Machine learning trains neural nets to simulate molecular motion

-

-

Diagram of the transfer learning technique evaluated in this work. Transfer learning starts from a pre-trained, lower-accuracy ANI-1x DFT model and then retrains to higher-accuracy CCSD(T)*/CBS data with some parameters fixed during training.

 

-

Computational modeling of chemical and biological systems at atomic resolution is a very valuable scientific tool; however, there is an inherent trade-off between accuracy, speed, and transferability. Researchers from Los Alamos National Laboratory, University of Florida, and University of North Carolina at Chapel Hill mitigated these issues and advanced the accessibility of computational modeling by employing machine learning, specifically transfer learning. The impact of their research is immense and will push the fields of computational biology and drug development to new frontiers. Their research was published in Nature Communications.

The researchers saw the limitations of current molecular modeling techniques, such as when a model is fine-tuned for one particular compound it makes it difficult to accurately transfer the simulation to a different compound. They wanted to develop a computational modeling technique that maintained accuracy when transferred to numerous compounds. To do this, the researchers trained a neural net first on a large amount of lower-accuracy data and then on a small amount of higher-accuracy data. The result was a general-purpose potential that they named ANI-1ccx. Written in a user-friendly code—Python—ANI-1ccx was made available to the public as open-source software on GitHub.

After extensive benchmarking, the researchers concluded that ANI-1ccx captures a broad range of organic chemistry, with accuracy comparable to quantum mechanics calculations at the coupled-cluster level of theory. This work offers a computationally efficient and accurate machine learning-based molecular potential for general use across a broad range of chemical systems. In short, ANI-1ccx is accurate and transferable.

Funding was provided by the LANL Laboratory Directed Research and Development (LDRD) Program. This research supports the Laboratory’s Global Security mission and the Information Science and Technology science pillar.

-

Watch the new ANI-2x (CHNOSFCl) potential in action