ACD — Anti-Cancer Peptide Dataset

ACD is a three-stage, cascading benchmark for anti-cancer peptide (ACP) research. It combines peptide sequences and provenance with precomputed conventional, molecular, sequence, and structure representations for 18,611 unique peptides.

The three prediction stages are:

1. Stage 1 — ACP identification: distinguish ACPs from non-ACPs.
2. Stage 2 — target-annotation identification: among ACPs, distinguish peptides with a retained cancer-site annotation from those without one.
3. Stage 3 — cancer-site prediction: assign one or more of six cancer-site labels: breast, cervix, colon, lung, prostate, and skin.

Stages are nested by design: all 2,326 Stage-2 peptides are Stage-1 positives, and all 733 Stage-3 peptides are Stage-2 positives. The same peptide ID is used across every table and precomputed modality.

Research-use warning: ACD is a computational research benchmark, not a clinical resource. Labels aggregated from databases and publications do not establish efficacy, selectivity, toxicity, pharmacokinetics, or clinical suitability. Predicted structures and learned representations are not experimental measurements.

Dataset summary

Stage	Task	Train	Test	Labels
Stage 1	Binary ACP vs non-ACP	8,524	10,087	4,262/4,262 train; 1,152/8,935 test (positive/negative)
Stage 2	Binary retained target annotation vs no retained annotation, within ACPs	1,208	1,118	604/604 train; 129/989 test (positive/negative)
Stage 3	Six-label cancer-site classification	604	129	Multi-label lists over six sites

Sequence lengths range from 5 to 50 standard amino-acid residues. The main Stage-1 tables contain 18,611 unique sequences in total. Train and test sets share no MMseqs2 cluster within any stage.

Stage-3 label frequencies

Counts are per label, so a multi-label peptide contributes to more than one row.

Label	Train	Test
breast	416	83
cervix	178	47
colon	229	50
lung	206	39
prostate	192	41
skin	192	33

Across the 733 Stage-3 peptides, 436 have one label and 297 (40.5%) have multiple labels.

Quick start

This repository is public but manually gated. Request access on the dataset page, then authenticate or pass a token that has been granted access.

from datasets import load_dataset

# Default configuration: Stage 1
stage1 = load_dataset("tanthinhdt/acd", "stage1", token=True)

# Multi-label Stage 3
stage3 = load_dataset("tanthinhdt/acd", "stage3", token=True)

# Leakage-sensitive, benchmark-free Stage-1 test set
stage1_bf = load_dataset(
    "tanthinhdt/acd",
    "stage1_benchmark_free",
    token=True,
)["test"]

Load a local Parquet table without the datasets library:

import polars as pl

train = pl.read_parquet("s1_train.parquet")

Download only selected precomputed artifacts:

hf download tanthinhdt/acd \
  --type dataset \
  --include "artifacts/biot5-plus-base/**/*.safetensors" \
  --local-dir acd

Configurations and evaluation subsets

Hub configuration	Backing file	Rows	Purpose
stage1	s1_{train,test}.parquet	8,524 / 10,087	Primary Stage-1 benchmark
stage2	s2_{train,test}.parquet	1,208 / 1,118	Primary Stage-2 benchmark
stage3	s3_{train,test}.parquet	604 / 129	Primary Stage-3 benchmark
stage1_benchmark_free	subsets/s1_test_bf.parquet	9,369	No exact sequence match to the external training-set union
stage3_benchmark_free	subsets/s3_test_bf.parquet	60	No exact sequence match to the external organ benchmark

The benchmark-free union was built from the training sequences of AntiCP2-Main, AntiCP2-Alternate, ACP-Mixed-80, mACPpred, mACPpred-2, MLACP-2, ACP240, ACP740, ACPred, ACP-ML, ACPredictor, CAPTURE, LEE, ACP500, AntiCP3, AI4ACP, and ACP-MLC-S2.

FASTA equivalents are included beside every Parquet table. Their headers use >{id}|{label}; Stage-3 labels are comma-separated.

Data fields

The six main Parquet tables have the following fields. Training files include fold; test files do not.

Field	Type	Description
id	string	Stable zero-padded six-digit identifier used to join all files and modalities. Treat it as a string.
sequence	string	Peptide sequence over the 20 standard amino-acid alphabet.
label	string or list[string]	positive/negative in Stages 1–2; one or more cancer-site labels in Stage 3.
cluster	int64	Dataset-wide MMseqs2 sequence-cluster identifier. The integer itself has no biological ordering.
is_natural	bool	Source-derived flag indicating a naturally occurring peptide.
is_disulfide	bool	Source-derived flag indicating disulfide bonding; such peptides were retained when their alphabet was otherwise supported.
is_amp	bool	Whether any source annotation marked the sequence as antimicrobial. This is independent of the ACP task label.
source	string	Retained source key: tpdb or apd6. When the same sequence occurred in multiple records, this is not necessarily exhaustive provenance.
reference	string	Retained or merged bibliographic/database/patent provenance text from the source records.
fold	int64	Cluster-aware cross-validation fold in [0, 4]; present only in training tables.

There are no null values in the released Parquet tables.

