Computers predict molecules' scents from their structures - C&EN

A team of researchers and volunteers from across the globe have trained computers to predict the way a molecule will smell based on its structure. The...
0 downloads 8 Views 142KB Size
Concentrates Chemistry news from the week

▸ Highlights First α-helical amyloid fibril observed Nanoparticles reduce nitrogen runoff Nanoscience pioneer Millie Dresselhaus dies at 86 Making colloids with molten salts U.S. grand jury indicts AK Scientific for smuggling Tronox plans to buy TiO2 rival Cristal EPA foe Scott Pruitt now leads the agency he sued Chemical manufacturers push for tax code rewrite

5 5 7 9 12 13 17 17

COMPUTATIONAL CHEMISTRY

Computers predict molecules’ scents from their structures Crowdsourcing project feeds odor perception data into machine-learning algorithms to sniff out compounds’ aroma profiles A team of researchers and volunteers from IBM-Rockefeller team provided across the globe have trained computers to a portion of these data to 22 volpredict the way a molecule will smell based unteer teams of computer scienon its structure. The feat may help scientists tists who used it to develop maunravel the still-mysterious relationship chine-learning algorithms that between molecular structure and odor percan predict a molecule’s scent. ception (Science 2017, DOI: 10.1126/science. Eventually, the volunteer teams aal2014). received the remaining data to This achievement in machine learning test their algorithms’ predictive could also be a boon to the fragrance induscapabilities. try, saving time and money that would otherThe models performed surHuman sniffers evaluated the odors of 476 wise be spent on laborious human sniff testprisingly well, and a final model, compounds in vials for a new study. ing. It even suggests that computers might which the IBM-Rockefeller one day identify molecules by their odor. group generated by combining perception generally involve the binding of To tackle the complex and poorly unvarious volunteer algorithms, was able to molecules to receptors in the nose. These derstood phenomenon of olfaction, an accurately identify a molecule’s “pleasanthypotheses are complicated by the fact that international team led by Pablo Meyer at ness” and “intensity,” as well as eight out of some very differently shaped molecules IBM and Leslie B. Voshall and Andreas Keller 19 other odor qualities initially assessed by can have similar smells, whereas similarly at Rockefeller University devised a crowdthe human sniffers. shaped molecules can sometimes smell sourced project, the DREAM Olfaction PreThe results show that computers can zero quite different from each other. diction Challenge. in on molecular characteristics responsible Scientists recognize that the use of maThey assembled a huge set of human odor for odors that humans aren’t able to discern chine learning to unravel these mysteries is perception data by arranging for 49 people to by looking at a molecule’s structure. in its infancy. The authors of the new study sniff 476 molecules in separate vials. These The IBM-Rockefeller effort is “a major envision future experiments involving odor human sniffers categostep forward in decodmixtures and human sniff data sets that rized each molecule acing how the brain inter- incorporate different cultural and genetic cording to various odor prets messages from populations. qualities such as “garthe nose,” says Eric Given that so much remains to be dislic,” “sweet,” “fruit,” Block, chemistry procovered about human odor perception, the ▸ 49: Number of human sniffers “spices,” “burnt,” fessor at the University identification of molecules based on their ▸ 476: Number of molecules sniffed “urinous,” “decayed,” at Albany, SUNY, and odor “may be on the horizon but is not yet in and “fish.” The group by each human an olfaction chemistry hand,” Block observes. ▸ 19: Number of “smells,” such as also collected 4,884 expert. Still, the authors note, this new strategy structural features garlic, fish, or flower, the humans rated Scientists have nevmay accelerate efforts to understand basic such as atom types and for each molecule er had a good handle on interactions between odor molecules and functional groups from ▸ 22: Number of volunteer teams that what makes a molecule their receptors and “test predictive models the molecules. devised machine-learning algorithms smell the way it does. of olfactory coding in both humans and aniThen, the to predict molecules’ scents Theories about odor mal models.”—ELIZABETH WILSON

CREDIT: ROCKEFELLER UNIVERSITY

Olfaction study by the numbers

FEBRUARY 27, 2017 | CEN.ACS.ORG | C&EN

3