Multi-dimensional QSAR Studies on Aurora-A kinase Inhibitors

Recently we reported a new tool for 3D-QSAR using R statistical program. This applicability and utility of this tool was explored on developing 3D-QSAR models with Aurora kinase inhibitors dataset. Aurora kinase inhibitors are promising anticancer agents and several Aurora kinase inhibitors are therapeutic agents in cancer. To gain deeper insights in to the structural traits and to establish a mathematical relationship, 2D-QSAR studies with relatively new descriptor generator method called PyDescriptor and 4D-QSAR with receptor independent LQTA-QSAR were carried out. In all of these multi-dimensional QSAR protocols, MLR method implemented in QSARINS and descriptor refinement in R interface was adopted. Dataset of 35 Aurora A inhibitors distributed in 8 test set and 27 training set was subjected to these QSAR studies. The dataset used was unique in the sense it diversity in structural features and many fold differences in the biological activity with IC50 values ranging from 0.6 to 85. All models showed good predictability and validity. The best 2D QSAR model showed R2 = 0.9347, Q2 (leave one out) = 0.9056, where as the best 3D-QSAR model showed R2 = 0.8768, Q2 (leave one out) = 0.8271 and the best 4D-QSAR model showed R2 = 0.9607, Q2(leave one out) = 0.9333 with 5 variables. The descriptors minus_ringC_4B, lan_all_5A, minus_lipo_4B, fNacc4A and lipoplus_MSA were found contributing in the 2D QSAR models. The scripts namely ‘QSAR-R-HB.sh’, ‘QSAR-R-QQ.sh’ and ‘QSAR-R-LJ.sh’ were used to generate hydrogen bond acceptor/donor descriptor, coulomb electrostatic potential energy descriptors and steric Leenard-Jones potential energy descriptors respectively. The placements of these descriptors around compounds gave CoMFA like clues to design the more active compounds. In case of 4D-QSAR, molecular dynamics simulations were carried out on Gromacs program and conformational ensemble profiles (CEP) of each compound was devised. These CEPs were aligned with most active compound and Coulomb and LJ descriptors were generated subsequently. All models were found to be the highly predictive, reliable model because of its significant internal as well as external predictive power. The details of methodologies, the results would be presented. Currently, we are devising our own in-house methods to build a consensus pharmacophore to be used in virtual screening.