Virtual screening 6 - Hybrid

Autoencoder

https://youtu.be/5WoItGTWV54

- Dimensionality reduction : input data의 dimension을 줄이기 위해 사용.

- 그 과정을 encoding 이라고 함.

- Dimensional reduction을 통해 핵심 feature를 잘 학습했는지 확인하기 위한 방법은, 원래의 data를 reconstruction할 수 있어야 함.

 

Classification with separate graphs

GCN with autoencoder for virtual screening

- Graph Convolutional Neural Networks for Predicting Drug-Target Interactions, https://doi.org/10.1021/acs.jcim.9b00628

- Ligand와 protein을 분리.

- Protein의 target structure를 정형화하기 위해 autoencoder를 활용.

- Interaction layer with 'Protein pocket graph' and 'Small molecule representation' -> Binding classifier

- Ligand fingerprint, Pocket fingerprint with same dimension.

https://pubs.acs.org/doi/10.1021/acs.jcim.9b00628

- DrugBank : 실재 임상에 활용되는 약들의 정보를 포함. High-resolution stuructures bound to the drug-like ligands.

- DUD-E : 102 binding pockets.

- Negative data set : IC50 > 50 μmol from ChEMBLE DB.

- MUV dataset : 실험적으로 증명되었지만 active와 inactive가 구조적으로 유사한 것.

 

GCN for virtual screening

- Combining Docking Pose Rank and Structure with Deep Learning Improves Protein–Ligand Binding Mode Prediction over a Baseline Docking Approach, https://doi.org/10.1021/acs.jcim.9b00927

https://doi.org/10.1021/acs.jcim.9b00927

- Pose prediction : PDBbind core v.2017 (4150 complexes)

- Virtual screening : DUD-E (102 complexes)

 

Regression with complex

Fusion of 3D CNN + GCN for virtual screening

- Improved Protein–Ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference, https://doi.org/10.1021/acs.jcim.0c01306

https://doi.org/10.1021/acs.jcim.0c01306

- Gaussian blur with σ = 1.

- Training : PDBbind v.2016 (9226 and 4057 complexes) 

- Test : core set (290 complexes)

- Structure properties : encoding of hydrophobic, aromatic, acceptor, donor, ring.

- R : refined set, G : general set.

- Concatenation : best performance.

 

Pros

- Utilizing flexibility of architecture of deep learning mehods

  : separating protein and ligand features - impossible for physics-based approach

- Combinatio of different architectures : simple average for physics-based approach

 

Cons

- Result performance is not convincing.

 

 

 

Reference

- Hybrid approach for virtual screening l 김우연, https://youtu.be/1qR2SkPVcqY

- Lecture 13 | Generative Models at Stanford University, https://youtu.be/5WoItGTWV54, http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture13.pdf

- Reducing the Dimensionality of Data with Neural Networks, https://www.cs.toronto.edu/~hinton/science.pdf

- Measuring Virtual Screening Accuracy, https://new.pharmacelera.com/science/measuring-virtual-screening-accuracy/