Read on Twitter
ML/MM is the new QM/MM! In our new @biorxivpreprint, we show how to break the 1 kcal/mol accuracy barrier for alchemical free energy calculations using hybrid machine learning / molecular mechanics (ML/MM) potentials!
https://www.biorxiv.org/content/10.1101/2020.07.29.227959v1
This @biorxivpreprint represents the awesome work of Dom Rufa, @hannahbruce, @maxentile, @wiedermc, and @grinaway from the lab along with collaborators @adrian_roitberg and @olexandr
https://www.biorxiv.org/content/10.1101/2020.07.29.227959v1
collaborators @adrian_roitberg and @olexandrhttps://www.biorxiv.org/content/10.1101/2020.07.29.227959v1
We used the same Tyk2 kinase:inhibitor benchmark system from the @Schrodinger @J_A_C_S paper, asking whether replacing ligand interactions with an ML potential like ANI-2x could significantly improve the accuracy for free energy calculations.
ANI-2x: https://chemrxiv.org/articles/Extending_the_Applicability_of_the_ANI_Deep_Learning_Molecular_Potential_to_Sulfur_and_Halogens/11819268
ANI-2x: https://chemrxiv.org/articles/Extending_the_Applicability_of_the_ANI_Deep_Learning_Molecular_Potential_to_Sulfur_and_Halogens/11819268
Many formulations of hybrid ML/MM potentials are possible---this is just the beginning! We considered the simplest case where ligand intramolecular interactions are replaced by ANI-2x, inspired by awesome work from @RowleyGroup:
https://pubs.rsc.org/lv/content/articlehtml/2020/sc/c9sc06017k
https://pubs.rsc.org/lv/content/articlehtml/2020/sc/c9sc06017k
The uber-talented @TPCB_NYC grad student Dom Rufa came up with a way to use short nonequilibrium simulations to easily correct MM free energy calculations in a post-processing step, inspired by work from @gavincrooks.
We can't just reweight MM to QM-like ML potentials because of poor overlap, but surprisingly, 10 ps nonequilibrium trajectories are sufficient to give great overlap for MM -> ML/MM reweighting!
This gives great overlap between MM->ML/MM and ML/MM->MM work distributions, allowing us to estimate MM to ML/MM correction free energies just like @gavincrooks demonstrated in his thesis!
A must-read: https://threeplusone.com/pubs/GECthesis.pdf
A must-read: https://threeplusone.com/pubs/GECthesis.pdf
We started with relative free energy calculations from
@hannahbruce using the @openforcefield 1.0.0 ("Parsley") force field, which achieves an RMSE of 0.97 [95% CI: 0.68, 1.22] kcal/mol on the Tyk2 benchmark set, statistically indistinguishable from the @Schrodinger JACS result.
![We started with relative free energy calculations from @hannahbruce using the @openforcefield 1.0.0 ("Parsley") force field, which achieves an RMSE of 0.97 [95% CI: 0.68, 1.22] kcal/mol on the Tyk2 benchmark set, statistically indistinguishable from the @Schrodinger JACS result.](https://pbs.twimg.com/media/EeMVV-GUwAAaFYQ.png "We started with relative free energy calculations from @hannahbruce using the @openforcefield 1.0.0 (\"Parsley\") force field, which achieves an RMSE of 0.97 [95% CI: 0.68, 1.22] kcal/mol on the Tyk2 benchmark set, statistically indistinguishable from the @Schrodinger JACS result.")
@hannahbruce using the @openforcefield 1.0.0 ("Parsley") force field, which achieves an RMSE of 0.97 [95% CI: 0.68, 1.22] kcal/mol on the Tyk2 benchmark set, statistically indistinguishable from the @Schrodinger JACS result.
After post-processing to ML/MM using the @olexandr @adrian_roitberg ANI-2x, the free energy calculation RMSE drops to 0.47 [95% CI: 0.32, 0.68] kcal/mol for the Tyk2 kinase:inhibitor benchmark set, and the R^2 goes way up!
![After post-processing to ML/MM using the @olexandr @adrian_roitberg ANI-2x, the free energy calculation RMSE drops to 0.47 [95% CI: 0.32, 0.68] kcal/mol for the Tyk2 kinase:inhibitor benchmark set, and the R^2 goes way up!](https://pbs.twimg.com/media/EeMVcGdUYAIbzsh.png "After post-processing to ML/MM using the @olexandr @adrian_roitberg ANI-2x, the free energy calculation RMSE drops to 0.47 [95% CI: 0.32, 0.68] kcal/mol for the Tyk2 kinase:inhibitor benchmark set, and the R^2 goes way up!")
ANI-2x seems to do a much better job of reproducing torsions and torsion-valence/torsion couplings! Way to go @olexandr @adrian_roitberg!
This is only the beginning, but it's exciting to see where ML/MM will take us in the world of alchemical free energy calculations!
preprint: https://www.biorxiv.org/content/10.1101/2020.07.29.227959v1
code: https://github.com/choderalab/qmlify
preprint: https://www.biorxiv.org/content/10.1101/2020.07.29.227959v1
code: https://github.com/choderalab/qmlify
