amp-plm: peptide corpus for ESM-C masked-flow training

Two tables of 8-50 residue peptides, each row carrying its source database and precomputed charge, length and hydrophobicity buckets. They are the training data behind the ESM-C 300M generator of the AMP Challenge 2027 submission (code, weights).

File	Rows	Size	Content
stage_a_prior.parquet	5,634,791	182 MB	the full corpus, a broad peptide background
stage_b_amp.parquet	889,293	29 MB	the rows of stage A from AMPSphere and the curated peptide databases

The generator is trained in three stages. This dataset holds stages A and B. Stage C (4,460 measured-potent peptides) is rebuilt from data/oracle/ by the code repository.

Who this is for

• Training or fine-tuning a peptide language model. Use stage A as a broad prior and stage B to adapt toward AMP-like sequences. The tables are plain sequences, so they fit masked-LM or any other sequence objective.
• Building a property-controlled generator. b_charge, b_length and b_hydro are ready-made control tokens (ids 0-6, edges documented below), so you can condition on net charge, length or hydrophobicity without recomputing them.
• Making bucket-matched decoys. Stage A minus stage B is 4.7M peptides that are not from AMP databases. Because every row has the same buckets, you can draw background sequences that match an AMP set on charge and length.
• Checking novelty for the AMP Challenge 2027. novel_ok marks rows that pass a screen against the challenge's antibacterial.fasta reference, so you can drop near-duplicates of it before training.
• Reproducing the submission. The training script in the code repository reads these two files directly.

It is not for predicting activity. There are no MIC labels here (see Limitations).

Quick start

Load a table with datasets. Stage A is 5.6M rows, so stream it if you only need a look:

from datasets import load_dataset

stage_b = load_dataset("eamag/amp-plm", "stage_b_amp", split="train")
stage_a = load_dataset("eamag/amp-plm", "stage_a_prior", split="train", streaming=True)
print(next(iter(stage_a)))

Or read a parquet file with pandas. This selects the clean cationic subset of stage B: peptides of 13-26 residues, net charge above 4, and no novelty flag (24,498 rows):

import pandas as pd

b = pd.read_parquet("hf://datasets/eamag/amp-plm/stage_b_amp.parquet")
subset = b[(b.b_charge >= 3) & b.b_length.between(1, 3) & b.novel_ok]

To train with the code repository, download the parquet files into data/plm/:

hf download eamag/amp-plm --repo-type dataset --include "*.parquet" --local-dir data/plm

Which table to use

• Stage A for a general peptide prior, or as the pool for decoys. It mixes AMP-like and non-AMP sources (next section), so filter on source.
• Stage B for AMP-like training data. It is 94% AMPSphere (836,962 of 889,293 rows), whose positives are classifier-predicted, not validated. The curated databases add about 52,000 rows.
• Stage B is skewed long: 49% of it is 35-44 residues (length bucket 5) and only 1.6% is 12 residues or shorter. If you want short peptides, select on b_length.

Columns

Both files have the same columns: sequence, source, b_charge, b_length, b_hydro, is_amp_like, novel_ok, max_lev_to_reference. All sequences use the 20 canonical amino acids, and neither file has a duplicate sequence.

• source: where the row came from. Stage B is a pure filter of stage A on this column: ampsphere, marlys, grampa, neuropep, conoserver, cppsite2, dbaasp, dramp_general, dramp_clinical, omegamp_pos.
• b_charge, b_length, b_hydro: control-token bucket ids 0-6. Charge is the Bjellqvist net charge at pH 7 with edges 0, 2, 4, 6, 8, 11. Length edges are 12, 16, 20, 26, 34, 44. Hydrophobicity is the mean Eisenberg value with edges -0.45, -0.25, -0.10, 0.05, 0.25, 0.50. A bucket is the number of edges the value exceeds, and a value equal to an edge falls in the lower bucket. The code is esmc_flow_amp.controls in the code repository.
• novel_ok: False for rows that fail the novelty screen against the AMP Challenge 2027 antibacterial.fasta reference (39,448 sequences): an exact match, or a similarity above the gate. Filter on it for the clean subset.
• max_lev_to_reference: the censored decision flag behind that screen. 0.0 means "did not exceed 0.80", never a distance. Do not use it as a feature or a similarity value.
• is_amp_like: a heuristic flag inherited from the first corpus build. It is not a label for antimicrobial peptides. In stage A it is True for 1,861,298 rows, and 719,836 of them are omegamp_neg (the OmegAMP negative set) and 297,764 are plain uniprot. Use source to select what you want. In stage B the build script sets it to True for every row.

Values in shared columns are identical in both files, except for is_amp_like.

Sources and licences

The source column names where each row came from. Licences are recorded, not resolved. Some sources are academic-use or have no licence file, so check the terms before redistributing or using them commercially.

source	Stage A rows	Stage B rows	Terms
ampsphere	836,962	836,962	CC-BY-4.0. AMPSphere v2022-03, Santos-Júnior et al., Cell 2024, Zenodo doi:10.5281/zenodo.4574468. Positives are classifier-predicted, not validated
marlys	39,767	39,767	CC0. MarLys/MLAMP merge, doi:10.17632/w4hb5grjwb.3
grampa	6,060	6,060	no licence file upstream. Included for completeness and flagged by source; filter it out if your policy needs an explicit grant
neuropep	3,541	3,541	academic research, cite the paper
conoserver	1,366	1,366	academic research, cite the paper
cppsite2	1,057	1,057	academic research, cite the paper
dbaasp	286	286	redistribution without restriction, with acknowledgement (below). Pirtskhalava et al., NAR 2021
dramp_general	246	246	CC-BY-4.0
dramp_clinical	4	4	CC-BY-4.0
omegamp_pos	4	4	MIT
omegamp_neg	722,332	0	MIT
smprot	1,502	0	academic research, cite the paper
toxprot	2,312	0	CC-BY-4.0, UniProt Consortium
protein_tiling	798,409	0	CC-BY-4.0, windows tiled from Swiss-Prot, UniProt Consortium
uniprot, uniprot_remaining, uniprot_transmem, uniprot_signal, uniprot_propep, uniprot_transit	3,220,943	0	CC-BY-4.0, UniProt Consortium

DBAASP acknowledgement: "Data were obtained from the DBAASP (https://dbaasp.org), an open-access AMP data resource supported by I. Beritashvili Center of Experimental Biomedicine (IBCEB), Tbilisi, Georgia and NIAID OCICB, Bethesda, MD."

Limitations

• There are no activity (MIC) labels here. The tables support unsupervised, physicochemically conditioned training only.
• No row is confirmed inactive. omegamp_neg, uniprot and protein_tiling are background, not experimentally tested negatives, so treat decoys drawn from them as unlabelled.
• Stage B still contains rows that fail the novelty screen: 40,677 of 889,293, mostly marlys (33,123 of 39,767) and grampa (5,649 of 6,060). Filter on novel_ok if you need sequences that are dissimilar to the reference.
• omegamp_neg is 12.8% of stage A. Stage B excludes it, and so should any training aimed at antimicrobial activity.