README.md

---
dataset_info:
  features:
  - name: name
    dtype: string
  - name: seq
    dtype: string
  - name: tertiary
    sequence:
      sequence: float64
  - name: valid_mask
    sequence: bool
  - name: metadata
    struct:
    - name: contact_density
      dtype: float64
    - name: has_valid_structure
      dtype: bool
    - name: invalid_residues
      dtype: int64
    - name: long_range_contacts
      dtype: int64
    - name: medium_range_contacts
      dtype: int64
    - name: seq_length
      dtype: int64
    - name: short_range_contacts
      dtype: int64
    - name: token_count
      dtype: int64
    - name: tokenization_ratio
      dtype: float64
    - name: total_contacts
      dtype: int64
    - name: valid_residues
      dtype: int64
  splits:
  - name: train
    num_bytes: 169570266
    num_examples: 25299
  - name: validation
    num_bytes: 1350998
    num_examples: 224
  - name: test
    num_bytes: 356465
    num_examples: 40
  download_size: 96822838
  dataset_size: 171277729
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---



## Background

### What is Protein Contact Prediction?

Protein contact prediction is a fundamental task in computational biology that aims to predict which amino acid residues in a protein sequence are in close spatial proximity (typically within 8Å) in the protein's 3D structure. This information is crucial for:

- **Protein structure prediction**
- **Understanding protein folding mechanisms**
- **Drug discovery and design**
- **Evolutionary analysis**

### Linear Probing for Protein Language Models

Linear probing is a technique used to evaluate the quality of learned representations from pre-trained language models. In the context of protein language models (PLMs):

1. **Freeze** the pre-trained model parameters
2. **Train** only a simple linear classifier on top of the frozen embeddings
3. **Evaluate** how well the linear classifier performs on downstream tasks

This approach helps assess the quality of protein representations learned by different PLMs without the confounding effects of fine-tuning the entire model.



## Task Definition

### Contact Map Generation

| Parameter | Value | Description |
|-----------|--------|-------------|
| **Distance Threshold** | 8.0 Å | Distance between Cα atoms |
| **Sequence Separation** | ≥ 6 residues | Minimum separation (\|i-j\| ≥ 6) |
| **Valid Contacts** | Required | Only consider residues with valid 3D coordinates |

### Evaluation Metrics

The evaluation follows standard contact prediction protocols with precision at different coverage levels:

#### Range Categories

| Range | Separation | Description |
|-------|------------|-------------|
| **Short Range** | 6 ≤ \|i-j\| ≤ 11 | Local contacts |
| **Medium Range** | 12 ≤ \|i-j\| ≤ 23 | Medium-distance contacts |
| **Long Range** | \|i-j\| ≥ 24 | Long-distance contacts |

#### Precision Metrics

- **P@L**: Precision at top L contacts (L = sequence length)
- **P@L/2**: Precision at top L/2 contacts
- **P@L/5**: Precision at top L/5 contacts

Each metric is calculated separately for each range category, resulting in **9 total metrics** per evaluation.



### Dataset Format

The system uses HuggingFace datasets with entries containing:

```json
{
    "seq": "MKTFFVLVLLPLVSSQCVNLTTRTQLPPAYTNSFTRGVYYPDKVFRSSVLHSTQDLFL",
    "tertiary": [[x1,y1,z1], [x2,y2,z2], ...],
    "valid_mask": [true, true, false, ...]
}
```

**Default Dataset**: `fredzzp/contact_data` (automatically downloaded from HuggingFace Hub)