Virtual screening 2 - AI

Molecular structure-property relationship by supervised learning

- Input : Structure or protein (X) 

 

- Method

  : Convention, 

$$ Y = f(X) $$

where f = Schrodinger equation or Hamiltonian

  : Modern, feature extraction (L),

$$ Y = f_θ(X)$$

where f = AI or machine learning (DNN, CNN, RNN, GNN, etc), θ = a set of learnable parameters

 

- Output data : Property (Y), biding affinity.

Modeling (θ) = Maximum likelihood estimate

- Assumption : Random error term follows Gaussian distribution.

 

θ : ML model

x : observation

y : target distribution

 

find θ is called machine learning training.

$$ L(θ) = L(θ; x,y) = p(y|x;θ) $$

 

Machine learnig vs. Deep learning

- Machine learning process (shallow learning)

  : Input -> feature extraction (hand-crafted features by experts) -> other ML -> Output

 

- Deep learning process

  : Input -> Deep Network (automated feature extraction by AI) -> Output

 

- They are different by data size.

 

Advantages using AI

- Physics vs. machine learning : cost efficiency.

- Inverse design : AI를 통해 원하는 binding affinity를 가지는 구조를 얻을 수 있다. (e.g. hit-to-lead)

 

 

 

Reference

- Why AI for virtual screening l 김우연, https://youtu.be/l2LfNIP82Yw