README.md

---
license: other
language:
- en
tags:
- biology
- protein
- peptide
- molecular-dynamics
- conotoxin
- toxicity
- protein-peptide-interaction
- thermodynamics
pretty_name: Pretraining and Downstream Benchmark Dataset
---

# ProtMMLM - datasets

## Overview
This dataset contains peptide resources for pre-training and subsequent evaluation. It includes amino acid sequences and structures in FASTA format, structure files derived from molecular dynamics simulations, and downstream benchmark datasets for cone snail toxin classification, peptide function prediction, toxicity prediction, protein-peptide affinity prediction, and thermodynamic feature modeling.

The dataset is divided into two main parts:

- `pretrain/`: Sequence and structure resources for self-supervised or multimodal pre-training.

- `downstream/`: Task-specific datasets for supervised evaluation.

```
datasets/
├── pretrain/
│   ├── all_sequences.fasta
│   ├── md.tar.gz
│   └── nature.tar.gz
└── downstream/
    ├── Conotoxin/
    │   ├── pos.fasta
    │   └── neg.fasta
    ├── PPIKB/
    │   └── Affinity Dataset(main).xlsx
    ├── PrMFTP/
    │   ├── train.txt
    │   └── test.txt
    ├── Thermodynamic/
    │   └── thermodynamic.csv
    └── Toxteller/
        ├── training_dataset.fasta
        └── independent_dataset.fasta
```

## Data Instances

### Pretraining Sequences

`pretrain/all_sequences.fasta` contains protein or peptide sequences in FASTA format.

Example:

```text
>1A11
GSEKMSTAISVLLAQAVFLLLTSQR
```

### Conotoxin Classification

`downstream/Conotoxin/pos.fasta` and `downstream/Conotoxin/neg.fasta` contain positive and negative examples for conotoxin-related binary classification.

Example:

```text
>sp|A0A068B6Q6|CA18_CONBE Conotoxin Bt1.8 (Fragment) OS=Conus betulinus OX=89764 PE=1 SV=1
PDGRNAAAKAFDLITPTVRKGCCSNPACILNNPNQCG
```

### PrMFTP Multi-Label Peptide Function Prediction

`downstream/PrMFTP/train.txt` and `downstream/PrMFTP/test.txt` use a FASTA-like format. Each header is a 21-dimensional binary label vector, followed by the peptide sequence.

Example:

```text
>000000000000000010000
ARRRRCSDRFRNCPADEALCGRRRR
```

### Toxteller Toxicity Prediction

`downstream/Toxteller/training_dataset.fasta` and `downstream/Toxteller/independent_dataset.fasta` contain peptide sequences with labels encoded in the FASTA header prefix:

- `pos_*`: toxic peptide
- `neg_*`: non-toxic peptide

Example:

```text
>pos_0
DLWQWGQMILKETGKLPFSYYTAYGCYCGWGGRGGKPKADTDRCCFVHDC
```

### Protein-Peptide Affinity

`downstream/PPIKB/Affinity Dataset(main).xlsx` contains protein-peptide interaction affinity records.

Columns:

- `ID`
- `Protein_Sequence`
- `UniProt_ID`
- `Peptide_Sequence`
- `Affinity`

### Thermodynamic Features

`downstream/Thermodynamic/thermodynamic.csv` contains frame-level thermodynamic or interaction-score features.

Columns:

- `sample`
- `frame`
- `total_score`
- `vdw`
- `hbbb`
- `hbsb`
- `hbss`
- `hp`
- `pc`
- `ps`
- `ts`
- `sb`

Example:

```csv
sample,frame,total_score,vdw,hbbb,hbsb,hbss,hp,pc,ps,ts,sb
1A11,2,17.947,-97.111,-18.912,0.0,0.0,-15.586,1.338e-06,0.0,0.0,-0.06789
```

## Intended Uses

This dataset collection can be used for:

- protein and peptide sequence representation learning;
- multimodal pretraining using sequence and structure data;
- binary peptide classification, including conotoxin and toxicity prediction;
- multi-label peptide function prediction;
- protein-peptide binding affinity prediction;
- frame-level thermodynamic or interaction-score modeling.

## Data Fields

### Sequence Files

- `header`: original FASTA header or encoded label string.
- `sequence`: amino-acid sequence.
- `label`: task-dependent label inferred from file name, FASTA header, or binary label vector.

### PPIKB

- `ID`: record identifier.
- `Protein_Sequence`: protein sequence.
- `UniProt_ID`: UniProt accession or identifier.
- `Peptide_Sequence`: peptide sequence.
- `Affinity`: protein-peptide binding affinity value as provided in the source table.

### Thermodynamic

- `sample`: sample or protein/peptide identifier.
- `frame`: frame index.
- `total_score`: total score for the frame.
- `vdw`: van der Waals contribution.
- `hbbb`: backbone-backbone hydrogen bond contribution.
- `hbsb`: sidechain-backbone hydrogen bond contribution.
- `hbss`: sidechain-sidechain hydrogen bond contribution.
- `hp`: hydrophobic contribution.
- `pc`: pi-cation contribution.
- `ps`: pi-stacking contribution.
- `ts`: T-stacking contribution.
- `sb`: salt-bridge contribution.

## Data Splits

The collection includes explicit train/test or independent splits for some downstream tasks:

- `PrMFTP`: `train.txt`, `test.txt`
- `Toxteller`: `training_dataset.fasta`, `independent_dataset.fasta`

For other tasks, users should define splits appropriate to their experimental protocol, avoiding leakage by sequence identity, source protein, or homologous family where relevant.