AI and human design are complementary because they fail in different places: AI is limited by judgment and by the need for experimental confirmation, while human design is limited by how many candidates can be physically built and tested.
Where each approach breaks down
AI pipeline
- Cannot decide which target or which trade-off matters
- Produces predictions, not measurements
- Least reliable on cases unlike its training data
Human pipeline
- Limited by physical rounds of building and assaying
- Cannot explore sequence space at computational scale
- Depends on individual expert experience and intuition
A high predicted binding score is a hypothesis, not a result. So is a well-reasoned human design. In both cases, expressing the antibody and measuring its binding is what turns a candidate into knowledge.