@ClausWilke @SashaGusevPosts @eperlste Yes, and having a pool of PhD-level numerate people to choose from is great when you end up needing them, regardless of specific expertise. (Just like teaching languages).

@vitaliikl As for (2), I don't think anybody forgets that one can have 80 years of non-enriching experience, because they're experiencing it. We've gotten to slightly longer lifespans but no closer to "best experience possible" so 🤷‍♀️the latter seems more impactful

@vitaliikl Get you where

@johnsontoddr4 @mbeisen Yes, and the exciting mysterious third thing (when authors say "Figure X shows Y" when no it doesn't)

RT @BrandesNadav: New preprint claims that most existing DNA language models perform just as well with random weights, suggesting that pret…

@H_Fotowaikha @rafalab A third thing altogether, with distinct consequences, from "includes superfluous DR not used for analysis" and "uses DR as a substrate for analysis without showing it is either necessary or sufficient" 🤷‍♀️ There are plenty.

@H_Fotowaikha @rafalab Then what is the point? What is the purpose? And this dismissal is the exact opposite of the nod to Monocle: there are entire subfields that explicitly rely on DR, and the implicit reliance is everywhere.

@rafalab @H_Fotowaikha It's also not great for the framing of DR as EDA, because EDA should probably not have very strong expectations of the outcome, or undergo severe and opinionated preprocessing, or be directly used as a substrate for downstream analyses

@vitaliikl @rafalab Problem is, just about any arbitrary partition is going to have sets of strong marker genes. And maybe they're even meaningful in that they represent distinct gene modules. But then it seems like a much better use of time to do module identification instead of viz

@H_Fotowaikha @rafalab big risk of turning expensive data into an exercise in confirmation bias!

@simocristea Absolutely. And when a method does, the answer will be more trustworthy if the field is first able to cogently treat questions like "how should I treat uncertainty," "are two expressions different," and "how do we compare results from experiments measuring different things."

@vitaliikl of course, someone would first need to develop those data-specific models, or at least theory to ask when a model might be informative (vs., say, the typical case of combination of unique experiment/no raw data/unusual analysis/errors in metadata). doesn't seem like that's coming

RT @vitaliikl: Most of data in biology is like that - what is measured, how and why - impacts the data. Rather than learning how cells work…

Every method is data-agnostic if you wrangle the data enough. And it is possible to invent and beat benchmarks based on ground truth (whatever got past reviewers in older papers). Perhaps there are cheaper and more interpretable methods that would also address known problems? 🤷‍♀️

@_avichawla K-means inference can't be any slower than knm (but could probably be far faster)

RT @tomgauld: My latest cartoon for @newscientist

RT @NASA: Keep your eye on the donut and not on the hole. In memory of David Lynch, we will continue to explore the otherworldly and the u…