Protein design model RFD4-Proteina puts chemistry and conformation in one latent space: 10.6% clustered pass rate on protein binding at 42-fold higher throughput
Synopsis
RFD4-Proteina encodes standard and modified amino acids, sugars, DNA and RNA in a single chemical and conformational latent space, and a flow-matching model generates molecular identity and conformation jointly, with recurrent refinement inside the network and constraint-guided search across trajectories. On four in silico benchmarks covering protein binding, small-molecule binding, enzyme active-site scaffolding and DNA binding, the high-thinking setting reaches higher clustered pass rates than any individual baseline evaluated while generating at higher throughput; experimentally it yields calcium-dependent calmodulin binders, phosphorylation-dependent binders, and de novo enzymes that hydrolyze paraoxon and form amide bonds.
Interpretation
A protein design model can now decide molecular identity and three-dimensional conformation together inside one shared chemical and conformational latent space, instead of splitting sequence design from structure generation. Earlier, RFD3 generated all-atom structures across diverse molecular contexts but did not design high-quality sequences for the structures it produced, while Proteina-Complexa co-generated sequence and structure but relied on a protein-specific representation with fixed-conformation interacting partners. On four in silico benchmarks, high-thinking RFD4-Proteina reached clustered pass rates of 10.6%, 14.3%, 12.1% and 3.5% for protein binding, small-molecule binding, atomic motif enzyme scaffolding and DNA binding, against 5.9% (BindCraft), 2.4% (RFD3), 5.7% (AlphaProtein-novo codesign) and 0.8% (BoltzGen) for the next-best baseline.
Generation is split into two reasoning layers: repeated internal updates refine the prediction inside the network, and external search filters trajectories by user-supplied energetic or geometric criteria. Earlier approaches placed recycling in an upstream feature-processing stage or relied on self-conditioning, rather than performing recycling within each denoising timestep. On one enzyme task cleaving a peptide-DNA bond, filtering trajectories in stages by interface area, hydrogen bonding and AlphaFold 3 interface confidence yielded three times as many unique passing designs per GPU-hour as naive sampling; raising internal updates at intermediate denoising steps from one to five improved interface hydrogen-bond energy, active-site geometry and interface confidence in AF3-refolded structures.
Perspective
The demonstrated scope covers four in silico benchmarks, specialized workflows such as side-chain packing and sequence design, experimental binders against calmodulin, peptide-MHC complexes, intrinsically disordered region peptides, the TrkA family, glycosylation and phosphorylation sites, and experimental characterization of a paraoxon hydrolase and an amide-forming ligase; the preprint has not been peer reviewed.
The authors note that the optimal reasoning schedule varies between modalities and tasks, so small-scale pilots are needed to fix a configuration; on easier tasks, or tasks whose objectives resolve late in the trajectory, search gains do not outweigh its computational overhead; and the calmodulin affinity estimate is conditional on the assay's lower off-rate limit.
