OWKIN

Why is progress against cancer so uneven? Because to make big gains we first had to understand the underlying biology💡 Identifying HER2 led to trastuzumab. Finding PD-1 led to pembrolizumab. And discovering a druggable pocket in a mutated form of KRAS, long considered “undruggable”, led to sotorasib. In each case it took decades from target discovery to drug on the market - most of that time went into understanding the disease. 🧠 At Owkin, we’re using AI to shorten that path, so the next breakthrough comes years sooner. Read more from the source: AACR Cancer Progress Report 2026, link in comments 👇🏼

Ce week-end en France, plus de 1 500 commerces ont été pillés ou vandalisés en marge des manifestations de samedi. Autour de nos bureaux à Paris, cela nous a touchés de près : des vitrines brisées, jusqu’au kiosque à journaux de notre rue, saccagé avec une réelle violence. C’est peu de chose, mais cela compte pour nous. Si vous connaissez des commerçants ou toute personne près de nos bureaux, au 14/16 boulevard Poissonnière, 75009 Paris, qui traversent cette épreuve en ce moment, dites-leur que notre porte leur est ouverte. Un café, un endroit chaleureux où s’asseoir un instant, quelqu’un à qui parler. Ni plus, ni moins, juste un peu de réconfort entre voisins. Take care --- This weekend in France, more than 1,500 shops were looted or vandalized on the sidelines of Saturday's demonstrations. Around our office in Paris, it hit close to home: shopfronts smashed, right down to the news stand on our street, damaged with real violence. It is a small thing, but it matters to us. If you know shopkeepers or anyone near our office, 14/16 Bd Poissonnière, 75009 Paris, who are dealing with this right now, please tell them our door is open. A coffee, a warm place to sit for a moment, someone to talk to. Nothing more, nothing less, just a bit of comfort from neighbours. Take care of each other.

“A French unicorn lands in Europe’s most digital country.” Thanks Marion Simon-Rainaud for this story. Read the full article in Les Echos (link in comment) 👇

Can we get AI to understand biology entirely by increasing models' general intelligence? That's what Rich Sutton's bitter lesson would argue: general methods that scale with computation beat methods built on human knowledge. Chess, Go, speech recognition and computer vision all went that way. But all those examples have one thing in common - they had training data that support accurate prediction of outcomes. Biology only follows the bitter lesson if the training data can support accurate prediction. But that might not be the case. Most large biological datasets were collected for reasons other than AI model training. They reflect care, administrative use, population research, hypothesis driven research. There's huge investment going into filling those gaps now. But although our instinct is that the denser the biological assays used, the better the AI predictions, that’s not always true. There’s so much individual variation, so much biological complexity, and so many batch effects going on that it’s easy to generate patterns, but hard to generate meaningful ones. The answer is in clinical cohorts that support biological predictions - where the patient’s individual treatment path itself denoises the biology (and vice versa!). That's what we do at Owkin: collect patient data from the largest network of partner institutions in the world (840+ hospitals from 100+ partner institutions). Our data is multimodal, follows the treatment path from baseline to resistance and remission, and provides the causal signals frontier AI models need to learn from. If the data is right, compute can do the rest. That's the bitter lesson.

Yesterday our Paris office handed its desks over to five teams from TOM Tikkun Olam Makers France for a one-day hackathon. Each team worked on a daily problem faced by someone living with a disability - with that person and their family in the room. One team built a learning app for a five-year-old. Another worked on extending what a wheelchair can do. 13 Owkin colleagues joined makers from the wider tech community. One team spent the morning in a hardware shop in the 11th hunting for the right wrench, with the parts kindly donated. At the end of the day, each team handed in something to help change someone’s life. Thank you to TOM Tikkun Olam Makers France and to the families who trusted us with their challenges. ❤️

🥰🥰🥰🥰🥰

How do you find out which AI system is better at addressing your use-case? You use a benchmark. But using a benchmark isn’t always simple. A benchmark is a defined standard for comparing a model against other models, earlier versions, or human performance. For a task with a known answer, like classifying a scan, one score is often enough. For an AI scientist, which chooses its own analyses and interprets noisy results, benchmarking can require numerous metrics. For example, accuracy on a single run doesn’t necessarily tell you whether the system is reliable. So what you choose to measure matters as much as the score. Read our 5-minute guide to benchmarking in healthcare AI 👇 Benchmarking in Healthcare AI→ https://lnkd.in/eziWHgQ6 Read our blog on K Pro vs Claude → https://lnkd.in/erY8ukvv

In a recent interview on the 632nm Podcast, Nobel Prize-winning scientist Eric Betzig makes a point that applies well beyond microscopy: you can’t model what you haven’t observed. Proteins we thought formed stable complexes turned out to touch DNA for seconds. The same holds for AI in medicine. Models built on incomplete data give confident answers to the wrong questions. That’s why at §Owkin we start from real patients (real tissue, real outcomes, real validation) before we ask AI to discover anything. Worth the listen. https://lnkd.in/gNTKj36t

That’s a wrap on BioTechX! In an ocean of blue & dark booths, suits and ties, we were hard to miss with our flashy white tee shirts and our lovely pink booth 🦩 Great conversations, inspiring keynotes (with maybe a few too many words per slide 😉), and lots of new connections. Thanks to everyone who stopped by, and huge kudos to our amazing crew: all smiles for hours, at the entrance by 8am handing out chocolates, and still there after closing pouring bubbles 🥂 Drug development doesn’t need more of the same. That’s why we’re building a state-of-the-art AI scientist, a true copilot for pharma and biotech teams, built on the world’s largest AI-ready real-world patient dataset, to accelerate drug development. Missed us at BioTechX? Don’t miss out twice. Book a demo now :) https://lnkd.in/enMB9mRv

📄 New publication in Genome Medicine: MOSAIC, the Multi-Omics Spatial Atlas In Cancer. Response and resistance in cancer depend on how tumour cells and their microenvironment are organized in tissue. Studying that at scale has been held back by a lack of large, consistent datasets. 🧬 2,700+ cancer samples profiled with six matched modalities: spatial transcriptomics, single-nuclei RNA-seq, bulk RNA-seq, whole exome sequencing, H&E and curated clinical data, processed through standardized pipelines that make the dataset AI-ready. 🔬 Four case studies show how integrating modalities reveals drug target insights that no single data type detects on its own. 📈 Most tumours contain several distinct malignant cell populations, each with its own signalling activity and surrounding microenvironment. 💊 Why this matters: this heterogeneity is where targets and patient populations are defined. We now use MOSAIC for ADC discovery and positioning in cancer. 🤖 The MOSAIC dataset is accessible as a data add-on in K Pro, and the MOSAIC Window sample (60 patients across 5 indications) is available on EGA (link in comments 👇🏼) Read the paper: https://lnkd.in/eN8pZtBX Thanks to the MOSAIC partners: CHUV | Lausanne university hospital, Charité - Universitätsmedizin Berlin, Universitätsklinikum Erlangen, Gustave Roussy and University of Pittsburgh. Thank you to the authors: Alex Cornish, Quentin Bayard, Andy Karabajakian, Elo Madissoon, Ginevra Ferrarini, Almoatazbellah Youssef, Cecile Badoual, Laurence de Leval, Devin Dressman, Eric Durand, Prof. Dr. med. Dr. rer. biol. hum. Ramona Erber., Stefan Florian, Ingrid Garberis, Carla Haigneré, Krisztian Homicsko, Ulrich Keilholz, Adrian Lee, Joseph Lehár, Vassili Soumelis, Markus Eckstein, MD PhD, Raphael Gottardo, Markus Morkel, Caroline Hoffmann