Changes

Wang, Renxiao, et al. "The PDBbind database: methodologies and updates." Journal of medicinal chemistry 48.12 (2005): 4111-4119.

In particular, on the 2015 release of PDBBind. The followignmodels are trained:

In particular, on the 2015 release of PDBBind. There are three

subsets of this dataset, "core", "refined", and "full." Core is

about a couple hundred structures, refined a couple thousand,

and "full" about 10 thousand.

You can load a version of this dataset by calling

`dc.molnet.load_pdbbind()`. Make sure to have environment

variable `DEEPCHEM_DATA_DIR` set to point somewhere meaningful

otherwise you'll download the full PDBBind dataset every time.

You can use the "grid" or "atomic" featurizations for this

datasets. We'll provide examples of both.

Note that computing the atomic featurizations is slow, so we recommend doing simple examples on "core" to avoid experiments taking too long.

- `dc.models.AtomicConvModel`: Trained in `pdbbind_atomic_conv.py`

import deepchem as dc

import numpy as np

from deepchem.molnet import load_pdbbind

# You'll want to see some of the logging statements to track progress

import logging

logging.basicConfig(level=logging.DEBUG)

# For stable runs

np.random.seed(123)

model = dc.models.AtomicConvModel(

    frag1_num_atoms=frag1_num_atoms,

    frag2_num_atoms=frag2_num_atoms,

    complex_num_atoms=complex_num_atoms)

    complex_num_atoms=complex_num_atoms,

    tensorboard_log_frequence=1)

# Fit trained model

print("Fitting model on train dataset")