SMELL U LATER

Dick Pountain /Idealog 362/ 05 Sep 2024 11:5

I have no qualms in claiming that I have a better (or least better trained) sense of smell than the average citizen. That’s partly, maybe mostly, because I studied organic chemistry in the 1960s. During my first few weeks of working in the cavernous Victorian college lab I was instructed to learn the odours of a dozen commonly encountered chemicals, and advised to employ smell as the first step in recognising any new compound. I can often tell an aldehyde from a ketone by sniffing, and became briefly addicted to ionone, cinnamaldehyde and menthol in succession, carrying little specimen tubes in my pocket. I’m sure this method is no longer taught, on health-and-safety grounds, as there are many substances nowadays that can kill at one sniff. 

In later life this training came in handy when my brother-in-law Pip founded The Scotch Malt Whisky Society and was writing a book that needed to categorise the nose of various famous spirits. Of course odour is now a huge business, not merely for perfumes as it has been for centuries but for those hundreds of flavourings contained in most supermarket foods which are manufactured in a huge chemical works in New Jersey. But smell has barely impinged upon the computer business so far, apart from the smell of burning insulation which most of us quickly learn to recognise (and investigate…)

I wrote semi-humorously in an earlier column about the possibility of a ‘sminter’ loaded with an assortment of smelly ‘inks’ that could be triggered via internet messaging, and I even got a letter some years later about an (unsuccessful) attempt at one. Even Hollywood attempted a brief stab at a Smell-o-Vision movie (‘Scent of Mystery’  by Mike Todd Jr.,1960 in case you’re interested) but the obstacles in both cases were the same, that smell is a chemical, not electronic, signal that moves at the speed of breeze rather than light – and you can’t just switch it off quickly either…

But a far more serious obstacle is that while the components of human light perception are threefold – red, green and blue retinal cells – the components of smell perception are vastly more numerous. Our noses contain at least 400 different chemical receptors and individual smells are recognised by trillions of combinations their outputs, which release a plethora of proteins that are still not entirely understood.  

But when you hear the word trillions nowadays, you’re usually either talking about a GPT (or at least about NVidia’s market cap). Understanding smell perception, like protein folding and DNA sequencing, is a perfect candidate for AI to analyse, so it comes as no surprise to learn (via an article in Nature https://www.nature.com/articles/d41586-024-02833-4) that many teams are working toward this end, with ample financing from industries. 

The problem has several aspects, which include: predicting the smell of a molecule from its structure; decoding the output of human odour sensors for particular compounds; and automating comparison of smells of different mixtures by identifying their components. The current hot variant of AI – the Generative Pre-trained Transformer (GPT) – works using the mathematics of parameter spaces: identify the important parameters of the subject to be analysed, apply tensor calculus to create a multidimensional space with a dimension for each parameter, and then map training examples into this space. For graphical AIs like Stable Diffusion and MidJourney such spaces already have trillions of parameters for identifying shapes in visual worlds.  

One immediate problem for applying this to smell is getting training data: odour receptors, whether human or animal, are hard to study, often won’t work outside the creature, fragile and the amount of protein released is minuscule. Two receptors from insects and two more from mice have been deciphered in the last year, leaving just 400+ more to go. A team at Duke University in North Carolina is using AlphaFold and machine learning to screen millions of chemicals for binding to two synthetic receptors they’ve engineered. A very important motivation for such work is to use smell recognition in diagnostic medicine by identifying odour molecules produced by disease processes (dogs are doing this already). Precisely how and where odour nerve signals are processed in the brain is perhaps the leading-edge study right now. 

Real progress is being made and AI may soon speed it enormously, but smell remains the least understood of our senses, and least amenable to digital manipulation. It’s so subjective that human tasters and perfumiers will retain an advantage over automated solutions for far longer than most professions, and I don’t expect to have to consult my laptop when I’m mixing our own custom bath oil from my little box of tubes of neroli, rose, ylang-ylang and sandalwood oils (plus several other secret ingredients).  

[Dick Pountain believes that a rose by any other name would smell mostly of β-phenylethanol]