Alisandra Denton (@[email protected])

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

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

What’s needed for a virtual cell to succeed in drug discovery? A new perspective paper from Recursion and our AI research engine @valence_ai lays out our vision for a virtual cell as a system that can reliably drive the discovery of new drugs via an iterative loop of: predict, explain, discover. We’ve spent more than a decade building many of the foundational elements of a virtual cell at Recursion – namely, massive fit-for-purpose datasets generated from millions of cell experiments each week in our automated lab; machine learning models that have been trained on that data; and industry-leading computing power. Our pipeline of potential medicines serves as the final proving ground for how these models translate to patients in the real world. 🔹 As described in the paper, we are building virtual cells that will: ▪️ Predict: Accurately predicting how cells respond to a broad range of perturbations — from genetic edits to novel chemical compounds – is critical to prioritizing hypotheses and de-risking early-stage drug discovery. ▪️ Explain: By explaining their reasoning, virtual cells will provide causal insights essential for building biological understanding and designing increasingly precise therapies. ▪️ Discover: We envision virtual cells evolving into powerful engines for biological discovery and serving as active guides in the search for new drugs. 👉 Read the paper: arxiv.org/abs/2505.14613

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 🧵

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 🧵

4️⃣ ”Benchmarking Transcriptomics Foundation Models for Perturbation Analysis: one PCA still rules them all” @ENoutahi @AlisandraDenton arxiv.org/abs/2410.13956

Yay! @Polaris_HQ is now hosting the RxRx3-core phenomics dataset and introduced a new standard for benchmarking compound-gene activity!

Missed the event yesterday? Our team is still at NeurIPS today and throughout the workshops. Come talk to us and learn more about open roles. We’re hiring for both full-time and internship positions: valencelabs.com/careers

Valence Labs will be co-hosting a TechBio social with @RecursionPharma and @nvidia at #NeurIPS in Vancouver. Join us on Thurs, Dec 12th. RSVP here: lu.ma/biikt7ox Our team will also be at NeurIPS throughout the week. See below for a summary of our papers👇

“Benchmarking Transcriptomics Foundation Models for Perturbation Analysis” Where: AIDrugX workshop in West Meeting Room 109, 110 When: Dec 15th at 8:15 AM PST arxiv.org/abs/2410.13956 @ENoutahi @AlisandraDenton

Benchmarking Transcriptomics Foundation Models for Perturbation Analysis: one PCA still rules them all 1. This study introduces a new, biologically relevant benchmarking framework for evaluating transcriptomics foundation models’ performance in perturbation analysis, comparing popular deep learning models with simpler techniques like PCA. 2. Despite the rise of transformer-based foundation models (Geneformer, scGPT), simpler models such as PCA and scVI excel in interpreting biological perturbations, particularly in real-world scenarios, showcasing a surprising lead over more complex architectures. 3. Key metrics introduced include “Structural Integrity,” a novel evaluation for gene activity structure preservation, ensuring the model’s latent space aligns with biological phenomena, reinforcing the practical utility of simpler models. 4. scVI and PCA outperform foundation models on tasks beyond batch effect reduction, providing consistent results in linear separability of known perturbations, perturbation consistency, and retrieval of known biological relationships across datasets. 5. Interestingly, foundation models, despite their batch reduction efficiency, lag in capturing complex biological insights essential for understanding perturbations, emphasizing the importance of biologically tailored objectives for these tasks. 6. Results highlight the adaptability of scVI in handling zero-shot scenarios and minimal data regimes, demonstrating robustness when trained on small-scale single-cell data, a critical feature for practical applications in gene knockout studies. 7. The study also reveals that gene distribution assumptions significantly impact model performance, with different distributions (e.g., Poisson vs. Negative Binomial) performing optimally on different datasets, underscoring the need for dataset-specific considerations in model training. 8. Future foundation models in transcriptomics must focus on biologically tailored objectives that capture the nuanced interactions in cellular responses to fully leverage the potential of these techniques. @valence_ai @AlisandraDenton @ENoutahi 💻Code: github.com/valence-labs/T… 📜Paper: arxiv.org/abs/2410.13956… #Transcriptomics #MachineLearning #Biology #PCA #scVI #GeneExpression #PerturbationAnalysis #DeepLearning

Kinda want this to be real. Would go. xkcd.com/3002/

In biology, unpaired data is the norm as experiments are often destructive. This limits our ability to build powerful multimodal models for drug discovery. To fuel a multimodal future in TechBio, we propose a simple algorithm for matching. 🧵 youtu.be/7a29mGz9LXI

@katharina_hoff Woohoo 🎉, excited to read this!!

Today's benchmarking landscape for ML in drug discovery is complicated. We want to help the community develop methods that matter!🌟 Join us for our launch party at ICML on July 25th to learn more: lu.ma/wj1agv8o

We are excited to host @NucleateHQ at our new London office for their sequencing technologies and applications event tomorrow night. There are a few spots left. RSVP to hear from @AlisandraDenton and others in the panel: lu.ma/xfa8h525

🧵1/3 Using LLMs as reasoning engines unlocks an exciting future where language agents can autonomously orchestrate complex drug discovery workflows. Tomorrow, @craigmichaelm will share more details about our vision at #AMLDEPFL2024 in the AI and the Molecular World track.

Not going to lie - I’m pretty excited about this one folks - @recursion is coming to LONDON and we are welcoming @mmbronstein as an advisor! London is 🔥 when it comes to Tech X Bio talent, which is why we are making it our formal footprint in Europe. It also happens to be one of my favorite cities! In fact, I’ll be there tomorrow to host a TechBio mixer with the incredible local community! Interested in joining our mission to decode biology to radically improve lives? We’ve got 20 new roles open at our King's Cross office - check them out here: recursion.com/careers and valencelabs.com/careers

Here’s why I’m in London this week 😀 Looking forward to catching up with people in the causal representation learning / active learning for biology space. If we haven’t already connected, let me know if you’d like to chat while I’m here.

Exciting #RXRX news this morning: 1. We’re opening an office in London! 🇬🇧 2. 20 new London job openings across @RecursionPharma and @valence_ai! 3. The one and only @mmbronstein is now a scientific advisor! recursion.com/careers valencelabs.com/careers