The authors propose some improvements of the weave graph convolutional neural network model (Molecular Graph Convolutions: Moving Beyond Fingerprints) for small molecules. In the weave model, molecular information is converted to features using two matrices. First, a 2d matrix is constructed whereby rows are assigned to atoms, while columns are features such as atom type, ring size, etc. Then a 3d matrix is constructed in which rows and columns correspond to atoms and bond information is added in pairwise fashion and stacked. Various convolutions and pooling operations are performed to ultimately feed a vector to a multilayer perceptron.
To improve the weave algorithm the authors propose:
… to make effective use of the distance features on the molecular graph in the Weave module, we considered three improvements: correction of the distance on the molecular graph with respect to atoms in the ring structure, convolution method of pair features, and assembling of the pair features.
As expected, the results are mixed compared to the original weave algorithm. Applying machine learning to small molecules using 2d information is inherently limited. We are unaware of a featurization type and algorithm that beats all other methods on benchmark datasets. This suggests that the best strategy is to combine methods via ensembling.
Machine learning is often used in virtual screening to find compounds that are pharmacologically active on a target protein. The weave module is a type of graph convolutional deep neural network that uses not only features focusing on atoms alone (atom features) but also features focusing on atom pairs (pair features); thus, it can consider information of nonadjacent atoms. However, the correlation between the distance on the graph and the three-dimensional coordinate distance is uncertain. In this paper, we propose three improvements for modifying the weave module. First, the distances between ring atoms on the graph were modified to bring the distances on the graph closer to the coordinate distance. Second, different weight matrices were used depending on the distance on the graph in the convolution layers of the pair features. Finally, a weighted sum, by distance, was used when converting pair features to atom features. The experimental results show that the performance of the proposed method is slightly better than that of the weave module, and the improvement in the distance representation might be useful for compound activity prediction.