Machine learning marched forward
Chemists demonstrated algorithms that could predict molecular properties and plan reactions
by Sam Lemonick
Intelligent robots may yet take our jobs; they definitely grabbed a lot of C&EN’s headlines this year. Researchers continued to explore applications of machine learning, a type of artificial intelligence used to make predictions or decisions through algorithms that can learn from large data sets. The technology powers self-driving cars and image-recognition software.
Scientists demonstrated numerous ways in which machine learning can help explore chemical space. For example, Heather Kulik of the Massachusetts Institute of Technology and colleagues identified inorganic molecules called spin-crossover complexes that could be useful as sensors or electronic switches (J. Phys. Chem. Lett. 2018, DOI: 10.1021/acs.jpclett.8b00170). Apurva Mehta of the SLAC National Accelerator Laboratory, along with collaborators, used machine learning to identify new alloys that are metallic glasses (Sci. Adv. 2018, DOI: 10.1126/sciadv.aaq1566). And the chemical company Symrise teamed up with IBM to search for new fragrances using machine learning.
Meanwhile, Thomas F. Miller and coworkers at the California Institute of Technology showed how machine learning can benefit chemical modeling, demonstrating how it can predict electronic properties of molecules with high accuracy and low computational cost (J. Chem. Theory Comput. 2018, DOI: 10.1021/acs.jctc.8b00636). In related work, Adrian Roitberg of the University of Florida showed off a machine-learning-based tool that calculates molecular forces and energies and offers high performance at a low computational cost.
Synthetic chemists also experimented with machine learning. Abigail G. Doyle of Princeton University worked with colleagues there and at Merck & Co. to optimize the yield of an amination reaction by asking their algorithm to vary the reagents used (Science 2018, DOI: 10.1126/science.aar5169). Bartosz Grzybowski of Ulsan National Institute of Science and Technology and the Polish Academy of Sciences put the synthesis-planning Chematica software, which he developed and MilliporeSigma acquired, to the test. Human chemists found that the computer program charted routes to products that were at least as good as those humans have developed (Chem 2018, DOI: 10.1016/j.chempr.2018.02.002). And Alán Aspuru-Guzik of the University of Toronto is one of several chemists who applied machine learning in software capable of independently running experiments and then using the results to improve the procedures.
Do these advances mean that machine learning is living up to its hype? Chemists have mixed feelings about the field. Most agree it’s a useful tool despite overheated enthusiasm from some corners.