Cell｜为什么机器还不会说“生物学语言”

TL;DR - A Cell perspective argues that scaling biological foundation models alone will not yield true biological understanding. It advocates process-aligned, multiscale world models that connect causal mechanisms across molecules, cells, tissues, space, and time.

• AlphaFold benefited from strong evolutionary priors and extensive structural data, but its success does not automatically transfer to dynamic, higher-scale biological systems.
• Models should learn repeatable biological processes—such as transcription, cell cycles, and embryonic development—rather than isolated components or static snapshots.
• Future datasets should emphasize native biological contexts, spatiotemporal alignment, multiple scales, and early responses to genetic, drug, or experimental perturbations.
• The proposed path to biological world models is to distill causal mechanisms, standardize reusable modules, and reconnect models across scales within shared process coordinates.