Generative diffusion model that samples biomolecular conformational ensembles for proteins, RNA, and ligands in hours instead of millisecond-scale MD.
No providers recorded yet. Browse all providers
Structure predictors return one answer per sequence, but biological function rarely comes from one structure. Enzymes catalyse through transition states, receptors signal through allosteric rearrangements, and many drug targets present pockets that exist only transiently. Recovering the underlying equilibrium distribution has meant molecular dynamics, which needs thousands of GPU-hours to reach biologically meaningful timescales for even a moderate protein.
OpenComplex2 (OC2), from the OpenComplex Team at the Beijing Academy of Artificial Intelligence and posted as a preprint in May 2025, is a generative model built to sample those ensembles directly. It applies score-based diffusion over a unified graph representation that standardises proteins, nucleic acids, and ligand SMILES into one format, with denoising handled by FloydNetwork — a graph neural network that passes messages simultaneously at the atomic, residue, and motif scales. Independent samples are drawn from the learned distribution rather than propagated along a trajectory, so a conformational landscape that would take weeks of simulation is covered in hours.
OC2 is positioned as a complement to static predictors such as AlphaFold 3 rather than a replacement. It matches them on single-structure accuracy for many system classes while recovering the conformational breadth they compress away, and it carries forward the same team's OpenComplex line of protein, RNA, and complex modelling tooling.
OC2 trains on curated Protein Data Bank structures with a 30 September 2021 cutoff, retaining native biological partners in each complex so the model learns conformational effects of the molecular environment. Sampling follows a score-based diffusion framework conditioned on polymer sequences and ligand SMILES; static predictions are extracted by averaging over the sampled ensemble.
On structure prediction the picture is mixed by system type. OC2 outperforms AlphaFold 3 on CASP15 RNA targets and on RNA-protein complexes (RMSD 2.91 Å vs. 5.68 Å; TM-score 0.94 vs. 0.85) and on DNA-protein complexes (RMSD 4.05 Å vs. 4.90 Å), matches it on protein monomers, and trails it on protein-protein interfaces (DockQ 0.58 vs. 0.61) and antibody complexes (DockQ 0.57 vs. 0.68). The authors attribute the antibody gap to the ensemble-centric objective, which prioritises heterogeneity over converging on one interface. On the 213-target PoseBusters v2 subset OC2 reaches 77% success at 2 Å with 25 samples, and its accuracy rises rather than falls as sequence similarity to the training set drops. On RNase P, OC2 tracks the experimental per-base RMSF profile at Pearson 0.77 while AlphaFold 3 collapses onto a single conformation.
The clearest use is drug discovery against dynamic targets: sampling cryptic pockets that never appear in a crystal structure, enumerating plausible binding modes early in a campaign before interactions are optimised, and characterising apo-to-holo transitions on a curated benchmark of 83 protein pairs. Beyond ligand work, OC2 suits mechanistic studies of allostery and induced fit, RNA conformational heterogeneity where flexibility governs specificity, and modelling of large symmetric assemblies such as viral capsids and encapsulins that strain conventional pipelines.
OC2 is one of the more ambitious attempts to move structural deep learning from single-state prediction toward thermodynamic ensembles, alongside work such as BioEmu-1, whose evaluation framework OC2 adopts for its fast-folding benchmarks. Its limitations are stated plainly by the authors: performance on heterogeneous systems including RNA-DNA hybrids, multicomponent nucleoprotein assemblies, and very large flexible machines is uncharted; thermodynamic fidelity for multi-domain proteins and for conditions such as temperature and pH is unproven; and rare substates behind high free-energy barriers remain hard to sample. The work is a preprint that has not been peer reviewed. The public GitHub repository currently hosts documentation only, with code and trained weights announced as forthcoming and no licence declared.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.