[alextheparrot]

I’m only vaguely familiar with Weinstein’s claims, but you might find the following news article about some research I remembered going to a talk to about 5 years ago [0]. There’s a lot of proposals in the stem cell community like this, though I lack the insight to vet whether that’s happening in practice.

> About 80 percent of experimental drugs fail in human clinical trials because they are unsafe or ineffective, with some 30 percent found to be toxic in people despite promising results in animal studies

Effectively, a huge percent of drugs are failing out in clinical trials — so a few false negatives where the phenotypes aren’t measured correctly or are on an incompatible time scale seems pretty expected.

I haven’t seen anyone showing that “fixing” lab mice would’ve resulted in different animal trial results — is Bret claiming this?

[0] https://madison.com/wsj/news/local/health_med_fit/tissue-chi...

[mlyle]

> I haven’t seen anyone showing that “fixing” lab mice would’ve resulted in different animal trial results — is Bret claiming this?

No one may have shown it, but it seems extremely unlikely to me that we are testing things in animals in the absolute optimum fashion to minimize failure in later human trials. It would be quite extraordinary if we happened to so luckily arrive at the ideal model organisms and procedures.

Presumably all the protocols we could use are not equally good, and good criticism of existing protocols towards improving their utility is important.

[alextheparrot]

This may be a conservative view point, but showing that it would increase the translatability of mice trials to humans seems like a prerequisite for the criticism to change anything, though. No one is saying that our current animal trials are optimal, one of the professors in the news article I linked is quoted as saying the exact opposite "The animal models used to understand the safety and efficacy of drugs are flawed". It just is the case that retraining, to draw analogy to computational models, our model animal is not free because the evaluation of its performance after that change is much more complex and expensive than an ROC curve produced from labeled data.

[mlyle]

Sheesh.

Well, sure. But even beginning to prove anything like this would be an extraordinarily expensive trial itself: what are you going to do, repeat trials for 100 substances in mice to see if you get a better alignment with later human results across potential substances than you do with current practices?

So instead, we talk about it, consider imperfect arguments, and do experiments that are somewhat silly until we start to become convinced that we've likely spotted some of the causal factors. That's what convinces us to eventually do the work that will show it.