Molecular design improved through machine learning

A) Parity plot comparing predicted and experimental log K values with mean error (ME = 0.032) lower than experimental accuracy and root mean squared error (RMSE = 0.629) approaching experimental errors. B) The top ranked ligands generated by linking pyridines and amides (blue functional groups).
A) Parity plot comparing predicted and experimental log K values with mean error (ME = 0.032) lower than experimental accuracy and root mean squared error (RMSE = 0.629) approaching experimental errors. B) The top ranked ligands generated by linking pyridines and amides (blue functional groups).

CMI researchers from Ames National Laboratory conducted the activity for this highlight

Innovation 
Developed software for metal complexing ligand design from molecular fragments and machine learning-predicted binding.

Achievement
We developed and published a fast computer-aided molecular design process that leverages machine learning to predict metal-ligand binding.

Significance and Impact

  • Accessible and expandable ligand design software. 
  • Fast, highly accurate predictions as compared to experiments or other computational methods.
  • In the works: a model with greater flexibility in ligand structures through an expanded and carefully curated data set, manuscript in progress. 

Hub Target Addressed 

  • Increasing the speed of discovery and integration
  • Predictive models for “real” materials
  • Workforce development