When machine learning models learn chemistry I: quantifying explainability with matched molecular pairs | Synapse