Iambic's second-generation generative model that predicts protein-ligand, protein-protein and protein-nucleic acid complex structures from sequence.
No providers recorded yet. Browse all providers
Classical docking places a ligand into a protein structure that has already been solved with a ligand bound. For most real drug targets that holo structure does not exist, and the pocket a drug occupies may only take shape once the drug is there. NeuralPLexer2 removes the need for it: given only protein sequences and the chemical structure of a ligand, it generates the three-dimensional coordinates of the whole complex, protein and ligand together, as one generative sample.
NeuralPLexer2 was developed by Iambic Therapeutics in San Diego, trained in October 2023, and announced on 12 February 2024 in a technical white paper on Iambic's website. The announcement came the same day the original NeuralPLexer, a collaboration between Iambic, Caltech and NVIDIA, was published in Nature Machine Intelligence. Where the first model focused on protein-ligand complexes, NeuralPLexer2 extends coverage to protein-protein complexes, protein-nucleic acid complexes, cofactors and post-translational modifications, spanning almost all structure types in the Protein Data Bank.
NeuralPLexer2 has no peer-reviewed paper or preprint of its own. Iambic's later NeuralPLexer3 report cites the NeuralPLexer2 white paper as its predecessor, uses NeuralPLexer2 as a comparison baseline, and draws part of its distillation data from NeuralPLexer2 predictions. The model runs inside Iambic's internal drug discovery platform, and Iambic has released no code, weights or public API for it.
Iambic's NeuralPLexer3 report describes NeuralPLexer2 as a diffusion-based generative model with an equivariant graph-based denoising decoder and a Gaussian prior, developed continuously from the original NeuralPLexer. For the second generation, Iambic built hardware- and memory-optimized geometric attention blocks that allow training directly on complexes of up to 2,400 residues. The training set grew to 134,000 refined PDB structures, nearly double that of the first model, after a processing pipeline removed artifact ligands based on their atomic context. Iambic reports that validation loss fell predictably as the generative head and training iterations were scaled, that this loss tracked interface LDDT on sampled structures, and that the final model size was chosen by extrapolating those scaling curves. The parameter count was not disclosed.
On the PoseBusters benchmark, with all related PDB entries removed from training, NeuralPLexer2 placed 76.8% of ligands within 2 Å RMSD of the experimental pose when the binding site was specified. On the 175 of 308 targets where the model was confident, the success rate was 75.4% from sequence and ligand alone and 93.2% with binding-site residues within 10 Å provided. On 32 recently determined G protein-coupled receptor structures, it reached a median receptor TM-score of 0.91 from sequence and ligand input alone; the white paper plots NeuralPLexer v1 on the same set for comparison. These figures come from Iambic's own white paper and have not been independently reproduced.
NeuralPLexer2 is built for structure-based drug discovery on targets without a usable ligand-bound structure. Iambic uses it to predict the bound structure of every compound in a program, to find cryptic binding sites, and to work out mechanisms of action for difficult targets, GPCRs in particular. Its inference speed also suits large in silico screens and proteome-wide studies of how proteins change conformation when ligands bind. Because the model is internal, these uses are available only through Iambic's own programs and partnerships.
NeuralPLexer2 sits between two published generations of Iambic's co-folding line: the original NeuralPLexer, one of the first deep generative models to fold protein and ligand jointly, and NeuralPLexer3, a flow-matching successor that uses it as a baseline. Its announcement came a few months before AlphaFold 3 made all-atom complex prediction across proteins, nucleic acids and small molecules a mainstream capability. With no paper, code or weights, its benchmark results rest on the company's own reporting, and outside researchers cannot use or evaluate it directly.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.