Decoding biology to radically improve lives. Décrypter la biologie pour radicalement améliorer les vies. Powered by | Propulsé par @recursionpharma.
Montreal
Joined April 2018
- Tweets995
- Following36
- Followers6.4K
- Likes1.2K
Pinned Tweet
To identify the most promising drug candidates, we need to understand not only whether a molecule is likely to bind to a target, but how strongly.
Today, we’re excited to introduce Nesso-1, an open-source, coarse-grained co-folding model that significantly accelerates binding-affinity predictions.
Valence Labs retweeted
Recursion Announces New Clinical Data Across Two Potential First-in-Class Programs, Sanofi Partnership Milestone, and Oversubscribed $200M Offering
Today, we announced important progress across our clinical pipeline, partnerships, and capital position—reflecting growing evidence of what our integrated AI-enabled drug discovery platform can deliver.
🚀 REC-4881 | Deep, durable polyp reductions in FAP
In an expanded adult population (18+), 87% of patients achieved polyp burden reductions after 12 weeks (47% median), with 85% remaining below baseline after 12 weeks off treatment.
In the upper GI—the area of greatest unmet need—92% achieved reductions (55% median), including two patients with near-complete reductions (≥98%). Of those with reductions, 92% remained below baseline after 12 weeks off treatment.
With no approved pharmacologic therapies for FAP, we are targeting registrational trial initiation in 2H 2027, pending FDA discussions. Full data will be presented November 2 at CGA-IGC.
🚀 REC-1245 | Encouraging early activity in heavily pretreated cancer patients
Our potential first-in-class oral RBM39 degrader demonstrated three partial responses (two confirmed in endometrial cancer; one unconfirmed in liver cancer) and 91.7% disease control (11/12) at doses ≥45 mg. No dose-limiting toxicities or treatment-related serious adverse events were observed at the selected 45 mg expansion dose.
🚀 Sanofi | First development candidate optioned
Two challenging targets, one oral molecule. A potential first-in-class candidate for inflammatory disease, enabled by Recursion’s AI platform and small-molecule chemistry capabilities alongside Sanofi’s expertise.
Phase 1 expected in 2027, marking the sixth progress-based milestone across our collaboration.
🚀 $200M Oversubscribed Follow-On Offering | Extending Runway Into 2029
Broad demand from leading healthcare specialists, long-only institutions, and growth investors, including Janus Henderson Investors, Invus, Longaeva Partners, ARK Invest, and Driehaus Capital Management.
The financing supports advancement of our clinical pipeline, continued AI engine innovation, and strategic partnerships.
Multiple independent proof points. Across diseases, programs, and partnerships. Powered by one integrated AI-enabled engine.
Thank you to our employees, collaborators, clinical investigators, patients, and investors who make this progress possible.
Links below 👇
Business Update: ir.recursion.com/news-releas…
Sanofi announcement: ir.recursion.com/news-releas…
Public offering: ir.recursion.com/news-releas…
To predict binding affinity, would you defer to lessons from physics or AI?
No matter which option you choose, you’re unlikely to get where you’re looking to go.
Check out Daniel Ramirez’s op-ed where he offers a third option, one that has the potential of going further than any one field alone.
Link to op-ed: valencelabs.substack.com/p/c…
Just as your two hands are mirror images of one another, so are many of the molecules in the human body.
But mirror images are not identical, and their differences are biologically relevant.
A fact that our models can silently fail to capture, even with chiral tags.
Enter CIP labels.
To learn how chemicals in the body are labeled so that AI tools can tell the difference between mirror images, check out Cristian's video below.
Full Video: youtu.be/0tT-UqhHbpI
A generative model can look like it's working and still fail to output the disease-relevant cell groups or small molecules you care most about.
Fortunately, we have techniques to diagnose and correct these failures.
In our most recent Valence at the Whiteboard, Liam explains how to evaluate generative models using Precision, Recall, Density and Coverage (PRDC) to spot subtle failure modes and know what to do about them.
Full video: youtu.be/teIoP92Aeow
Physics-based simulation tools like DFT and FEP are being repurposed for dataset development.
Once manually run and reviewed by experts, these tools are now producing tens of millions of synthetic data points without close human oversight.
These techniques are not without their issues. Studies have shown that in certain computational datasets as much as 58% of DFT generated data is not physically possible.
In Ihor's op-ed, he describes several cheap and effective methods of avoiding subtle failure modes in physics simulation as a reminder to be careful with these methods at scale.
