Why can't we mimic a dog's ability to smell COVID?

Question

As far as I can tell, we have invented tools and algorithm to:

• Detect a wider range of colors at a larger range than humans or any other animals on the planet
• Detect sound with wavelengths inaccessible to humans or most animals on the planet

But why is it that dogs can smell COVID or Cancer and we can't produce a similar tool to "smell diseases"? Why can't we mimic the dog's sense of smell: is it a hardware limitation or a software one? Am I mistaken in thinking that this sense is the hardest to mimic?

Answer 1

Addressing the hardware side of your question:

A dog's sense of smell was developed through millions of years of evolution. The dog's nose is powered by hundreds of millions of organic nanomachines (olfactory receptors) working in concert to detect the faintest traces of odors, in the form of individual molecules floating among an endless sea of nitrogen, oxygen, and molecules from other nearby sources producing orders. I don't think we have hard numbers, but some estimates say that a dog's nose can distinguish molecules in the parts-per-billion or even parts-per-trillion (or higher) range.

When these millions of finely crafted organic nanomachines detect molecules of something besides oxygen and nitrogen floating in the air, the signal is sent to the brain, which then cross-references this data against an exhaustive library of known molecules (some instinctive and some learned), and interprets it as a "smell". Different concentrations of different molecules will be interpreted as different smells, and we know from numerous practical use cases that dogs can be trained to seek out specific smells.

It takes extremely sophisticated sensors to even begin to approach a dog's ability to detect and classify those stray molecules floating through the air that make up an odor.

In contrast, measuring sound is child's play (since it's just vibrations through air or another medium) and even imaging is comparatively simple (measure the wavelength and intensity of light striking the image sensor).

An astronomically greater amount of R&D has gone into light and sound because these have the greatest number of commercial applications. We can use light sensors to record and share photos and videos. We can use audio sensors to record and share music, speech, and more. We can combine these two technologies to produce movies, television, and more.

On the other hand, the best you could do with an odor sensor is produce a chart or graph of relative molecular concentrations. We don't have any kind of consumer-level technology that can reproduce arbitrary odors from digital recordings for other audiences to smell. Hollywood isn't likely to be investing millions of dollars per year into odor sensors. There aren't enough practical applications.

When we do have a practical need for odor detection outside of scientific research (e.g. for security or police purposes), do we typically issue some multi-million dollar precision gadget that can detect all of the relevant molecules in the air? No, we do what humans have probably done for thousands of years before the first electronics: use a trained dog.

Answer 2

We can actually detect some diseases via smell, and the term to search for is olfaction. The general problem is known as breath analysis.

However, the research into olfaction and machine learning is rather new (perhaps even surprisingly new). As Lötsch et al. point out, little research (prior to the very recent research) on olfaction and machine learning has been performed, with a few exceptions:

1. Quantifying olfactory perception: mapping olfactory perception space by using multidimensional scaling and self-organizing maps, Mamlouk et al., Neurocomputing, 2003.
2. Relationships between molecular structure and perceived odor quality of ligands for a human olfactory receptor, Sanz et al., Chem Senses, 2008.
3. Diagnosis and Classification of 17 Diseases from 1404 Subjects via Pattern Analysis of Exhaled Molecules, Nakhleh et al., ACS Nano, 2017.
4. And the one mentioned above, Machine Learning in Human Olfactory Research, Lötsch et al, Chemical senses, 2019.

I don't know whether the problem in general is harder, but as you are touching on in your question, the problem is much harder from a hardware perspective. Where imaging only needs a simple camera, and hearing only need a simple microphone, to detect smell you need a so-called as chromatography–mass spectrometry instrument. As the Wikipedia article mentions:

Breath gas analysis consists of the analysis of volatile organic compounds, for example in blood alcohol testing, and various analytical methods can be applied.

Here are some pointers from popular science that should assist you in getting into the literature:

• Scientists Invent An AI That Can Smell 17 Diseases From Your Breath, Including Cancers

• Innovative AI Breath Analyzer Diagnoses Diseases by “Smell”

• AI is acquiring a sense of smell that can detect illnesses in human breath

Answer 3

One of the major players in this question is adaptation. The brain is a marvel of adaptation which software developers are still desperately trying to unlock. So from that perspective, it's a software issue. However, unlike many engineering efforts, it isn't "hardware leads to software." In the brain, hardware and software intermingle to the point where it is sometimes hard to distinguish them.

In the case of smell, the brain adapts to the particular sensors it has. This permits a different kind of sensor which is less a-priori specialized and more "we'll work with what we get." This can be a difficult kind of sensor to work with in engineering, requiring complex calibration and training. At some point, it simply becomes more cost and time effective to train a dog, leveraging a few million years of evolution.

From an engineering perspective, it's not worth spending \$50,000 to build a smelling machine and half a man year of training and calibration when you can put a female dog and a male dog together, wait for a free smelling machine to be produced, spend a few weeks training them (the smell training phase for drug sniffing dogs is around 3-4 weeks), and be ready to roll. Glendale Police claim the whole price of a K-9 unit dog is \$22,500. The typical engineering process is to turn up the precision in the process, so that you can roll out units with very little calibration/training. And while calibration/training is a software thing, it also has a major effect on the choices that the hardware team makes!

We simply haven't found a cheap, high precision, reproducible way of handling smells the way dogs do.

---

You mention the amazing things we have done with detecting light, spanning many wavelengths. But evolution is amazing too. Consider this Single Photon Detector.^† It can detect single photons in a range of wavelengths just a little larger than "visible wavelengths." It costs $4,600 at the time I wrote this answer.

Our Mark-I eyeballs have detectors too. In particular the rods are very effective at sensing light. If given 30-45 minutes of pure darkness, the rods will adapt to the low light, becoming more and more sensitive. Once fully adapted, a rod is capable of detecting single photons striking it. This is done through a remarkable active feedback system involving several inhibition loops.

We have over 100 million rods in the human eye. That many Single Photon Counter Modules would cost over $4 billion, and take up roughly the volume of an Olympic swimming pool! Of course the data shows the human eye can detect a single photon roughly 1.6% of the time while the spec for that detector is closer to 20% on average and 35% at its peak, but since 90% of the light that reaches the eye never actually hits a photodetector, the photodetectors themselves have to be at least 16% effective at detecting this single peaks. Not bad for a detector no larger than the diameter of a human hair!

^†Full disclosure: I have no relationship with Thor Labs nor this product. It was the first product recommendation for a google search for "Single Photon Detector."

Answer 4

Well, given this is a computing forum, I would say that all computing requires input. You have to have a way to "input" the smell. So far, as one of the answers described, we have very few ways of translating a smell which is a physical world thing, into data, and even so far, those methods mentioned have arguable limits on how they translate to a "smell".

Once that happens in a cost effective and available manner, everything else will just translate over, just like it did with imaging. As the old saying goes, Garbage in, Garbage Out. I'll add, Nothing in, Nothing out.