<p>Utilizing a genetic algorithm training of the atom dipole interaction model was performed to arrive at C,H, N, & O atomic polarizabilities that constitute a correction to semiempirical molecular polarizability calculations increasing the accuracy of these calculations to near parity with DFT at a fraction of the computational load.</p>

Indiana University-Purdue University Indianapolis (IUPUI) / Molecular polarizability plays a significant role in chemistry, biology, and medicine. Classical prediction of polarizability often relies on atomic-type specific polarizability optimized for training set molecules, which limits the calculations to systems of similar chemical environment. Although ab initio (AI) quantum mechanical (QM) methods are more transferable in predicting molecular polarizability, their high computational costs especially when used with large basis sets for obtaining quantitatively reliable results make them less practical. To obtain accurate QM polarizability in an efficient manner, we have developed a dual-level approach, where the polarizability (α) obtained from the efficient semiempirical QM (SE) method is corrected using a set of element-base atomic polarizabilities derived from the atomic dipole interaction model (ADIM) to reproduce the density functional theory (DFT) results. We have optimized the atomic polarizability correction parameters for CHON-containing systems using a small training set of molecules and tested the resulting SE-ADIM model on the neutral drug-like molecules in the QM7B database. SE-ADIM corrected AM1 showed substantial improvement with its relative percent error (RPE) compared to B3LYP reduced from 33.81% to 3.35%. To further test its robustness for larger molecules in broader chemical bonding situations, we applied this method to a collection of drug molecules from the e-Drug3D database. For the 1004 molecules tested, our SE-ADIM model, which only contains four empirical parameters, greatly reduces the RPE in AM1 polarizability relative to B3LYP from 26.8% to 2.9%. Error decomposition shows consistent improvements across molecules with diverse bond saturations, molecular sizes, and charge states. In addition, we have applied AlphaML, a promising machine learning (ML) technique for predicting molecular polarizability, to the e-Drug3D dataset to compare its performance with our SE-ADIM correction of AM1. We found SE-ADIM performs competitively with AlphaML bolstering our confidence in the value of our method. Errors distinct to AlphaML were also discovered. We found four molecules for which AlphaML predicts negative molecular polarizabilities, all of which were peroxides. In contrast, SE-ADIM has no such issue with these molecules or this chemical type. Finally, to improve performance of SE-ADIM when correcting AM1 molecular polarizability calculations for charged molecules, we introduce a charge dependent polarizability (CDP) enabled SE-ADIM. Training the CDP enabled SE-ADIM with a single additional parameter, B, we were able to reduce error in AM1 molecular polarizability calculations of charged molecules relative to B3LYP from 29.57% to 5.16%. By contrast, SE-ADIM without CDP corrected AM1 relative to B3LYP had an RPE of 8.56%. The most benefit of CDP was evident within negatively charged molecules where AM1 error relative to B3LYP fell from 32.20% to 3.77% while SE-ADIM without CDP enabled error for these same negative molecules was 10.06%.