Diffusion model that predicts protein-ligand complex structures and ligand-induced conformational changes from protein sequence and a ligand graph.
No providers recorded yet. Browse all providers
When a drug binds a protein, the protein often changes shape to accommodate it. A pocket opens, a loop swings, a domain closes over the ligand. Structure predictors such as AlphaFold2 return one structure per sequence and never see the ligand, so they cannot say which of those states the bound complex adopts. Docking programs see the ligand but usually treat the protein as rigid. NeuralPLexer addresses both problems at once. It is a generative model that takes a protein sequence and a ligand's molecular graph and samples the 3D structure of the whole complex, including the conformational change that binding induces.
NeuralPLexer was developed by Zhuoran Qiao, Weili Nie, Arash Vahdat, Thomas F. Miller III and Anima Anandkumar at Caltech and NVIDIA. Miller was also affiliated with Entos, Inc., now Iambic Therapeutics, where Qiao and Miller later worked. The preprint appeared on arXiv in September 2022, and the peer-reviewed paper was published in Nature Machine Intelligence in February 2024.
The model is the first release of a line that Iambic continued with NeuralPLexer2 and NeuralPLexer3. This entry covers the original academic model only.
NeuralPLexer has three components. A pretrained graph encoder, the Molecular Heat Transformer (51.6 million parameters), embeds ligands and amino acids from their molecular graphs. The contact prediction module (8.3 million) combines these with ESM-2 650M sequence embeddings and a template structure, typically an AlphaFold2 prediction. The equivariant structure denoising module (5.0 million) runs a reverse-time stochastic differential equation over a heterogeneous graph of atoms and local coordinate frames. The main network has 65.8 million trainable parameters, plus 5.0 million in the confidence head. Training used PL2019-74k, a set of 74,477 apo and holo chains from the Protein Data Bank deposited before January 2019, cross-referenced to BioLiP and GPCRdb. During training, templates alternated between experimental structures and AlphaFold2 models sampled from subsampled MSAs. Training took 16 days on six NVIDIA V100 GPUs.
On blind docking against the PDBBind2020 test split, 39.5% of NeuralPLexer's top-ranked poses fell within 2 Å ligand RMSD, the highest fraction among the compared methods, which included DiffDock, GNINA and Uni-Dock pipelines. Pretraining on end-to-end structure prediction raised docking success by about 20% over a model trained on PDBBind alone. On 33 apo-holo pairs from the held-out PocketMiner set, the model reached an average TM-score of 0.93 and outperformed AlphaFold2. On recently deposited ligand-binding proteins it averaged 0.89. For binding-site recovery from truncated AlphaFold2 scaffolds, it recovered up to 46% of sites under a relaxed criterion.
NeuralPLexer suits structure-based drug discovery problems in which the target's bound conformation is unknown or differs from its apo structure. Examples include cryptic and allosteric pockets such as the Switch-II pocket of KRAS G12C, where it predicted an open-like pocket that AlphaFold2 missed. It also serves enzyme engineering and the study of substrate-cofactor complexes. The paper's case studies covered ketol-acid reductoisomerase and purine-binding GPCRs. Once auxiliary features are computed, the model generates a structure in seconds on a single GPU.
NeuralPLexer was among the first deep learning models to predict protein-ligand complexes end to end from sequence while modeling the protein's conformational response. That framing, co-folding protein and ligand rather than docking into a fixed receptor, is also the approach of later all-atom predictors such as RoseTTAFold All-Atom and AlphaFold 3. Its benchmarks are computational, and the case studies compare predictions against existing experimental structures rather than new measurements. The code is released under the Clear BSD license, but the checkpoints on Zenodo are licensed CC BY-NC-SA 4.0 for non-commercial use only. The model's commercial development continued at Iambic Therapeutics in NeuralPLexer2 and NeuralPLexer3.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.