Op-ed: valencelabs.substack.com/p/s…
@ineporozhnii
To decode biology, we want to understand how interventions (genetic or molecular) alter cellular states.
The challenge is the scale: There are ~10⁶⁰ drug-like molecules and we want to understand their effects across many different cellular contexts.
Even among the most advanced labs, it is intractable to cover such a large search space.
But sufficient data diversity is essential to train perturbation prediction models that can generalize.
So, what do we do?
Rather than learning the underlying biology, models trained on wet lab data can instead learn about the quirks of a research lab.
"Batch effects" is when biological samples say more about who ran the experiment, when, and on which machine than it does about the biology of interest.
Watch Zeinab explain how labs account for batch effects to ensure AI models learn the underlying biology, not the noise.
Valence Labs retweeted
🎉We won!
Proud to announce that Recursion has won a @FiercePharma AI Innovation Award for AI Innovation in Drug Discovery.
💠 The award recognizes the impact of the Recursion OS which connects biology, design, and clinical development into a single AI product engine.
They note that this engine is more than a collection of point solutions. “Instead of stitching together fragmented public datasets, Recursion generates its own proprietary data at industrial scale and feeds it directly back into the system.”
As CEO Najat Khan says: “It enables our scientists to generate and test hypotheses at a scale and resolution that simply wasn’t possible before, uncovering novel biology and designing more precise medicines faster for patients who are waiting. And it has delivered tangible proof points – including multiple drugs advancing in our internal clinical and pre-clinical pipeline, and ongoing milestones for pharma partners.”
Those proof points include REC-4881 for FAP (familial adenomatous polyposis). “REC-4881 marks the first potential treatment for the condition and the first clinical validation of the Recursion OS, backed by encouraging Phase 2 data,” they write. “The same platform also uncovered a novel mechanism for treating FAP – MEK 1/2 inhibition – and used real-world data to broaden patient eligibility for the trial.”
We’ll be presenting additional Phase 2 data from the TUPELO trial at the Collaborative Group of the Americas on Inherited Gastrointestinal Cancer (CGA-IGC) Annual Meeting in November.
🚀 We’re thrilled to have won this recognition from Fierce and look forward to continuing to demonstrate proof points from both partnerships and clinical programs.
👉 Read more: fiercepharma.com/ai-and-mach…
#FierceAIInnovationAwards
There are many ways to measure what happens inside our cells.
On their own, each measurement is incomplete. Taken together, we can start predicting cell response, learning about mechanisms and generalizing those findings across cell types.
Unfortunately, combining data modalities isn't as straightforward as it sounds.
To find out more, check out Ali Denton's Valence at the Whiteboard.
🎥 Full video link below
Valence Labs retweeted
I ran Boltz-2 100 million times 🤯
A cell is shaped by billions of interactions between proteins, RNA, DNA and small molecules. Co-folding can now simulate some of these interactions at scale.
So what can we do with all this data?
Could it open a bottom-up path to Virtual Cells?
Can it help us understand the limitations of co-folding models? And learn when they actually translate to biology?
🧵
A well-known challenge with generative AI models is that they're rarely humble.
These tools don't naturally calibrate their conviction according to the available evidence.
In drug discovery, false confidence can waste critical time and money. Especially when evaluating new drug molecules.
To help scientists know which predictions to act on, and which to send back to the research lab for further data collection, confidence scores are vital.
🎥 Tristan breaks down new work at Valence on confidence prediction. See the full video link below.
Virtual cells promise to revolutionize drug discovery by simulating human biology.
But the technology, while necessary, is not sufficient for a revolution.
The revolution comes from pointing the technology in the right direction.
Find out more in our researcher, Ihab Bendidi's blog post, “Where in the Drug Discovery Pipeline Does Virtual Cell Have the Most to Contribute?”
valencelabs.substack.com/p/w…
Proteins and drug compounds should fit together like a lock and key.
In this context, machine learning models can act as our analyst, assessing whether the two match.
However, since there are billions of purchasable molecules to choose from, it would be impossible to evaluate all of them one by one. That's why Valence Labs built Synflownet.
Synflownet acts as our locksmith. It's a model capable of generating synthesizable compounds that are most likely to act as desired, making the search for active molecules much more efficient.
Learn more with @juleroy13 at our most recent "Valence at the Whiteboard".
Link to the full video below.
