henripal

Multimodal LLMs for Drug Discovery @ D.E. Shaw Research. 🇫🇷🇺🇸

Previously led ML@Anagenex, Founder@PoissonAI, PhD@Columbia, Head Trader@SGCIB

[henripal @henripal]

@owl_posting @FoundationOAI congrats!!!

[Patrick Boyle — e/🦀 @p_maverick_b]

https://t.co/jjDkTq50an

[Jake Wintermute 🧬/acc @SynBio1]

I choose to believe this is satire

[henripal @henripal]

@prairial_75 @Briviagra C'est pas ce que je vois statistiquement parmi les centaines d'entretiens et les milliers de CV que je vois. Et le côté "usine a papiers" - oui c'est vrai il y a ça aux US, mais ça crée des profils distincts, avec des étudiants qui sont encouragés à explorer leurs intérêts tot

[henripal @henripal]

@prairial_75 @Briviagra le potentiel est la, et certains l'exploitent. typiquement il y a eu la voie Meta/FAIR qui a ouvert des portes. Mais un X tout frais, sans le moindre papier de recherche a BAC+5 se fait manger tout cru par les MIT qui ont trois papiers first author a NeurIPS au meme age

[henripal @henripal]

@brolysw @Briviagra @FlipFlap2024 non (source je suis un X aux US en IA)

[henripal @henripal]

@finbarrtimbers think this is true in 2 cases
1) science for which timelines match VC timelines (uncommon)
2) cos which can shield research behind a parallel growth narrative

Otherwise "let's not do Y, it would take too long and we have to show something by next board meetng" effect dominates

[owl @owl_posting]

it feels like the the lottery-like nature of drug development success leads some people to believe that it is, in fact, a genuine, bona-fide lottery. all you need is to throw random ideas in and hope for the best and sometimes, against all odds, you win bigly. this is not correct and has been responsible for an astronomical amount of waste in both human lives and capital. have the courage to ask your local drug developers what their theory of change is and have the courage to ponder whether it is dumb

[henripal @henripal]

"What wins from here is physics-informed AI, augmenting and accelerating molecular dynamics, not replacing it"

[Sylvain Gariel @SylvainGariel]

https://t.co/S6xX2C0TuQ

[henripal @henripal]

🤘

[Nathan Benaich @nathanbenaich]

when i see someone write "ai/ml" i think "boomer"

[Jake Wintermute 🧬/acc @SynBio1]

In the coming years we’re going to see a lot of claims about “autonomous labs” accompanied by videos of cool-looking robots.

I’m extremely bullish about this space and I’m rooting for all these companies to succeed. But right now it’s hard to know who is succeeding because the word “autonomous” has no fixed meaning.

It’s lowkenuinely a problem that robots look so cool. It's easy to see a video of robots in motion and be convinced that the future has arrived. If you haven’t worked on an automated lab floor, frontier tech looks about as cool as useless arm-flailing.

So here are some quick heuristics you can use to get a sense for how “autonomous” the lab you see in a video really is.

- Compact form factors. Physical space is at a premium in the lab. Mature automation systems use it efficiently. If a robot has lots of open space around it, that’s a sign that it requires a lot of human support.

- Cold storage. Almost all biological protocols need some kind of refrigeration. Automating sample retrieval from freezers is particularly annoying, because frost interferes with mechanical gripping, barcode reading, etc. If you don’t see a freezer near the robot, it means humans are doing the sample management off-camera.

- Sample transfer. Real lab protocols require the operations of many different devices (PCR machines, incubators, liquid handlers, centrifuges etc) - too many to fit in a single workstation. An autonomous lab needs a way to shuttle samples around. If you don’t see plates moving around the room, humans are doing that.

- Inventory and waste. Biotech eats a lot of reagents and makes a lot of plastic waste. Managing inventory is labor intensive but unglamorous - the last thing most buzzy startups want to care about. A true autonomous lab demo will include robots doing boring and unsexy things.

With these tips, you too can become a cynical jerk who looks at a startup’s tech demo and says “that’s not REALLY an autonomous lab!”

But don’t do that. Instead, root for success and watch as more automation teams check off more items from this list over time. I’m expecting fully autonomous labs, even by my high standards, before 2030.

[Kalyan Palepu @kalyanmpalepu]

Many of the most important drug classes in modern history were nearly abandoned by their financial backers. If we can solve the structural risk-aversion that almost prevented these drugs from getting to patients, then we can dramatically accelerate medical progress.

[Andreas Kirsch 🇺🇦 @BlackHC]

Every single time 😅

I get excited about an LLM paper that shows amazing reasoning performance with super simple methods...

And ofc it's Qwen-based 😭

(This unrelated paper reports detailed results on Llama 3.1 and OLMo3 in the appendix, too, so do check it out! 😊)

[henripal @henripal]

@_Evan_Boyle Session branching!

[henripal @henripal]

@owl_posting The pinnacle of the owl

[henripal @henripal]

Great post! What if these issues weren't only found in protein-ligand binding but also on other ML-DD tasks? 🔎

[owl @owl_posting]

An ML drug discovery startup trying really, really hard to not cheat

https://t.co/SVcgYNeBBn

on the 12-person, Utah-based startup @leashbio, their culture of rigor, and the many ways small molecule models accidentally learn the wrong thing

[henripal @henripal]

🔥

[Alex Zhavoronkov, PhD (aka Aleksandrs Zavoronkovs) @biogerontology]

Hm. Someone said it, not me. AlphaFold is a super valuable research tool but I have not seen a single drug reach developmental candidate stage in 7 years since publication. If you have money for drug discovery, you have $50K for the crystal. Tag me on LN @WilsonLab2 I can repost

[henripal @henripal]

neglected but important development for LLMs in our field - can frontier LLMs understand the basics of molecules?

[Nicholas Runcie @NicholasRuncie]

GPT-5.2 has impressive molecular understanding. ⚗️

SMILES → IUPAC translation:
~0% last year (GPT-4o).
67% today (GPT-5.2 high).

This is a significant improvement!
@sama @gdb

[James Roney @jamesproney]

I'm super excited to announce the first preprint of my PhD, together with Chenxi Ou and
@sokrypton!

ML has revolutionized protein modeling, but key challenges remain. For example, we can't predict complicated protein structures without MSAs, which limits what we can design.

[Jeffrey Low @JeffreyLowMD]

1. AI sped up one part of the stack, but biology pushed back everywhere else.

Some teams nailed the chemistry but picked the wrong targets.

Others nailed the biology but bottlenecked on chemistry.

And AI didn’t help in clinical execution.

3/