Design of phosphoryl containing podands with Li/Na selectivity using machine learning.

In this work we demonstrated, that machine learning opens a way for real design of ligands with required metal ion selectivity. We performed the ensemble QSPR modelling of the Li/Na complexation selectivity and the stability constants for the LiL and NaL complexes of phosphoryl podands in nonaqueous solvent THF/СНCl (4:1 v/v). The models were built and cross-validated using MLR with the ISIDA QSPR program and SVM with the libSVM package. The program SVMsmf was implemented to fulfil an ensemble modelling using libSVM and the Substructural Molecular Fragments (SMF) descriptors. SMF were used as descriptors for the ensemble modelling, properties predictions by consensus models and design of combinatorial library of new ligands. SMF such as the P=O group, the ether and P=O groups bound through the aromatic ring contribute significantly to the Li/Na selectivity. The developed models were applied for the prediction of the studied properties for a focused virtual library of 3057 phosphoryl podands generated using SMF contributions promising for selective binding of lithium. Consensus models selected hits for a synthesis by combinatorial library screening. Among the constructed selective ligands - hits, three new podands were synthesized, for which the experimentally estimated selectivity is in satisfactory agreement with that predicted.