README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- regression
tags:
- protein-protein interaction
- binding affinity
- protein language models
- drug discovery
- bioinformatics
- multi-chain proteins
- ppb affinity
pipeline_tag: regression
---

## Dataset Description

This repository provides several enhanced versions of the **PPB-Affinity dataset**, ready for both sequence and structure-based modeling of multi-chain protein-protein interactions. The original PPB-Affinity dataset was introduced in the paper "[PPB-Affinity: Protein-Protein Binding Affinity dataset for AI-based protein drug discovery](https://www.nature.com/articles/s41597-024-03997-4)".

This version of the dataset was prepared for the study: "*[Beyond Simple Concatenation: Fairly Assessing PLM Architectures for Multi-Chain Protein-Protein Interactions Prediction](https://arxiv.org/pdf/2505.20036)*."

Code: https://github.com/Proteinea/ppiseq

The primary enhancements in this repository include:
* Various levels of data filtration and processing.
* The addition of pre-extracted "Ligand Sequences" and "Receptor Sequences" columns, making the dataset ready for use with sequence-based models without requiring PDB file parsing. For complexes with multiple ligand or receptor chains, the sequences are comma-separated.

## Dataset Configurations

This dataset offers four distinct configurations:

### 1. `raw`
* **Description**: Minimally processed data from the original PPB-Affinity dataset. Only annotation inconsistencies have been resolved (see Section 2.1.1 of "*Beyond Simple Concatenation...*" for details).
* **Size**: 12,048 entries.
* **Splits**: Contains a single `train` split encompassing all entries.
* **How to load**:
    ```python
    from datasets import load_dataset

    raw_ds = load_dataset(
        "proteinea/ppb_affinity",
        name="raw",
        trust_remote_code=True
    )["train"]
    ```

### 2. `raw_rec`
* **Description**: Similar to the `raw` version, but with an additional step to recover missing residues in the protein sequences (see Section 2.1.2 of "*Beyond Simple Concatenation...*" for details).
* **Size**: 12,048 entries.
* **Splits**: Contains a single `train` split.
* **How to load**:
    ```python
    from datasets import load_dataset

    raw_rec_ds = load_dataset(
        "proteinea/ppb_affinity",
        name="raw_rec",
        trust_remote_code=True
    )["train"]
    ```

### 3. `filtered`
* **Description**: This version includes additional cleaning and filtration steps applied to the raw with missing residues recovered data (see Section 2.1.2 of "*Beyond Simple Concatenation...*" for details on filtration). It comes with pre-defined train, validation, and test splits (see Section 2.1.3 of "*Beyond Simple Concatenation...*" for splitting methodology).
* **Size**:
    * Train: 6,485 entries
    * Validation: 965 entries
    * Test: 757 entries
* **Splits**: `train`, `validation`, `test`.
* **How to load**:
    ```python
    from datasets import load_dataset

    dataset_dict = load_dataset(
        "proteinea/ppb_affinity",
        name="filtered",
        trust_remote_code=True
    )
    train_ds = dataset_dict["train"]
    val_ds = dataset_dict["validation"]
    test_ds = dataset_dict["test"]
    ```

### 4. `filtered_random`
* **Description**: This version uses the same cleaned and filtered entries as the `filtered` configuration but provides random 80%-10%-10% splits for train, validation, and test, respectively. The shuffling is performed with a fixed seed (42) for reproducibility.
* **Size**: Same total entries as `filtered`, split as:
    * Train: 6,565 entries
    * Validation: 820 entries
    * Test: 822 entries
* **Splits**: `train`, `validation`, `test`.
* **How to load**:
    ```python
    from datasets import load_dataset

    dataset_dict = load_dataset(
        "proteinea/ppb_affinity",
        name="filtered_random",
        trust_remote_code=True
    )
    train_ds = dataset_dict["train"]
    val_ds = dataset_dict["validation"]
    test_ds = dataset_dict["test"]
    ```

## Data Fields

All configurations share a common set of columns. These include columns from the original PPB-Affinity dataset (refer to the original paper for more details), plus two new sequence columns:

* **`Ligand Sequences`**: `string` - Comma-separated amino acid sequences of the ligand chain(s).
* **`Receptor Sequences`**: `string` - Comma-separated amino acid sequences of the receptor chain(s).

**Note on Sequences**: When multiple ligand or receptor chains are present in a complex, their respective amino acid sequences are concatenated with a comma (`,`) as a separator in the "Ligand Sequences" and "Receptor Sequences" fields.