SMBp Group Meeting: Erik Henning Thiede (CCM)

Thursday Feb 3, 2022, 10:00 AM → 11:00 AM America/New_York

3rd Floor Conference Room/3-Flatiron Institute (162 5th Avenue)

Topic: Autobahn: Constructing Neural Networks from Molecular Substructures

Abstract: In the last half-decade, neural networks have shown transformative results in fields such as computer vision and natural language processing. There has been considerable interest in achieving similar results for machine learning on molecular structures, most commonly through the use of graph neural networks. However, message-passing neural networks (MPNNs) – the most commonly used family of graph neural networks – have known theoretical limitations and often struggle to outperform non-neural approaches. To address these limitations, we introduce a new family of graph neural networks that we call Automorphism-based graph neural networks (Autobahn). In contrast to most MPNN architectures where neurons correspond to single nodes, in Autobahn neurons correspond to subgraphs and operate by convolving over their Automorphism group. The resulting neurons reflect the natural way that parts of the graph can transform, preserving the intuitive meaning of convolution without sacrificing global permutation equivariance. To demonstrate our formalism we introduce a new graph neural network that decomposes graphs into path and cycle subgraphs. We apply the resulting architecture to standard benchmarks for learning on small organic molecules, where we achieve results competitive with state-of-the-art MPNNs.