1. All ANI models were are fitted to wB97x  DFT functional data *minus* D dispersion term. This is done because dispersion is an analytical ad hoc correction. The intention that at the run time dispersion should be added back. D3 could be easily computed with ASED3 code referenced in our GitHub.
We did not find that the ANI models correlated perfectly with ωB97X, and have concerns about applying an empirical correction to ANI.
post hoc dispersion D3 correction to the ANI-1x and ANI-2x models slightly improve the performance over the non-dispersion corrected models, although the results were not statistically significant. On the other hand, it is not possible to calculate ωB97X dispersion corrections using any of the standard calculators (e.g. ASED3) as suggested by the reviewer, since no existing D3 calculator has parameters for ωB97X. In our manuscript, we can provide the corrections, since we have run the same molecules / geometries with ωB97X-D3 and can extract the dispersion correction directly.
In the case of the new D4 dispersion correction, we can use the dftd4 calculator to apply the correction for ωB97X to the ANI-1x and ANI-2x models - but the results surprisingly get worse. Again, the results are not statistically significant.
In the revised manuscript, we added a new table summarizing these findings, as well as a short section discussing the challenge of doing post hoc dispersion correction to a ML model.