README.md

metadata

license: mit
tags:
  - protein
  - structural-biology
  - thermodynamic-stability

DeepEF Datasets

Protein structure datasets used for training and evaluating DeepEF — a deep learning framework for predicting protein thermodynamic stability.

Contents

Path	Size	Description
casp12_data_30/	~110GB	CASP12 structures with ProtT5 embeddings; per-protein .pt tensors split into train/test/valid-*
Processed_K50_dG_datasets/	~2.2GB	K50 ddG mutation datasets with AlphaFold PDB models
megascale_proteins.csv	small	Megascale protein list

Data Format (casp12_data_30)

Each protein is a folder under train/, test/, or valid-*/:

PROTEIN_ID/
├── crd_backbone.pt       # Backbone coordinates [seq_len, 4, 3]
├── ang.pt                # Dihedral angles
├── mask.pt               # Valid residue mask
├── seq_one_hot.pt        # One-hot amino acid encoding [seq_len, 20]
├── seq.pt                # Raw sequence string
├── seq_mut.pt            # Mutant sequence string
├── proT5_emb.pt          # ProtT5 embeddings [seq_len, 1024]
├── proT5_emb_cycle1-4.pt # Cyclic permutation embeddings
└── proT5_emb_mut.pt      # Mutant ProtT5 embeddings

Usage

from huggingface_hub import snapshot_download, hf_hub_download
import torch

# Download a single protein
path = hf_hub_download(
    repo_id="shaharec/deepef-data",
    repo_type="dataset",
    filename="casp12_data_30/train/1A0C_1_A/crd_backbone.pt",
)
coords = torch.load(path)

# Download an entire split (large!)
snapshot_download(
    repo_id="shaharec/deepef-data",
    repo_type="dataset",
    allow_patterns="casp12_data_30/train/*",
    local_dir="./data",
)