Manjari Narayan | @Neurostats
LightNews LightNews
neurostats.org
Manjari Narayan | @Neurostats
@neurostats.org
7.3K followers 860 following 1.6K posts
AI in Bio & Health & Therapeutic Development

Bio: https://linktr.ee/mnarayan
Substack: https://blog.neurostats.org
Peek into my brain: notes.manjarinarayan.org

Previously @dynotx @StanfordMed PhD@RiceU_ECE | BS@ECEILLINOIS

🧪🧮⚕️🧬🧠🖥🤖📈✍️🩺👩‍📈📉
Posts Replies Media Videos
neurostats.org
5 years later, it is finally on youtube. Thanks @ohbmofficial.bsky.social

youtu.be/X5syQN9ZkQE?...

#neuroskyence #CausalSky
June 17, 2025 at 9:07 PM
neurostats.org
FDR metrics only offer error-control of the false discovery proportion under expectation. It is already a liberal standard of evidence when you realize that. Most don't.

If one wants to defend a paper in nature, you might find AI reasoning agents quite useful to check your work.

#StatSky 🖥️ 🧬 📈💻
June 17, 2025 at 1:08 AM
neurostats.org
Yeah or you take classes from a nice applied math department that teaches you matrix & spectral analysis even if you are in EECS.

Nice reference I didn't know about.

Forever grateful to Mark Embree who is one of the best lecturers I had.
June 15, 2025 at 12:57 PM
neurostats.org
I've had to reference Daniel Carpenter's book so often I just finally got a physical copy.
April 21, 2025 at 1:11 PM
neurostats.org
Just read section 3:
> "It will be critical to continuously rigorously validate
NAMs to build confidence in their reliability."

There has never been a continuous evaluation leaderboard of translational models ever!
April 13, 2025 at 8:18 PM
neurostats.org
Omigod. I knew about this case but didn't know the details till now.
March 29, 2025 at 4:30 PM
neurostats.org
I think science is an entirely different bucket. I mean the really creative stuff, not the workhorse stuff. If we were to put an agent into context circa Einstein as a boy, what would it be capable of?

I think about Polanyi / Strevens a lot these days. It puts my own experiences into perspective
But the historical facts are different. Einstein had speculated already as a schoolboy, at the age of sixteen, on the curious consequences that would occur if an observer pursued and kept pace with a light signal sent out by him. His autobiography reveals that he discovered relativity after ten years’ reflection . . . from a paradox upon which I had already hit at the age of sixteen: If I pursue a beam of light with the velocity c (velocity of light in a vacuum), I should observe such a beam of light as a spatially oscillatory electromagnetic field at rest. However, there seems to be no such thing, whether on the basis of experience or according to Maxwell’s equations. From the very beginning it appeared to me intuitively clear that, judged from the standpoint of such an observer, everything would have to happen according to the same laws as for an observer who, relative to the earth, was at rest.6
March 25, 2025 at 1:01 AM
neurostats.org
Don't really buy into non-standard tests like this. Even researchers haven't fully tested the construct validity of different notions of biological age. The neuroimaging methods folks have dissected problems with brain aging as a construct. But the proteomics crowd hasn't caught on yet, for instance
March 19, 2025 at 1:27 AM
neurostats.org
Apparently my psychometrician friends agree
Nathaniel Haines reposts Eugene Yan commenting on Andrej Karpathy, both worrying about LLM evaluations
March 5, 2025 at 12:56 PM
neurostats.org
How do I use AI reasoning agents to generate richer information?
One formula I use is exploiting Asimov’s model of creativity. There are so many topics in the world where area X has not been explored via lens Y. You can probe different ways of getting AI reasoning agents to explore this. The ticket to consuming richer information is to start off with an expansive primary knowledge-base and get reasoning agents to fill in the gaps for you. You have to bake in a lot of checks along the way though. So long as you know enough about each fields so that can you bake some positive and negative controls into your questioning, you can also call bullshit on obviously wrong stuff. If you have some directional hot takes, it is too easy for the more sophisticated reasoning agents to craft a story to align with that, so you have to get them to elicit adequate criticism. But also, you need to be cautious about when “both-sidism” is factually wrong in a field too.
February 27, 2025 at 3:35 PM
neurostats.org
What do you think of the idea of placebo responders and placebo no responders?

@stephensenn.bsky.social @f2harrell.bsky.social @eturnermd1.bsky.social
February 21, 2025 at 10:15 PM
neurostats.org
A well known CRO makes statements like this.

cc: @eikofried.bsky.social @statsepi.bsky.social
Effect sizes have declined over time. Have you come across better reasoning about it other than blaming placebo arm?
February 14, 2025 at 10:55 PM
neurostats.org
Aside from @dshaywitz.bsky.social has anyone else brought up handling the kinds of labor market failures that James Besson talks about as modern AI systems takeoff?
February 14, 2025 at 2:09 PM
neurostats.org
"Opinion that runs hot-blooded ahead of established fact is the life force of scientific inquiry."
Were Kelvin, Tait, and other advocates of a “youngish earth” being unscientific? Not at all; there were good reasons and great uncertainty on both sides. The physicists saw their assumptions about the geological structure of the earth as natural and reasonable compared with the largely speculative theory of evolution; if something had to go, it was biological guesswork, not careful physical extrapolation. The biologists saw their grand explanation of life’s intricacy, though hardly proven, as a breathtaking achievement based on extensive observation of nature across the globe, a breakthrough that could scarcely be discarded on the grounds of pure conjecture about the unseen goings-on thousands of miles beneath their feet. Such contrary attitudes are business as usual in science, an unexceptional manifestation of the subjectivity that swirls through all scientific reasoning, planning, and debate. Popper might perhaps have counseled both physicists and biologists to keep an open mind, to refrain from taking sides, until more was known and a definitive falsification of either biological theories or physical hypotheses was achieved. Such a prescription is hardly realistic: it is precisely the sort of stricture that scientists, being also humans, will consistently fail to follow. And in any case it is, as we saw in the case of Wegener and continental drift, bad advice. Science is driven onward by arguments between people who have made up their minds and want to convert or at least to confute their rivals. Opinion that runs hot-blooded ahead of established fact is the life force of scientific inquiry.
January 31, 2025 at 2:09 AM
neurostats.org
I didn't know this about Eddington disregarding a conflicting plate in 1919. It isn't mentioned in many Phil. Sci accounts about the growth of knowledge.

