ChaoHou/protein_dynamic_properties

Protein Sequences and Dynamic Properties for Training SeqDance and ESMDance

This dataset contains protein sequences, dynamic properties, and feature weights used for training SeqDance and ESMDance, two protein language models designed to learn and represent protein dynamic properties.

Dataset Files

training_test_data_sequence.csv

This file contains 64,403 protein sequences with associated metadata for training and testing.

Columns:

• name – Unique identifier for each protein or protein complex.
• source – Data source from which the sequence was obtained:
  • ATLAS
  • GPCRmd
  • IDRome
  • PED
  • mdCATH
  • proteinflow_pdb
  • proteinflow_sabdab
• sequence – The amino acid sequence of the protein.
  • For complexes with multiple sequences, "<linker>" is inserted between them.
• modify_seq – Modified version of sequence, where "<linker>" is replaced with "<eso><cls>" for compatibility with the tokenizer.
• label – Indicates whether the sequence is used for training, testing, or is a duplicate.
• cluster_100 – MMseqs clusters at 100% sequence identity, 95% coverage (almost identical sequences).
  • Only one representative sequence is retained per cluster.
  • Other sequences in the cluster are labeled as duplicate.
• cluster_50 – MMseqs clusters at 50% sequence identity, 80% coverage of representative sequences from cluster_100.

---

feature_mdCATH_ATLAS_GPCRmd_PED_IDRome.h5, feature_proteinflow_1.h5, feature_proteinflow_2.h5

These files contains residue-level and pairwise dynamic properties for training and testing. All properties are stored as numpy arrays (dtype=float16). For a protein with name 1ab1_A in training_test_data_sequence.csv, the corresponding properties are saved as 1ab1_A_res_feature and 1ab1_A_pair_feature in the .h5 file.

Residue-Level Properties

For a protein with a structure ensemble of length L, the residue-level properties are stored in a numpy array of shape (L+2) × 47. The index of each feature in the array is as follows:

Index	Feature
0	sasa_mean
1	sasa_std
2	rmsf (not normalized)
3 – 10	ss (secondary structure, rank: B, E, H, I, G, S, T, ; see mdtraj.compute_dssp)
11 – 22	chi (torsion angles, 12 angle ranges from -π to π)
23 – 34	phi (torsion angles)
35 – 46	psi (torsion angles)

For a protein with normal mode analysis (NMA), the residue-level properties are stored in a numpy array of shape (L+2) × 3, corresponding to normal modes in three frequency ranges.

---

Pairwise Features

For a protein with a structure ensemble of length L, pairwise interaction features are stored in a numpy array of shape (L+2) × (L+2) × 10, indexed as follows:

Index	Feature
0	vdw (van der Waals interactions)
1	hbbb (backbone-backbone hydrogen bonds)
2	hbsb (sidechain-backbone hydrogen bonds)
3	hbss (sidechain-sidechain hydrogen bonds)
4	hp (hydrophobic interactions)
5	sb (salt bridges)
6	pc (pi-cation interactions)
7	ps (pi-stacking interactions)
8	ts (T-stacking interactions)
9	corr (residue movement correlation)

For a protein with normal mode analysis (NMA), the pairwise properties are stored in a numpy array of shape (L+2) × (L+2) × 3, corresponding to normal modes in three frequency ranges.

---

Notes:

• Two padding values (-1) are added at the start and end of sequences.
• For complexes, where "<eso>" and "<cls>" appear in the middle, all feature values are set to -1.
• All interaction values are pow(1/3) of interaction frequencies in structural ensembles.

---

Merging HDF5 Files

For training SeqDance, you need to merge three .h5 files, which is simple in Python:

import h5py

# Input files and the output merged file
input_files = [
    'feature_mdCATH_ATLAS_GPCRmd_PED_IDRome.h5',
    'feature_proteinflow_1.h5',
    'feature_proteinflow_2.h5'
]

output_file = 'feature_mdCATH_ATLAS_GPCRmd_PED_IDRome_proteinflow.h5'

# Create the merged HDF5 file
with h5py.File(output_file, "w") as out_f:
    for file_path in input_files:
        with h5py.File(file_path, "r") as in_f:
            for key in in_f.keys():
                in_f.copy(key, out_f)

---

Feature Weights

The weights for each feature in the training loss are based on the standard deviation of each feature. Please refer to our paper for further details.