Cell types and diagnoses lose information, and AI inherits these simplifications, Morgan Levine argues, proposing that models learn to describe biology from observations themselves

Cell types and diagnoses lose information, and AI inherits these simplifications, Morgan Levine argues, proposing that models learn to describe biology from observations themselves

On October 26, aging researcher Morgan Levine published an essay titled “Biology Beyond Human Intuition” on X. In August, she announced her departure from Altos Labs, a cellular rejuvenation company where she was vice president of computation.

Biological behavior is difficult to predict even when the components are known: the same intervention can have different effects in different cells and genetic backgrounds. Levine suggests that the human mind cannot keep track of thousands of interacting variables at once, so scientists simplify living systems by organizing them into categories: genes, cell types, diseases. These categories made molecular biology and medicine possible, but, Levine writes, “our desire to make everything interpretable or simple may be one of our main limitations.”

Making biology understandable means losing some information, and we rarely know exactly what is lost. A cell’s state may depend on a combination of many features and fit poorly into a list of cell types, while a diagnosis may give an incomplete account of a disease. The choice of measurements also determines what science can see: “Our measurements are shaped by our models of biology, and our models of biology are shaped by our measurements.” Accepted views influence the evidence available to revise them.

Experiments, too, often focus on familiar subjects: laboratories measure the same things in a small number of cell lines (cells grown in laboratories for years) and model organisms. This supports reproducibility, but a model trained on these data may predict outcomes well under those conditions and poorly elsewhere. For a model to learn general principles of biology, Levine argues, varied conditions may be as important as repeated measurements.

Drug development relies on a chain of proxies: binding to a target (the molecule a drug is intended to act on) stands in for drug action, animal experiments stand in for human physiology, and a biomarker (a measurable indicator) stands in for a patient outcome. Uncertainty accumulates at each step: a molecule may bind to its target and have an effect in animals yet fail to help a patient. Levine considers this one reason for frequent failures.

Language models, she says, have inherited many of these simplifications because they were trained on scientific texts, and scientific language is a compressed account of reality, filtered through hypotheses and publication conventions.

“Reasoning more effectively within an existing conceptual framework is a different task from discovering a better framework, which may be what we actually lack.”

A model reasons from an existing description of a system. If that description leaves out important properties, even skilled reasoning may fail to recover them. Models trained on experimental data may inherit the same simplifications if their training objective is to reproduce existing measurements.

Levine proposes instead training models to find representations within the observations themselves: descriptions that retain the information needed to predict a system’s behavior. Her basis for this proposal is that physical processes connect measurements of the same system at different scales. The structure of the observations should therefore help determine how the system is described, alongside established concepts; context remains necessary.

In another post that day, Levine suggested that “life” might be the appropriate subject to model in place of “biology”: “Biology is our human framework for understanding life, distinct from the phenomenon itself.”