After noticing wonky pKa prediction results from the current MarvinBeans version (6.1.6), I ran cxtrain on the same training set with archived versions and compared validation files. Same command, same training file - just reinstalling different archived versions. Between 6.0.6 and 6.1.0 four of my compounds see their fit error increase from 1 to 3 units to ~10. All other fit values are identical between versions.
The attached graphic plots fit-versus-experiment for 6.0.6 (red) and 6.1.0 (blue). Pairs of fit/experiment values are connected by a line. For most compounds, the fit values are identical, hence their points perfectly overlap, so points were jittered for presentation only. However, four pKa's see their the fit values going way south.
I'd love to have the best of the newest versions. How do we get this fixed in the next release?