@Dorialexander @cortisquared Et ils poussent le curseur très loin. Leurs équipes life science nous ont demandé des idées de bench pertinents pour des tasks ultra-précises en bio, j’étais assez impressionné.
Français
Gustave Ronteix
2.7K posts
Gustave Ronteix
@ronfleix
From the country of turtlenecks | Quant. biology | Biobuilding @orakldotbio | Former @institutpasteur @Cambridge_Eng @Polytechnique | 🇫🇷 🇸🇪 🇪🇺
Paris, Frankreich Katılım Mayıs 2012
690 Takip Edilen353 Takipçiler
@cortisquared Oui. Je connais un peu les équipes derrière : StepFun a entraîné sur un nos sets.
Français
La réalité c’est qu'on est *en ce moment* en train d'assister à une forte commodification des modèles. N'importe quel boite avec de la data et des thunes pour le compute peut sortir un modèle pertinent. Xiaomi vient de sortir un modèle à 1T paramètres apparemment très pertinent
Français
We're bringing 50 brilliant scientists and engineers on a startup tour of UK deep tech startups. Who should we visit?
English
@anshulkundaje But then the question becomes what is the "right" way to vet a virtual cell model?
English
Strongly recommend folks be extremely skeptical of "virtual cell" models until they are independently vetted. This shud be the case for all scientific claims but the virtual cell literature in particular has an unfortunate history of misleading & massive overhyping.
English
Gustave Ronteix retweetledi
Excellent post by Rachel Thomas in the ways AI hype is outrunning reality in life sciences
linkedin.com/feed/update/ur…
English
Gustave Ronteix retweetledi
Early data (small sample size, grain of salt) suggests that AI-designed drugs are beating industry averages at Phase I by a lot. But at Phase II, they fail at the same rate as everyone else.
I don't see why Phase I success rates won't keep improving, these models are genuinely good at designing well-behaved molecules. But Phase II success is all about picking the right targets, which means better understanding of biological mechanisms, and thats a much harder problem. It's also where the real alpha is.
AI can surface candidates from the ether but surfacing a target isn't the same as validating it. To actually get better at predicting which targets will work in humans, we need more human data. To get more human data, we need more trials. Once again, while we're getting great at chemistry, biology humbles us.
Apoorva@apoorvasriniva
revived my substack because what people think AI will do to drug development timelines vs what it actually can do was driving me crazy. new essay on the two kinds of slow, why the timeline has a floor, and where the real value is: apoorvasrinivasan.substack.com/p/how-much-wil…
English
@Dremontjones i interviewed with them forever ago! iirc they were doing 3D cell culture stuff?
knownmed.com/approach
cancer organoids sourced from patients
wonder how well it worked
English
insanely cool ideas for a startup. i dont know anybody involved in these, but it just like such an retrospectively obvious play given the personalized cancer therapy trendlines
kernis.health
specicare.com
English
Gustave Ronteix retweetledi
People are just making these models with no concern for usefulness just because it had become a thing in the AI world to build a model off a large dataset
Biology is plagued with this problem of having lots of data but its not the data we need to use AI to answer the questions we want
So instead of collecting the data that is needed to train useful models in biology people build up shitty models using pre-existing data in hopes it gets people's attention or a paper published or funding. No one will use this just like no one uses the ARC foundational models
People need to stop using the wrong data to build useless models that no one will use
English
@JeffreyLowMD Yes, dual AI x Bio natives exist, but it’s mostly been useful in academia. Bio is not pharma and vice versa.
English
@ronfleix Investors have been talking about "dual natives" in biology and AI for a decade. What we actually need are dual natives in drug discovery and/or development and AI
English
Everyone said AI would eat biology.
A decade later, AI is the one getting indigestion.
Here’s what actually happened - and what the new wave of GenAI biotechs need to learn:
🧵 1/
English
@JeffreyLowMD There is a high likelihood that success will come to drug dev companies with an experienced pharma/bio team that learnt to use AI rather than the opposite.
English
What do you think the first AI-driven outlier success will look like?
7/7
English
Gustave Ronteix retweetledi
Recent advancements in one-shot AI protein / antibody development by Chai, Nabla, AI Proteins, Generate, and a few others are accelerating the *main* theme in biotech:
Value of building the molecule is going down. The value of novel targets, novel translational ideas, AND also the value of clinical execution is going UP
Here's where the value graph is moving towards:
The twin forces of AI and China are quickly driving down price of mlc dev across many modalities:
For AI - mainly Ab right now, emerging for genetic medicines, small mlc, ADCs, cell therapy;
For China - Abs, cell and gene therapy, small mlc, and soon genetic medicines
Having a "best in class" mlc is no longer enough - many tech platforms will soon offer you a mlc priced on metered compute (getting cheaper) and China CROs / biotechs will continue to eat the world with (over)capacity (continued involution).
To make a valuable drug, you must differentiate on either:
a) Novel translational ideas.
Novel targets, novel mechanisms, but not just that - connecting targets with diseases; novel application of certain targets in new disease settings, new intuition on which patient pops have widest therapeutic index for a drug, etc
OR
b) Clinical execution.
