	MIC; spectrum; selectivity	Cai et al. 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 Modeling 65(7), 3186-3202 (2025).
YADAMP	MIC; spectrum; selectivity	Piotto et al. YADAMP: yet another database of antimicrobial peptides. International Journal of Antimicrobial Agents 39(4), 346-351 (2012).
DADP	MIC; spectrum; selectivity	Novkovic et al. DADP: the database of anuran defense peptides. Bioinformatics 28(10), 1406-1407 (2012).

Source names, evidence URLs, and source-term status are documented in metadata/source_databases.json. Some historical assay rows do not have a recoverable public assay-source name and therefore leave source_databases_public and source_ids blank. These blanks mean that no public assay-source label is available for the row; they do not indicate a missing endpoint target. In the selectivity table, these source fields describe the MIC component, while toxicity_hc50_label_id links the HC50 component. Missing endpoint evidence is not converted into a negative, low-risk, or non-hemolytic label. A peptide appears in an endpoint only when the corresponding evidence satisfies the documented endpoint-specific inclusion rules.

Acknowledgments

AMPBench-MT gratefully acknowledges the curators, maintainers, authors, and contributors of the source resources and collections listed above. Their work in collecting, standardizing, and documenting peptide sequences and assay evidence made this benchmark possible. Users should cite both AMPBench-MT and the original resources relevant to their analyses, consult metadata/source_databases.json, and comply with each source's current terms. AMPBench-MT does not claim ownership of source database records, and attribution does not replace permission where a source requires it.

License and Use

The Hugging Face metadata uses license: other because the benchmark is a compiled derivative package from multiple public sources with mixed, restrictive, or unclear source terms. The local package uses the AMPBench-MT Research and Review Use License in LICENSE.

The compiled benchmark is intended for non-commercial academic research, manuscript review, reproducible evaluation, and citation-backed benchmarking. Commercial use, broad redistribution, or relicensing of the compiled package should not be assumed from this release. Users must also comply with the terms of the original source databases listed in metadata/source_databases.json.

Limitations

AMPBench-MT is an in silico benchmark. It is not a wet-lab validation set and should not be interpreted as evidence that a peptide is a therapeutic candidate. The release is restricted to unmodified canonical sequences and does not cover D-amino acids, terminal modifications, cyclization, non-canonical residues, amidation, salt forms, formulation, delivery, pharmacokinetics, or clinical utility. The split protocol controls sequence identity at the cluster level but is not a source-held-out, species-held-out, publication-held-out, assay-held-out, or external-database validation design. Protein language model and large language model pretraining corpora are not audited for overlap with AMPBench-MT records.

Citation

If you use AMPBench-MT, cite this dataset release. A paper citation can be added after the manuscript record is available.

@misc{ampbenchmt2026,
  title = {AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction},
  author = {Zhou, Ziheng and Luo, Huiyu and Zhu, Xiaohu and Wang, Nan and Qin, Xuebiao and Zhang, Chaoyan and Yan, Jun},
  year = {2026},
  url = {https://huggingface.co/datasets/ZihengZhou06/AMPBench-MT},
  note = {Dataset release dated 2026-07-08}
}