Repository layout

.
├── README.md
├── s1_train.parquet / s1_train.fasta
├── s1_test.parquet  / s1_test.fasta
├── s2_train.parquet / s2_train.fasta
├── s2_test.parquet  / s2_test.fasta
├── s3_train.parquet / s3_train.fasta
├── s3_test.parquet  / s3_test.fasta
├── subsets/                         # benchmark-free test views
├── confeat/macppred2.parquet        # 163 conventional descriptors per ID
└── artifacts/                       # first-three-ID-digit sharding for Hub limits
    ├── structure/<prefix>/<id>.pdb  # ESM-3-predicted structures
    ├── protein/<prefix>/<id>.pkl    # serialized TorchDrug Protein objects
    ├── biot5-plus-base/<prefix>/<id>.safetensors
    └── esmgearnet-siamdiff/<prefix>/<id>.safetensors

The release occupies approximately 6.4 GB on disk and contains one item per modality for each of the 18,611 unique IDs. Artifact files are sharded by the first three digits of the six-digit ID to stay below the Hugging Face Hub per-directory file limit. For example, peptide 000006 is stored under prefix 000. The ID sets in the main sequence tables, conventional-feature table, PDB collection, pickle collection, and both embedding collections were checked for exact equality before upload.

Precomputed modalities

Conventional peptide descriptors

confeat/macppred2.parquet contains 18,611 rows and 164 columns: id plus 163 numeric features.

Family	Features	Description
AAC	20	Amino-acid composition
APAAC	26	Amphiphilic pseudo-amino-acid composition
PAAC	23	Pseudo-amino-acid composition
QSOrder	46	Quasi-sequence-order descriptors
PDE	48	Peptide descriptor encoding / physicochemical features

The AAC, APAAC, PAAC, and QSOrder implementations reproduce iFeatureOmegaCLI 1.0.2 defaults (nlag=3, lambda=3, weight=0.05); PDE values were computed with peptidy.

Molecular embeddings

Each artifacts/biot5-plus-base/<prefix>/<id>.safetensors file contains:

Key	dtype	Shape	Description
embedding	float32	[768]	BioT5+ encoder representation

Peptides were converted with RDKit from sequence to a molecule, canonicalized to SMILES, encoded as SELFIES, and passed through QizhiPei/biot5-plus-base. Long SELFIES streams were encoded in overlapping windows and combined by a token-count-weighted mean.

Sequence and structure embeddings

Each artifacts/esmgearnet-siamdiff/<prefix>/<id>.safetensors file contains four float32 tensors, where L is sequence length:

Key	Shape	Description
seq_graph_feat	[1280]	Mean-pooled ESM-2-650M sequence representation
seq_node_feat	[L, 1280]	Per-residue ESM-2-650M representation
struct_graph_feat	[3072]	Mean-pooled six-layer GearNet representation
struct_node_feat	[L, 3072]	Per-residue concatenated GearNet representation

The sequence encoder was ESM-2-650M. The structural encoder was a six-layer GearNet initialized from the SiamDiff checkpoint distributed through Zenodo record 10034578.

Read the safe tensor files without loading arbitrary Python objects:

from safetensors.torch import load_file

mol = load_file("artifacts/biot5-plus-base/000/000006.safetensors")
joint = load_file("artifacts/esmgearnet-siamdiff/000/000006.safetensors")
print(mol["embedding"].shape)          # torch.Size([768])
print(joint["struct_graph_feat"].shape)  # torch.Size([3072])

Predicted structures and TorchDrug objects

artifacts/structure/<prefix>/<id>.pdb contains a structure predicted from the sequence with the open ESM-3 model using structure-track generation. These are predictions, not experimentally solved structures, and should not be interpreted as validated conformations.

artifacts/protein/<prefix>/<id>.pkl contains the corresponding PDB converted to a TorchDrug Protein object through RDKit. The object format depends on compatible Python, PyTorch, and TorchDrug versions.

Pickle safety: Python pickle files can execute code when deserialized. Load artifacts/protein/**/*.pkl only if you trust this repository and its provenance. Prefer the PDB or SafeTensors representations when interoperability or untrusted-input safety matters.

Construction and preprocessing

Source aggregation

The source snapshots were drawn from:

• TPDB — Therapeutic Peptide Database
• APD6 — Antimicrobial Peptide Database

Rows were deduplicated using sequence and annotation identity. Duplicate provenance records were resolved by removing known placeholder references where possible, preferring a unique PMID-bearing row, or merging distinct reference strings. Records were then collapsed to one row per sequence.

A sequence was marked ACP-positive when any retained record had a first-degree anticancer annotation. The AMP flag similarly records whether any retained record had a first-degree antimicrobial annotation. Cancer-site labels came from the distinct second-degree annotations attached to anticancer records. antiangiogenesis and antilivercancer were excluded from Stage 3; the latter was too sparse for a reliable label.

Only sequences of 5–50 residues were retained. Cyclic peptides, D-peptides, and sequences with non-standard amino acids were excluded because the molecular feature pipeline could not represent them faithfully. Standard-amino-acid disulfide peptides were retained and marked with is_disulfide.

