🧬 We are building increasingly detailed maps of human cells. How do we make them useful for everyone generating new data? Our new paper in Nature Genetics introduces ArchMap: a free web platform for mapping single-cell datasets onto reference atlases, without writing code. Built scArches, ArchMap brings reference mapping, cell-type annotation, uncertainty estimates and collaborative exploration into one workflow. One aspect I particularly like: understanding where a new dataset differs from the reference. In a pulmonary fibrosis example, we show how annotation uncertainty, followed by gene-expression analysis, can help reveal disease-associated cell states. For me, this is useful for Human Cell Atlas and related efforts: making the maps we build accessible as everyday tools for biological discovery. Congratulations to Mo , Chelsea, Ronald, Malte and the whole team! Try ArchMap: archmap.bio Read the paper: nature.com/articles/s41588-0… #SingleCell #HumanCellAtlas #ComputationalBiology

How much biology can we recover from single-cell latent representations? 🧬 Latent spaces have become the backbone of modern single-cell analysis. They power atlas integration, foundation models, and increasingly perturbation prediction and Virtual Cell approaches. Yet there is a surprisingly basic question that has received little attention: 👉 How well can we reconstruct gene expression from these latent representations? In our new work, we introduce ReconEval, a benchmark for gene expression reconstruction across more than 100 million cells. We systematically evaluate three settings: 🔹 End-to-end reconstruction (PCA, autoencoders, VAEs) 🔹 Foundation model reconstruction (a new benchmark task for single-cell foundation models, asking how much gene-level information can be recovered from pretrained embeddings) 🔹 Latent shift reconstruction (evaluating reconstruction after perturbation prediction models operate in latent space) A few takeaways: 📊 Autoencoders (not VAEs) achieve the strongest stand-alone reconstruction performance. This is perhaps not surprising, as reconstruction is exactly their training objective. 🧠 Foundation model embeddings retain substantial recoverable gene-level information, although performance depends strongly on the decoder and pretraining strategy. 🔬 The optimal latent representation depends on the downstream task. What works best for reconstruction is not necessarily what works best after perturbation prediction. As the field moves toward increasingly sophisticated Virtual Cell models, reconstruction deserves more attention as an evaluation axis ie another task to evaluate. Ultimately, biological interpretation happens in gene expression space, not in latent coordinates. Congratulations to Xiaotong Fu for driving this project and to all collaborators involved in making this benchmark possible. 👏 📄 Preprint: biorxiv.org/content/10.64898… 💻 Code: github.com/theislab/ReconEva… #SingleCell #ComputationalBiology #VirtualCell

🗺️ Patient Landscapes Looking back, I think this idea has been with us longer than I realized. More than a decade ago we started working on DeepFlow, an AI method to reconstruct trajectories from single-cell data. The goal was simple: understand how cells move through biological state spaces and make decisions. What surprised me was that the same idea also worked for human disease data. In 2017, we showed that patients with diabetic retinopathy could be organized into a continuous disease landscape rather than a few discrete categories. Patients naturally arranged themselves along trajectories from healthy to severe disease. At the time, we were mainly interested in understanding how machine learning models make decisions. Looking back, I think something deeper was already there. Over the years, from cellular trajectories to patient trajectories and, more recently, our work on tracking patient journeys across healthcare records, where patients move through different healthcare events over time, I have become increasingly interested in what I would call patient landscapes. At some point, all of us enter this landscape. The entrepreneur. The factory worker. The farmer. The teacher. The scientist. We all become patients. Yet our journeys can be very different. Some people receive the right diagnosis quickly. Others spend years searching. Some benefit from the best expertise available. Others do not. This makes me wonder whether one of the most important goals for AI in healthcare is not simply prediction or automation. Perhaps it is navigation. Can we help people navigate the patient landscape more successfully? 🔹 Earlier diagnosis 🔹 Better prevention 🔹 More personalized treatment 🔹 Fairer access to expertise If AI succeeds in healthcare, perhaps its greatest contribution will be helping every person benefit from the best medical expertise available, regardless of who they are or where they live. A question I keep coming back to: If we could truly map the patient landscape, how should we use it to improve patient journeys? 📖 References DeepFlow: Reconstructing cell cycle and disease progression using deep learning (Nature Communications, 2017) EHRapy: An open-source framework for end-to-end analysis of electronic health record data (Nature Medicine, 2024) #DigitalHealth #PatientLandscape #AIHealth