AI predicts odor

Artificial Intelligence Predicts Odors Like Humans

Humans detect smells by inhal­ing air that con­tains odor mol­e­cules, which then bind to recep­tors inside the nose, relay­ing mes­sages to the brain, where­as, AI inter­prets the sig­na­tures and clas­si­fies them based on a data­base of pre­vi­ous­ly col­lect­ed smells.

Unlike the human eye, which has only three sen­so­ry recep­tors for sens­ing the pho­tons of red, green, and blue col­or, the human nose has over 300 recep­tors, mak­ing it more dif­fi­cult to pre­dict smell than color.

A 2014 study showed that humans can dis­tin­guish at least 1 tril­lion dif­fer­ent odors. But you may won­der how an arti­fi­cial intel­li­gence will be able to han­dle the impos­si­ble. These are just ini­tial steps in AI sys­tems, which are being con­vert­ed more into human-like social indi­vid­u­als all over the world in dif­fer­ent areas.

In the recent times, Google has built an AI mod­el with a human-like capac­i­ty to pre­dict odors. Sci­en­tists have suc­cess­ful­ly for­mu­lat­ed a “Prin­ci­pal Odor Map” (POM) with the prop­er­ties of a sen­so­ry map. The map devel­oped by the team of Google AI links mol­e­c­u­lar struc­ture to the aro­ma of sub­stances and can even pre­dict smells that are still unno­tice­able by humans.

The Google researchers began the research in 2019 using a deep learn­ing algo­rithm. The type of smell inter­act­ed with the mol­e­c­u­lar struc­ture. Var­i­ous sam­ples of spe­cif­ic mol­e­cules were trained to be iden­ti­fied by a graph neur­al net­work (GNN) mod­el along with the smell labels they evoke, such as beefy, flo­ral, or minty.

Researchers looked into whether the GNN mod­el could learn to pre­dict the odors of new chem­i­cals that peo­ple had nev­er smelled before and that were dif­fer­ent from the mol­e­cules used to train it. The researchers referred to the study as an “impor­tant test” in their Google post. Many mod­els work well with data that resem­bles data they have pre­vi­ous­ly seen, but often fail when test­ed with new data.

The Google mod­el was suc­cess­ful and demon­strat­ed excep­tion­al intel­li­gence in pre­dict­ing smell from mol­e­cule structure.

The mod­el was also test­ed to see whether it could pre­dict how ani­mals would per­ceive odors. They found that the map was capa­ble of pre­cise­ly pre­dict­ing the activ­i­ty of sen­so­ry recep­tors, neu­rons, and behav­ior in most of the ani­mals that olfac­to­ry neu­ro­sci­en­tists have stud­ied, includ­ing mice and insects.

AI detect­ing odors can be used for a vari­ety of tasks, includ­ing iden­ti­fy­ing scents in the envi­ron­ment to aid peo­ple who have lost their sense of smell; and cre­at­ing new arti­fi­cial scents.

The research team at Google found that the com­mon appli­ca­tion of the sense of smell may be to detect and dis­tin­guish between var­i­ous meta­bol­ic states, such as know­ing when some­thing is ripe vs rot­ten, nutri­tious vs inert, or healthy vs sick.

They had gath­ered data about meta­bol­ic reac­tions in dozens of species across the king­doms of life and found that the map cor­re­spond­ed close­ly to metab­o­lism itself.

The sci­en­tists retrained the mod­el to tack­le the issue of the spread of dis­eases trans­mit­ted by mos­qui­toes and ticks while killing hun­dreds of thou­sands of peo­ple each year.

The team improved the orig­i­nal mod­el with two new sources of data. The first set was a long-for­got­ten set of exper­i­ments con­duct­ed by the USDA on human vol­un­teers begin­ning 80 years ago and recent­ly made dis­cov­er­able by Google Books. Sec­ond­ly, a new dataset was col­lect­ed by their part­ners at TOPIQ, using their high-through­put lab­o­ra­to­ry mos­qui­to assay.


With the help of the POM, researchers hope to pre­dict ani­mal olfac­tion to bet­ter respond to the dead­ly dis­eases trans­mit­ted by mos­qui­toes and ticks. Both datasets mea­sure how well a giv­en mol­e­cule keeps the mos­qui­tos away. Togeth­er, the result­ing mod­el can pre­dict the mos­qui­to repel­lence of near­ly any mol­e­cule, enabling a vir­tu­al screen over huge swaths of mol­e­c­u­lar space.

“Less expen­sive, longer last­ing, and safer repel­lents can reduce the world­wide inci­dence of dis­eases like malar­ia, poten­tial­ly sav­ing count­less lives,” said the researchers.

Accord­ing to their find­ings, a Prin­ci­pal Odor Map might be pro­duced using the researchers’ method of smell pre­dic­tion in order to tack­le odor-relat­ed issues more wide­ly. The key to mea­sur­ing smell was in the map. It pro­vid­ed answers to a vari­ety of ques­tions regard­ing new odors and the mol­e­cules respon­si­ble for them. The mod­el also linked the evo­lu­tion of odors to the nat­ur­al world.

To sum up, AI algo­rithms do have a poten­tial to effec­tive­ly pre­dict smell. The Google mod­el was one of the first to demon­strate this.

Leave a Reply

Your email address will not be published.

Join our NewsletterDaily Glimple of Future

Our blog, "Daily Glimpse of Future", strives to make the future much clearer than it is today. Join our newsletter for free now.