Hi everyone, I’m back with another announcement of sorts, and a request.
Firstly, thank you to everyone who gave my TMF archive ( Update and Introduction ) the old college try. It’s obvious from the logs that I am not very good at designing agentic experiences, which, you know, my bad. RaplhCramden gets top marks for trying the hardest to make it work. There seems to be something with individual message retrieval, so you can view lists of authors and things, but Maple always reports an error or that she couldn’t find anything when you ask for something specific. I can’t promise I will work on it any time soon, though, because I’ve started this new Project.
So… TL;DR: My wife was Dxed with Stage 4a colon cancer in early 2024. She has responded well to the standard of care (improvements to which have been amazing even just since her diagnosis), but the only thing her doctors can say for sure is that eventually these treatments will stop working. She’s 40, so the timing just does not work for us.
Inspired by David Fajgenbaum’s story (he cured his own “incurable” terminal disease by repurposing an already-approved drug that nobody had ever thought to try on his disease), I decided to try and follow in his footsteps and try to find new treatments that might work for her after we exhaust our current options (or who knows? work better than what we have now?). So I have started to build a system to surface new treatments, tailored to Maureen’s specific genetics, disease state, treatment history, etc. But how?
Fortunately, there is brand-new work in this area! I am starting by reproducing the results of two papers published in March 2025, April 2026 (just a few weeks ago). They both use graph neural networks to rank drug candidates, which is nothing new, but this team uses a variety of ways to personalize the output to a patient (not just as a pre- or post-filter, but part of the training process). Their “target disease” was Alzheimers, for lots of good reasons, but I need to adapt it to Colorectal Cancer (for one reason).
And… So far so good! I posted my first writeup last week, in which I partially reproduced the results of the first paper (PDR1). You can read it here if you’re interested: https://whafa.substack.com/p/phase-0-gp-kg-repro (the whole substack is a dev log for this project, more or less).
I’m ready to start the personalization process, and I have access to so much more granular data than the paper studied. I’m not a data scientist by training or trade, but I have been profitably employed in IT for almost 30 years in e-discovery and healthcare, so I’m not exactly a stranger to the domains. And I wouldn’t say that I’m stuck, either, but I don’t have any other humans to talk things over with, just Claude, which is working great, but obviously fraught. It would be nice to have other wet brains looking at this when I am unsure about the right direction. Or just to be working next to other people who are doing similar things.
Anyway, there’s lots of smart people here, and if you’re an ML person, or even just interested in the idea, I could use new subscribers (and literally anyone posting any kind of comment). Or if you know where ML hobbyists go to talk about stuff with other humans, I’d love to know where that is.
This does have macroeconomic importance. I believe that in 20 years or less, everyone will have a personalized biomedical graph, fine-tuned to their specific genetics, disease state, and treatments to date, which will help to diagnose and treat individuals in new and unique ways. But that’s not a time scale that works for us, so I need to homebrew hers, which is fine. Wish me luck!
