Artificial Intelligence Could Bring Us the New Iso E Super. Will It?
Or will newly-discovered fragrance molecules continue to be hoarded by flavor and fragrance companies like Firmenich and Givaudan?
Every year, new fragrance molecules are designed in chemistry labs and released into the eager hands of perfumers. Incorporated into products, these newly synthesized smells make their way into the fragrance market, waiting to greet curious consumers who have never smelled anything exactly like this before.
Many of these are what are called captive molecules: proprietary fragrance molecules protected by patents. By patenting a fragrance molecule, a company reserves an exclusive legal right to use it in its perfumes. Examples of captive molecules include Petalia (Givaudan) in Burberry’s Burberry Brit Rhythm for Women, Hedione (Firmenich) in Dior’s Eau Sauvage, and Dynascone in Davidoff’s Cool Water.
Sometimes these patents expire, and the molecule becomes available for public use. One example of this is Ambermax, a previously-captive Givaudan molecule representing a woody-amber note that sticks very well to fabric.
But is the discovery of these new fragrance molecules an innately human task? A paper published in ScienceDirect in January of 2022 says otherwise.
In “Design of Fragrance Molecules Using Computer-Aided Molecular Design with Machine Learning,” Dr. Yee Jia Ooi et al. propose a system for using artificial intelligence to create new fragrance molecules. The work falls under the domain of computer-aided molecular design (CAMD), the field of study concerning computational approaches to molecule creation.
The paper describes a machine learning system able to identify both the physical properties and the estimated odor characteristics of a given fragrance molecule. The system was trained with a set of fifty molecules and tested with a set of twenty.
Each molecule had been annotated with one of twenty mutually exclusive category labels. Each of these categories was further grouped into either “pleasant” or “unpleasant” scents, as determined by surveyed humans.
The scent categories were the following: edible, bakery, sweet, fruit, fish, garlic, spices, cold, sour, burnt, acid, warm, musky, sweaty, ammonia, decayed, wood, grass, flower, and chemical.
Of the seventy samples, 33 were identified as pleasant, and 37 as unpleasant.
In addition to predicting the qualities of a presented molecule, the model was able to propose feasible new molecules that meet structural requirements to be synthesized in a stable form and are estimated to have a pleasant scent. It provided descriptions of the odors of these molecules, including scores representing their pleasantness and strength.
Machine learning systems like this one could greatly speed up the process of discovering new aroma molecules. It’s an exciting prospect: refinement of such tools could bring dozens of new smells and innovations in fragrance upon us.
In the age of captive molecules, however, many of these are likely to be declared proprietary and reserved for sale by a single patent-holding company. This raises the question: Are we on the brink of an AI-aided fragrance molecule arms race between fragrance and flavor companies like Firmenich and Givaudan?
Furthermore, is that patent culture going to keep these new molecules from really making a mark on the fragrance industry? The restriction of captive molecules makes it unlikely that they’ll be experimented with, examined, and used to their full potential as they would be were they available to the public.
Consider the molecules that have captivated the fragrance industry lately.
The most in-vogue right now is arguably the diaphanous Iso E Super, also called Timbersilk. Discovered by John B. Hall and James M. Sanders from International Flavors and Fragrances (IFF) in 1973, the molecule was patented but has not been held captive. It can be bought online by anyone, and seemingly uncountable perfume houses big and small are experimenting with its elusive, warm, skin-smell properties.
Its close cousin, Iso Gamma Super, is a captive molecule held by IFF, used only in a few compositions like Escentric Molecules’ Molecule 01 and Ellis Brooklyn’s aptly-named Iso Gamma Super. It’s similar, but noticeably different from Iso E to the trained nose, such that the average home perfumer can’t simply mix up a copy of Molecule 01 from Iso E and alcohol.
Sure, the captive molecule status keeps Molecule 01 and other fragrances like it from being easily copied. But at what cost to creativity and innovation in the industry as a whole?
Had Iso E Super been a captive molecule, it never would have captivated audiences the way it does today. It would remain inimitable but little-known, lost in the notes of a single fragrance or two. It wouldn’t have left a substantial mark on the world of perfumery.
Consider also ambroxan, the molecule whose synthetic production was developed in the 1950s as a replacement for ambergris. It’s oft-bemoaned these days for its liberal use in so many fragrances, but there’s no question ambroxan has had a profound impact on the industry.
By providing a far more affordable (and ethical) alternative to natural sperm-whale-produced ambergris, ambroxan allows the proliferation of a scent that wouldn’t otherwise be accessible to most perfumers. True ambergris now cannot be legally sold in many countries, including the United States. Without ambroxan, the entire tradition of that unique warm and salty-fresh marine accord would be entirely lost.
Computer-aided molecule design could help us find more ambroxans, so to speak. Better and more affordable replacements to the disappearing perfume ingredients of yesterday are always welcome and useful in the community. Every year, more commonly-used fragrance materials become unavailable due to legislation, health concerns, ethical concerns surrounding animal-sourced molecules, and simple extinction. Most recently, the European Union banned the use of lilial, a fragrance molecule approximating the smell of lilies very common in makeup, hair, and skin products.
What if we could create a better fake oakmoss? A more convincing pseudo-Santal sandalwood? A safer lilial? Artificial intelligence could help us sort through all the many molecular options out there faster to approximate these disappearing scents more closely.
Between approximating the scents of yesterday and discovering those of tomorrow, artificial intelligence is a powerful tool for sifting through the molecular noise and bringing new materials to perfumery.
It is for this reason that I am both thrilled and very cautious at the prospect of computer-aided systems for fragrance molecule design. Making it easier to discover new aroma molecules should be a boon to the industry and art of perfumery. Given the current atmosphere of fragrance and flavor companies eagerly bottling up every new smell they find, though, I fear that won’t be the case. Systems such as this one add to the tools in the discovery arsenal of large fragrance brands without ensuring that the benefit trickles down to the art and community at large.
If the authors of this paper (or anyone else) wish to issue a machine learning system for fragrance molecule design to consumers, they ought to carefully consider how these systems would be used. A stipulation that all molecules produced by the system cannot be kept captive would ensure that the new molecules can be used by all, not just by the deep-pocketed flavor company that gets to them first.
It’s a well-known fact that open source work — which is free for anyone to read, copy, and build off of — is the backbone of computer science. Though the term “open source” is rarely used in this context, the same applies to perfumery. Tools and systems such as this one for computer-aided fragrance molecule design are an opportunity to discover beautiful new molecules. They’re unlikely to make a truly meaningful impact on the industry, however, unless they’re ensconced in an open-source philosophy of releasing these molecules for public use.