• Pallatom, a dual-track diffusion-based model (residue-level and atom-level) that can co-generate protein structures and sequences simultaneously
• Introduce the atom14 representation as 14 atom coordinates of a residue with non-existing atoms using the Cα atom position. Traverses atomic representations across decoder layers
• In addition to AF3, trains the model with cross-entropy for sequence prediction and local relative distance distribution at the atomic level
• Uses new evaluation metrics: DES-aa (mean pLDDT > 80 & all-atom RMSD < 2 Å), DIV-str (FoldSeek), DIV-seq (MMSeq), NOV-str (TM-score), NOV-seq