How do we engineer, screen and optimise complex multi-specific antibodies in the age of AI?
01 Mar 2027
Formats & Scaffolds
- Optimisation acts on the whole molecule, not the binder — ML models are focused on single binding entities against specific epitopes. However, affinity, epitope, valency and geometry interact non-additively, and good arms often fail once assembled.
- Maximal affinity is often the wrong property — avidity-driven selectivity needs deliberately tuned, often weak arms; generative models are built to maximise binding.
- Architecture drives function, and there is no data for it — arm spacing, linker length and approach angle decide agonism, yet no public data on architecture-to-function exists.
- The design space outruns the assays — thousands of possible assemblies which require functional screening (not just binding).
- Developability is emergent — chain pairing, aggregation, titre and PK are properties of the assembly, not of the individual de novo binders.
