Skip to content


Repository files navigation

Predicting Lewis Acidity: Machine Learning the Fluoride Ion Affinity of p-Block Atom-based Molecules

This repository corresponds to
L. M. Sigmund, S. Sowndarya S. V., A. Albers, P. Erdmann, R. S. Paton, L. Greb, Angew. Chem. Int. Ed. 2024, 63, e202401084, DOI: 10.1002/anie.202401084.
The FIA49k dataset and all other data files can be downloaded from the corresponding figshare project and should be placed in the "data" folder of this repository to run the provided Jupyter notebooks and Python scripts.

Use FIA-GNN via the Web App

The FIA-GNN models can be used free of charge at this homepage.

Use FIA-GNN within a Jupyter Notebook

The deploy_fia_gnn.ipynb notebook demonstrates how to use the trained models of FIA-GNN as published.


If model training is intended, run the script. This will create the preprocessed graph inputs.


Preprocessing of the data is done simultaneously for FIAgas and FIAsolv. For the preprocessing to be successful, the provided data file must contain a column named "la_smiles" with the SMILES strings of the Lewis acids, as well as columns named "fia_gas-DSDBLYP" and "fia_solv-DSDBLYP", with the respective FIA values for gas and solution phase.

Next, the model can be trained with the script. Training is either done for FIAgas or FIAsolv.


A respective conda environment for training can be obtained with

cd fia_gnn/
conda env create -f tf2_gpu_fia_gnn.yml -n fia_gnn_tf_env
conda activate fia_gnn_tf_env