HydroNet: Benchmark Tasks for Preserving Long-range Interactions and Structural Motifs in Predictive and Generative Models for Molecular Data, at the 34th Conference on Neural Information Processing Systems (NuerIPS), Workshop on Machine Learning and the Physical Sciences [https://arxiv.org/abs/2012.00131]
We could build a classifier using a few different approaches to create a classification engine that distinguishes graphs which are impossible to invert into 3D structures.
We have plenty of data for "invertable" graphs because we have a bunch of 3D geometries. The challenge will be to generate those graphs with do not have valid 3D structures:
Generate random graphs. If we assume that most graphs do not have a valid inversions, we can just label all of these as "invalid."
Attempt to invert graphs and label "invalid" as those which do not reconstruct to something close to the ground state. Our inverters are bad, so me might say "if only 50% of the original bonds are re-created"
We can train these classifiers using either an MPNN or the kind of "ring distribution" statistics that @jenna1701 has done before.
We could build a classifier using a few different approaches to create a classification engine that distinguishes graphs which are impossible to invert into 3D structures.
We have plenty of data for "invertable" graphs because we have a bunch of 3D geometries. The challenge will be to generate those graphs with do not have valid 3D structures:
We can train these classifiers using either an MPNN or the kind of "ring distribution" statistics that @jenna1701 has done before.