Towards reasoning in virtual cells Bridging predict and discover with autonomous mechanistic explanations

Want to learn more?

Blog: valencelabs.substack.com/p/towards-re...
Paper: arxiv.org/abs/2604.11661
Github: github.com/valence-labs...

Applying our VCR-Agent to the Tahoe-100M perturbation-cellular context pairs, we construct a dataset of 18,950 reasoning traces, called VCR-TRACES.

We have publicly released this dataset on Github: github.com/yunhuijang/V...

We propose VCR-Agent: a multi-agent framework that generates structured reasoning traces from external biological sources such as StarkPrimeKG and Harmonizome.

We use verifiers, like Boltz-2, for filtering and quality-control by evaluating the reliability of each step in the structured explanation.

Biological reasoning is often hard to verify since ground truth is scattered, context-dependent, and expensive to annotate.

Inspired by pathway diagrams, we structure reasoning traces as DAGs of biological actions where each step can be independently falsified.

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

We propose VCR-Agent; a multi-agent framework that generates structured reasoning traces from external biological sources such as StarkPrimeKG and Harmonizome.

We use verifiers, like Boltz-2, for filtering and quality-control by evaluating the reliability of each step in the structured explanation.

To realize our vision of better and more effective drug discovery, we must turn predictions into testable hypotheses and experimental designs.

Dive into the Explain pillar of our Virtual Cell framework: why it matters and how we adapt LLM reasoning to biology.

open.substack.com/pub/valencel...

Attending @iclr-conf.bsky.social in Rio? Interested in AI for drug discovery? We’d love to meet you.

Valence Labs x LMRL Workshop are hosting a TechBio social event on April 26th. Rooftop bar, waterfront views, and engaging conversation.

Register here: luma.com/9pyf9k90

GitHub - valence-labs/mars-fm Contribute to valence-labs/mars-fm development by creating an account on GitHub.

Want to learn more?

We've published the code: github.com/valence-labs...

You can read the paper: arxiv.org/abs/2509.24779

Or come talk to us at ICLR: ​​https://iclr.cc/virtual/2026/poster/10007880

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/valencel...

Want to learn more?

Come talk to us at ICLR! We will be presenting MarS-FM on April 24th at Poster Session 2.

iclr.cc/virtual/2026...

Our MSM Emulators are a class of generative models that directly learn to sample the rare but key transitions between states.

We instantiate this class with MarS-FM; using flow-matching, we outperform MIT's MDGen and Microsoft's BioEmu, while being 100x faster than MD.

MD is a powerful computational microscope, but its simulated trajectories contain high-frequency detail that is often irrelevant.

In this work, we instead start from a coarse-grained representation that clusters conformations in distinct states using Markov State Models (MSM).

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

3/ We only have 8 presentation slots and sign-ups close on November 3rd.

Make sure to RSVP below 👇

RSVP here: luma.com/n47bshdx

2/ We want to connect the people reimagining how drug discovery happens, to share ideas, spark collaborations, and explore ways to build together.

If you want to present a “Show & Tell”, we review submissions on a rolling, first-come, first-served basis.

1/ Valence Labs is bringing together next generation of Montréal researchers and innovators in Bio-AI.

We are hosting a Bio-AI Show & Tell — a quick and fun showcase of ideas, demos and projects in ML.

RSVP here: luma.com/n47bshdx

5/ If you’re attending ICML, make sure to connect with @jhartford.bsky.social and Ihab to learn more about the research happening at Valence Labs and @recursionpharma.bsky.social.

4/ “Towards Scientific Discovery with Dictionary Learning: Extracting Biological Concepts from Microscopy Foundation Models”

Link: arxiv.org/abs/2412.16247

3/ “A Cross Modal Knowledge Distillation & Data Augmentation Recipe for Improving Transcriptomics Representations through Morphological Features”

Link: arxiv.org/abs/2505.21317

2/ ”ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy”

Link: openreview.net/forum?id=niy...

1/ Valence Labs, @recursionpharma.bsky.social's AI research engine, is at ICML this week!

Two of our scientists, @jhartford.bsky.social and Ihab Bendidi are presenting posters on work featured in their papers 👇🧵

3/ By combining SynFlowNet with Boltz-2, this new release aims to accelerate the design of high-affinity binders for more efficient early-stage drug discovery.

Check out the GitHub repository here: github.com/recursionpha...

Learn more about Boltz-2 here: www.rxrx.ai/boltz-2

2/ Originally introduced by Valence Labs, Recursion’s AI research engine, SynFlowNet helps ensure generative models propose molecules that are not only novel but also synthetically feasible.

1/ We are excited to open-source the SynFlowNet-Boltz-2 trainer today! 🧵

Pleased to have hosted the 2025 Molecular Machine Learning Conference last week at @mila-quebec.bsky.social

Thank you to everyone who joined us and shared their valuable insights!

We hope to see you all again soon.

Looking forward to seeing everyone at MoML this Wednesday, June 18th, at @mila-quebec.bsky.social in Montréal.

We anticipate a day of engaging discussions, talks and poster presentations, exploring research at the intersection of machine learning and drug discovery.

Agenda Here: portal.ml4dd.com

Check out some of the early media coverage here: www.forbes.com/sites/alexkn...

Proud to share @recursionpharma.bsky.social's exciting announcement on Boltz-2! Valence Labs, Recursion’s AI research engine, contributed to the foundational research supporting the @mit.edu team led by Regina Barzilay.

Learn more about Boltz-2 here: www.rxrx.ai/boltz-2

3/ MoML is an excellent opportunity to engage with leading minds, explore cutting-edge research, and connect with the community in TechBio.

We encourage everyone to secure their spot before it’s too late!

Register now to be part of MoML 2025: portal.ml4dd.com/events/molec...