Determining the appropriate endpoints in a trial. Recruiting the right patients. Appropriate relationships with the right PIs / clinical sites. Ability to finance registrational studies in US markets ($10s to 100s of Ms)
Either be a translational target discovery engine / tech platform that unlocks new modalities (which unlocks new translational hypotheses) OR get a team of grizzled clin dev / CMO vets and go raise $X00M+ to validate a clinical hypothesis
Living in the middle (ie being "full stack") is dangerous work (at least for a startup)
David Li@davidycli
**A grand unified theory on what will happen in biotech in the next 10-20 years** the two major forces reshaping industrial biotech in the next decade are: 1. China 2. AI - and they're critically linked how? China's low R&D cost basis democratizes execution by providing infrastructure to more drug developers (similar to how AWS helped cloud apps explode in 2010s) AI makes scientific information much more freely available; agents & lab automation increase R&D productivity as well as throughput, further deflating development costs What happens when many more translational ideas can be tried much more cheaply? Value starts accruing in the best ideas to try ie the value shifts earlier in the value chain if the cost of everything from preclinical R&D to clinical trials are dropping significantly due to combo of AI and China, the disparity between clinical stage vs early pipeline assets shrinks dramatically from the current order of magnitude difference The premium on true creativity, novel scientific insight, fundamentally new biology will 100x In a few years the top-of-industry drug hunters / translational biologists will command a hefty premium (maybe not $100m a year like current top AI scientists but ... maybe??) Even more provocatively, foundational models in translational biology that surface / accelerate novel biological hypotheses will suddenly capture outsized value When will a translational foundational model be worth more than a top 10 pharma co? sounds crazy... but like everything else --> slowly, then all at once
English
@RuxandraTeslo And (like often) this is driven by a whole community of researchers. In our case teams at Gustave Roussy who have spent years optimizing the processes. AI helps catalyse all this biological and medical knowledge and it’s super cool to see!
English
@RuxandraTeslo 👋 we’re using AI and patient tumor material to predict patient and clinical trial outcomes in cancer. There’s a publication in prep with more results but these methods are getting mature!
English
Would love to hear more from those using AI with a deliberate focus on in-human predictive validity in mind.
Ron Alfa@Ronalfa
We are literally doing this right now. Testing clinical drugs using our Foundation Models to select the best candidates for clinical optimization to increase probability of success.
English
@owl_posting The culture part is key. I think this is a classic tech-push situation where the AI community is struggling to align their capabilities with the needs of the clinical scientists. But we’re getting there as a community!
English
this is a big reason why i joined the startup that i did
all the virtual cell stuff seems kinda nonsense unless its (at least trying to be) a simulation of human biology. i wrote more about our efforts to do that here: owlposting.com/p/how-do-you-u…
but maybe an addendum to all this is that the culture of the people involved in clinical trial design should be taken into account. i was interested in the role of ai in clinical trial interpretation last year, talked to a lot of folks involved in it, and found that people were pretty negative on the short-term utility there: owlposting.com/p/drugs-curren…
it's a long article, most of it doesnt need to be read, the important bit is in the attached picture. tldr: very few people involved in green-lighting clinical-stage seem to currently believe ai can help make a pivotal clinical decision. they just dont trust it enough. they are happy to rely upon ai at the preclinical stages, but once a chemical has touched a human, most people are too paranoid to let anything else guide their decision-making other than historical precedent
it does feel like this is almost entirely a cultural problem, one that will be fixed with more education + better models. my revealed preference is that i strongly believe biotechs will slowly grow more comfortable with AI aiding clinical-stage decisions, especially as the FDA continues to signal their approval of it (see: FDA approval of the ArteraAI prostate test)
Ruxandra Teslo 🧬@RuxandraTeslo
Would love to hear more from those using AI with a deliberate focus on in-human predictive validity in mind.
English
@simocristea « We do AI » (someone is using a jupyter notebook to run a regression)
English
AIxBio’s best kept secret:
computational biology = virtual cells
Català
Gustave Ronteix retweetledi
Everyone talks about context windows in techbio, but no one talks about the data into the transformer.
@MathSRIsh and I have been cooking for a bit and have done a series of posts presenting the different data generation protools, their cost per datapoint, and why you want them
English
And if you have an increasing stock of drugs on the trial waiting list, how do you prioritize which one you’ll be sending to what sub population? Interesting problems!
English
Indeed! The choke point for drug development is clinical trials. Increasing the flow of drugs upstream of that is great, but tilting the scales requires figuring out how to make drugs work in patients!
English
@SpeakingBee @hsu_steve Since he buys physical theatres in Paris and funds theatre groups as a hobby I guess he could.
English
@hsu_steve When someone tells me markets aren’t efficient, I expect him to invite me on his private jet to explain why over champagne and caviar.
English
Fun interview with Jean-Philippe Bouchaud, who chairs Capital Fund Management (CFM), one of Europe’s large hedge funds, AUM $20 billion. He was formerly a physics professor and still teaches at ENS.
"Bouchaud is both baffled and frustrated by how economics has fallen in thrall to theorists — mathematical, social or political — and become bereft of the rigour of proper sciences, where real-world experiments and experiences actually matter. Nowhere is that clearer than when it comes to the belief that markets are somehow efficient. “It’s all wrong. It’s not weakly wrong — it’s badly wrong,” he argues.
In the interview Bouchaud mentions recent empirical results supporting the INELASTIC MARKETS HYPOTHESIS...
English
English
@ElliotHershberg @arjunrajlab I assume you saw the OG: Can a biologist fix a radio, which this riffs on
cell.com/cancer-cell/pd…
English
Throwback to the iconic "flightosome" diagram by @arjunrajlab
"here's the mol bio mechanistic model of an airplane"
Say what you will about "virtual cells" but I think it's good people are trying to build general and quantitative models of cellular dynamics.
English
Gustave Ronteix retweetledi
Remarkable that US spending on pharmaceuticals is just ~1.6% of GDP. Amazingly good value for money considering the welfare gains. (And especially given the extent to which the US subsidizes the rest of the world.)
English