Tutorial for Running Atomic Convolutions (ACNNs)

[rbharath]

We currently lack a tutorial for running atomic convolutions. Here's some of the material it would be good to cover in such a tutorial:

• Introduce the basic structure of the Atomic Conv model. The code for this model is here: https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/atomic_conv.py.
• Create an IPython notebook in https://github.com/deepchem/deepchem/tree/master/examples/notebooks for the atomic conv tutorial. This notebook can be modeled on https://deepchem.io/docs/notebooks/protein_ligand_complex_notebook.html, but with ACNNs rather than grid fingerprints.
• The tutorial should featurize a selection of the pdbbind dataset and train a simple ACNN model. You might find the code in https://github.com/deepchem/deepchem/tree/master/contrib/atomicconv useful.

[nitinprakash96]

Can I work on this?

[rbharath]

@nitinprakash96 Yes please go for it! Could you open a "WIP" pull request to track your on-going progress?

[nitinprakash96]

#1180

[nitinprakash96]

Hi @rbharath sorry I've been caught up with exams and semester project reports. So took a little late to implement this. But its over now and I can start writing this tutorial.
Just to get my doubt cleared, I wrote the structure of ACNN as described by the paper(I thought it explained the best). And as for the featurization, how exactly do we need to present the tutorial? I mean the featurization part?

[rbharath]

Great to hear you can work on this again!

For the featurization part, I think it can be presented as a method of processing 3D structures to extract the neighbor information.

[nitinprakash96]

I tried some of the ways to do this. I tried constructing a Nx3 matrix of cartesian coordinates and compute neighbour list using that. But I keep running into ValueError: Bad Conformor Id whenever I try to convert the Atomic units from Angstrom to Bohr.

Is there a way to tackle this? I'm using https://github.com/deepchem/deepchem/blob/master/deepchem/feat/atomic_coordinates.py as my reference.

[rbharath]

@nitinprakash96 Can you post a minimal failing example? Will take a look

[nitinprakash96]

import numpy as np
import os
import tensorflow as tf
import deepchem as dc
from deepchem.feat import Featurizer
import rdkit

data_dir = os.path.join(dc.utils.get_data_dir())
dataset_file= os.path.join(dc.utils.get_data_dir(), "pdbbind_core_df.csv.gz")
raw_dataset = dc.utils.save.load_from_disk(dataset_file)
df = raw_dataset['smiles'][:20]

for i in range(20):
    #print(df[i])
    mol = rdkit.Chem.MolFromSmiles(df[i])
    #print(mol)
    N = mol.GetNumAtoms()
    print(N)
    coords = np.zeros((N, 3))
    #print(coords)
        
    coords_raw = [mol.GetConformer(0).GetAtomPosition(i) for i in range(N)]
    for atom in range(N):
        coords[atom, 0] = coords_raw[atom].x
        coords[atom, 1] = coords_raw[atom].y
        coords[atom, 2] = coords_raw[atom].z
    coords = [coords]

And the error is ValueError: Bad Conformer Id

[rbharath]

I think we deprecated pdbbind_core_df.csv.gz a while back. Try instead using the function featurize_pdbbind from dc.molnet.load_function.pdbbind_datasets.featurize_pdbbind to featurize an atomic conv dataset.

[nitinprakash96]

Hi, So while the featurization part is okay and I got it working. But in the same pdbbind_dataset module. I had a query.
Supposing if we run the following snippet,

split = "random"
subset = "full"
pdbbind_tasks, pdbbind_datasets, transformers = load_pdbbind_grid(
    split=split, subset=subset)
train_dataset, valid_dataset, test_dataset = pdbbind_datasets

It returns NoneType object is not iterable. Is this a bug or I'm missing something here?

[nitinprakash96]

So looking into it a little more, if we use grid featurizer with load_pdbbind_grid we get caught into the above error. And doing it manually is no problem.

[nitinprakash96]

@rbharath Is there a way to use TensorflowFragmentRegressor for acnns?

[nitinprakash96]

@rbharath can you please help me out here? The following example should train a simple acnn model on the pdbbind dataset. But it throws the following error:

Exception in thread Thread-9:
Traceback (most recent call last):
  File "/home/nitin/anaconda3/envs/deepchem/lib/python3.5/threading.py", line 914, in _bootstrap_inner
    self.run()
  File "/home/nitin/anaconda3/envs/deepchem/lib/python3.5/threading.py", line 862, in run
    self._target(*self._args, **self._kwargs)
  File "/home/nitin/Documents/deepchem/deepchem/deepchem/models/tensorgraph/tensor_graph.py", line 1006, in _enqueue_batch
    for feed_dict in generator:
  File "/home/nitin/Documents/deepchem/deepchem/deepchem/models/tensorgraph/models/atomic_conv.py", line 238, in feed_dict_generator
    num_features = F_b[0][0].shape[1]
IndexError: tuple index out of range

Example code:

pdbind_dataset, tasks = featurize_pdbbind(feat='grid', subset='core')

split = 'random'
splitters = {
        'index': dc.splits.IndexSplitter(),
        'random': dc.splits.RandomSplitter(),
        'time': dc.splits.TimeSplitterPDBbind(pdbbind_dataset.ids)
    }
splitter = splitters[split]
train_dataset, valid_dataset, test_dataset = splitter.train_valid_test_split(pdbbind_dataset)
transformers = []

y_train = train_dataset.y

y_train *= -1 * 2.479 / 4.184
train_dataset = dc.data.DiskDataset.from_numpy(
    train_dataset.X,
    y_train,
    train_dataset.w,
    train_dataset.ids,
    tasks=tasks)

tg, feed_dict_generator, label = atomic_conv_model()
tg.fit_generator(feed_dict_generator(train_dataset, batch_size=24, epochs=10))

[rbharath]

Not sure what's happening here. Will need to find some time to sit down and take a crack at this code. Hope to do so later this week.

[rbharath]

Closing since implemented in #2431