Human Cell Atlas co-founder Sarah Teichmann: the single-cell reference is near a first draft, and drug development uses it first for targets and safety

The atlas is the layer between genome and physiology

She places the Human Cell Atlas after the Human Genome Project and the Cancer Genome Project: single-cell and spatial genomics that measure each cell's molecular fingerprint and where it sits in tissue. Her example is variants that raise the risk of atrial fibrillation: read so far from averaged heart-tissue samples, they now appear to act in electrically active myocardial sleeve cells in the pulmonary vein, and in the opposite direction from the bulk data, raising expression of an enzyme that makes polyamines.

“Before we had kind of a Google at the level of cities and towns and villages, now we have a Google Street Maps at the level of individual houses.”Sarah Teichmann

The healthy reference alone already informs targets and safety

Even from healthy tissue alone, she says, the atlas shows which cells express a drug target and how much, which matters for safety and toxicity. Disease atlases come mostly from separate cohort studies using the same technology; set against the healthy reference, they show which cell changes drive a disease early and which follow, and they help stratify patients. Citing her group's preprint, she says such data predicts which targets pass Phase I, II and III, and that the Zou lab's Virtual Biotech at Stanford found something similar with AI agents.

From prediction to lab is a climb, one level at a time

Asked what must happen before an AI-suggested target earns lab money, she describes a back and forth between data science and wet lab. Most of the atlas is inferred from RNA, so the first test is whether the protein is really on the cell surface; then perturbation experiments in the cell type concerned, then organoids, engineered tissues and perhaps animals. The high-throughput lab in the loop is already clear for antibody development; she thinks the same loop can run across kinds of experiment, each step earning the confidence for a more complex and costly one.

Disease is the real perturbation, and scientists still choose

The host mentions a frontier AI company that has built its own robotic wet lab. She says the cell perturbations robots run are systematic, but human disease is the real perturbation, and data from it is not systematic, so it has to be integrated and modelled with care to say what happens inside the body. Scientists still build the models, design the robots and cell systems and choose the perturbations; much of antibody discovery is already robotised and in a cycle with AI, and those teams have not gone, they work with different skills. Models also go down 'rabbit holes', she says, so benchmarks, evaluations and checks are needed.

Two things limit open data

Progress would be fastest, she says, if all models and data were open, but commercial interest and intellectual property keep companies' data and models in-house, and medical data on living people is bound by privacy rules. One open question is how much of the training data sits in a model's weights, and so whether such a model can be released.