Recursion

“This is a once-in-a-lifetime opportunity.” In a Fireside Chat with Recursion co-founder and CEO Chris Gibson at Download Day, NVIDIA founder and CEO Jensen Huang shared his vision for the future of healthcare, and how the companies are working together on a data-driven approach to drug discovery that has the potential to radically transform medicine.  ◾ Why he’s excited about this collaboration: “If you look at the foundation of Recursion, three things are in play – the invention of a new algorithm, a new family of algorithms that we call deep learning; supercomputing capabilities that you’ve used here and that we’ve partnered with you to create; and the know-how of processing and extracting the meaning of biology that’s embedded within life, within that data.” ◾ With accelerated computing and generative AI, we are changing “the type of problems we can solve,” he added. ◾ NVIDIA has been an incredible partner as we push to develop the next generation of medicines for patients in need -- including their support in building BioHive-2, the largest supercomputer in the pharma industry, which allows us to run more experiments, build more data, and develop even more powerful foundation models. Watch the full Fireside Chat here: https://lnkd.in/evqAm3QB The full day of presentations and slides from Download Day, featuring experts from across the Recursion ecosystem sharing the latest developments on our platform, pipeline and partnerships, is available here: https://lnkd.in/e9bRKvqf  #techbio #future #ai #health #medicine #pharma #tech

How we use phenomics and computational modeling for chemical design. At the upcoming Gordon Research Conference (GRC) on Computational Chemistry (July 21-26 in Portland, Maine), a premier, international scientific conference focused on advancing the frontiers of science through the presentation of cutting-edge and unpublished research, Senior Scientist Ivan Franzoni will share insights into how we use our operating system to find new therapeutic targets and quickly produce and test new drug candidates. ◾ It starts with building a purpose-built dataset using high-content microscopy, arrayed CRISPR genome editing methods, and machine learning – amassing phenotypes from millions of perturbations in multiple cell types. ◾ This leads to promising new biological insights and novel therapeutic targets with unique mechanisms of action. ◾ We then introduce a broad but targeted method for expanding the chemical space of interest, identifying compounds related to the desired phenotypic profile. ◾ Using computational techniques, we then search chemical libraries and use compound scoring to deliver crucial insights and employ rapid GPU-enabled shape screening in our internal chemical library to discover novel chemical structures. Learn more about the GRC on Computational Chemistry here: https://lnkd.in/e2FxZEBZ #ai #ml #techbio #platform #chemistry #biology #drugdiscovery #crispr #grc #grc2024 Gordon Research Conferences

From data to drugs. Great insights from our investor Kinnevik on the "seismic shift" happening in drug discovery thanks to 3 key elements: ▪ The industrialisation of data collection and curation ▪ An explosion of compute power ▪ Advancements in AI algorithms This shift involves the true convergence of technology and biology and includes tech powerhouses like NVIDIA partnering with biotech companies. Ala Alenazi writes: "Recursion has become synonymous with TechBio. Through its industrialisation of research, Recursion’s robotised labs can carry out 2.2 million experiments each week for up to 50 weeks per year. Today, they boast 50 petabytes of high-dimensional data. In a strong sign of faith, NVIDIA has both invested in Recursion and partnered with them to build BioHive-2, the largest system in the industry and the 35th fastest globally according to the TOP500 list. BioHive-2 will allow Recursion to double the insights generated from wet-lab work, with that ratio improving over time according to CTO Ben Mabey." Read more: https://lnkd.in/eWn4vNyX #ai #partner #techbio #biotech #drugdiscovery #tech #venture #investment

Congrats to Berton Earnshaw, AI Founding Fellow and Director of Valence Labs at Recursion who has been named College of Science Senior Fellow at the University of Utah College of Science. “I first met Berton in the math department during his PhD studies,” said Dean Peter Trapa. “It’s great to see him come full circle with the U as a Senior Fellow in the College of Science. Currently, he’s at the top of his game in machine learning as it relates to drug development and will add appreciably as an executive advisor to the College and its research priorities.” #ai #ml #research #drugdevelopment

