#2 · Structural bio & protein design
Conditioned generative models for antibody, small-molecule and peptide design
The dominant thread is conditioning generative or predictive models on structural, target, or property information. AbRefine fuses backbone-conditioned structural predictions with framework-conditioned antibody language model distributions for CDR design, and pocket contrastive learning with reference-conformation fusion generates 3D molecules robust to receptor conformational variation. DesignMaster applies an E(3)-equivariant diffusion model conditioned on linker length for PROTAC design, while PolyTamGen generates multitarget ligands from ESM3 target encodings. What is new across the set is explicit multi-constraint steering: M-GoRT uses goal-conditioned return blocks for multi-objective de novo generation, and DeepSAP pairs conditional generation with property prediction for self-assembling peptides. The thread is not uniform; several items, including FunFOLD5, Seq2Pocket and RCP-DTA, are predictive rather than generative.
Preprints
A deep learning-generated hASIC1a miniprotein inhibitor confers neuroprotection in stroke
Deep learning-designed hASIC1a miniprotein Denasin1 gives nanomolar, subtype-selective channel inhibition and cuts infarct volume over 50% in a stroke mouse…
Frustration Quenching and Network Topology of the Energy Landscape as Primary Determinants of Protein-Ligand Binding Pose Prediction by Deep Learning Models
Links deep learning ligand pose prediction accuracy to frustration quenching and network topology, showing allosteric sites underperform due to flatter energy…
Efficient detection of cryptic pockets on challenging drug targets using MD-free co-vibrational modes
Applies a new MD-free COVIB variant to detect cryptic pockets on KRAS and STAT5B N-terminal domain, validated by photoaffinity fragment MS mapping and NMR.