Vijai Joseph

Cancer Geneticist @sloan_kettering, NYC. Fxnl Genomics, Bioinformatics, GWAS, and automation of pathogenicity in cancer genes. Views my own.

[nature @Nature]

One of the world’s top professional societies for scientists and engineers has retracted 1,417 conference papers in one go

https://t.co/GVoa57kyiC

[Eric Topol @EricTopol]

An unexpected and potentially big finding:
A single mitochondrial protein has an outsized role, acting as a central node, coordinating our circadian clock, temperature, dietary cues, sleep/wake cycle to control brown fat metabolism @ScienceMagazine
https://t.co/jBbmaSjkaQ

[Andrej Karpathy @karpathy]

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:

Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:

Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:

Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!

In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

[Vijai Joseph @JVJAI]

@SashaGusevPosts Insightful

[Sasha Gusev @SashaGusevPosts]

A searchable mapping between NIH study sections and program officers that doesn't suck. (Claude one-shot)

https://t.co/TjFZ0mT3S5

[Vijai Joseph @JVJAI]

@EricTopol @TheLancet What about muscle preservation?

[Eric Topol @EricTopol]

A very rich new issue @CellCellPress on AI and biology
https://t.co/ZOSpUbzt9H
aging @biogerontology
world models @marinkazitnik
virtual cell @ericxing
common failures @MarzyehGhassemi
autonomous science @lucieguojam
and much more

[Vincent, MD @vincent_toxins]

We have trained biological language models to learn what sequences look like. What happens if we train them on what molecules actually do?

Most RNA foundation models learn from enormous collections of unlabeled sequences. Hide a nucleotide, predict it. Repeat this billions of times, and the model gradually learns the statistical grammar of RNA.

PARNET, starts from a very different source of supervision: experiments.

RNA does not function alone. RNA-binding proteins continuously interact with it to control splicing, stability, localization and translation. PARNET was trained directly on 223 eCLIP experiments covering 150 RNA-binding proteins, learning at base resolution where those proteins physically interact with RNA.

So instead of learning:

sequence → predict missing sequence

it learns:

sequence → predict molecular interaction

And something interesting emerges. Without task-specific fine-tuning, its frozen RNA representations match or outperform larger self-supervised RNA and genomic language models across tasks ranging from splice-site recognition and intron retention to translation efficiency and non-coding variant effects. It also generalizes to unseen cell types and an independent CLIP technology.

But the part I like most is interpretability.

When PARNET predicts a regulatory consequence, that representation can be traced back to specific RNA-binding proteins and sequence motifs contributing to it. The embedding therefore carries something closer to a mechanistic vocabulary of post-transcriptional regulation, not merely sequence similarity.

This raises a broader question for biological foundation models:

Do we always need more sequence or do we need better biological supervision?

Evolution gives us an enormous record of what sequences survived. Experiments give us a much smaller record of what those sequences actually do.

Perhaps the strongest biological representations will eventually need both.

https://t.co/FrZFINQfKu

[Marios Georgakis @MariosGeorgakis]

Through my posts here and elsewhere, I've tried to make the point that, like any other evidence layer, human genetic data supporting or arguing against a therapeutic hypothesis can be very nuanced.

This is why, for example, I've been critical of popular efforts to binarize genetic evidence for a target as simply "present vs. absent", even if they have served the field of human genetics in terms of visibility.

Following the same logic, the recent phase 3 results for Lp(a)-lowering and IL-6 inhibition have often been lumped into the same bin of "genetic evidence betrayed us". But the reality is that the two examples are very different, both in terms of the human genetic data supporting them, as well as the biological roles that Lp(a) or IL-6 may play in atherosclerosis.

Bringing drugs targeting these mechanisms into trials adds another layer of complexity, as findings need to be interpreted in the clinical context in which the treatment will ultimately be used. Patient selection, stage of disease, competing medications, and duration of treatment can all influence the results, and what a trial ultimately tells us about a therapeutic hypothesis.

Our discussion with @EricTopol on the two recent trials, the genetic evidence supporting them, and atherosclerosis more broadly

[Iannis Aifantis @iannisaifantis1]

Excited to share our brand new work today @CellStemCell ...the first single-cell transcriptomic + epigenomic map of the human spleen in myelofibrosis. Using scRNA/CITE-seq + scATAC-seq from patient splenectomies, we show how the spleen becomes an aberrant niche for malignant extramedullary hematopoiesis. @NYUGSOM_Path @Perlmutter_CC
https://t.co/q1otl6QFpW

[Sarah Urbut @tigerstatdoc]

In summary, we have a model that discovers and predicts: 151 genome-wide significant loci, biological subtypes inside single diagnoses, and 348-disease that outperforms existing clinical tools. Years in the making, SO thankful for an incredible team!!! Paper, code, interactive results: https://t.co/dm9ySZtyOf @pnatarajanmd @SashaGusevPosts @g_parmigiani @MGHHeartHealth @harvardmed @broadinstitute @BWFUND

[Vijai Joseph @JVJAI]

@amarderstein @soumyakundu_ @anshulkundaje Congratulations 🎉

[Vijai Joseph @JVJAI]

@pnatarajanmd @Amgen Congratulations 👏🏽

[Vijai Joseph @JVJAI]

Me deciding on which LLM model to use this week.

[Nature is Amazing ☘️ @AMAZlNGNATURE]

Just a dog having the best day of its life.

[Andrew White 🐦‍⬛ @andrewwhite01]

Read the tech report: https://t.co/3QKoC9E0Fz

Use it: https://t.co/GpbBzTIyfl

See our new website with much more info: https://t.co/Z1CwTn6SfU

[Vijai Joseph @JVJAI]

@skathire @EliLillyandCo Congratulations!

[Vijai Joseph @JVJAI]

@nickmmark @threadreaderapp unroll

[Vijai Joseph @JVJAI]

Creators beware?

[Vijai Joseph @JVJAI]

@KobeissiLetter @threadreaderapp unroll

[Vijai Joseph @JVJAI]

@topazlabs Nice