Do machines dream of atoms? Crippen's logP as a quantitative molecular benchmark for explainable AI heatmaps

While there is a great deal of interest in methods aimed at explaining machine learning predictions of chemical properties, it is difficult to quantitatively benchmark such methods, especially for regression tasks. We show that the Crippen logP model (J. Chem. Inf. Comput. Sci. 1999, 39, 868) provides an excellent benchmark for atomic attribution/heatmap approaches, especially if the ground truth heatmaps can be adjusted to reflect the molecular representation. The “atom attribution from finger prints”-method developed by Riniker and Landrum (J. Chem. Inf. Comput. Sci. 2013, 5, 43) gives atomic attribution heatmaps that are in reasonable agreement with the atomic contribution heatmaps of the Crippen logP model for most molecules, with average heatmap overlaps of up to 0.54. The agreement is increased significantly (to 0.75) when the atomic contributions are adjusted to match the fact that the molecular representation is fragment-based rather than atom-based (the finger print-adapted (FPA) ground truth vector). Most heatmaps and the corresponding FPA overlaps are relatively insensitive to the training set size and the results are close to converged for a training set size of 1000 molecules, although for molecules with low overlap some heatmaps change significantly. Using the “remove an atom” approach for graph convolutional neural networks (GCNNs) suggested by Matveieva and Polishchuk (J. Cheminform. 2021, 13, 41) we find an average heatmap overlap of 0.47 for the atomic contribution heatmaps of the Crippen logP model. Like the simpler attribution benchmarks for classification tasks that have come before it, this work sets the bar for regression tasks.