Official implementation of NeuralPLexer, a deep generative model to jointly predict protein-ligand complex 3D structures and beyond.

Qiao Z, Nie W, Vahdat A, Miller III TF, Anandkumar A. State-specific protein-ligand complex structure prediction with a multi-scale deep generative model. Nature Machine Intelligence, 2024. https://doi.org/10.1038/s42256-024-00792-z.

Pretrained model checkpoints described in the published manuscript, downstream evaluation datasets, and predicted structures are available at the following Zenodo repository for non-commercial usage under the CC BY-NC-SA 4.0 license: https://doi.org/10.5281/zenodo.10373581.

A GPU machine with CUDA>=10.2 support is required to run the model.

We recommend setting up the libmambda solver for conda before installation. For a Linux environment, run the following commands to install the package:

make environment
make install

Example usage for the base model with a template structure in pdb format:

neuralplexer-inference --task=batched_structure_sampling \
                       --input-receptor input.pdb \
                       --input-ligand <ligand>.sdf \
                       --use-template  --input-template <template>.pdb \
                       --out-path <output_path> \
                       --model-checkpoint <data_dir>/models/complex_structure_prediction.ckpt \
                       --n-samples 16 \
                       --chunk-size 4 \
                       --num-steps=40 \
                       --cuda \
                       --sampler=langevin_simulated_annealing

NeuralPLexer CLI supports the prediction of biological complexes without ligands, with a single ligand, with multiple ligands (e.g. substrate-cofactor systems), and/or with receptors of single or multiple protein chains. Common input options are:

• input-receptor and input-ligand are the input protein and ligand structures;

  • input-receptor can be either a PDB file or protein sequences. In case the input is a multi-chain protein in the primary sequence format, the chains should be separated by a | sign; in case the input is a PDB file, no coordinate information from the file is used for generation unless the file itself is separately provided as a template structure via input-template.
  • input-ligand can be either sdf files or SMILES strings. In case the input is a multi-ligand complex, the ligands should be separated by a | sign;

• use-template and input-template are the options to use a template structure for the input protein;
• out-path is the output directory to store the predicted structures;
• model-checkpoint is the path to the trained model checkpoint;
• n-samples is the number of conformations to generate in total;
• chunk-size is the number of conformation to generate in parallel;
• num-steps is the number of steps for the diffusion part of the sampling process;
• separate-pdb determines whether to output the predicted protein structures into dedicated PDB files;
• rank-outputs-by-confidence determines whether to rank-order the predicted ligand (and potentially protein) output files, where outputs are ranked using the predicted ligand confidence if available and using the predicted protein confidence otherwise;

Expected outputs under <output_path>:

• prot_all.pdb and lig_all.sdf contains the output geometries of all n_samples predicted conformations of the biological assembly;

  • prot_0.pdb, prot_1.pdb, ... stores the individual frames of the predicted protein conformations;
  • lig_0.sdf, lig_1.sdf, ... stores the individual frames of the predicted ligand conformations.

In benchmark_tiny.sh we also provided minimal example commands for running complex generation over many distinct input sets using data provided in in the Zenodo repo, analogous to the process used to obtain the benchmarking results but with reduced number of samples, denoising steps, and template choices.

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.