From Michael Streven's The Knowledge Machine.
In the methodist’s dream of science, the bodies of data from the three telescopes, the three measurements of gravity’s power to bend light, would be assessed by a procedure that evaluated the evidential weight of each as carefully and as coldly as Eddington had in the first place calculated the numbers. The method would act, in effect, like a high-minded tribunal, objective and authoritative, sorting truth from falsehood without playing favorites or allowing personal or moral or self-aggrandizing considerations to enter into its deliberations. If the Eddington story is any guide, this is pure mythology. There was no tribunal, no method, to sort the good photographic plates from the bad. The matter was settled the old-fashioned way, by a mix of partisan argument, political maneuvers, and propaganda. If the systematicity and objectivity of science can be seen in the painstaking measurement and calculation and the transparent presentation of the astrographic bending number, the subjectivity and unruliness of science can be seen in what happened next: the number, with its Newtonian implications, was brushed off in a few sentences by Eddington, declared unimportant by his allies in the British scientific establishment, and ultimately dropped from the textbooks altogether, leaving the more Einsteinian numbers supplied by the other Brazilian telescope and the Príncipe telescope to decide the issue conclusively against Newton and in favor of Einstein’s theory of relativity.
January 11, 2025 at 12:29 PM
neurostats.org
I liked this paper from Lancet Public Health. It's just one type of research design but it is quite stark in showing that the risk of 17 cancers are on the rise in populations born 1990 relative to 1920.
Birth cohort incidence and mortality rate ratios trends from 1920 to 1990 for nine cancers with reversing trends in incidence compared with older birth cohorts in the USA, 2000–19. The birth cohort with the lowest incidence rate was used as the reference to calculate the relative risk. The panels are sorted by the magnitudes of the IRR for the latest birth cohort. IRR=incidence rate ratio.
January 4, 2025 at 1:44 AM
neurostats.org
True, it was somewhat like an ancient Wikipedia, to which anyone who knew Sanskrit, or who knew someone who knew Sanskrit, could add a bit here, a bit there. But the powerful intertextuality of Hinduism ensured that anyone who added anything to the Mahabharata was well aware of the whole textual tradition behind it and fitted his or her own insight, or story, thoughtfully into the ongoing conversation. However diverse its sources, for several thousand years the tradition has regarded it as a conversation among people who know one another’s views and argue with silent partners. It is a contested text, a brilliantly orchestrated hybrid narrative with no single party line on any subject.
January 3, 2025 at 1:33 PM
neurostats.org
Recently discovered Carole Satyamurti's translation of the Mahābhārata

#BookSky
January 3, 2025 at 1:33 PM
neurostats.org
Here are examples from genomics and the medicine feeds
December 27, 2024 at 5:20 PM
neurostats.org
Michael Strevens on the "scientific method".
science is a machine for motivating disputatious humans to carry out tedious measurements and perform costly and time-consuming experiments that they would otherwise not care to undertake. It is these empirical minutiae, so painful to collect, that single out the truth among the plausible falsehoods. Eventually, enough such evidence accumulates that just about every scientist, whatever their quirks, biases, and prejudices, agrees on its significance: one theory stands well above the rest as the best explainer and predictor of it all.
December 26, 2024 at 11:02 PM
neurostats.org
How Life Works
December 24, 2024 at 4:02 AM
neurostats.org
This reminds me of Peter Huber's criticisms of optimality. All of these guarantees about uncertainty quantification likely hide one or more pitfalls of optimality

#StatSky #MLSky
December 20, 2024 at 10:56 PM
neurostats.org
Well the kinds of "fibers" one gets from diffusion imaging doesn't correspond to any entity that can be analyzed in those terms.

pmc.ncbi.nlm.nih.gov/articles/PMC...
December 17, 2024 at 1:51 AM
neurostats.org
@kevinlala.bsky.social makes the case that language evolved to teach close kin life skills.

Could medical knowledge for reproduction also have been one of those skills?

Inspired to think of this from a review of Cat Bohanen's Eve. thenoosphere.substack.com/p/why-reprod...

🧪👩‍🔬🧬♀️
December 16, 2024 at 4:36 PM
neurostats.org
My god, this is pretty devastating. It was returning incorrect gradients!
I remember all too un-fondly a time in which one of my Julia models was failing to train. I spent multiple months on-and-off trying to get it working, trying every trick I could think of. Eventually – eventually! – I found the error: Julia/Flux/Zygote was returning incorrect gradients. After having spent so much energy wrestling with points 1 and 2 above, this was the point where I simply gave up. Two hours of development work later, I had the model successfully training… in PyTorch.
December 14, 2024 at 7:41 PM