SPICE-Protein/spice_protein

SPICE-Protein: A Curated Set of Protein Structures with Environmental Annotations

SPICE-Protein provides high-quality experimental protein structures with environmental metadata (pH, temperature, ionic strength) extracted from crystallographic conditions.
It is designed as the pretraining foundation for SPICE (Sequence-Protein Interaction under Conditional Environments) — a framework that maps proteins to their stability boundaries under varying pH/temperature conditions.

---

📊 Dataset Overview

Property	Value
Total structures	~45,000
Resolution	≤ 2.5 Å
Protein lengths	40–400 residues
Chain composition	Single-chain monomeric proteins only (no nucleic acids, no complexes)
Format	Parquet (entries + atoms)
License	CC-BY 4.0

Why this dataset?

Most PDB-derived datasets provide only structure + sequence. SPICE-PDB is distinct because it also preserves the experimental conditions under which the structure was determined — specifically, the pH and temperature of the crystallization buffer. These environmental labels are essential for training models that can generalize across different physical conditions, rather than just predicting a single static structure.

---

🧬 Data Sources

All structures are sourced from the RCSB Protein Data Bank, filtered via the RCSB Search API with the following criteria:

• exptl_crystal_grow.pH exists
• exptl_crystal_grow.temp exists
• rcsb_entry_info.resolution_combined ≤ 2.5 Å
• rcsb_entry_info.polymer_entity_count_nucleic_acid = 0 (protein only)
• rcsb_entry_info.polymer_entity_count = 1 (single-chain monomer)
• 40 ≤ rcsb_entry_info.deposited_polymer_monomer_count ≤ 400

Ionic strength is also extracted where available (from exptl_crystal_grow.pdbx_details), using a heuristic pattern-matching parser.

---

📁 File Structure

spice_protein/
├── entries_*.parquet        # One row per structure
├── atoms_*.parquet          # One row per atom (long format)
└── README.md

entries_*.parquet Schema

Column	Type	Description
pdb_id	str	4-character PDB accession
method	str	Experimental method (X-ray, EM, NMR)
resolution	f64	Structure resolution (Å)
ph	f64	Crystallization pH
temperature	f64	Crystallization temperature (K)
rfree	f64	R-free value (if available)
ionic_strength_m	f64	Extracted ionic strength (M)
has_ionic	bool	Whether ionic strength was found
seq	str	Amino acid sequence (1-letter)
n_residues	i32	Number of residues
has_env	bool	True if pH, temperature, and ionic strength are all valid

atoms_*.parquet Schema

Column	Type	Description
pdb_id	str	4-character PDB accession
chain_id	str	Chain identifier
res_seq	i32	Residue sequence number
res_name	str	Residue name (3-letter)
atom_name	str	Atom name (e.g., CA, CB, N, C)
element	str	Element symbol
x, y, z	f32	Atomic coordinates (Å)
occupancy	f32	Occupancy
b_factor	f32	B-factor
is_ca	bool	True for Cα atoms

---

🔧 Usage

Load entries table

import polars as pl

entries = pl.read_parquet("entries_0001.parquet")
print(entries.head())

Filter by environment

# Get all structures at low pH
acidic = entries.filter(pl.col("ph") < 5.0)

# Get all structures with ionic strength reported
with_ions = entries.filter(pl.col("has_ionic") == True)

Reconstruct a protein structure

import polars as pl
import numpy as np

pdb_id = "1L2Y"  # Trp-cage example
atoms = pl.read_parquet("atoms_0001.parquet")
ca_atoms = atoms.filter(
    (pl.col("pdb_id") == pdb_id) & (pl.col("is_ca") == True)
)

coords = np.stack([
    ca_atoms["x"].to_numpy(),
    ca_atoms["y"].to_numpy(),
    ca_atoms["z"].to_numpy(),
], axis=1)

---

🧪 Preprocessing Pipeline

The dataset was generated using the spice_protein script, which:

1. Queries RCSB Search API for eligible PDB entries
2. Downloads full-entry mmCIF files (not assemblies — assemblies lack environmental metadata)
3. Parses each file using gemmi to extract:
   • Sequence (from _entity_poly or reconstructed from Cα residues)
   • Atomic coordinates (filtered to ATOM records only)
   • pH, temperature, and ionic strength (heuristic extraction)
4. Validates environment values (pH 0–14, temp 150–400 K)
5. Writes Parquet shards (5000 structures per shard)
6. Resumes on interruption (checkpoints converted IDs)

---

📊 Statistics

(These numbers will be filled after your full run completes)

Metric	Value
Total entries	~45,000
Average resolution	~2.0 Å
pH range	3.0 – 9.5
Temperature range	180 – 340 K
Average sequence length	~150 residues
Structures with ionic strength	~15-20%

---

🔗 Relationship to SPICE

SPICE-PDB is the pretraining dataset for the SPICE framework. During pretraining, the model learns to predict Cα coordinates from (sequence, pH, temperature, ionic_strength). This learned representation is then used as the initial backbone for the RL fine-tuning stage, where SPICE explores stability boundaries under environmental perturbations.

In the SPICE paper, this dataset is referenced as the foundation for all subsequent RL-based adaptation.

---

📝 Citation

If you use this dataset, please cite this dataset. You can get BibTex via Hugging Face.

Additionally, the underlying structures are from the PDB; please cite the original PDB entries.

---

⚠️ Notes and Limitations

1. Environmental labels are crystallization conditions, not physiological conditions. pH/temperature in the crystallization drop may differ from cellular environments.
2. Ionic strength extraction is heuristic — it is not guaranteed to be accurate for all entries, but has_ionic flag indicates whether a value was found.
3. Only monomeric single-chain proteins are included. Multi-domain proteins, complexes, and membrane proteins are excluded for v1 compatibility.
4. Resolution-based filtering excludes low-quality structures but may also remove some valid large proteins > 400 residues.

---

🙏 Acknowledgements

• RCSB PDB for maintaining the structural database
• The gemmi and polars teams for excellent parsing and data tools
• The SPICE organization contributors