> What is the metric for "working well" here? Model fit. Validation on a holdout set. All of the usual standards for ML performance. > My point is a graph con

> What is the metric for "working well" here? Model fit. Validation on a holdout set. All of the usual standards for ML performance. > My point is a graph convolution will have a hard time distinguishing the haouamine atropisomers, because they have the same graph, but very different activities. Isomers can be represented. The typical problem with isomers is that you don't have good data on them, because real-world laboratory solutions are rarely pure. > Like some molecule that isn't the actual active form but needs cyp450 to be activated. Drug metabolism is a common set of endpoints for ML models. Nothing you're describing is so mysterious or difficult that the problems "break" machine learning in this domain. > In the end your corpus is going to have to fall back to experiment All ML models in this space depend on experiment.