Virtual screening 4 - 3D CNN

Deep Neural Network (fully connected)

- Large number of parameters -> easy to be overfitted when data is small. or large memory consumption (GPU)

- Does not enforce any structure, e.g., local information (local feature를 찾아내는게 어렵다.)

 

Convolution Neural Network (weight sharing and convolving)

- Reduce the number of parameters (less overfitting)

 

3D CNN for virtual screening

- Grid representation

- Binding site

Atom type 별 channel을 만든다.

 

Atom position -> Atom density representation

https://doi.org/10.1021/acs.jcim.6b00740

 

Binary classification (active vs. inactive)

- AtomNet, ChEMBL-20 PMD: DUDE-like dataset derived from ChEMBL version 20.

- PAINS rule.

- Inactive : measured activity is higher than 30μmol.

- AutoDock Vina scoring.

- Adjustment logAUC

- Deep Learning은 Generalization이 잘 안된다.

 

Regression (scoring)

- Pearson's correlation (R) among predicted and experimental fiinities on blind targets.

- Force field energy 

$$ f(r) = 1 - exp(-(r_{vdw}/r)) $$

- ResNet : skip connection with zero-padding.

- 4 molecular descriptors (4HTMD)

  1) aromatic carbon 2) hydrogen bond acceptor 3) positive ionizable 4) negative ionizable

 

pros

- Protein-ligand 결합의 원천이 되는 3차원 결합구조를 반영할 수 있음.

- 3차원 구조를 반영하기 때문에 pose prediction이 가능함.

- Torsion angle과 같은 복잡한 3차원적 특징을 스스로 학습가능.

 

Cons 

- Grid를 사용하기 때문에 grid 간격에 따라 정보의 손실이 있음.

- 3차원 구조를 사용하지만 protein residue와의 해석에 한계가 있음.

- Rotational 및 translational invariant 하지 않음.

 

 

Terminology

- Receptive field : 이미지의 local한 영역.

 

 

 

Reference

- 3D CNN for virtual screening l 김우연, https://youtu.be/eWz-nwDrgws

- Binary classification: active vs. inactive

  : AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery, https://arxiv.org/abs/1510.02855

  : Protein–Ligand Scoring with Convolutional Neural Networks, https://doi.org/10.1021/acs.jcim.6b00740

- Regression - binding affinity scoring

  : KDEEP: Protein–Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks, https://doi.org/10.1021/acs.jcim.7b00650

  : RosENet: Improving Binding Affinity Prediction by Leveraging Molecular Mechanics Energies with an Ensemble of 3D Convolutional Neural Networks, https://doi.org/10.1021/acs.jcim.0c00075

- GNINA 1.0: molecular docking with deep learning, https://doi.org/10.1186/s13321-021-00522-2

  : https://github.com/gnina/gnina

- Convolutional Neural Networks - Ep. 8 (Deep Learning SIMPLIFIED), https://youtu.be/JiN9p5vWHDY

- PySCF: the Python-based simulations of chemistry framework, https://doi.org/10.1002/wcms.1340

- Fast Rescoring Protocols to Improve the Performance of Structure-Based Virtual Screening Performed on Protein–Protein Interfaces, https://doi.org/10.1021/acs.jcim.0c00545