scRNAseq cell type annotation is notoriously messy. Despite so many algorithms, most researchers still rely on manual annotations using marker genes
In a new preprint accepted at ICML GenAI Bio Workshop, we ask if reasoning LLMs (DeepSeek-R1) can help with cell type annotation🧵
Very excited to share the most important discovery of my career yet: your brain switches between a conscious and an unconscious mode ~15 times every second. Consciousness may depend not on where in the brain activity occurs but when.
Find out more below.
https://t.co/23dyMGnaLX
theres many details here being glanced over, for example there are some cancer types which are essentially a death sentence, while others have 90% 10yr survival with minimal treatment, so can’t treat all equal.
point is this has little to do with drug MoA understanding. people understand very little of how a drug works (if at all), and strong explainable MoA is not a proxy for clinical efficacy.
everybody cares for AI to be safe of course.
this example is just inappropriate because it’s a very far off parallel with maths where things are spelled out explicitly. it’s also insensitive.
thats why people are picking on it, not because they think AI in medical applications should not be safe/verified etc.
“passes a stage 3 trial” is what gets every single drug into use.
absolutely nobody fully, mathematically understands the mode of action for any drug.
TT must know that drugs get approved via trials, and not by having expert X map out all the biochemical reactions of the drug.
This is a slide from Terence Tao's 'Math 2.0' public lecture yesterday evening at Caltech. I am highly confident that, if the cure actually worked, almost any Stage 4 cancer patient would tell you that human comprehension of its mechanism is completely irrelevant to them.
@JohnBryBry@proofchecker Lean here is the parallel to positive clinical trial phase3 result; Lean verification shows it’s correct. the OpenAI maths theorems are also correct (his parallel), just not understandable
@Styo28183449@proofchecker of course nobody claims that clinical trials are perfect or that all drugs approved are great. but this is again not TT’s point (he seemed to have changed the slides later), which was simply about lack of drug MoA, despite drug being effective
@LabListon@SimonLermenAI i don’t think anybody would be happy to have people die in an AI-drug trial; anybody at all.
historically, mechanistic understanding behind the drug has been a poor proxy for clinical efficacy and trial success
@JohnBryBry@WAWoloszyn@Silimund trials are not infallible at all, they are truly imperfect instruments showing a potentially tiny average effect.
the point was about the mechanism of action for a drug and whether that would make a person more confident for taking the drug.
@WAWoloszyn@Silimund@JohnBryBry if that happens, we'd probably be killed by AI in more efficient ways than this.
he's not talking AI safety though. rather, this is a response to open maths problems being actually solved but not by mathematicians, and with non-elegant solutions.
@GoodFaithOnly@JohnBryBry yeah exactly, so the analogy (quite weird and totally unfortunate in my view) metaphorically refers to people with no other options (even though not explicitly mentioned). hence the irony. quite insensitive also, come to think about it.
the cancer analogy was chosen especially because there is no cure for cancer, so the thought experiment is metaphorically referring to people who have no other alternative (even though he doesn't mention what type of cancer, alternative therapies, survival etc). it's a weird and unfortunate analogy, quite insensitive as well.
making drugs is hard. cancer drugs is really really hard. could at some point Claude tinker with the excel sheet and put in wrong values of the blood tests such that it falsifies the results of the trial? (even though he does say it is verified in 'Lean'). maybe. but then we'll be surrounded by an AI who wants to kill us, so probably we would have overall bigger problems than AI cancer trials.
(ignoring the offensive language in the message and engaging with the point)
@ADavs79 then what is the point of the analogy? he chose cancer as an analogy especially because no drug actually cures cancer (with very few exceptions).
most drugs do exactly what you say, kill 99% of cancer cells so it's not detectable anymore, yet the cells that remain are deadly
@Silimund@NextTokenParrot right, but maths is as exact and as accurate as it gets. we all agree real world and esp biology is not similar to math proofs at all
@BaumMensch16 I see; in that case, the AI would find multiple ways to kill us anyway, such as reply everybody’s questions in a biased way to convince people to help models takeover, incentivize people to do harm etc. not sure he meant such a far-fetched scenario, rather parallel to maths