Arthur B.
Arthur B. Arthur Breitman. Machine learning, functional programming, applied cryptography, and these days mostly #tezos. Husband of @breitwoman, oligocoiner.

Crowdsourcing supplements stacks, part 1

Part 1: LLM ideation

nasal spray

This is the first part of a three-part post exploring ideas around quantified self, citizen science, black-box optimization, and graph algorithms for Sybil resistance. If you’re here for Tezos content, this isn’t about that; I recommend you check @tezos for daily news.

No one welcomes the first symptoms of a cold. Still, lately, as I began taking singing lessons, the dread of the upcoming two weeks of sniffling misery has become even more salient to me. And that is assuming it’s not something worse like a bad covid infection!

My teacher suggested I try Vick’s First Defence. It’s a nasal inhaler meant to be taken at the first symptoms of the common cold or after potential exposure. I was (and remain) skeptical of how effective this could be. Surely, if this worked, the product would be far more popular, wouldn’t it? But I did try it when my wife returned from a trip with a cold. Surprisingly, I did not develop symptoms, even though I typically would in such circumstances.

I set out to understand how it worked. And, for fun, I decided to do so using GPT-4. Yes, large language models are prone to confabulation, and I know it would be naive to ask random questions and blindly trust the output. Yet, rejecting the tool for that reason misses the forest for the tree:

  1. The type of questions I’m asking are very likely to be answered correctly because they are specialized and technical, yet not obscure.
  2. Having to double-check the answer does not mean the answer is useless or not a time saver.

Thus I copied the ingredient list and asked GPT-4 what they might all do.

Aqua, Hydroxypropyl Methylcellulose (1%), Succinic Acid (1%), Disodium Succinate (0.44%), PCA (0.35%), Phenethyl Alcohol, Zinc EDTA, Zinc Acetate, Polysorbate 80, Menthol, Camphor, Sodium Saccharin, Eucalyptol.

I learned that some ingredients might provide a physical barrier, some lower the pH, creating a less favorable environment for the virus to replicate, and some have antiviral properties (Zinc, menthol, eucalyptol).

Let me stop at this point to stress that I have no medical or pharmaceutical training whatsoever. Worse, I am using GPT-4 for research. You are bonkers if you take medical advice from this blog post. What is the point, then? Well, people who actually know their stuff might weigh in and say: “yeah, that makes sense,” in which case, this won’t be for naught. Or they will have a good laugh and explain why it doesn’t make sense. In all cases, the rest of us will have learned something. Ready?

Ok, back to Vicks’ First Defence. “Can we do better?” — I wondered. I asked what stronger antivirals could be used. GPT-4 suggested interferon. No surprises here; intranasal interferon is indeed a treatment for viral infections. The top search result for “intranasal interferon” is, in fact: “Intranasal interferon-alpha 2 for prevention of natural rhinovirus colds.”

Nice! But Interferon has many side effects; it’s a serious medication and a prescription drug. What if I could make a DIY defense spray with more commonly available substances with antiviral activity? I asked GPT-4 to come up with a list of candidate compounds. Did a bit of research on them, and they all seemed interesting.

One, in particular, caught my attention, Iota-carrageenan. A Google search for intranasal iota-carrageenan brings up this paper: “Efficacy of a Nasal Spray Containing Iota-Carrageenan in the Postexposure Prophylaxis of COVID-19 in Hospital Personnel Dedicated to Patients Care with COVID-19 Disease”. In this study, medical personnel was given prophylactic intranasal iota-carrageenan and a relative risk reduction of 80%. It’s not a huge sample of people, but… well done, GPT-4.

After talking things through a bit with GPT-4, we devised the following DIY recipe for a prophylactic intranasal spray made of readily available ingredients. While I suspect GPT-4 is likely to choose sensible compounds — which all seemed backed up by at least tentative studies — I have little faith in its ability to pick good concentrations, so please: do not make this as is. If you’re excited about it (as I am!), please wait until people with actual pharmaceutical knowledge take a look.

  1. Start with 100 ml of distilled water or saline solution (0.9% sodium chloride). If using distilled water, add 0.9 grams of non-iodized salt to create a saline solution. Stir until the salt dissolves completely.

  2. Add 1% to 2% of an emulsifier like polysorbate 20 or polysorbate 80 to the saline solution.

  3. In a separate container, mix the following ingredients:

    - Tea tree oil: 0.1 ml (0.1% concentration)

    - Menthol: 0.1 gram (0.1% concentration)

    - Camphor: 0.1 gram (0.1% concentration)

    - Eucalyptol: 0.1 gram (0.1% concentration)

  4. Slowly add the oil mixture to the saline solution while stirring to ensure an even distribution of the oils in the solution.

  5. Add the following ingredients to the mixture:

    - Iota-carrageenan: 1 gram (1% concentration)

    - Elderberry extract: 1 ml (1% concentration)

    - Echinacea extract: 1 ml (1% concentration)

    - N-acetylcysteine (NAC): 1 gram (1% concentration)

    - Zinc acetate: 0.5 gram (0.5% concentration) or another zinc compound

    - Succinic acid: 1 gram (1% concentration) or another suitable acidifier

    - Hydroxypropyl methylcellulose: 1 gram (1% concentration) as a thickening agent

  6. Add a preservative like grapefruit seed extract or phenoxyethanol at a concentration of 0.5% to 1% or as the manufacturer recommends.

  7. Mix the solution thoroughly to ensure all ingredients are evenly dispersed. You may need to shake or stir the solution occasionally to prevent the ingredients from settling.

  8. Transfer the solution to a clean, sterile nasal spray bottle. You can purchase empty nasal spray bottles online or at a pharmacy.

  9. Before using the nasal spray, shake the bottle to ensure the ingredients are well-mixed.

I am not a trained medical professional (I did mention that, didn’t I?), but I do know that spraying things inside your nose can be dangerous. People have contracted fatal brain amoeba infections using neti pots with unsterilized tap water. Very, very few people, but it still gives you a healthy pause. I do not think doing DIY sprays safely is beyond the reach of laypeople — pharmacies wouldn’t recklessly sell empty nasal spray bottles if that were the case — but please do not blindly trust the above.

Can we do even better? This little experiment with GPT-4 showed that many readily available compounds with antiviral activities could be tried. But can we find the best formula? And can we ensure it actually works? In the next part, we’ll explore large-scale experimentation with Gaussian processes.

comments powered by Disqus