Leash-Biosciences/ARC-Bind-Pose

ARC-Bind Pose Prediction

Overview

ARC-Bind is a synthetic 2D binding dataset designed to study how machine learning models learn to predict molecular interactions. The goal is to understand model reasoning about binding in a setting small enough to fully inspect and debug.

Inspiration: ARC-AGI Format

Inspired by François Chollet's ARC-AGI benchmark, ARC-Bind uses 16×16 pixel grids with discrete color tokens to represent binding:

• Protein: Light blue (token 8) - large, contiguous region with binding pockets
• Ligand: Light gray scaffold with functional groups
• Chemistry:
  • Red / Blue pixels = complementary polar interactions (charges, H-bonds)
  • Yellow pixels = hydrophobic regions

The format emphasizes abstraction and systematic generalization - the tasks are easy for humans but challenging for conventional deep models, especially in the low-data regime.

Dataset Description

This dataset contains 1,000 protein-ligand binding examples for pose prediction, analogous to X-ray crystallography structure prediction.

Task: Molecular Docking

Given an initial "apo" (unbound) configuration where the ligand floats somewhere near the protein, predict the "holo" (bound) configuration where the ligand has found its optimal binding pose.

Splits

• Train: 900 examples
• Validation: 100 examples (randomly sampled)

The splits are fixed to enable reproducible benchmarking across different models and papers.

Data Fields

• protein: 16×16 grid representing the protein (color-coded: 0=empty, 1=red, 2=blue, 3=yellow)
• ligand: 5×5 grid representing the ligand molecule
• input: 16×16 grid showing initial (unbound) binding pose
• output: 16×16 grid showing optimized binding pose (ground truth)
• protein_representation: Canonical string for deduplication
• ligand_representation: Canonical string for deduplication

Molecular Similarity

The protein_representation and ligand_representation fields are canonical strings designed so that Hamming distance reflects molecular similarity. This enables:

• Clustering analysis of protein/ligand families
• Train/val/test splits based on molecular diversity
• Study of structure-activity relationships

Example usage:

def hamming_distance(rep1, rep2):
    return sum(c1 != c2 for c1, c2 in zip(rep1, rep2))

# Find similar proteins
dist = hamming_distance(protein1_rep, protein2_rep)

Dataset Statistics

• Total examples: 1,000
• Distinct proteins: 289
• Distinct ligands: 298
• All examples: delta_energy > 25 (detectable binding or better)

Usage Example

from datasets import load_dataset
import numpy as np

# Load dataset
dataset = load_dataset("username/arc-bind-pose-prediction")

# Access splits
train_dataset = dataset['train']  # 900 examples
val_dataset = dataset['validation']  # 100 examples

# Get an example
example = train_dataset[0]
protein = np.array(example['protein'])  # (16, 16)
ligand = np.array(example['ligand'])    # (5, 5)
input_pose = np.array(example['input'])  # (16, 16)
output_pose = np.array(example['output'])  # (16, 16) - prediction target

Evaluation Metrics

TODO

Citation

TODO

License

MIT License - See repository for details

Source

Generated from the ARC-Bind project: https://github.com/Leash-Labs/arc-bind