We spoke with Yann Flereau, Co-founder & CEO of Blossom Lifesciences, about what turns hypotheses into biological truth. The hardest part isn’t generating ideas — it’s finding the right signal and having the data to validate it. That’s where rigorous validation matters most.

Over the past few weeks we shared notes on building target ID into K Pro and testing it on our bladder cancer program: the honest metric (PR-AUC), the bias trap (ERBB2 for everything), the attrition funnel (314 → 1). Here's the full blog tying it together: www.owkin.com/blogs-case-s...

🔬 New preprint: PRDX5, an atypical peroxiredoxin, is upregulated in epithelioid malignant pleural mesothelioma. Using ML to identify cryptic allosteric pockets + a 3.6M-compound virtual screen: 36 non-covalent inhibitors, 8% hit rate, sub-µM potency. www.biorxiv.org/content/10.6...

📢 Announcing our new deal to build AI agents as part of a multi-year K Pro license agreement with @astra-zeneca.bsky.social Read the full press release here - www.owkin.com/newsfeed/owk...

Peer-reviewed in BMC Med Research Methodology: our benchmark of 9 subgroup analysis methods for time-to-event RCTs finds interaction tests best for detecting heterogeneity, ML-based CATE estimators best for identifying good responders. link.springer.com/article/10.1...

AACR 2026 confirmed AI's central role in cancer research. From foundation models like Apollo to agentic systems, the field is moving from passive tools to active collaborators. Our recap on the three waves of AI in oncology — and where the real bottleneck now lies: www.owkin.com/blogs-case-s...

🎤 We're at AACR 2026 — stop by booth 4616 to meet the team and try K Pro demos. On Monday April 20, join Dr. Caroline Hoffmann and Jonas Béal for our spotlight: "AI Agent for Pharma R&D: Powering nextgen drug discovery & development". See you there!

90% of ADC trials fail because we've been asking the wrong question. Not "is the target there?" but "WHERE is it, WHAT is it interacting with, and HOW will the TME respond?" Owkin's spatial biology-first approach uses MOSAIC data + K Pro to design next gen ADCs www.owkin.com/blogs-case-s...

New preprint: AI agents that learn from experimental feedback for iterative perturbation discovery. Agents receiving results after each round found 166% more gene-phenotype hits than random (29.3 vs 11, p=0.003). 📄 arxiv.org/abs/2603.26177 📄 www.owkin.com/blogs-case-s...

Preprint: CytoSyn, a foundation diffusion model for histopathology. Generates diverse and highly realistic synthetic tissue images using H0-mini’s embeddings as conditioning. Benchmark shows our models reach state-of-the-art FID. Paper: arxiv.org/abs/2603.18089 Model: huggingface.co/Owkin-Biopti...

Preprint on bioRxiv: Multi-modal benchmarking of Ultima UG100 vs Illumina across snRNA-seq, WTS, WES, WGS on FFPE samples. UG100 shows high concordance with Illumina for oncogenic/immune signals, with characteristic indel profile but enhanced precision Preprint: www.biorxiv.org/content/10.6...

Published in Nat Commun: MISO uses local attention transformers to predict spatial gene expression from H&E at near single-cell res. Trained on MOSAIC samples. Outperforms existing methods. In collaboration with FFCD. Paper: www.nature.com/articles/s41... Code: github.com/owkin/miso_c...

We're donating PyDESeq2 to @scverse.bsky.social for long-term maintenance & development. Python reimplementation of DESeq2, ecosystem member since 2023, now in the hands of one of the best open source communities in omics. Good science requires good open source. github.com/scverse/PyDE...

Our new blog explores OwkinZero, a language model fine-tuned on 300,000+ biological Q&A pairs using RLVR. Compact models that outperform larger commercial models on biological inference. Teaching AI to understand mechanisms: www.linkedin.com/posts/owkin_...

Spatial omics generates rich multimodal data, but extracting biological insights requires infrastructure at scale. Our new blog covers the 2025 data engine. All building toward BASI. Read more: www.linkedin.com/pulse/buildi...

We introduce IMILIA: interpretable Multiple Instance Learning for IBD inflammation prediction from H&E WSIs. Combines H0-mini, HistoPLUS, and EpiSeg to automatically derive quantitative inflammatory markers. Preprint on arXiv: arxiv.org/abs/2512.13440

Published in Scientific Reports: SarcNet uses federated AI to forecast metastatic relapse in STS, achieving 0.797 AUC. Dive deeper: www.nature.com/articles/s41...

We’re releasing CytoSyn, our new diffusion model for synthetic histopathology image generation. REPA-E architecture. H0-mini conditioning. Model: huggingface.co/Owkin-Biopti... Supporting article: www.owkin.com/blogs-case-s...

Discover OKN4395 at #SITC2025—Poster #1169. A triple inhibitor overcoming tumor immunosuppression by targeting EP2, EP4, DP1. Learn more about OKN4395 at epkin.ai

Our new multimodal ML models achieve high accuracy in predicting iCCA outcomes (c-index 0.70). This collaborative study with @mskcancercenter.bsky.social enhances understanding of the disease and may guide treatment selection. Check out the paper: www.sciencedirect.com/science/arti...