Are we entering a new era of phenotypic discovery? At Download Day, CMO David Mauro spoke to cancer genomics expert Deepak Nijhawan, Associate Professor in the Departments of Internal Medicine and Biochemistry at UT Southwestern Medical Center, about discovering the druggability of RBM39 and why he predicts a new era of phenotypic discovery. ◼ Could you talk about the history of this target in your lab? What we've done over the years is to build genetic tools to give us a single mutation that takes us right to the key binding site of a molecule. With this program, we built our first-generation forward genetic system. And we started to put a lot of orphan compounds through. One of those compounds, we found mutations in RBM39 leading to resistance. This compound turned out to be a molecular glue. We did a kind of deductive reductionist science to get to the mechanism for how that chemical target is being turned on. RBM39 would be a conventionally undruggable protein. Unless you’re approaching it phenotypically, no team of medicinal chemists or structural biologists could imagine how you would get a small molecule to target RBM39. That's one of the real benefits of phenotypic discovery with small molecules — you end up finding things that people couldn't predict. ◼ Can you comment on Recursion’s approach and the data that you've seen? I think that's a novel insight that your group has come up with. I just sent you this manuscript where people had done phenotypic screens of our RBM39 degraders across multiple PDX models and came to the same conclusion that your algorithms predicted. That is the key element – once you have a small molecule targeting this now undruggable protein, is there a clinical problem that you can solve with that, and who are those people? ◼ Why has your lab decided to use a phenotypic approach? Without knowing the target, it was really difficult to advance drugs that were discovered phenotypically. As a result, it became expensive and uncertain. With the target-based approach, you have a precise protein you want to target, but you have an entirely new challenge: How do you get a small molecule to target that protein? As a consequence, you have an intense amount of competition on a small number of targets, and there's this vast biological space that's unexplored. I think phenotypic screening can now come back because things can be done in parallel. So you can look for compounds and targets and patients all at the same time. With the tools that you guys have, now you can find the molecules, the targets, and potentially the patients much quicker and start to explore this space. So we're not targeting the same proteins over and over again. I think we're leaving a lot of potential clinical opportunities on the table by focusing on just a small number of targets. Watch the full conversation: https://lnkd.in/e39aYmSz #ai #techbio #rbm39 #pharma #cancer #phenomics #drugdiscovery #downloadday #science

Last year at the International Conference on Machine Learning (ICML), we launched Valence Labs – a machine learning research lab in Montreal focused on developing the next generation of cutting-edge methods and models for drug discovery, with a focus on open science and academic partnerships. This year, we're showcasing discoveries that have emerged from these collaborations and helping to launch Polaris - Benchmarks for methods that matter – a new platform for ML in drug discovery that aims to provide standardized, domain-appropriate datasets, guidelines, and tools for method evaluation and comparison. 🎉 Polaris Launch Party is 7/25, register here: https://lu.ma/wj1agv8o Discoveries include: ◼ Learning to Scale Logits for Temperature-Conditional GFlowNets: In this paper, researchers including Yoshua Bengio demonstrate Logit-scaling GFlowNets (Logit-GFN) – a new generic architecture design of temperature-controlled Generative Flow Networks (GFlowNets) which offer a training framework for learning generative policies that sequentially construct compositional objects to be sampled according to a reward function and also adjust based on temperature, a limitation of current models. GFlowNets can uncover a multitude of highly related samples from a diverse set of modes, making them ideal for drug discovery. They also introduced an online discovery algorithm that uses Logit-GFN alongside a temperature control policy, inducing superior exploration-exploitation trade-offs. In experiments, they found that the Logit-GFN architecture significantly enhances training stability, characterized by a smooth and rapid loss convergence. 📎 https://lnkd.in/ekDVhV47 ◼ Graph Positional and Structural Encoder: In this paper, researchers present the Graph Positional and Structural Encoder (GPSE), the first-ever graph encoder designed to replace hand-crafted positional and structural encodings (PSE) by learned PSE in any graph neural network (GNN). Hand-crafting PSEs that work optimally for all graph prediction tasks has been challenging and prone to overfitting. By learning an efficient common latent representation from multiple PSEs, the proposed GPSE learns an efficient common latent that is highly transferable: The encoder trained on a particular graph dataset can be used on datasets drawn from markedly different distributions and modalities. Across a wide range of benchmarks, GPSE-enhanced models significantly outperformed those that employ hand-crafted PSEs, paving the way for the development of foundational pre-trained graph encoders for extracting positional and structural information, and highlighting their potential as a powerful, efficient alternative to explicitly computed PSEs. 📎 https://lnkd.in/efhBCiXH #icml #openscience #academic #ai #ml #drugdiscovery #techbio #gnn Berton Earnshaw Joohwan Ko Emmanuel Bengio Woo Chang Kim Semih Cantürk Renming Liu Vincent Létourneau Guy Wolf Dominique Beaini Ladislav Rampasek

