Jan Řezáč

#CompChem scientist at @IOCBPrague

[Jan Řezáč @RezacChem]

Machine learning is doing wonders for correcting issues in PM6 that we could not fix any other way. Interface to MOPAC available at https://t.co/FyPAs7ZbKt

[Jan Řezáč @RezacChem]

Our first paper of 2025:

PM6-ML: The Synergy of Semiempirical Quantum Chemistry and Machine Learning Transformed into a Practical Computational Method | Journal of Chemical Theory and Computation https://t.co/9bydQq7G40

[Jan Řezáč @RezacChem]

@Zubatyuk It's as good as the implementation of PM6. We use MOZYME in MOPAC, so its a couple of minutes for 1-2k of atoms on one CPU core. We do lots of these in parallel. We'll show some timing in the paper later.

[Jan Řezáč @RezacChem]

Our take on Δ-ML: Semiempirical quantum chemistry with a machine learning correction works better than either of these approaches alone, and scales well to large systems.

Preprint at ChemRxiv: https://t.co/kEJE8PPIXt

[Jan Řezáč @RezacChem]

@adrian_roitberg We briefly tested that by recomputing some structures from SPICE, and the difference was negligible. Fine grid and tighter convergence limits were used in Orca.

[Jan Řezáč @RezacChem]

@spiwokv To je špatný překlad, to mají být uhlovodany.

[Jan Řezáč @RezacChem]

This is how we automate our calculations on benchmark data sets:

Working with benchmark datasets in the Cuby framework https://t.co/cFSFBOJD83

[Jan Řezáč @RezacChem]

Completely agree. Why so many names for the same physics?

Aerogen Bond, Halogen Bond, Chalcogen Bond, Pnictogen Bond, Tetrel Bond, Triel Bond ... Why So Many Names? | Crystal Growth & Design https://t.co/bizZhEGo5d

[Jan Řezáč @RezacChem]

@AdamBask Thanks, I will check MACE out. The PhysicsML package looks interesting too.

ANI-2x reference DFT level is lower than what we are looking for.

[Jan Řezáč @RezacChem]

We're benchmarking our new method and want to add more general ML potentials to the comparison. The requirements are:
- code and model available
- trained on quality data (DFT-D or better)

So far we have AIMNet2, TorchMD-NET and ANI1-ccx.
Any other tips?

[Jan Řezáč @RezacChem]

@guillemsimeon I know - but there's the model trained on the SPICE data set available, one of the few which cover reasonable chemical space. And we also built another one ourselves.

[Jan Řezáč @RezacChem]

I wanted to test some #CompChem QM calculations on the ligands from the PDBbind 2020 database using the provided SDF files.

It turned out that 2975 of them (15 %) are open-shell molecules... Or their charge/protonation is off by 1.

[Jan Řezáč @RezacChem]

We just published a preprint on our tools for benchmarking #CompChem methods and on the latest extensions of our collection of ready to use datasets:

Working with benchmark datasets in the Cuby framework | ChemRxiv - https://t.co/fqlJ7peuCH

[IOCB Prague @IOCBPrague]

A universal physics-based scoring function from @RezacChem @madAniceP @Lepsik_science & J.Fanfrlík now in @NatureComms:
SQM2.20: Semiempirical quantum-mechanical scoring function yields DFT-quality protein–ligand binding affinity predictions in minutes: https://t.co/PjhadwiLRf