---
tags:
- proteins
- molecules
- chemistry
- SMILES
- complex structures
---

## How to use the data sets

This dataset contains more about 80,000 unique pairs of protein sequences and ligand SMILES, and the coordinates
of their complexes from the PDB. 

SMILES are assumed to be tokenized by the regex from P. Schwaller.

Every (x,y,z) ligand coordinate maps onto a SMILES token, and is *nan* if the token does not represent an atom

Every receptor coordinate maps onto the Calpha coordinate of that residue.

The dataset can be used to fine-tune a language model, all data comes from PDBind-cn.

## Ligand selection criteria

Only ligands with

- at least 3 atoms,
- a molecular weight >= 100 Da,
- that don't occur more than 75 times in different PDB complexes (this includes common additives like PEG, ADP, ..)

are considered.

### Use the already preprocessed data

Load a test/train split using

```
from datasets import load_dataset
train = load_dataset("jglaser/pdb_protein_ligand_complexes",split='train[:90%]')
validation = load_dataset("jglaser/pdb_protein_ligand_complexes",split='train[90%:]')
```

### Manual update from PDB

```
# download the PDB archive into folder pdb/
sh rsync.sh 24 # number of parallel download processes

# extract sequences and coordinates in parallel
sbatch pdb.slurm
# or
mpirun -n 42 parse_complexes.py # desired number of tasks
```