Using AI to find & correct a CRISPR bias. Recursion’s operating system picks up on patterns that human scientists can’t see – modeling changes that are happening both in and around a cell, telling us how biological events are connected. In a recent paper in Nature Genetics, researchers at Recursion revealed a critical finding discovered via our AI-driven operating system. CRISPR editing of genes – “cutting” a DNA sequence to make precise gene knockouts that mimic disease – was leading to a proximity bias in other parts of the genome. In other words, knockouts showed unexpected similarities to unrelated genes on the same chromosome. The finding has significant implications for drug development and AI modeling. Imran Haque, SVP of AI and Digital Sciences at Recursion, explains.    ◼ Talk about the proximity bias finding.  Knocking genes out using CRISPR makes cells behave differently under a microscope, and cataloging those differences helps us build our maps of biology. But we can’t interpret the changes in those images by eye. Not only do we take too many (over 2 million experiments per week in our labs), most of these pictures would look the same to a person. But we can pick out consistent meaningful changes using the deep learning methods we've developed. Across the genome, we found not only that images were highly similar for knockouts of functionally-similar pairs of genes but that images of knockouts that were near each other on the genome looked similar – what we call proximity bias. In particular, we saw a striking block similarity pattern on chromosome arms. Most of the genes on each chromosome arm looked similar to each other, and less so to genes on other arms. This told us that something important was going on that is spatial, not functional. ◼ Why is this bias problematic?  Bias like this might lead you to attribute effects to genes that are not actual effects. As described in the paper, on the Cancer Dependency Map dataset we got false signals for genes that seemed to be key to cancer but were actually the bleeding through of a proximal gene that was the true driver. It was a leakage from a strong nearby signal. Discovering this effect was transformative for how we train and develop our machine learning models - optimizing only those models that are picking up on real signals. ◼ What was the response from the scientific community? There has been a lot of positive feedback, particularly around the scale at which we did this. We replicated the findings both in our data and externally. The DepMap folks have done lots of work to correct for various biases. When we put out the preprint they later announced a new data release that implemented exactly the correction we described. Full paper: https://lnkd.in/gCrSvQSQ #ai #ml #data #drugdiscovery #dna #gene #cancer Nathan Lazar, PhD, Safiye C., Lu Chen, Marta Fay, Jonathan Irish, James Jensen, Conor Tillinghast, John Urbanik, William Bone, Chris Gibson

Unlocking a lifetime of genetic insights. A story by Brian K. Buntz in Drug Discovery & Development highlights our partnership with Helix in which we are combining Helix’s growing clinico-genomic dataset with more than 25 petabytes of our proprietary biological and chemical data. ◾ This genomic information about patients, he writes, provides "insights for years to come" and "can help unearth the hidden connections between genes, diseases, and potential treatments." ◾ The story notes that the partnership will utilize a "secret weapon" -- the BioHive-2 supercomputer -- "the fastest, wholly-owned supercomputer in the pharmaceutical industry globally." ◾ He writes: "Over the past decade, Recursion has also amassed one of the world’s largest collections of biological and chemical data. Now that its supercomputer, BioHive-2, is operational, it plans to use the horsepower to create more advanced AI models that can help streamline the drug discovery process." Read more: https://lnkd.in/eQtBYYfj #ai #drugdiscovery #data #techbio #supercomputer #biotech #pharma #patients

🎤 Speaker Announcement: Estefanía Barreto-Ojeda, PhD, ML Engineer at Recursion Join Estefania on her presentation on "Industrializing ML Workflows in Drug Discovery" where she'll share insights into Recursion's approach to predicting compound properties crucial for drug discovery using ML. 🌟 Session Highlights: - Overview of ML workflows for predicting compound properties like ADME and toxicity. - Strategies for managing large and diverse chemical assays datasets. - Deployment techniques using cloud and supercomputing resources. - Automation solutions to accelerate drug discovery processes. 👩💼 Estefania is an ML Engineer at Recursion Pharmaceuticals with a PhD in Biophysical Chemistry from the University of Calgary. She specializes in automating ML pipelines for drug discovery and has presented at major data conferences including PyData. 📅 Register Now: https://lnkd.in/g6EEsEc

The AI transformation of drug discovery is underway. A new story from Rachel Brazil in The Pharmaceutical Journal looks at how we are leveraging AI capabilities, computing power, and purpose-built data to identify novel biomarkers and targets, and accelerate the drug discovery pipeline end to end. ◾ She writes: "The pharmaceutical industry has been successfully using computers and mathematical models to identify and design new drugs for several decades, but the development of generative AI — based on deep neural networks and large language models capable of understanding and generating text — has brought big changes." ◾ In the article, Imran Haque describes how Recursion's approach involves collecting microscopy images from human cell assays, taking pictures to compare disease and healthy states, and feeding those images into trained AI models that identify 1,000 different features in each image. “This turns out to be an incredibly powerful capability,” he says. Read the full article: https://lnkd.in/gBSNG_at #ai #techbio #pharma #drugdiscovery #biotech #biopharma Pharmaceutical Journal Publications