Our current retrospective analysis will now shift back to comparing the accuracy of individual microstate transformations with the experimental data differences between compounds, now that we have more accurate microstate stereochemistry assignments.
When we return to benchmarking the accuracy of compounds, we currently use the estimate derived from running DiffNet with all available experimental data. This is not sensible for assessing the accuracy of predictions for compounds whose experimental data is available, since the data for that compound will contaminate the benchmark.
Instead, we should consider a few different variants:
We can use the DiffNet solution where no experimental data has yet been included, and all computed DeltaGs are shifted by a constant to minimize the RMSE
We can use the DiffNet solution where a single experimental measurement has been included and no constant is added to the computed DeltaGs
We can use the DiffNet solution where all other experimental measurements (except the compound in question) have been included, and no constant is added to the computed DeltaGs.
Our current retrospective analysis will now shift back to comparing the accuracy of individual microstate transformations with the experimental data differences between compounds, now that we have more accurate microstate stereochemistry assignments.
When we return to benchmarking the accuracy of compounds, we currently use the estimate derived from running DiffNet with all available experimental data. This is not sensible for assessing the accuracy of predictions for compounds whose experimental data is available, since the data for that compound will contaminate the benchmark.
Instead, we should consider a few different variants: