Structure-Based Virtual Screening

Contents

• Introduction

Introduction

This repo contains all Structure-Based Virtual Screening source codes and descriptions.

• Docking-Based
• NonDocking-Based

Docking-Based

Contents

• Introduction
• Methodology
• Dataset
• Performance

Introduction

Graph-based algorithm is implemented for docking SBVS problem.

Methodology

1. Docking is performed via 3rd party software (e.g. VINA).
2. Ligand atoms and selected protein atoms formed a graph.
   • selection is based on euclidean distance with an arbitrary cutoff distance.
3. Atom embedding is performed for HAG-Net.
   • randomly initiated embedding is utilized.
4. A Graph-Based classification network based on HAG-Net is utilized for this classification problem.

• TODO:
  • Atom pretraining

Dataset

NoDecoy: An in-house dataset excluding any decoy samples

Dataset	Datasize	Positive Sample Number	Positive Sample Ratio
NoDecoy	309581	245890	79.43%

Performance

All models are trained with a random 5-Fold cross validation, and mean results are listed.

Model	Dataset	AUC	EF @ 2%	EF @ 20%	Accuracy
HAG-Net	NoDecoy	0.89	1.26	1.25	85.80%

• EF: Enrichment Factor
  
  
  
  

NonDocking-Based

Contents

• Introduction
• Methodology
  • NonDockingSG
• Dataset
• Performance

Introduction

Non-docking models are implemented for SBVS problem. Ligand is modeled by graph-based model(HAG-Net), while protein/pocket is modeled by both sequence-based and graph-based models.

• NonDockingGG: graph-based(pocket) + graph-based(ligand)

• NonDockingSG: sequence-based(pocket) + graph-based(ligand)

Methodology

1. Representations of ligand and protein pocket are provided by two different neural network.
2. Interaction between two representations are modeled by a MLP classification network.

NonDockingSG

• Protein/Pocket sequences are modeled using amino acids as basic units.
• Sequence pretraining is performaned with utilizing LM Mask task.

Dataset

NoDecoy: An in-house dataset excluding any decoy samples

Dataset	Datasize	Positive Sample Number	Positive Sample Ratio
NoDecoy	284792	222687	78.19%

Performance

All models are trained with a random 5-Fold cross validation, and mean results are listed.

Model	Dataset	AUC	EF @ 2%	EF @ 20%	Accuracy
NonDockingGG	NoDecoy	0.937	1.279	1.277	89.0%
NonDockingSG	NoDecoy	0.923	1.279	1.276	87.6%

• EF: Enrichment Factor
• Models are trained with 50 epoches, further training MIGHT enhance performance.