#34 · Structural bio & protein design

Search and attention layered on structure models for biomolecular design and docking

The dominant thread is procedural control wrapped around biological models rather than a new base architecture. MCTH runs Monte Carlo Tree Search over hallucinated states from frozen folding and inverse-folding models for all-atom sequence-structure co-design, while NACraft is training-free and backpropagates through structure-model feedback to design RNA and DNA aptamers. MG2Act imposes a sequential cross-attention structure that mirrors two-step molecular glue logic to predict CRBN-mediated degradation. Lacuna and PandaDock are tool-oriented contributions: Lacuna detects and clusters cryptic pockets across conformational ensembles, and PandaDock combines flexible-ligand search with fast affinity grids and an SE(3)-equivariant GNN scoring function. The new element is the use of search or feedback loops and mechanism-shaped attention operating on top of pretrained or physically grounded models.