README.md

---
dataset_info:
  features:
  - name: seq_name_a
    dtype: string
  - name: seq_name_b
    dtype: string
  - name: seq_a
    dtype: string
  - name: seq_b
    dtype: string
  - name: score
    dtype: float64
  - name: label
    dtype: int64
  splits:
  - name: train
    num_bytes: 925231810
    num_examples: 800000
  - name: validation
    num_bytes: 115759871
    num_examples: 100000
  - name: test
    num_bytes: 115544832
    num_examples: 100000
  download_size: 1078584544
  dataset_size: 1156536513
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# STRING PPI Human 1M

This dataset contains 1 million human protein–protein interactions (PPIs) derived from STRING v11.5.

**Columns**:
- `seq_a`, `seq_b`: Amino acid sequences of the interacting proteins (≤2048 AA).
- `seq_name_a`, `seq_name_b`: Protein names from STRING.
- `score`: Combined score from STRING (0–1000, normalized to 0–1). This score integrates various evidence channels (experimental data, text mining, co-expression, etc.) into a single confidence metric.
- `label`: Binary interaction label.
  - `1`: High-confidence positive interaction (STRING experimental score > 0.7).
  - `0`: Low-confidence negative interaction (STRING experimental score < 0.2).

**Processing Steps**:
- Downloaded human interactions from STRING (`9606.protein.links.detailed.v11.5.txt.gz`).
- Filtered by experimental score thresholds for positives/negatives.
- Retrieved protein sequences from UniProt by matching STRING protein names.
- Excluded sequences longer than 2048 amino acids.
- Balanced top and bottom 500k interactions by combined score.

Useful for training models for binary PPI prediction or embedding-based methods.

Source: [STRING Database v11.5](https://string-db.org).