Sabitlenmiş Tweet
The thing I feel most about coding agents is tremendous gratitude. Every half-finished project that has haunted me for a decade gets to actually come to life
English
alex rubinsteyn
22.2K posts
alex rubinsteyn
@iskander
Genomics + immunology + ML = personalized cancer immunotherapy | https://t.co/nReVwtVHPq | https://t.co/8DWibdfDWa
Durham, NC Katılım Haziran 2007
4.8K Takip Edilen6.9K Takipçiler
@_Dave__White_ @andrewbadr Like we’re not yet at the point of “Claude, here’s my cancer’s data, make a therapy to kill it” but personalized cancer vaccines point in a plausible direction where that might be real sooner than later
English
@_Dave__White_ @andrewbadr Patrick is kinda right, there’s still some technical parts in need of optimization for this stuff to work in general BUT I think he’s undercounting how qualitatively different this kind of therapy is compared with one-off mouse cures, it really is a big deal that it’s algorithmic
English
am I correct in understanding that we now have technology to make personalized mRNA cancer vaccines relatively easily?
if so… are we starting a company to do that? I understand the major barriers are regulatory but surely there’s a jurisdiction where it makes sense?
English
@_Dave__White_ @andrewbadr There's been occasional coverage of the Moderna & BioNTech trials (eg nytimes.com/2023/05/10/hea…) but I think people zone out "promising cancer therapy" news articles bc they rarely pan out, not broadly appreciated how different from other medicines personalized cancer vaccines are
English
@_Dave__White_ @andrewbadr "does it work?" is the big question and answer used to be "not really" and now is "for a minority of patients if used early enough" but there are clear directions for improvement to get it to "hell yes!"
English
Reminder: the dog had aPD1 + cKITi, the “Covid vaccines help aPD1” study is hopelessly confounded by the timeline of PD1 approvals for lung cancer, and the PDAC mRNA vaccine is just showing stronger Tc response in the half of patients w/ favorable baseline characteristics
English
@agingroy An easy analysis that this paper didn’t try: subset down to just patients who (1) were vaccinated either within 100 days of aPD1 or some time before AND (2) got aPD1 as first line therapy (3) grouped by stage
The stage II/III adjuvant/neoadjuvant approvals came in 2020s btw
English
@agingroy Even just the move from eg third line for a patient in 2016 to first line in 2023 can explain most if not all of the effect, then there’s all the other confounders!
English
The unvaccinated group are mostly pre-Covid, when lung cancer patients getting ICI had later stage disease (ICI was new) and AE management was less established. Subset of them are during Covid who refused vaccine (pulling in a million other correlated confounders)
Avi Roy@agingroy
Your COVID mRNA vaccine may have accidentally helped fight cancer. MD Anderson studied 884 lung cancer patients on immunotherapy. 180 of them got a COVID mRNA vaccine within 100 days of starting treatment. Results (AACR 2026): - Vaccinated group: 37.3 months median survival - Unvaccinated group: 20.6 months - Nearly doubled. The mechanism: mRNA acts like a siren for your immune system. It floods the body with type 1 interferon and upregulates PD-L1, the exact protein that checkpoint drugs target. The vaccine didn't fight COVID here. It supercharged the cancer treatment. This wasn't planned. It was discovered by accident in retrospective data. But 884 patients is not a small number.
English
@andrewbadr @_Dave__White_ That precedent is pretty useful beyond just cancer vaccines: it’s the first instance of FDA reviewing medicine-generating computational processes instead of specific molecules. I think this is the wedge through which other generative / AI designed medicine can get into the clinic
English
@andrewbadr @_Dave__White_ The FDA has been surprisingly friendly to personalized cancer vaccines. When we started some trials in mid 2010s I thought they’d be a bottleneck but FDA just asked about basic physical manufacturing integrity then let us YOLO output of Python scripts into patients.
English
@iskander Basically, exactly what you described! Biggest bottleneck is raising funds for experimental validation.
English
Missing ingredient for fully generative AI designed medicine is a single platform that can be safely targeted to different tissues with de novo payloads.
Make that and a lot of wishful noise about AI broadly curing diseases starts to seem less fanciful.
English
@MathSRIsh Nah you need to assemble it end to end — then build up safety priors as you broaden
English
@iskander If you mean what's missing is data to train the AI then that's very solvable.
Any cell type in any organ except the brain and the germ cells can be solved in mice and NHP for $20m each.
Maybe $100m for the B and T cells because those are a pain to get anything into.
English
@MathSRIsh AI can’t really design them in a vacuum though, at least to the point where any one will feel confident using the platform. this is the bespoke human effort / animal model limited step that then unlocks a lot of generality
English
@iskander It's part of AI's job to design those targeted delivery systems.
And between LNP, peptides, Ab, aptamers, and branches glucides there's a bunch of approaches to make progress there.
Aerska's approach with hijacking metal recognition is also interesting.
English
I want a full family tree of Israelite religions.
Karaites, Samaritans, whatever was going on at Elephantine &c
Who’s got the full set with when they branched off?
English