Compound function prediction method based on neural network and connectivity map algorithm

Patent metadata

Field	Value
Publication number	US-12176074-B2
Application number	US-202118278411-A
Country	US
Kind code	B2
Filing date	Apr 25, 2021
Priority date	Feb 23, 2021
Publication date	Dec 24, 2024
Grant date	Dec 24, 2024

Abstract

The present disclosure provides a compound function prediction method based on a neural network and a connectivity map (CMAP) algorithm. The compound function prediction method is used to predict an efficacy of a compound, and the compound function prediction method includes the following steps: constructing a compound molecule-encoding vector neural network; constructing and training an encoding vector-marker gene expression variation deep neural network; constructing and training a marker gene expression level or gene expression variation-whole genome gene expression level or gene expression variation neural network; constructing upregulated and downregulated gene sets of a disease or a phenotype; and evaluating a correlation between the compound and the disease or the phenotype.

First claim

What is claimed is:

1. A compound function prediction method based on a neural network and a connectivity map algorithm, wherein the compound function prediction method is used to predict an efficacy of a compound, and the compound function prediction method comprising the following steps:
(1) acquiring, by a processor, a molecular formula of the compound from a public database; constructing a compound molecule-encoding vector neural network according to a molecular fingerprint generated based on a molecular structure or a molecular fingerprint acquired from another representation of the molecular structure; and training an autoencoder or a variational autoencoder based on the molecular formula of the compound for outputting an encoding vector of the compound molecule or outputting an encoding vector and an encoding radius vector of the compound molecule;
(2) acquiring, by the processor, compound-gene expression variation data from a public database; connecting a multi-layer deep learning network to the compound molecule-encoding vector neural network constructed in the step (1); and constructing and training an encoding vector-marker gene expression variation deep neural network for outputting a marker gene expression variation;
(3) acquiring, by the processor, marker gene expression level or gene expression variation-whole genome gene expression level or gene expression variation data from a public database; connecting a multi-layer deep learning network to the encoding vector-marker gene expression variation deep neural network constructed in the step (2); fitting with a linear equation or a nonlinear equation; constructing and training a marker gene expression level or gene expression variation-whole genome gene expression level or gene expression variation neural network; and calculating a linear or nonlinear relationship between a marker gene expression level or gene expression variation and a whole genome gene expression level or gene expression variation;
(4) acquiring, by the processor, a variation in a disease expression profile from a public database; defining upregulated and downregulated gene sets of a disease or a phenotype; and constructing, based on the compound molecule-encoding vector neural network, the encoding vector-marker gene expression variation deep neural network, the marker gene expression level or gene expression variation-whole genome gene expression level or gene expression variation neural network, and the upregulated and downregulated gene sets of the disease or the phenotype, a compound and disease or phenotype correlation evaluation system, comprising the compound molecule-encoding vector neural network, the encoding vector-marker gene expression variation deep neural network, the marker gene expression level or gene expression variation-whole genome gene expression level or gene expression variation neural network, and an evaluation system for a correlation between the whole genome gene expression level or gene expression variation corresponding to the compound and the upregulated and downregulated gene sets of the disease or the phenotype; and
(5) inputting, by the processor, a molecular formula of a virtual molecule to be evaluated into the compound and disease or phenotype correlation evaluation system to evaluate a correlation between the compound and the disease or the phenotype, specifically comprising, inputting the molecular formula of the virtual molecule to be evaluated into the compound molecule-encoding vector neural network, and outputting an encoding vector; inputting the encoding vector into the encoding vector-marker gene expression variation deep neural network, and outputting the marker gene expression variation; inputting the marker gene expression variation into the marker gene expression level or gene expression variation-whole genome gene expression level or gene expression variation neural network, and outputting the whole genome gene expression level or gene expression variation; and inputting the whole genome gene expression level or gene expression variation and the upregulated and downregulated gene sets of the disease or the phenotype into the evaluation system for the correlation between the whole genome gene expression level or gene expression variation corresponding to the compound and the upregulated and downregulated gene sets of the disease or the phenotype, and outputting a score based on a scoring formula for a probability of the compound in treating or exacerbating the disease;
wherein the scoring formula is as follows:

a
=

max

j
=

1
⁢

tot

[

j
t

-

V
⁡
(
j
)

n

]

b
=

max

j
=

1
⁢

tot

[

j
t

-

V
⁡
(
j
)

n

]

score
=

{

a
-
b

when
⁢

b
*
a

<
0

0

when
⁢

b
*
a

>
0

wherein t denotes a number of upregulated (or downregulated) genes, n denotes a total number of genes in a whole genome expression variation prediction system, and V(j) denotes a sequence number of a j-th upregulated (or downregulated) gene in a descending order of gene expression variations among the n genes; and
(6) screening out compounds

Assignees

• Beijing Gigaceuticals Tech Co Ltd
• Univ Beijing

Inventors

• Xie Zhengwei
• Zhu Jie
• Wang Jingxiang
• Gao Mingjing
• Liu Zhurui

Classifications

• G16C20/70

  Machine learning, data mining or chemometrics · CPC title

• G16C20/90 Primary

  Programming languages; Computing architectures; Database systems; Data warehousing · CPC title

• G16C20/10

  Analysis or design of chemical reactions, syntheses or processes · CPC title

• G16C20/30 Primary

  Prediction of properties of chemical compounds, compositions or mixtures · CPC title

• Y02A90/10

  Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation · CPC title