Method for generating small molecule based on pharmacophore model, device, and medium

A method for generating a small molecule based on a pharmacophore model, comprising:

• step 1, using a fully connected graph to represent a pharmacophore model;
• step 2, using a feature extraction model based on a graph neural network to extract a feature vector of the pharmacophore model from the fully connected graph of the pharmacophore model;
• step 3, performing random sampling on a specified prior distribution of latent variables to obtain a latent variable;
• and
• step 4, inputting the latent variable and the feature vector of the pharmacophore model to a pre-trained decoder to generate a molecule matching the pharmacophore model;
• wherein the pre-trained decoder is a decoder obtained by training a variational autoencoder with training samples;
• and the variational autoencoder is expressed as:
• P
• ⁡
• (
• x
• ❘
• c
• )
• =
• ∫
• z
• ∼
• P
• ⁡
• (
• z
• ❘
• c
• )
• P
• ⁡
• (
• x
• ❘
• c
• ,
• z
• )
• ⁢
• P
• ⁡
• (
• z
• ❘
• c
• )
• ⁢
• dz
• ;
• wherein P(x|c) represents the variational autoencoder, and P(z|c, x) and P(x|c, z) represent an encoder and a decoder forming the variational autoencoder, respectively;
• x represents a molecule, which is uniquely encoded as a combination of the feature vector c of the pharmacophore model and a latent variable z;
• and the latent variable z represents a way of the pharmacophore model mapping to the molecule, namely how chemical components in the pharmacophore model express and are combined into the molecule;
• wherein the variational autoencoder comprises an encoder, a decoder, a prior distribution of latent variables, and a loss function;
• the training samples for training the variational autoencoder comprise the feature vector of the pharmacophore model and the molecule matching the pharmacophore model;
• upon training the variational autoencoder with the training samples:
• the feature vector of the pharmacophore model and the molecule matching the pharmacophore model are input to the encoder to obtain a latent variable distribution of the molecule with respect to the pharmacophore model;
• a latent variable is sampled from the latent variable distribution output from the encoder, and the feature vector of the pharmacophore model and the sampled latent variable are input to the decoder to generate a molecule by an iterative loop;
• the prior distribution of latent variables is used for performing minimal optimization on a difference between the latent variable distribution obtained by the encoder during training and a given prior distribution;
• and
• the loss function is used for quantizing a difference between the molecule generated by the decoder and the molecule in the training samples.