---
license: other
license_name: dbaasp-derived
license_link: https://dbaasp.org/
tags:
  - antimicrobial-peptides
  - bioinformatics
  - protein-sequences
size_categories:
  - 10K<n<100K
---

# Controllable AMP Design — Curated E. coli MIC Dataset

10,044 curated antimicrobial peptide (AMP) sequences paired with a continuous minimum
inhibitory concentration (MIC) activity score against *E. coli*, cleaned and deduplicated from
[DBAASP v3](https://dbaasp.org/). Used to train the CVAE generator and Judge predictor in
[Sloudis/controllable-amp-design](https://huggingface.co/Sloudis/controllable-amp-design)
([GitHub repo](https://github.com/Sloudis/controllable-amp-design)).

**This is the cleaned/derived dataset only.** The raw DBAASP bulk exports it was built from are
not redistributed here — their redistribution terms are unclear, so if you want to reproduce
the cleaning pipeline from scratch, pull your own exports from
[dbaasp.org](https://dbaasp.org/) and run `src/data/clean.py` from the GitHub repo.

## Curation pipeline

Starting from raw DBAASP activity + peptide-metadata exports:
1. Filter to MIC assays against *E. coli*, monomeric linear peptides with a known sequence
2. Parse free-text MIC values (numbers, inequalities, ranges, ± notation) into a single float
3. Standardize units to µM (µg/mL converted via computed monoisotopic molecular weight)
4. Restrict to the 20 standard amino acids, length 5–50
5. Deduplicate by sequence, taking the geometric mean of MIC across repeated measurements
6. log10-transform MIC

See `src/data/clean.py` in the GitHub repo for the exact implementation.

## Columns

| Column | Description |
|---|---|
| `sequence` | Peptide sequence (uppercase, standard 20 amino acids) |
| `mic_uM` | Minimum inhibitory concentration against *E. coli*, in µM |
| `log_mic` | log10(mic_uM) |
| `length` | Sequence length (residues) |
| `n_measurements` | Number of raw DBAASP measurements averaged (geometric mean) into this row |
| `stage` | `train` / `valid` / `test` split (stratified by log_mic decile, 80/10/10) |

## Stats

- 10,044 unique sequences — 8,044 train / 1,000 valid / 1,000 test
- Sequence length: min 5, mean 19, max 50
- MIC: min 0.005 µM, median 13.5 µM, max 18,000 µM
- log10(MIC): mean 1.15, std 0.76 (train split)

## Usage

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "Sloudis/controllable-amp-design-dataset",
    "amp_dataset.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)
```

## Citation

Cite the source database:

```
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.
```

And, if you use this specific curated/cleaned version:

```
Stavros Loudis. "Controllable Antimicrobial Peptide Design via Conditional Variational
Autoencoders." Technical University of Crete, 2026.
```