co-founder @valence_ai, now powered by @recursionpharma
Joined July 2017
- Tweets428
- Following1.1K
- Followers772
- Likes3.4K
Clinical AI should be right by design, not just get better at being right.
Love the thinking behind what Compiled Health is building. Huge congrats to the team on the launch! 🚀
Should we build clinical AI to be provably right, or to look more plausibly right?
The late Dr. Lawrence Weed (also behind the SOAP note framework) put it perfectly in a medical grand rounds he gave in 1971:
“Better that you not write anything at all than something that’s not absolutely reliable.”
We believe current architectural paradigms have gotten so good that models are being mistaken for clinical practice-ready. Pace will still only look like progress so long as the onus of perpetual oversight (still the primary safeguard) remains with the physician.
At Compiled Health, we are ushering in an era of determinism and verifiability for clinical AI, brought to the world by a dedicated group of clinicians, engineers, researchers, operators, and partners.
We can’t wait to see what this unlocks for healthcare professionals, administrators, and patients worldwide.
Therence retweeted
Last month, we released AquaGen.
A model that can help us simulate the physics of water 100x faster than classical physics simulations of molecular dynamics.
The more efficiently and accurately we can understand the movement of water and organic molecules, the better we can decode our biological processes and, ultimately, identify promising therapeutics.
If you’re interested in learning more, check out our 2 minute video explainer.
Two weeks ago, we released Nesso-1, and the response from the community has been pretty incredible to watch!
Today, Nesso was highlighted during @RecursionPharma earnings call. It was a pretty special moment to see the work led by @NikhilShenoy12 , David Errington, and @Francesco_dgv recognized—and to share that the model is already being used internally to support drug discovery programs.
We often talk about research releases as milestones. This one feels much more like a starting point.
Very proud of this team and very excited to share more of the cool work coming out of @valence_ai soon.
Watch today’s earnings call: youtube.com/watch?v=kJMGIjhR…
On Nesso-1:
Technical report: biorxiv.org/content/10.64898…
GitHub: linkedin.com/feed/update/urn…
– Shot on a Canon AE-1, Francesco presenting Nesso at all-hands a few weeks ago
Therence retweeted
Earlier this month, we released AquaGen - a model that can help us simulate the physics of water 100x faster than classical molecular dynamics simulations.
To learn more about this work, and what it unlocks in terms of biochemical understanding and drug research, check out the presentation by Sanjeev Raja and Emmanuel Bengio.
@folinoid
Links to full video, technical report and blog below.
Therence retweeted
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.
Today, we’re releasing Nesso-1, the fastest open-source co-folding binding-affinity model available.
The idea behind the work is simple: coarse-grained co-folding representations can match the affinity-prediction performance of full-atom models at a fraction of the computational cost. By leveraging @NVIDIAHealth cuEquivariance, we further accelerated both training and inference by an additional 2-3×.
Nesso-1 delivers strong performance across public benchmarks and challenging internal medicinal chemistry datasets while running more than 10× faster than other leading open-source models.
The work is only possible because of the people behind it. @NikhilShenoy12, David Errington, and @Francesco_dgv brought an unusual mix of scientific judgment, technical depth, and persistence to this line of research. It’s been a real privilege to work alongside them.
GitHub: lnkd.in/d6XcZHZT
HuggingFace: lnkd.in/dEjCc8gy
Technical Report: lnkd.in/d5egyCG5
- Shot on Pentax K1000 during one of the final Nesso sprints
@valence_ai @RecursionPharma
A week ago, we released AquaGen: the first all-atom, explicit-solvent, periodic-boundary-condition-aware generative model that produces molecular configurations from the Boltzmann distribution at a fraction of the cost of molecular dynamics.
The science is exciting but the release also reminded me of something I’ve come to believe after almost 8 years in AI for science: frontier research is a craft. The person behind the work matters. A lot.
The taste, intuition, persistence, and occasional stubbornness it takes to push into genuinely new territory aren’t interchangeable.
@folinoid led this work, and getting to work with him over the past few years has been a real privilege. He’s one of those rare people who is both an exceptional researcher and a genuinely great human.
He also wrote a great blog post about AquaGen and the thinking behind it. Well worth a read: valencelabs.substack.com/p/t…
Technical report: valencelabs.com/wp-content/u…
- Shot on Pentax K1000
@valence_ai @RecursionPharma
In Seoul for ICML this week!
I spent the weekend catching up with collaborators and meeting people across the local AI for science community, which was a good reminder of how much amazing work is happening well beyond the usual handful of labs everyone talks about and how fast things are moving here.
@kaistpr in particular continues to impress me year after year, genuine depth of research talent and ambition coming out of here.
If you’re working on AI for science or anything adjacent, I’d love to connect while I’m here!
And I'm not alone! A few of us from @valence_ai/ @RecursionPharma are around this week, reach out to @folinoid, Frederik Wenkel, Alina Selega, @jasonhartford, Rik Knowles, Allison Cohen, @Francesco_dgv, Andrew Pang and David Errington.
- Shot on Minolta X-700
AI doesn't replace physics, it works alongside it.
That's one of the ideas that excites me most about AquaGen!
Rather than treating molecular properties as a black-box prediction problem, AquaGen generates physically meaningful molecular ensembles at 𝗮𝗹𝗹-𝗮𝘁𝗼𝗺 𝗿𝗲𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻, with 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁 𝘄𝗮𝘁𝗲r and 𝗽𝗲𝗿𝗶𝗼𝗱𝗶𝗰 𝗯𝗼𝘂𝗻𝗱𝗮𝗿𝘆 𝗰𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻𝘀, making them directly compatible with the same force fields used in molecular dynamics. The result is a framework that combines the speed of generative AI with the interpretability and rigor of physics-based modeling.
