MoINN

MoINN, short for Moiety Identification Neural Network, is designed to automate the identification of chemical moieties (fundamental building blocks of molecules) from machine learned representations. It has been shown that MoINN is suitable for the automatic construction of coarse-grained force fields, selecting representative entries in chemical databases, as well as the identification of reaction coordinates. The design of MoINN (differantiable, transferable w. r. t. molecule size) makes it very versatile and paves the way for many other interesting applications.

Requirements:

• python 3.8
• Atomic Simulation Environment (ASE)
• NumPy 1.23.3
• PyTorch 1.13.1
• H5py
• tensorboard
• tqdm
• RDKit
• NetworkX
• SchNetPack 1.0

Note: We recommend using a GPU for training the neural networks.

Installation

You can install the most recent code from our repository:

git clone https://github.com/jnsLs/MoINN.git
cd MoINN
pip install .

Getting started

In the following, we show the workflow of training and evaluating MoINN. Here, we focus on using the CLI to train on the QM9 dataset, but the same procedure applies for other datasets as well. First, create a working directory, where all data and runs will be stored:

mkdir moinn_workdir
cd moinn_workdir

Then, the training of a pretrained MoINN model with default settings for QM9 can be started by:

moinn_train.py --model_dir ./run0 --datapath ./data/qm9.db --tradeoff_warmup_epochs 0 0 --split 1000 100 --clustering_mode pretrained --rep_model_dir /directory/of/MPNN/model --manual_seed 3 --cuda

The dataset will be downloaded automatically to spk_workdir/data, if it does not exist yet. Then, the training will be started. We provide a pretrained SchNet model, which can serve as a representation model for the training of a pretrained MoINN model. It is stored in trained_models/schnet_model.

Training of an end-to-end MoINN model is performed with the following command:

moinn_train.py --model_dir ./run1 --datapath ./data/qm9.db --tradeoff_warmup_epochs 100 130 --split 1000 100 --clustering_mode end_to_end --manual_seed 3 --cuda --lr 0.0002

Finally, we can evaluate the respective MoINN model:

moinn_eval_qm9.py --model_dir ./run0 --datapath ./data/qm9.db

The evaluation results will be stored in the model directory in a folder named eval.

References

• [1] Lederer, J., Gastegger, M., Schütt, K. T., Kampffmeyer, M., Müller, K.-R., & Unke, O. T.
  Automatic identification of chemical moieties. Physical Chemistry Chemical Physics (2023). (https://doi.org/10.1039/D3CP03845A)