-- Machine-learning code for predicting molecular properties --

Choose the path for the virtual environment according to your preferences:

VENVPATH=$HOME/venvs/maltose-pip

Set up a new python environment:

python3 -m venv $VENVPATH

Activate the new environment:

source $VENVPATH/bin/activate

Upgrade the pip package manager:

python3 -m pip install --upgrade pip

To install the requirements, execute the commands described below in the folder where also this readme file is located. The requirements.txt contains at least one package (torch-scatter) that depends on torch already being installed. Therefore, if you do not have torch installed yet, install the basic requirement (including torch, but not torch-scatter) first:

python3 -m pip install -r requirements0.txt

Then, installed all the other requirements:

python3 -m pip install -r requirements.txt

Build the maltose package:

python3 -m build

Install the maltose package:

pip install dist/maltose-1.0.2-py3-none-any.whl

You're ready to go!

Some users prefer conda over pip because it does not only install python packages, but also allows installing a different python version. Somtimes conda is better at resolving dependencies than pip.

Create and activate a new conda environment:

VENVNAME="maltose-conda"
conda create -n "${VENVNAME}" -c conda-forge python=3.9.7
conda activate "${VENVNAME}"

Install dependencies:

conda install -c conda-forge --file requirements_conda.txt

Build the maltose wheel:

python3 -m build

Install the maltose wheel via pip:

pip install --no-dependencies dist/maltose-1.0.2-py3-none-any.whl

You're ready to go!

Now, the contents of the library src/maltose can be imported and used in python. The package contains:

• Some helper functions for format conversions of molecular data (maltose.conversions).
• Some helpers for preparing the primary data (maltose.primary_data).
• The Set2Set output module (in maltose.output_modules).
• The functionality for multitask training (maltose.atoms.MultitaskAtomsData, maltose.loss.build_gated_mse_loss, maltose.metrics.MultitaskMetricWrapper).

All experiments are executed via python scripts in the ./scripts folder. These scripts write to and read from the following directories, by default placed next to the sripts directory:

• scratch/ directroy that is used only during the preparation of the training/validation/test data. Can be deleted afterwards.
• data/ is the data directory that is written during the preparation of the training/validation/test data.
• models/ is the diectory for saving trained models. Also some derived data, like accuracy measures of the model on the test sets, are saved here.
• figures/ folder is written to by skripts/make_plots.py.

Remark on file locations: Some scripts accept command-line parameters to specify these folder locations.

python3 scripts/primary_data/prepare_qm9.py
python3 scripts/primary_data/prepare_alchemy.py
python3 scripts/primary_data/prepare_oe62.py
python3 scripts/primary_data/prepare_hopv.py

Next, prepare the unified train/validation/test split by running:

python3 scripts/primary_data/unified_split.py

python3 scripts/train_model.py --config <conig_name>

where, for example, conig_name = "multitask_model_v01".

Note that this may take some days or even weeks, depending on your hardware. If available, run the training on a GPU:

python3 scripts/train_model.py --config <conig_name> --device cuda

python3 scripts/compute_errors.py --config <conig_name> --device <device>

and:

python3 scripts/make_plots.py --config <conig_name> --device <device>

There are also some trained models available in the ./trained-models folder. To eveluate these, add the option --model-base-dir ./trained-models to the above calls.