MTENN

Modular Training and Evaluation of Neural Networks

Copyright

Copyright (c) 2022, Benjamin Kaminow

Minimal usage example

Building models should be done using the mtenn.config API. A small example for a SchNet model is shown below, but more details for SchNet and other models can be found in the respective class definitions.

We will construct a SchNet model with default parameters and a delta G strategy for combining our complex, protein, and ligand representations. We will leave our predictions in the returned implicit kT units (ie no Readout block).

from mtenn.config import SchNetModelConfig

# Create the config using all default parameters (which includes the delta G strategy)
model_config = SchNetModelConfig()

# Build the actual pytorch model
model = model.build()

The input passed to this model should be a dict with the following keys (based on the underlying model):

• SchNet
  • z: Tensor of atomic number for each atom, shape of (n,)
  • pos: Tensor of coordinates for each atom, shape of (n,3)
• E3NN
  • x: Tensor of one-hot encodings of element for each atom, shape of (n,one_hot_length)
  • pos: Tensor of coordinates for each atom, shape of (n,3)
  • z: Tensor of bool labels of whether each atom is a protein atom (False) or ligand atom (True), shape of (n,)
• GAT
  • x: Tensor of input atom (node) features, shape of (n,feats)
  • edge_index: Tensor giving source (first row) and dest (second row) atom indices, shape of (2,n_bonds)

The prediction can then be generated simply with:

import torch

# Using random data just for demonstration purposes
pose = {"z": torch.randint(low=1, high=17, size=(100,)), "pos": torch.rand((100, 3))}
pred = model(pose)

Installation

mtenn is now on conda-forge! To install, simply run

mamba install -c conda-forge mtenn

Acknowledgements

Project based on the Computational Molecular Science Python Cookiecutter version 1.6.