Split construction and leakage control

All retained sequences were clustered together with MMseqs2 easy-cluster using minimum sequence identity 0.4, coverage 0.8, cov-mode=1, and cluster-mode=0. Whole clusters—not individual sequences—were assigned to train or test. The Stage-3 universe was multilabel-stratified by cancer-site set; the remaining universe was stratified by ACP status. A single cluster assignment was propagated through all three nested stages.

Before negative sampling, CD-HIT-2D removed negative examples similar to positives at 90% identity for Stage 1 and 75% for Stage 2. Stage-1 and Stage-2 training sets were then balanced 1:1 with length-bin-matched negative sampling. Stage-1 training negatives were additionally sampled with an antimicrobial-annotation cap of 50%. Test sets retain their natural post-filter imbalance.

Five training folds were assigned at cluster level: stratified by the binary label in Stages 1–2 and multilabel-stratified by the cluster-level union of Stage-3 labels. No cluster spans folds.

Recommended evaluation

• Use the provided train/test split; random row-level splitting can reintroduce sequence-similarity leakage and break the cascading design.
• Use the supplied fold values for five-fold cross-validation instead of assigning new row-level folds.
• Report class-aware metrics for the imbalanced test sets: MCC, balanced accuracy, sensitivity, specificity, ROC-AUC, and AUPRC for Stages 1–2; per-label and macro/micro multilabel metrics for Stage 3.
• For comparisons with previously published classifiers, prefer the common stage1_benchmark_free or stage3_benchmark_free subset.
• Do not train on subsets/; they are derived evaluation views of the test data.

Biases, limitations, and appropriate use

• Database and literature bias: examples reflect what TPDB/APD6 and their cited literature contain, including publication, assay, organism, and curation biases.
• Annotation is not assay normalization: binary and site labels aggregate heterogeneous evidence. They do not encode assay protocol, dose, cell line, potency, toxicity, or confidence in a standardized form.
• Negative labels are task-relative: a Stage-1 negative means no retained ACP annotation; it is not proof that the peptide lacks anticancer activity. A Stage-2 negative is an ACP without a retained target label, not proof of target-independence.
• Multi-function overlap: many ACPs are also annotated as AMPs. In Stage 1, 4,379 of 5,414 ACP positives are AMP-annotated; models may learn biology shared across these activities.
• Label imbalance: the full test sets, especially Stages 1–2, are imbalanced. Accuracy alone is not informative.
• Sequence scope: peptides are limited to 5–50 residues and a standard alphabet; excluded chemistries may behave differently.
• Structural uncertainty: ESM-3 PDBs and all downstream structure features are computational predictions. No per-peptide experimental validation is implied.
• Representation dependence: precomputed embeddings inherit the assumptions, training data, versions, and licensing terms of BioT5+, ESM-2, GearNet, SiamDiff, ESM-3, and their software stacks.
• Provenance compression: source and reference are retained summary fields after deduplication and may not enumerate every database occurrence of a sequence.
• Temporal snapshot: upstream databases evolve. This release does not automatically incorporate later corrections or additions.

Appropriate uses include benchmark development, sequence- and structure-aware ACP modeling, representation learning, ablation studies, and method comparison. It is not appropriate as the sole basis for clinical, diagnostic, therapeutic, or laboratory safety decisions.

Reproducibility

Processing and feature-extraction code is maintained in the Therapep repository, notably notebooks/process_acd.ipynb, scripts/extract_features.sh, and the extraction tools under tools/.

Randomized split and sampling procedures used seed 42. Feature files are keyed by the stable id column, so modalities can be joined without relying on table row order.

Licensing and terms of use

No dataset-wide license has been declared for this ACD release. Absence of a license does not grant permission to redistribute or reuse every component without restriction. Users are responsible for reviewing and complying with the source-database terms, cited publications, and the licenses of the models and software used to produce the derived artifacts. In particular, APD6 publishes its own download terms.

Citation

An ACD-specific paper citation has not yet been supplied. If you use this release, cite the dataset repository and the applicable upstream resources and representation models. The per-row reference field preserves source-level citation text where available.

Suggested access citation until a paper citation is available:

Tanthinhdt. ACD — Anti-Cancer Peptide Dataset. Hugging Face Datasets.
https://huggingface.co/datasets/tanthinhdt/acd

Relevant resources include:

• Wang G, Schmidt C. APD6: the antimicrobial peptide database is expanded to promote research and development by deploying an unprecedented information pipeline. Nucleic Acids Research. 2026;54(D1):D363–D374. doi:10.1093/nar/gkaf860
• Pei Q, et al. BioT5+: Towards Generalized Biological Understanding with IUPAC Integration and Multi-task Tuning. arXiv:2402.17810
• Steinegger M, Söding J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature Biotechnology. 2017;35:1026–1028. doi:10.1038/nbt.3988
• Li W, Godzik A. Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics. 2006;22(13):1658–1659. doi:10.1093/bioinformatics/btl158

Contact

For questions or corrections, open a discussion on this Hugging Face dataset repository or an issue in the Therapep code repository.