An Arb Research project

You can't buy melatonin in the UK.^[1] In the US, it's hard to avoid: it's in gas station soft drinks. And: you can legally get bupropion online in the UK, while in the US it's prescription only.

How general is this kind of thing?

Our new beige site looks at every case where medical regulators disagree about whether to approve a medicine. We found 101 instances just between the US and UK.

Background

A while back I noticed a bunch of "discordances" (medicines banned or unapproved in one country, but approved, even OTC, in another)

Despite being broadly sceptical of the health impacts of microplastic, I found myself admiring PlasticList and noticing that one could easily do this for all medicines.

Moreover, various countries have "regulatory recognition agreements", where they can reuse each other's vetting work and so get things approved quickly if it's already approved in a sensible jurisdiction. So THEN (one thought) one could use these agreements to get hundreds of good meds approved elsewhere.

Surely (one thought), AI is good enough now to make this a weekend job. Unfortunately not. It's nearly unbelievable how messy and ambiguous pharma data is. e.g. One pharma company, Novartis (75000 employees), needs around 3000 analysts, statisticians, data guys to function.

Still, over a full damn year we had a go with huge amounts of manual work on top of Claude's virtual work. We only got around to doing US, UK, and EU. Caveat aegrotus.

What's the point?:

• I was looking to get ambroxol and melatonin approved in the UK, and dreamt that there would be many others. It's on their desk and we'll see.
• Use cross-country data to draw attention to the arbitrariness of regulatory decisions
• Help people look where to get meds they can't get at home
• If you've just moved countries, you can also look up the equivalent of your favourite meds

Limitations

• We manually inspected a big sample of the outputs of the pipeline, but there are 200,000 rows here and we're not pharmacists.
• The 2026 models would do a lot better but we're out of patience.
• It's not obvious what fraction of discordances *should* resolve to universal approval
• None of the databases list things by the active ingredient alone, which is what outsiders usually mean by "drug". The LLM's job is to try and remove the salt names, the compounding identifiers... But absurdities remain in our final dataset, like unmarked child formulations.
  

Links

• Repo
• Notebook (with various manual steps mocked out)

Acknowledgments

Thanks for huge amounts of tedious work by Rian O'Mahoney, Paul Crowe, Phil Harrison, and Jake Slosser. Huge thanks to Emergent Ventures for funding this work. Grudging thanks to Ben Southwood for bullying me into doing something concrete to improve the world for once

We did all the data work before Claude Code existed; I hope to trick one of you into doing better than us. Should take a weekend.

1. ^{^}

   20 people on Twitter kindly informed me that you totally can buy melatonin, you just need to break the law or get a doctor to allow you to buy it or import it from a country where it can be sold, demonstrating the reading skill for which that beloved site is renowned.