Network-based prediction of drug combinations

Nature communication 2019 / 03 IF=15

https://doi.org/10.1038/s41467-019-09186-x

image-20220429151826294

文章使用数据

1、PPI

high-quality protein-protein interactions (PPIs)

原文说有243,603 PPIs connecting 16,677 unique proteins ，但是根据文章附件链接只有217160 PPIs, 涉及15970个蛋白质

https://static-content.springer.com/esm/art%3A10.1038%2Fs41467-019-09186-x/MediaObjects/41467_2019_9186_MOESM3_ESM.xlsx

2、药物数据

（1）药物组方：DCDB

1,363 clinically reported drug combinations for 904 distinctive components

http://public.synergylab.cn/dcdb/index.jsf

在本文，作者仅关注 pairwise drug combinations ：681 combinations connnecting 362 drugs

（2）TTD数据库

http://db.idrblab.net/ttd/

drug, target, and drug-target pathway information

3、疾病靶点

（1）OMIM：The OMIM database (Online Mendelian Inheritance in Man)

http://www.omim.org/

literature-curated human disease genes with various high-quality experimental evidences.

（2）CTD：The Comparative Toxicogenomics Database

http://ctdbase.org/

only manually curated gene-disease interactions from the literatures were used.

（3）ClinVar：relationships among sequence variation and various human phenotypes

https://www.ncbi.nlm.nih.gov/clinvar/

cardiovascular[Disease/Phenotype]

（4）GWAS

https://www.ebi.ac.uk/gwas/

unbiased SNP-disease associations with genome-wide significance

p <5.0×10 -8

（5）GWASdb

http://jjwanglab.org/gwasdb NOT USED

SNP-trait associations from GWAS for PubMed and other resource

p <1.0×10 -3

（6）PheWAS Catalog

https://phewascatalog.org/phewas

phewas.mc.vanderbilt.edu

SNP-trait associations identified by the phenome-wide association study (PheWAS) paradigm within electronic medical records

p <0.05

（7）HuGE Navigator

https://phgkb.cdc.gov/PHGKB/phgHome.action?action=home

an integrated disease candidate gene database based on the core data from PubMed abstracts using text mining algorithms

literature-reported disease-gene annotation data with known PubMed IDs

（8）DisGeNET

16,000 genes and 13,000 diseases

only expert-curated data

附件上传的数据

The publicly available human protein–protein interactome (Supplementary Data 1)

experimentally validated drug–target interactions (Supplementary Data 2)

experimentally validated drug combinations (Supplementary Data 3)

分析思路

1、药物靶点群与疾病群距离

  • 假设：Disease proteins are not scattered randomly in the interactome, but tend to form localized neighborhoods, known as disease modules

  • 计算药物与疾病的Network-based proximity

$$ d(X,Y) = \frac{1}{||Y||}\sum_{y\in{Y}}min_{x\in{X}}d(x,y) $$

  • 计算相同数量size与连接度degree的靶点群与疾病靶点群的距离，拟合高斯分布，进行Z值转换。由此判断药物距离与疾病距离是否足够近(z<0)。 $$ z = \frac{d-\mu}{\sigma} $$
image-20220429163822810

2、计算药物靶点群与药物靶点群的距离

  • 药物靶点的数目通常比较少（FDA批准的1978个药物中，平均靶点为3。由上公式计算的药物与药物间距离的随机化不符合正态分布。
  • 进一步改进计算公式，考虑两个药物的影响范围。同第一点，进行Z值转换。

$$ s_{AB} = d_{AB} - \frac{d_{AA} + d_{BB}}{2} $$

image-20220429164754123

通过药物组合实验数据、药物性质以及药物距离的关系

（1）两个药物靶模块之间的拓扑关系也反映了生物和药理学关系。

（2）FDA批准的两两药物组合之间的距离比较近。如下图A

image-20220429165335033

3、药物-药物-靶点群关系

  • 文章分为6种药物-药物-靶点群关系
  • 结合已知药物组合数据发现：only drug pairs that have Complementary Exposure relationship to the disease module(下图B) show a statistically significant efficacy for drug combination therapies

image-20220429165833622

  • 最后文章利用Complementary Exposure relationship，预测、筛选具有抗高血压的药物组合。发现预测显著的结果有相应的研究数据支持，表明这种药物组合发现模式的准确性。

前期准备好数据很重要。不仅仅是用于预测的数据，还要有支持验证的数据。