This deepchem tutorial introduces Atomic Convolutional Model. We'll see the structure of the Atomic Conv Model and write a simple program to run Atomic Convolutions.
%% Cell type:markdown id: tags:
### Structure
%% Cell type:markdown id: tags:
ACNN’s directly exploit the local three-dimensional structure of molecules to hierarchically learn more complex chemical features by optimizing both the model and featurization simultaneously in an end-to-end fashion.
The atom type convolution makes use of a neighbor-listed distance matrix to extract features encoding local chemical environments from an input representation (Cartesian atomic coordinates) that does not necessarily contain spatial locality. Following are the methods use to build ACNN architecture:
- #### Distance Matrix
The distance matrix R is constructed from the Cartesian atomic coordinates X. It calculates distance from the distance tensor D. The distance matrix construction accepts as input a (N, 3) coordinate matrix C. This matrix is “neighbor listed” into a (N, M) matrix R.
The output of the atom type convolution is constructed from the distance matrix R and atomic number matrix Z. The matrix R is fed into a (1x1) filter with stride 1 and depth of Na , where Na is the number of unique atomic numbers (atom types) present in the molecular system. The atom type convolution kernel is a step function that operates on neighbor distance matrix R.
- #### Radial Pooling layer
Radial Pooling is basically a dimensionality reduction process which down-samples the output of the atom type convolutions. The reduction process prevents overfitting by providing an abstracted form of representation through feature binning, as well as reducing the number of parameters learned.
Mathematically, radial pooling layers pool over tensor slices (receptive fields) of size (1xMx1) with stride 1 and a depth of Nr, where Nr is the number of desired radial filters.
- #### Atomistic fully connected network
Atomic Conolution layers are stacked by feeding the flattened(N, Na x Nr) output of radial pooling layer into the atom type convolution operation. Finally, we feed the tensor row-wise (per-atom) into a fully-connected network. The
same fully connected weights and biases are used for each atom in a given molecule.
%% Cell type:markdown id: tags:
Now that we have seen the structural overview of ACNNs, we'll try to get deeper into the model and see how we can train it and what do we expect as the output.
For the training purpose, we will use the publicly available PDBbind dataset. In this example, every row reflects a protein-ligand complex, and the following columns are present: a unique complex identifier; the SMILES string of the ligand; the binding affinity (Ki) of the ligand to the protein in the complex; a Python list of all lines in a PDB file for the protein alone; and a Python list of all lines in a ligand file for the ligand alone.
%% Cell type:code id: tags:
``` python
%load_extautoreload
%autoreload2
%pdboff
# set DISPLAY = True when running tutorial
DISPLAY=True
# set PARALLELIZE to true if you want to use ipyparallel
