keras-io/MPNN-for-molecular-property-prediction

Model description

This repo contains the model and the notebook for implementing Message Passing Neural Network (MPNN) to predict a molecular property known as blood-brain barrier permeability (BBBP). Message-passing neural network (MPNN) for molecular property prediction.

Full credits go to Alexander Kensert

Reproduced by Vu Minh Chien

Motivation: Until now, more traditional methods, such as random forests, support vector machines, etc., have been commonly used to predict molecular properties. In contrast to GNNs, these traditional approaches often operate on precomputed molecular features such as molecular weight, polarity, charge, number of carbon atoms, etc. Although these molecular features prove to be good predictors for various molecular properties, it is hypothesized that operating on these more "raw", and "low-level", features could prove even better. In this tutorial, MPNN was used to prove that GNN is a good method to predict molecular properties by using the undirected graph of molecules

Training and evaluation data

Information about the dataset can be found in A Bayesian Approach to Silico Blood-Brain Barrier Penetration Modeling and MoleculeNet: A Benchmark for Molecular Machine Learning. The dataset will be downloaded from MoleculeNet.org.

The dataset contains 2,050 molecules. Each molecule comes with a name, label, and SMILES string.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

name	learning_rate	decay	beta_1	beta_2	epsilon	amsgrad	training_precision
Adam	0.0005000000237487257	0.0	0.8999999761581421	0.9990000128746033	1e-07	False	float32

Model Plot

View Model Demo

View Model Plot