NIM for DiffDock

Type: Microservice Tags: NVIDIA, NIM, BioNeMo, DiffDock, scalable molecular docking, protein-ligand docking, pose prediction, drug discovery, diffusion model Related: NVIDIA-BioNeMo, NVIDIA-NIM, NIM-for-MolMIM, NIM-for-GenMol, NIM-for-Boltz2, NIM-for-OpenFold3, NIM-for-ALCHEMI-Batched-Geometry-Relaxation, cuEquivariance, TensorRT, NVIDIA-Clara, NVIDIA-AI-Enterprise, NGC Sources: https://docs.nvidia.com/nim/bionemo/diffdock/latest/overview.html Last Updated: 2026-04-29

Summary

NIM for DiffDock is NVIDIA’s BioNeMo NIM for scalable molecular docking and protein-ligand pose prediction. Current NVIDIA docs describe DiffDock as a generative model that predicts the three-dimensional structure of a protein-ligand complex and ranks sampled poses with confidence estimates.

Detail

Purpose

DiffDock supports drug-discovery workflows that need to predict how a small molecule ligand binds to a protein target. It can sit downstream of molecular generation and upstream of ranking, refinement, or simulation.

Current scope

  • Takes protein and molecule 3D structures as input.
  • Does not require prior binding-pocket information.
  • Uses an equivariant geometric/diffusion approach where ligand position, orientation, and torsion angles can change during denoising.
  • Outputs multiple sampled protein-ligand poses with confidence estimates.
  • Current docs describe training data from PLINDER and SAIR-derived protein-ligand structure/activity sources.
  • Current docs position DiffDock for accurate, computationally efficient docking in AI drug discovery pipelines.

NVIDIA context

DiffDock is the current Scalable Molecular Docking NIM in the NVIDIA NIM index. It connects molecular generation NIMs such as NIM-for-MolMIM and NIM-for-GenMol to structure/affinity workflows such as NIM-for-Boltz2 and atomistic refinement with ALCHEMI NIMs.

Connections

Source Excerpts

  • NVIDIA docs describe DiffDock as predicting protein-ligand complex structures for molecular docking or pose prediction.
  • The current docs state that DiffDock samples poses and ranks them through a confidence model.

Resources