Pallatom2: Navigating the De Novo Protein Universe with Atom-to-Atom Modeling

Protein binder design requires precise coordination of molecular structure, binding geometry, and experimentally measurable performance. We introduce Pallatom2, an efficient, precise Atom-to-Atom generative model that jointly models protein sequence, backbone conformation, side-chain geometry, and target-interface contacts. A unified Atom14 representation and residue-guided sparse all-atom attention enable bidirectional information exchange between residue-level and atom-level representations, coordinating global structure with local atomic detail in a compact model of just 65.9 million parameters. Its conditioning system provides fine-grained control over binder-target engagement and global and local protein conformations, supporting diverse target-specific design requirements. In silico evaluation demonstrates broad generative capability across long-chain monomers, multi-domain proteins, and homomultimers, with monomer sampling extending to 1,600 residues. In binder design, Pallatom2 achieves state-of-the-art in silico screening performance across all nine targets with cross-method benchmark data under the reported protocols. These in silico results translate into substantial experimental success: biolayer interferometry identifies 112 active designs among 210 tested candidates (53.3%), with binders obtained for all ten experimentally evaluated targets. Six targets yield picomolar-scale apparent affinities, including K_D values of 3.96 pM for SC2RBD and 33.8 pM for PD-L1, while TNFα reaches 1.166 nM. Beyond de novo discovery, starting-interface-guided BCMA design preserves association kinetics despite substantial sequence divergence; partial diffusion improves TNFα affinity by up to 7.57-fold; and CDR-associated interface remodeling converts an inactive de novo nanobody into binders with apparent K_D values of 4.31 and 5.93 nM. Together, these results establish Pallatom2 as a versatile model for protein design, combining efficient atomic modeling, leading in silico screening performance, and experimentally demonstrated capabilities in high-affinity binder discovery, interface transfer, affinity maturation, and recovery of inactive designs.