Emmanuel Noutahi

One step closer to replacing the 'Labs' in our name with 'Simulations'

We’re launching “Inside Valence”—a new series taking you behind-the-scenes of our research, exploring new ways to predict, explain, and ultimately decode biology. In our inaugural post, we discuss why Virtual Cells are making a comeback and why there’s reason to be excited.

Mauricio Alvarez and I have a fully funded PhD position open to study representation learning and abstraction questions. Deadline next Friday (5 December). Please reach out if you're interested. findaphd.com/phds/project/c…

@apsarathchandar Nicely done, congrats to the team! My experience with FCD has been the same (useless metric). FYI, we explored improving SAFE generative models to get the best of both worlds (out of box fragment-constrained design, and SMILES-level efficiency) here arxiv.org/abs/2410.20232

We just made NovoMolGen easy to play with: Transformers-native checkpoints on the Hub and small notebooks that let you load, sample, and fine-tune in minutes. The few lines of code that load the model, plug in a reward, run a short RL finetune, and plot the curve.

@byersblake We’re actually digging into that question as part of our Virtual Cell work at @valence_ai. Happy to chat more about it.

Has anyone shown whether perturbation models trained on cell lines transfer well to predicting primary cells perturbations?

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

Guess what??? Gen AI works for chemistry! 🤯💥 Generate co-folding → Generate ligand ensemble → Predict affinity → Generate synthesizable strong binders

Today, we’re announcing an important step forward in modeling binding affinity alongside @MIT and a research team led by Regina Barzilay. Boltz-2 is the first open-source co-folding model to combine protein structure and binding affinity predictions – measuring how strongly a small molecule drug will bind to a target protein. 🔹 Highlights: ▪️ Boltz-2 goes beyond AlphaFold 3 and Boltz-1 by jointly modeling complex structures and binding affinities ▪️ It’s the first deep learning model to approach the accuracy of physics-based free-energy perturbation (FEP) methods, while running 1000x faster ▪️ It’s particularly useful in hit discovery and hit-to-lead, computationally predicting binding affinities directly from molecular structure and enabling teams to rapidly prioritize which molecules to synthesize and test Boltz-2 code, model weights, and processing pipelines will be released under an open source MIT license, making this advance accessible to every scientist. 👇

RT @RecursionChris: In my JPM presentation in January, I talked about the work happening at @RecursionPharma to help build a virtual cell,…

Excited to unveil Boltz-2, our new model capable not only of predicting structures but also binding affinities! Boltz-2 is the first AI model to approach the performance of FEP simulations while being more than 1000x faster! All open-sourced under MIT license! A thread… 🤗🚀

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 🧵

🤯 We can now directly interact with biology via this new foundational knowledge graph. A must read 🧵

Virtual cells shouldn’t just be predictive models, they should be testable theories of cell biology. They should predict cellular responses, explain them mechanistically, and discover new biology and therapies. Excited to share our perspective paper: arxiv.org/pdf/2505.14613

Can discrete diffusion-based generative models outperform GPT-based next-token predictive models in molecular tasks? Learn about GenMol, NVIDIA’s new #DrugDiscovery Swiss Army Knife! paper: arxiv.org/abs/2501.06158 blog: developer.nvidia.com/blog/evaluatin… demo: build.nvidia.com/nvidia/genmol-…

(3/4) Our second work "Graph Classification Gaussian Processes via Hodgelet Spectral Features" has been accepted as an oral presentation at the NeurIPS 2024 Workshop on Bayesian Decision-Making and Uncertainty. We have derived a graph representation that combines the topology of the graph and its features, permitting graph classification using conventional Gaussian processes. A follow-up is already being written! Link: arxiv.org/abs/2410.10546 @epomqo @Pseudomanifold @ENoutahi @valence_ai

(2/4) Our first work "Gaussian Processes on Cellular Complexes" has been accepted to ICML. We have extended Gaussian processes to a class of topological spaces called cellular complexes. The results of this work will be crucial for some of next year's projects! Link: arxiv.org/abs/2311.01198

(1/4) Hello everyone! I'd like to tell you about the research I've done this year. I'm very happy with the results, but it hasn't been an easy year. I've struggled most of the time with depression caused by my inability to work due to long covid, which started in mid-2023. Although things are improving, it's a slow process and my advice to everyone is to follow your body and go at your own pace. Coming back to Canada at the start of this year to rest was the second best decision I made during my PhD. Nevertheless, we managed to produce three papers and I'm very excited about next year's projects, some of which are already underway!

I am happy to have presented my work as a contributed talk at the NeurIPS workshop on Bayesian Decision-making and Uncertainty. The paper is available here: arxiv.org/abs/2410.10546

🥇 Inside the Best Paper from the @NeurIPSConf Foundation Models for Science Workshop. ▪️At NeurIPS, @phil_fradkin of @valence_ai powered by Recursion presented MolPhenix, a foundation model that can predict the effect of any given molecule & concentration pair on phenotypic cell assays and cell morphology. It was awarded best paper at the NeurIPS FM4Science Workshop. ▪️ As Fradkin explains, the model relies on our massive dataset of high resolution images of cells which are captured under various forms of molecular perturbation at different concentrations. ▪️ “The thing we’re interested in is called PhenoMolecular Retrieval,” he says. “To retrieve the identity of the perturbation and the concentration that was used to perturb the phenomics data” with the goal of “zero-shot generalization of new molecules at new concentrations.” ▪️ MolPhenix is one of the first such models, he says, to integrate concentration & that's critical to AI drug discovery. 👉 Learn more about MolPhenix: recursion.com/news/learning-…

@miniapeur is now giving the first contributed talk for our Bayesian decision making and uncertainty workshop #NeurIPS2024! Join us to hear about how to convert graphs into Hodgelet spectral features and do classification using Gaussian processes!