Molecular graph generation with Graph Neural Networks

Introduction

• Unlike recurrent generators that typically use prior decisions, GNNs can use the entire partial graph; DeepGMG is the cited graph-net-based approach of this type.The authors motivate GNNs by their ability to process graphs without losing connectivity information.Source: §1 · pp. 2–3
• MG2N2 adds one node at a time, expands its neighbors before moving on, uses BFS ordering, and generates edges in parallel to avoid disconnected components and reduce computation.Three independently trained GNN modules control generation; Gumbel softmax output layers support training over discrete stochastic distributions.Source: §1 · p. 3
• MG2N2 outperforms competitive baselines for unconditional generation on QM9 and Zinc, and generates molecules with chemical characteristics similar to the original datasets.The paper identifies its evaluation on these two benchmarks as a contribution.Source: §1 · p. 3

2 The GNN Model

• Vertex states start from vertex attributes and update iteratively in parallel using each vertex’s previous state and messages from its neighbors.The updates continue until convergence or a maximum iteration count is reached.Source: §2 · pp. 3–4
• The model learns the state-update function with an MLP, while messages concatenate source-node states with edge attributes and aggregation uses a mean or element-wise sum.A hyperparameter selects mean aggregation by scaling messages by 1/|Ne(vi)| or sum aggregation by scaling them by 1.Source: §2 · p. 4
• Generation is divided into three classification subtasks—one node-based and two edge-based—with output functions defined for the relevant graph elements.State updating stops at convergence, when successive states differ by less than ϵ, or at the maximum iteration count kmax; ϵ and kmax are hyperparameters.Source: §2 · p. 4
• An MLP output network follows the state network, and training compares predictions with supervision using a loss and Backpropagation Through Time.Source: §2 · p. 5

3.1 Generation Algorithm

• Generation expands one focus node at a time, adding a new node and its connecting edge before predicting any additional links, which keeps every generated graph connected.The process starts from a single node and stops when all nodes are expanded or the maximum graph size is reached.Source: §3.1 · p. 5
• Three functions decide whether and how to add a labeled node, classify its first edge, and predict links to other nodes in parallel.Parallel link predictions avoid edge-ordering bias and reduce time cost, but ignore dependencies between edges.Source: §3.1 · p. 5

3.3 Graph Preprocessing

• Training, validation, and test examples are built from incomplete graphs in each molecule’s generation sequence: M1 uses 2|V| − 1 graphs, while M2 and M3 use |V| − 1.All graphs from the same sequence are kept in the same set and batch to avoid evaluation bias from near-replicas crossing between training and evaluation.Source: §3.3 · pp. 8–9

3.4 Node Ordering

• Node orderings use breadth-first search, expand lower-centrality node types first at the same depth, and fix the starting node to obtain a unique sequence.Remaining ties are randomized; fixing the start and resolving ties this way reduces ordering possibilities at the cost of generality.Source: §3.4 · p. 9

4.1 Dataset Description

• The experiments used QM9 and Zinc molecular-graph benchmarks with different molecule sizes and atom and bond vocabularies.QM9 contains 133,885 molecules of up to 9 heavy atoms, split into 120,000 training, 10,000 test, and 3,885 validation graphs; Zinc contains 249,455 molecules of up to 38 heavy atoms, split into 230,000, 10,000, and 9,455 graphs, respectively.Source: §4.1 · pp. 9–10

4.2 Experimental Setup

• The modules were initially tuned independently using validation accuracy, then trained with Adam and cross-entropy loss.Gumbel-softmax temperature was linearly annealed from 5.0 to 1.0; tests below 1.0 caused unwanted repetitive behavior.Source: §4.2 · p. 11

4.3 Evaluation

• Evaluation measured validity, uniqueness, novelty, and their product VUN, alongside molecular-weight, logP, and QED distributions against the test set.Each evaluation used 10,000 generated graphs; RDKit assessed validity and chemical properties, while NetworkX graph isomorphism supported uniqueness and novelty checks.Source: §4.3 · pp. 11–12

4.4 Results and Discussion

• Raising QM9’s maximum graph size increased valid and unique outputs but shifted molecular-weight and QED distributions, so the authors retained |V|max = 29.The generated compounds at higher thresholds were heavier and had lower QED; the authors judged these low-drug-likeness compounds unsuitable for potential drug generation.Source: §4.4 · p. 12, Table 2 · p. 13, Figure 2 · p. 13
• On QM9, M3 class weighting brought no performance advantage and aggregation methods were equivalent; C3 gained validity but worsened QED matching, making C2 the preferred configuration.The M3 balancing experiment addressed a highly imbalanced supervision distribution, with roughly 97% of examples in the disconnected class.Source: §4.4 · pp. 13, 15, Table 3 · p. 14, Figure 3 · p. 16
• On QM9, MG2N2 achieved the highest VUN score, although GraphVAE and MolGAN had higher validity and MG2N2 had higher uniqueness.MPGVAE also achieved good uniqueness and novelty while nearly matching MolGAN’s validity.Source: §4.4 · p. 14, Table 5 · p. 15
• On Zinc, MG2N2 outperformed standard VAE baselines in validity; CGVAE’s hard chemical constraints distinguish it from methods that learn these rules.The authors suggest sequential atom generation and repeated graph examination become more important as the number of atom types grows.Source: §4.4 · p. 17, Table 7 · p. 17

• MG2N2 feeds each generated graph into the next step, exploiting graph neural networks’ native graph processing rather than relying mainly on prior decisions.The authors identify this as an advantage over other sequential methods.Source: §5 · p. 17
• The modular design makes learning easier and less resource demanding, while the generation sequence exposes where errors occur and which atoms or bonds are created.The authors say this interpretability simplifies improving or repurposing the model.Source: §5 · pp. 17–18
• On QM9, generated descriptor distributions retraced the held-out test set and graph quality was very high, outperforming competitive baselines; Zinc showed the same performance level against similar approaches.Source: §5 · p. 18
• Future work includes generalizing to other molecular graph generation problems and conditional generation, potentially by concatenating desired-property vectors to each module’s input.Source: §5 · p. 18
• The authors propose studying constrained or fragment-based MG2N2 for larger datasets such as Zinc, alongside a theoretical mathematical formulation of sequential generation.A constrained or fragment-based version could improve performance on larger molecules.Source: §5 · p. 18