Graph networks for molecular design

Abstract

Deep learning methods applied to chemistry can be used to accelerate the discovery of new molecules. This work introduces GraphINVENT, a platform developed for graph-based molecular design using graph neural networks (GNNs). GraphINVENT uses a tiered deep neural network architecture to probabilistically generate new molecules a single bond at a time. All models implemented in GraphINVENT can quickly learn to build molecules resembling the training set molecules without any explicit programming of chemical rules. The models have been benchmarked using the MOSES distribution-based metrics, showing how GraphINVENT models compare well with state-of-the-art generative models. This work compares six different GNN-based generative models in GraphINVENT, and shows that ultimately the gated-graph neural network performs best against the metrics considered here.

Practical notes on building molecular graph generative models

“…This section contains technical details on each of the specific workflows in GraphINVENT, including details on ineffective methods that were not discussed in the original work. 25 However, we believe other researchers might still benefit from knowing what we tried and did not work. Throughout this section, the following notation is used: G = V, ℰ ð Þ is a molecular graph, where V is the set of nodes and ℰ is the set of edges, and G n ⊆G is a subgraph of G.…”

“…Although initial methods were focused on the generation of small molecular graphs (ie, <10 atoms) due to computational limitations, 17,18 developments in the field have made it such that recently published methods can easily generate molecular graphs with 10 to 100 nodes. 12,16,23,25 Developments in the field of graph-based molecular design happen very quickly, as they closely follow developments in graph representation learning, just as string-based molecular generation quickly follows developments in natural language processing.…”

“…Below we summarize a few models which follow this scheme: NeVAE, 17 a model from DeepMind, 10 MolMP and MolRNN, 11 JT-VAE, 12 HierVAE, 16 and GraphINVENT. 25 For additional examples of iterative molecular graph generation, see GCPN, 14 MolecularRNN 48 (extension of GraphRNN 15 to molecules), CGVAE, 13 RL-VAE, 49 GraphAF, 24 and ScaffoldVAE. 19 Additionally, while Graph Recurrent Attention Networks (GRANs) 50 have not yet, to our knowledge, been used to generate molecular graphs, they have been used with great success to generate protein graphs, where each node represents an individual amino acid.…”

“…Molecular generative models have emerged as promising methods for exploring the chemical space through de novo molecular design. [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16] Although molecular generative models have largely focused on string-based approaches, graphbased approaches have also emerged in the last 3 years, [10][11][12][13][14][15][16][17][18][19][20][21][22][23][24] including a recent approach, GraphINVENT, 25 from our group. Like GraphINVENT, many graph-based generative models are built using graph neural networks (GNNs), while many others are built using variational autoencoders (VAEs), generative adversarial networks (GANs), and combinations thereof.…”

“…All of the aforementioned methods can be powerful architectures for modeling patterns in graph-structured data. GNN-, VAE-, and GAN-based models have shown recent promise in molecular design applications, 10,11,14,16,25 just as recurrent neural networks (RNNs) transformed string-based molecular generation in 2016. 26 However, when it comes to development tools, there is a lack of practical information surrounding the construction of molecular graph generative models; for example, things like why a specific architecture/method was chosen, or if the authors tried any other methods unsuccessfully, are rarely, if ever, explained.…”