Flow-matching co-folding model that predicts protein-ligand, protein-protein and nucleic acid complexes in seconds from sequence and ligand topology.
No providers recorded yet. Browse all providers
Co-folding models such as AlphaFold 3 predict a protein and its bound ligand together, but in drug discovery a pose that is nearly right is often not good enough. A ligand with an inverted chiral centre, an impossible torsion or a clash with the pocket wall misleads a medicinal chemist, and a model that needs minutes of A100 time per complex is hard to run across a screening library. NeuralPLexer3 (NP3) targets both problems. It generates all-atom structures of biomolecular complexes from sequences and molecular topology alone, and it does so with a flow-matching sampler that starts from a physically sensible guess instead of random noise.
NP3 was developed by a team at Iambic Therapeutics, led by Zhuoran Qiao and Matthew Welborn. It was previewed as a beta in November 2024, described in a technical report posted to arXiv in December 2024, and published at NeurIPS 2025. It is the third generation of Iambic's line after NeuralPLexer, a diffusion model built with Caltech and NVIDIA, and NeuralPLexer2. The NP3 paper cites the NeuralPLexer2 white paper as its direct predecessor and uses it as a baseline.
The central change from its predecessors is the generative framework. Where NP2 denoised from a Gaussian prior, NP3 learns a continuous normalizing flow by flow matching. Its prior is a globular polymer model that keeps bonded atoms and residues of each chain at realistic distances, produced by a short Langevin relaxation of random coordinates. Because the path from prior to answer is short and nearly straight, the model needs far fewer integration steps to reach a valid structure.
NP3 is an encoder-decoder model. The 350M-parameter encoder embeds atom, bond and stereochemistry features, processes paired multiple sequence alignments, and applies AlphaFold 3-style PairFormer blocks on a budget of anchor atoms. It also conditions on embeddings from ESM-2 (3B) for proteins and RiNalMo (650M) for RNA, in place of RNA MSAs. The 350M-parameter decoder consists of diffusion-transformer flow blocks that attend densely over anchors and with sliding-window attention over all heavy atoms. A 50M-parameter confidence module completes the system. Training used all PDB structures deposited before September 2020, plus 270,000 OpenFold monomer predictions, 2,801 NP2 predictions of disordered multimers, and DFT-refined small-molecule conformers from OrbNet Denali, and took 24 days on a multi-node H100 cluster. On PoseBusters V2, given only sequence and ligand graph, the NeurIPS version reports 77.9% of 308 complexes both within 2 Å ligand RMSD and physically valid, against 73.1% for AlphaFold 3, with RMSD-only accuracy at parity (80.2% versus 80.5%). Ligand stereochemistry was correct in 98.8% of predictions. On NPBench, built from low-homology PDB entries released after 2023, NP3 reached 60.7% success on noncovalent ligands, 69.6% on covalent ligands, 82.6% on modified residues, and 85.1% on protein-peptide interfaces, compared with 76.6% for AlphaFold-Multimer v2.3. Protein-protein interface accuracy matched AlphaFold-Multimer (52.7% versus 52.3% DockQ > 0.23). On ConfBench, NP3 predicted the correct conformational state for 67.4% of apo targets, with a similar rate on holo targets.
NP3 is built for structure-based drug design where no ligand-bound experimental structure exists. It is designed to predict the bound structures of candidate compounds, to model how ligands reshape binding pockets, and, for kinases, to anticipate whether a ligand stabilizes an inactive conformation. Its speed makes structure-aware virtual screening practical, and its coverage of covalent ligands and modified residues suits covalent inhibitor programs. The model runs only inside Iambic's platform.
NP3 replaces the diffusion sampler used by most co-folding models with flow matching from an informative physical prior, and its authors report that this change can match AlphaFold 3's pose accuracy while producing more physically valid structures at a fraction of the inference cost. NPBench fills a practical gap, since AlphaFold 3's recent-PDB evaluation code was never released. The model itself is closed. Iambic has released neither code nor weights, and the NeurIPS paper states that weights will not be released, citing dual-use concerns. Its results cannot be reproduced independently, and its evaluation is computational rather than experimental.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.