Virtual Screening Workflows: AutoDock Vina and DiffDock

This repository contains two complementary workflows demonstrating both classical and AI-based molecular docking approaches for structure-based virtual screening.
The goal of this project is to establish an end-to-end reproducible pipeline — from protein and ligand preparation to post-docking analysis — using two paradigms:

• AutoDock Vina: Traditional physics-based scoring and docking
• DiffDock: Deep learning–driven generative docking via diffusion models

---

Repository Structure

autodock-workflow/      # Classical docking workflow using AutoDock Vina
DiffDock-workflow/      # AI-based docking workflow using DiffDock

---

Workflow Overview

Workflow	Tool	Description
AutoDock Vina	Physics-based	End-to-end molecular docking pipeline including receptor/ligand preparation, docking with AutoDock Vina, post-processing, ADME triage, and clustering analysis.
DiffDock	Machine learning–based	Equivalent workflow built using DiffDock for fast and accurate docking via diffusion models, including setup scripts, CSV generation, and post-docking clustering analysis.

• Both workflows use CDK2 (PDB ID: 1H1Q) as a model system, docking analogs of the co-crystal ligand 2A6 (NU2058).
• Ligands are sourced from a subset of the Enamine Hinge Binder Library and close-in analogs (80% similarity) with co-crystal ligand 2A6, from PubChem.
  • the co-crystal ligand was also used in the workflow, for a quick validation of the process - although not include in hit triaging/candidate selection.

---

Environment Setup

Due to compatibility differences between the toolchains, the workflows are currently maintained in separate conda environments:

• autodock_env — includes AutoDock Vina, Meeko, Open Babel, and RDKit
• diffdock_env — includes DiffDock, PyTorch, RDKit, Open Babel, and torch_geometric

# For AutoDock Vina
cd autodock-workflow
conda env create -f environment.yml
conda activate autodock_env

# For DiffDock
cd ../DiffDock-workflow
conda env create -f environment.yml
conda activate diffdock_env

Note:
With additional dependency reconciliation (especially between PyTorch and RDKit versions), these workflows could ideally be merged into a single unified environment to streamline reproducibility.

---

Suggested Execution Order

1. AutoDock Workflow (autodock-workflow/)

   • Prepare receptor and ligand libraries.
   • Perform docking using AutoDock Vina.
   • Analyze poses and cluster results.

2. DiffDock Workflow (DiffDock-workflow/)

   • Reuse same receptor and ligand datasets, by converting the formats (bash scripts included)
   • Prepare DiffDock input CSVs and run inference.
   • Analyze DiffDock poses and compare with AutoDock outputs.

---

Outputs

• AutoDock: Top poses (SDF/CSV), ranked binding energies, clustering and ADME triage notebooks.
• DiffDock: AI-predicted docking poses with confidence scores and ranked summaries.

The combined analysis provides a side-by-side comparison of physics-based vs AI-based docking performance on the same target and ligand set.

---

Final List of Candidates

Autodock was completed and produced selections; however, due to low confidence in hinge-binder poses, I did not advance candidates from those. The list below reflects compounds selected via DiffDock.

Selected Compounds

Catalog_ID	Source	Method
Z16226232	Enamine	DiffDock
Z3071631534	Enamine	DiffDock
49863196	PubChem	DiffDock
Z56854611	Enamine	DiffDock
Z1013695804	Enamine	DiffDock
Z1514091612	Enamine	DiffDock
5329551	PubChem	DiffDock
Z2636563025	Enamine	DiffDock
Z442313918	Enamine	DiffDock
5329506	PubChem	DiffDock

Notes

• Autodock runs completed successfully, but hinge-binder pose confidence was insufficient to move those candidates forward.
• DiffDock selections prioritized pose plausibility near the hinge region.
• The PubChem compounds are already CDK2 inhibitors and part of patents, but the data could be explored for further evaluation, and possibly analysis on lead optimization.

Citation

If referencing this work, please cite:

• Trott & Olson, AutoDock Vina: Improving the speed and accuracy of docking, J. Comput. Chem. (2010)
• Corso et al., DiffDock: Diffusion steps, twists, and turns for molecular docking, NeurIPS (2022)
• O’Boyle et al., Open Babel: An open chemical toolbox, J. Cheminf. (2011)