I'm incredibly proud of the @valence_ai for pushing the boundaries of what's possible at the intersection of AI, physics, and drug discovery. A special shout-out to @folinoid and Sanjay Raja for leading this effort, and to everyone who contributed behind the scenes to make this release possible.
Excited to finally share this with the community!
Our bodies’ molecules aren’t static so why do many AI models treat them that way?
Without considering the physical movement of molecules, many AI tools fail to robustly capture what makes drug candidates promising.
AquaGen is a generative AI framework built to close that gap: an all-atom, explicit-solvent model that runs an order of magnitude faster than traditional MD methods.
We’re excited about the implications of this work on early-stage drug discovery.
Technical report and blog below.
Feeling incredibly grateful to be building alongside this team. Excited to finally share some of what we've been working on over the past few months!
- Shot on K1000
@valence_ai @RecursionPharma
Therence retweeted
Kicked off day 1 at @iclr_conf with our MarS-FM poster presentation.
We're introducing a new class of generative models delivering 600x speedup compared to traditional molecular dynamics simulations, without sacrificing structural accuracy.
If you want to connect with the team and others in the AI for drug discovery space, register for our TechBio social: luma.com/9pyf9k90
👉 MarS-FM code: github.com/valence-labs/mars…
👉 MarS-FM paper: arxiv.org/html/2509.24779v3
@KKapusniak1 @CristianGabell1 @mmbronstein @TOSSOUPrudencio @Francesco_dgv
Prediction scales. Explanation doesn't.
New preprint: VCR-Agent turns perturbation data into falsifiable reasoning traces, DAGs where every step can be independently tested, verified by Boltz-2.
18,950 traces on Tahoe-100M. Releasing as VCR-TRACES!
LLMs are strong at reasoning over unstructured text. But in biology, that reasoning is hard to verify.
In our new preprint, we propose a framework for generating structured, falsifiable biological hypotheses from perturbation data.
arxiv.org/abs/2604.11661
@yunhuijang_ @LuZhu66 @FawkesJake @AlisandraDenton @dom_beaini @ENoutahi
Code is out!
Learn more about MarS-FM in the blog.
github.com/valence-labs/mars…
Traditional MD requires long simulations in which interesting events are dominated by irrelevant dynamics.
MarS-FM is our generative model that rethinks protein simulation by learning the transitions between biological states, achieving a 600x speedup.
open.substack.com/pub/valenc…
MarS-FM accepted at ICLR 🚀
Modeling biomolecular state transitions, ~100× faster than MD
Go @valence_ai team!
We're excited to announce that MarS-FM has been accepted at ICLR!
We propose a new class of generative models that learn to sample state transitions of biomolecular systems, reproducing the statistics of Molecular Dynamics (MD) with drastic speedups.
arxiv.org/abs/2509.24779
Therence retweeted
1/ Introducing TxPert: a new model that predicts transcriptional responses across diverse biological contexts
It’s designed to generalize across unseen single-gene perturbations, novel combinations of gene perturbations, and even new cell types 🧵
Therence retweeted
1/ At Valence Labs, @RecursionPharma's AI research engine, we’re focused on advancing drug discovery outcomes through cutting-edge computational methods
Today, we're excited to share our vision for building virtual cells, guided by the predict-explain-discover framework 🧵
Have you been waiting for a chance to show off how well your model performs on a fresh, blind, and consistent test set?
Now’s your moment! 🎉
Super excited to share that we’re launching our very first competition at @Polaris_HQ !
Join us and let’s have some fun! 🚀
🦠 We’re excited to announce our first competition in partnership with @asap_discovery and @openmsf!
Test your skills across three sub-challenges revolving around SARS-CoV-2 and MERS-CoV Mpro:
🧩 Ligand Poses: Given a training set of SARS-CoV-2 Mpro X-ray structures, predict poses of a test set of compounds for SARS-CoV-2 and MERS-CoV Mpro
💥 Potency Prediction: Given a training set of dose-response fluorescence potency data, predict potencies for a blind set of compounds for both targets
🧪 ADMET Prediction: Given a training set of endpoints, predict the same endpoints for a blind set of compounds
The competition will officially start on Jan 13th and end on Mar 10th
Learn more about the competition and access to the sample data: polarishub.io/competitions
Read the blog: polarishub.io/blog/antiviral…
Therence retweeted
News! @modal_labs is supporting @RAAIS fellowships with $10,000 credits on their platform, alongside the $5,000 GPU credits from @LambdaAPI.
Working on AI-first, open source research outside big institutions?
Apply at airstreet(dot)com/fellowships
News! We’re delighted to announce that applications are open for a new batch of @RAAIS fellowships - provided support to AI-first, open source research outside big institutions.
Apply at airstreet(dot)com/fellowships
Many thanks to our partners @LambdaAPI for their support.
Therence retweeted
Designing better cancer therapeutics using AI - Round 2!
🧬 In Round 1 of our competition, 90 designers submitted 700+ protein designs with the goal of developing a new therapeutic by blocking a cancer-associated receptor called EGFR.
🔎 We then validated 200 proteins in our lab, found some promising designs and open-sourced the results.
It’s now time for Round 2 - same goal, more data: Protein designers can use the new dataset to inform their design strategy and create even better proteins!
