Surrogate Science , Auditing AIxBio, AI epistemics. Prev @Dyno_Tx | @StanfordMed | PhD@RiceU_ECE | BS@ECEILLINOIS Feedback: https://xitter.cf/t.co/wR1ooEvxzP
Pasteur's Quadrant
Joined November 2014
- Tweets10.2K
- Following5.4K
- Followers5.4K
- Likes20.6K
Manjari Narayan retweeted
Predicting the answer to interventional "what if?" questions — the outcome of an action you never took — need a *mechanistic* model, not a curve fit. And you can only learn one by *experimenting*. Experiments are costly, so the real game is **data efficiency**.
Meet the Model Discovery Agent (MDA). 🧵
Manjari Narayan retweeted
🔥 Exactly! Genetic variation is comparatively clean to interpret, but dynamic biological measurements (transcriptomic, epigenetic, imaging, physiological) are deeply entangled with life-course factors: environment, development, health status, social, & reverse causation...
1/3
Manjari Narayan retweeted
The recent very welcome @Arcinstute challenge made this painfully clear: defining evaluation metrics is hard. In some cases, trivial data transformations—and even random data—can score astonishingly high.
Great AI performance ≠ biological meaning. 4/6
Manjari Narayan retweeted
You also don’t need to be a gold medalist of the IMO to be a great mathematician. He isn’t drawing a contrast between smart and stupid. It’s a contrast between shallow and deep thinking.
Manjari Narayan retweeted
I did a postdoc at CMU. I was grateful that @ericxing took a chance on me, and was supportive of the work I wanted to do. I was given space and trust, which made all the difference.
Manjari Narayan retweeted
But why is this relationship so clean? A and G basically define K-FAC, which is an approximation the loss curvature w.r.t. weights, which prior work connects to memorization! i.e., K-FAC->curvature->memorization
Manjari Narayan retweeted
I would say to any PhD student not to take the time for granted. It is really special. Also don't get influenced by herd mentality, fomo, or thinking on short time-scales. Pursue fundamental questions that people aren't asking, but should be asking. Then go all in.
Manjari Narayan retweeted
Nobody has seriously solved the enhancer & promoter design problem except for trivial cases like discriminating glaringly different cell lines and cell types.
Manjari Narayan retweeted
Empirically false. The largest study on natural history of BPD is called CLPS which shows that *most* (>50%) improve over the course of 10 years. pubmed.ncbi.nlm.nih.gov/1627…
Manjari Narayan retweeted
PTSD is a cognitive disorder, not a disorder of facial morphology or expression. Facial morphology and expression will never provide a reliable inferential basis for a brain condition.
Manjari Narayan retweeted
The benchmarking is also thorough using difficult and strong benchmarks introduced by several papers. Also they do extensive ablations to help understand what modeling and data inclusion choices are most helpful for model performance & discovery. More on this later. 4/
Manjari Narayan retweeted
Pfizer (the authors here) is the end user, they shouldn't be the ones having to call out papers from @NatureBiotech
What is the point of academic research if the hurdle of "trusting it enough to try it" gets so big no one uses academic methods at all?
Manjari Narayan retweeted
An unexpected consequence of mass llm use: fake citations of me show up in Google scholar that seem hallucinated from a prompt like "write me a related work". Showing up means the SAME hallucination was published twice! I find this hilarious but this seems a bad sign for academia
Manjari Narayan retweeted
I was reading "Breaking Through" by Katilin Kariko, and the book ends with this note relevant in this context--"I believe my work made a difference. But for this to have been the case, so many things had to come together. There was so much luck involved. "
Manjari Narayan retweeted
Innovation is not linear — by @jasoncrawford
The more I read science history, the more I think this piece was brilliant.
worksinprogress.co/issue/inn…
Manjari Narayan retweeted
This is a must-read for everyone involved in metabolomics.
Personally I love this article because I no longer have to write a long paragraph when I say 'please don't use pathway analysis' during peer review 🤣
nature.com/articles/s42255-0…
Manjari Narayan retweeted
Look, I work with the things, and they are awesome. There’s a number of publications on their potential applications in the clinic.
But we’re not there yet: they tend to be too variable and don’t allow for measurements of off-target effects and lack the systems biology.
Manjari Narayan retweeted
In honor of international women's day, let's take a moment to remember the most famous statistician in history.
You've definitely heard of her, but you probably have no idea she was a statistician.
It's Florence Nightingale.
Manjari Narayan retweeted
Nobody would dream of writing an article like this if the exact same model was produced by @MistralAI @allen_ai or @huggingface. Instead, the lead would be about how cool it was that a lower resource company innovated training methods that enabled them to use bad GPUs.
Manjari Narayan retweeted
A single injection of a drug, called lenacapavir, protects 96-100 percent of recipients from HIV for up to six months.
It's the best HIV preventive ever made. To close out Issue 05, @KLNahas explains how Gilead made it & why getting it to those who need it will not be simple.🔻