We dived into cancer cell interactions with COMPOTES. Using MOSAIC data, we uncovered 45 programs across major cancers, shedding light on immune dynamics and therapeutic paths. Pre-print at: doi.org/10.1101/2025...

We used ML & multimodal data to find glioblastoma biomarkers linked to long survival. Our model (AUC 0.85), trained on @ap-hp.bsky.social data, reveals tissue regions predicting survival. Model interpretation & biological insights from MOSAIC’s GBM cohort. Pre-print: www.biorxiv.org/content/10.1...

📢 New on bioRxiv: We developed SpaMIL, a novel Multiple Instance Learning (MIL) framework for predicting patient-level outcomes from spatial transcriptomics data. Read the preprint: www.biorxiv.org/content/10.1...

K Navigator launches on May 6th - introducing the scientific literature tool, K-literature, tailored to scientific research, featuring: 📑 26.5+ million papers 🧠 Natural language input 👥 Hallucination limitation K Navigator is free to use for academic researchers. www.owkin.com/k-os/owkin-k...

🕹️ How fast can you reach K? Play the game as we get ready to launch K Navigator, a free agentic AI co-pilot boosting the productivity of biology research and breakthroughs across the world. Try it now: Google Play → play.google.com/store/apps/d... App Store → apps.apple.com/fr/app/first...

Excited to share our latest research on a novel approach for controlling the sensitivity of whole-slide image classification models in digital pathology, using the TSM (Tile Score Matching) method. Read the full pre-print: arxiv.org/abs/2502.20144

💥 Owkin K Navigator launches on the 6th of May. Our free-to-use agentic AI co-pilot for researchers across the world, to enhance the productivity of biological research and drive breakthrough discoveries. Kountdown to K begins now. Sign up here → www.owkin.com/k-os/owkin-k

We're launching a comprehensive multimodal patient data discovery program with 20 leading healthcare institutions across France, Spain, Germany, Italy, the US, the UK, and Israel to accelerate medical research by breaking down data silos. Read more at → www.owkin.com/newsfeed/owk...

Join us for a webinar, co-hosted with @owkin.bsky.social, on "Harnessing spatial biology & AI to develop targeted therapeutic strategies." Learn how Owkin is using spatial & AI to optimize development of antibody-drug conjugate strategies & generate novel hypotheses: bit.ly/4jVbQZr

After an eventful #AIActionSummit, our SVP, Head of Public Affairs & Impact Yedidia Levy-Zauberman unpacks Owkin’s key moments in our latest blog - from a high-impact glioblastoma hackathon to recognition by President Emmanuel Macron. 📕 Read the blog → www.owkin.com/blogs-case-s...

A comprehensive evaluation of histopathology foundation models for ovarian cancer subtype classification www.nature.com/articles/s41... @owkin.bsky.social

Can AI help Pharma speed up drug development? World Street Journal features our CEO Thomas Clozel alongside Aviv Regev from Genentech, AstraZeneca's Chief Data Scientist Jim Weatherall, and GSK's Senior Director of ML and AI Patrick Schwab. 📰 Read the WSJ article → www.wsj.com/tech/ai/phar...

Owkin in the news 👀 Emmanuelle Macron references Owkin’s work, speaking on France’s AI boom 💥 for the #AIActionSummit. Les Echos dives deeper into Owkin’s mission to redefine healthcare through AI: “AI in healthcare - France’s trump card”. Read more here → www.lesechos.fr/start-up/eco...

Le Figaro features our CEO, Thomas Clozel, as one of the top seven French ⭐️s in the AI industry alongside @yann-lecun.bsky.social, Jérôme Pesenti - Owkin board member - and Luc Julia, creator of Apple’s Siri. 👀 Read the article → www.lefigaro.fr/secteur/high...

📢 4 hours in, 100s of lines of code already written, our Glioblastoma moonshot hackathon is off with a bang. 💥 Learn more at: www.owkin.com/connect/glio... #GBM #Hackathon

3 days to go until 100 experts meet at Palais Brongniart, Paris, for our AI Moonshot Hackathon in Glioblastoma Research! Sponsored by Servier and Owkin, with support from AWS, 10x Genomics, Bioptimus, and La Place Fintech. Learn more: www.owkin.com/connect/glio... #gbm #neuroskyence #spatialomics

We’re excited to announce the first patient has been dosed in our Phase 1 clinical trial of OKN4395, a first-in-class EP2/EP4/DP1 triple inhibitor for solid tumors⁠. Learn more in STAT’s latest interview with Andrew Pierce, Owkin’s SVP Discovery & Development: www.statnews.com/2025/01/30/o...

Check out our explainer video to see how our new AI model predicts YAP1/TEAD activity directly from routine H&E-stained histology slides. Watch the video: vimeo.com/1035641635?s... #pathsky #aisky #medai