Alexander Donchev

Many proteins can be partially or completely disordered under physiological conditions. Structural characterization of these disordered states using experimental methods can be challenging, since they are composed of a structurally heterogeneous ensemble of conformations rather than a single dominant conformation. Molecular dynamics (MD) simulations should in principle provide an ideal tool for elucidating the composition and behavior of disordered states at an atomic level of detail. Unfortunately, MD simulations using current physics-based models tend to produce disordered-state ensembles that are structurally too compact relative to experiments. We find that the water models typically used in MD simulations significantly underestimate London dispersion interactions, and speculate that this may be a possible reason for these erroneous results. To test this hypothesis, we create a new water model, TIP4P-D, that approximately corrects for these deficiencies in modeling water dispersion interactions while maintaining compatibility with existing physics-based models. We show that simulations of solvated proteins using this new water model typically result in disordered states that are substantially more expanded and in better agreement with experiment. These results represent a significant step toward extending the range of applicability of MD simulations to include the study of (partially or fully) disordered protein states.

Advances in computational chemistry create an ongoing need for larger and higher-quality datasets that characterize noncovalent molecular interactions. We present three benchmark collections of quantum mechanical data, covering approximately 3,700 distinct types of interacting molecule pairs. The first collection, which we refer to as DES370K, contains interaction energies for more than 370,000 dimer geometries. These were computed using the coupled-cluster method with single, double, and perturbative triple excitations [CCSD(T)], which is widely regarded as the gold-standard method in electronic structure theory. Our second benchmark collection, a core representative subset of DES370K called DES15K, is intended for more computationally demanding applications of the data. Finally, DES5M, our third collection, comprises interaction energies for nearly 5,000,000 dimer geometries; these were calculated using SNS-MP2, a machine learning approach that provides results with accuracy comparable to that of our coupled-cluster training data. These datasets may prove useful in the development of density functionals, empirically corrected wavefunction-based approaches, semi-empirical methods, force fields, and models trained using machine learning methods.

Noncovalent interactions are of fundamental importance across the disciplines of chemistry, materials science, and biology. Quantum chemical calculations on noncovalently bound complexes, which allow for the quantification of properties such as binding energies and geometries, play an essential role in advancing our understanding of, and building models for, a vast array of complex processes involving molecular association or self-assembly. Because of its relatively modest computational cost, second-order Møller-Plesset perturbation (MP2) theory is one of the most widely used methods in quantum chemistry for studying noncovalent interactions. MP2 is, however, plagued by serious errors due to its incomplete treatment of electron correlation, especially when modeling van der Waals interactions and π-stacked complexes. Here we present spin-network-scaled MP2 (SNS-MP2), a new semi-empirical MP2-based method for dimer interaction-energy calculations. To correct for errors in MP2, SNS-MP2 uses quantum chemical features of the complex under study in conjunction with a neural network to reweight terms appearing in the total MP2 interaction energy. The method has been trained on a new data set consisting of over 200 000 complete basis set (CBS)-extrapolated coupled-cluster interaction energies, which are considered the gold standard for chemical accuracy. SNS-MP2 predicts gold-standard binding energies of unseen test compounds with a mean absolute error of 0.04 kcal mol−1 (root-mean-square error 0.09 kcal mol−1), a 6- to 7-fold improvement over MP2. To the best of our knowledge, its accuracy exceeds that of all extant density functional theory- and wavefunction-based methods of similar computational cost, and is very close to the intrinsic accuracy of our benchmark coupled-cluster methodology itself. Furthermore, SNS-MP2 provides reliable per-conformation confidence intervals on the predicted interaction energies, a feature not available from any alternative method.