Dear Caspar,
The issue is probably in how ORCA handles external programs. The QM/MM interface to ORCA supports electrostatic embedding but this was developed and tested for ORCA’s built-in HF and DFT models. The key is that the QM code includes the point charges in the Hamiltonian and computes a point charge gradient (or electric field due to QM charge density at point charge positions). It appears from your description that ORCA does not handle point charges with external tools.
PM6-ML applies a D3 dispersion and Delta-ML correction to the PM6 Hamiltonian. Hence this should work just like PM6 for electrostatic interactions in a QM/MM calculation. However, it’s not currently supported in Amber and I am not aware of anybody working on it.
Since you mentioned xTB, this software is coupled to Amber and can be used for QM/MM simulations with GFN1-xTB and GFN2-xTB Hamiltonians. This also supports proper QM/MM Ewald electrostatics while all interfaces to external QM codes such as ORCA use a cutoff.
All the best,
Andy
—
Dr. Andreas W. Goetz
Associate Research Scientist
San Diego Supercomputer Center
Tel: +1-858-822-4771
Email: agoetz.sdsc.edu
Web: www.awgoetz.de
On Mar 25, 2026, at 1:21 AM, Schattenberg, Caspar via AMBER <amber.ambermd.org> wrote:
Dear Amber Mailing List,
I am using the AMBER program for QM/MM of Peptides and Proteins, and recently came across the machine learning corrected PM6 model as discussed in this paper:
https://urldefense.com/v3/__https://pubs.acs.org/doi/10.1021/acs.jctc.4c01330__;!!Mih3wA!Bt7XZiqvlBTqNglrnpfhDcsH4urZUJgE1J6fHma2GWXvlPzvN2jWpqSJA6xLDbCwHLOM7t6Dub3Vgw$
I have tested it for amino-acids in the gas-phase and it appears to provide extremely accurate structures. For this reason I tried to run the approach in AMBER with MOPAC as an external program used via ORCA.
(AMBER as the QM/MM driver, ORCA as the external QM program, and in ORCA I used external tools:
https://urldefense.com/v3/__https://github.com/faccts/orca-external-tools/__;!!Mih3wA!Bt7XZiqvlBTqNglrnpfhDcsH4urZUJgE1J6fHma2GWXvlPzvN2jWpqSJA6xLDbCwHLOM7t4U3uUjvQ$
which allowed to link AMBER to the PM6ML-MOPAC version provided here:
https://urldefense.com/v3/__https://github.com/Honza-R/mopac-ml__;!!Mih3wA!Bt7XZiqvlBTqNglrnpfhDcsH4urZUJgE1J6fHma2GWXvlPzvN2jWpqSJA6xLDbCwHLOM7t7AX350eg$
)
Unfortunately, the computations break down, as no point-charge gradient are provided.
Is there maybe a different way to use such models?
Can anyone tell me if the reason is that point charges are not supported (yet) by such models?
(for gXTB tested via the same approach I get the same result)
On the other hand, it there any plans for the future to include such models into the program?
Any hints or suggestions would be highly appreciated.
Thank you very much, and best regards,
Caspar
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
https://urldefense.com/v3/__http://lists.ambermd.org/mailman/listinfo/amber__;!!Mih3wA!Bt7XZiqvlBTqNglrnpfhDcsH4urZUJgE1J6fHma2GWXvlPzvN2jWpqSJA6xLDbCwHLOM7t4974BVOw$
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Mar 25 2026 - 17:00:02 PDT
