Prediction of Bioconcentration Factors (BCF) using Graph Neural Networks

This paper presents a Graph Neural Network (GNN) approach for the prediction of bioconcentration factors (BCF). We show that GNNs are able to exploit structural information of molecules to regress BCF values that are comparable to commonly used quantitative structure-activity relationship (QSAR) models in terms of the coefficient of R2 determination. However, the main advantage of GNNs is that molecular descriptors do not need to be determined and pre-selected by the user. Instead, they are learned directly by the GNN using backpropagation. A database of 473 molecules was used to train and test the present model. The results obtained suggest that GNNs might be useful for the prediction of other types of sustainable indicators of molecules, which is subject of our further research.