Fig. 2: The proposed supervised variational graph autoencoder (VGAE) presented schematically in a typical drug discovery scenario with a large and low-fidelity (LF) high-throughput screening dataset, and a sparse and high-fidelity (HF) confirmatory screening dataset.

Graph convolutions are used to propagate and learn atom-wise representations according to the connectivity imposed by the bonds, which are then aggregated into a single molecule-level representation or embedding (a fixed-dimension vector). The readouts are standard pooling functions, e.g., sum, mean, max, or neural networks (adaptive aggregators). The symbol ∥ denotes concatenation, μ(x) and σ(x) denote the mean and standard deviation learnt by the VGAE, and ‘Dense NN’ is a multi-layer perceptron. The four workflows presented in this figure are listed in the top right and correspond to the first four experiments presented in “Results”. A low-fidelity model is first trained with supervised information to produce latent space embeddings z*(E1). A separate model with the same architecture can then be trained to predict high-fidelity values, by concatenation with either the actual labels (E2) or embeddings/predictions generated by the LF model (E3). A strategy unique to graph neural networks with adaptive readouts is that of pre-training a model on LF data as in (E1) and then fine-tuning exclusively the adaptive readouts with the VGAE layers being frozen (E4). We also emphasise that the learnt low-fidelity embeddings/predictions can be integrated into any machine learning algorithm (e.g., support vector machines, random forests, etc.).