Sometimes it is fun to write about an article that is completely different from one’s specialty.
Here is a summary of: Improving the accuracy of Møller-Plesset perturbation theory with neural networks, The Journal of Chemical Physics 147, 161725 (2017); https://doi.org/10.1063/1.4986081
Bigger and faster is better, right? Well, let’s add accuracy into the equation. The D. E. Shaw Research group recently published a paper that describes a new method to model noncovalent interactions, spin-network-scaled Møller-Plesset perturbation (SNS-MP2). This new model has a 6- to 7-fold increase in accuracy. That is the equivalent of someone shooting a basketball from the 3 point line, now just as precisely throwing a ball into the net from standing beyond the opponent’s net, over 90ft away.
These scientists model intermolecular interactions by combining quantum chemistry and machine learning. To build the model, instead of using geometry and atomic composition information, the model uses quantum chemical features to focus on the nature of the interaction. The input for the model includes 22 features and 10 layers, each with 50 nodes, to produce this high level of accuracy. Furthermore, over 200,000 unique chemical data points were used to train the system with approximately 20,000 to validate and then an additional 20,000 to test the model. In its design, SNS-MP2 is built to augment, rather than circumvent, traditional quantum chemical calculations, encoding features in the calculations that can easily identify commonality between geometrically and chemically distinct fragments.
The implications of this research are far reaching in biology, material science, and chemistry whereby the modeling method may help with drugs binding to proteins, construction of nanomaterials, and the properties of liquids. The MP2 theory is one of the most widely used methods for modeling processes involved in molecular association, however, described in this recent publication is now a method that uses MP2 combined with modern big-data machine learning techniques to predict interaction energies with error bars as no other methods can perform. This revolutionary method, with open source availability, will greatly increase accuracy of modeling for intermolecular interactions.