#28 · ML / AI methods

Geometric interface conditioning is extending biomolecular design beyond sequence alone

The dominant thread is conditioning learned design on explicit geometry at the level of biomolecular partners rather than sequence alone. RiboBridge separates RNA conformation generation from protein-RNA assembly using cross-frame interaction geometry, evaluated on a 124-complex benchmark. RiboDFM applies backbone-conditioned discrete flow matching to RNA sequence design, reporting high sequence recovery and self-consistency. Astrolabe extends this logic to all-atom design through closed-loop coordination of sequence proposals with structural intervention. FragDockRL and AbOptimizer are reinforcement learning search frameworks over building-block assembly with tethered docking and multisite sequence editing, so they share the design-search framing more than the geometric conditioning. What is new is geometric representation serving as the conditioning interface, with Astrolabe adding closed-loop structural feedback.