How to DX corona detector dog to e-Nose

I read an article in May or so that said corona-sniffing dogs can sniff out infected people with over 90% accuracy. It seems that the virus itself does not have a smell, but responds to something in the sweat of the infected person.

As is well known, there are only five types of taste, but mammals are said to be able to distinguish about 1000 different smells. That is not to say that we have a thousand different sensors, but that we can distinguish different output patterns from several different sensors. It is the same way that light can represent a million colors with three primary colors.

For example, odors that humans perceive as bad smells are regulated by the Offensive Odor Control Law.
What kind of patterns can be found are as follows.

offensive odorCAS
Ammonia7664-41-7ammoniacal
methyl mercaptan74-93-1decomposing cabbage garlic
Hydrogen Sulfide7783-06-4odor of rotten eggs
dimethyl sulfide75-18-3cabbage-like smell
dimethyl disulfide624-92-0sulfurous vegetable cabbage onion
trimethylamine75-50-3strong “fishy” odor
acetaldehyde75-07-0Pungent, fruity odor
propionaldehyde123-38-6pungent odor
butyraldehyde123-72-8pungent odor
isobutyraldehyde78-84-2pungent odor
pentanal110-62-3Diffusive, fermented, bready, fruity with berry nuances
3-methylbutanal590-86-3ethereal aldehydic chocolate peach fatty
isobutanol78-83-1sweet musty odor
ethyl acetate141-78-6Etherial, fruity, sweet, grape and rum-like
4-methyl-2-pentanone108-10-1Fruity, Ethereal, Spicy
toluene108-88-3sweet
styrene100-42-5sweet balsam floral plastic
xylene
propanoic acid79-09-4Pungent acidic and dairy-like
butyric acid107-92-6Sharp, dairy-like, cheesy, buttery with a fruity nuance
pentanoic acid109-52-4Acidic and sharp, cheese-like, sour milky, tobacco, with fruity nuances
3-methylbutanoic acid503-74-2Cheese, dairy, acidic, sour, pungent, fruity, stinky, ripe fatty and fruity notes
Skatole83-34-1very strong animal fecal indole civet
Cadaverine462-94-2cadaverous
1,4-Butanediamine110-60-1cadaverous

With words, it can be difficult to convey the smell.
But if we use the Hansen solubility parameters (HSP), we get the following.

Odor typedDdPdHdodHacname
Smells like rotten onions.17.65.806methyl mercaptan
Rotten onions17.54.504.3methyl sulfide
Rotten vegetables, garlic177.804.5methyl disulfide
Odor typedDdPdHdodHacname
Pungent green odor1613.709.5acetaldehyde
Pungent, sweet and sour burnt smell.15.710.807.2propionaldehyde
Pungent, sweet and sour burnt smell.15.98.605.9butyraldehyde
Pungent, sweet and sour burnt smell.15.68.405.1Isobutyraldehyde
Rancid, sweet, sour, burnt smell.167.505.5valeraldehyde
Rancid, sweet, sour, burnt smell.15.57.104.8iso-valeraldehyde
Odor typedDdPdHdodHacname
Pungent, paint thinner-like odor15.85.807.5ethyl acetate
Pungent, paint thinner-like odor15.75.903.8MIBK
Smells like gasoline.17.62.804.4toluene
Smells like city gas.17.52.403.8styrene
Smells like gasoline.17.82.703.5xylene

In other words, if the HSPs are similar, the smells are similar.
For more information, see my web page

Hansen Solubility Parameters (HSP) and Offensive Odors
Fragrance and Hansen Solubility Parameters (HSP)

Therefore, we believe that if the HSPs are similar, the solubility in olfactory cells will also be similar.

In fact, it is this smell that got me into the development of HSP.
A friend of mine at CALTECH was considering an e-Nose (artificial nose). (This was in 2008)

Their method is to coat the electrodes with seven different polymers.
When an odorant substance dissolves in it, it swells and provides a sensor signal. After a period of time, it desorbs.
The signals from the seven different sensors are used to train a neural network to identify the odor.

What kind of polymer should be used for coating?
It is best to use polymers with characteristic Hansen solubility parameters.

This is how I started my relationship with Dr. Hansen and prof. Abbott.

When you educate a corona detection dog, you probably teach it to sniff and sniff until it correctly recognizes that this is an infected person and that this is a non-infected person. This is completely different from using a gas chromatograph and saying that the presence of this ingredient indicates an infected person.
The method of teaching the e-Nose is exactly the same as that of teaching corona detection dogs.

When I don’t know the causal relationship, I tend to use qualitative SOM.
In this article, let’s learn how to educate AI using Shionogi’s oral drug S-217622 for corona as an example.

I read in today’s newspaper that Shionogi’s oral drug S-217622 for corona was also effective against the Omicron strain. It is said to suppress the action of an enzyme called “protease” in the cells.
At that time, HSP researchers think that some HSP of this drug and enzyme may be similar.

Since such a large compound is not likely to have a smell, we consider a substructure divided into three parts.
If the temperature is high enough, there may be an odor.

Let’s assume that the Hansen Solubility Parameters (HSP) of the left (L), middle (C), and right (R) substructures of S-217622 are contained in sweat, and let e-Nose actually learn them.

Press the “read” button to read the data, then press the “Start” button to start training the self-organizing neural network (SOM). The data except for S-217622 is the data of the HSP of the bad odor as mentioned earlier. When the motion stops, press the Stop button.

What this SOM has learned to do is to map a 4-dimensional vector of HSPs to a similar vector to a similar 2-dimensional position.
Since it uses random numbers, the answer will move with each calculation, but the relative position will not change much.

When such SOM is obtained, it is close to the odor of compounds that are near S-217622 (close in HSP). For example, if S-217622 is hydrolyzed and mixed with sweat, it will grow into an AI that smells like a mixture of pungent bluish odor, pungent thinner, rotten meat, and fermented fruit.

If you put seven odor sensors at appropriate positions in these two dimensions, they will become the California Institute of Technology (CALTECH) e-Nose.
The sweat of a normal person can be smelled by the e-Nose. Since body odor varies from person to person, we can get signaled results from a group of sensors with some range.
The odor of a person infected with corona has a pattern somewhere that is not present in the sweat of a normal person. So, for example, sensors in the S-217622, -L, -C, and -R regions show unique patterns.

I used a narrow HSP with a bad smell this time, so S-217622 came in out of the box. I probably should have used the Amoore list. However, I think it is very interesting to note that both AI and corona detection dogs do the same thing in learning.

If we want to DX a corona detection dog to an AI equipped with e-nose, we need to collect sweat from normal people and sweat from infected people now. Then, at the entrance of the concert hall, we will be able to sort out the entrants. If we make one, we can make as many copies as we want, so we can think of airports, restaurants, dog-shaped, cat-shaped, and so on.

I wonder if anyone has the budget to build something like that.


Leave a Reply

Your email address will not be published. Required fields are marked *