Figure 7.

Step plots showing improvement in RMSD and TFD of optimized conformer geometries and a closer match of ΔΔE with previous generations of force fields. The error bars are bootstrapped errors for each bin. The force field Sage 2.0.0 is highlighted with a bold line, while other force fields are shown with narrower lines. Overall, Sage appears to do substantially better than previous releases based on geometric measures (RMSD, TFD), while performing only marginally worse than the best prior force field (Parsley 1.2.0) on ΔΔE. The slight drop in ΔΔE values can likely be attributed to the reduced set of opt-geo training targets used to train Sage.