What happened
This is a survey paper that maps out how AI systems are being used to design new proteins — not just predict their shape, but generate entirely new sequences and structures from scratch. The real contribution isn't a new AI method, but a diagnosis of a fragmentation problem: researchers are using wildly different evaluation standards, making it impossible to know if one approach actually works better than another or if the designed proteins function in the real world.
Why it matters
For the past five years, AI protein design has been moving fast and getting published a lot, but there's been no agreed-upon way to measure whether a designed protein actually does what it's supposed to do — the field has been testing itself in isolation rather than against real biological function, which means investors, biotech companies, and drug developers can't reliably tell which approaches are genuine breakthroughs and which are just good at gaming benchmarks.