Chemical-induced gene expression ranking and its application to pancreatic cancer drug repurposing

Abstract

Chemical-induced gene expression profiles provide critical information of chemicals in a biological system, thus offering new opportunities for drug discovery. Despite their success, large-scale analysis leveraging gene expressions is limited by time and cost. Although several methods for predicting gene expressions were proposed, they only focused on imputation and classification settings, which have limited applications to real-world scenarios of drug discovery. Therefore, a chemical-induced gene expression ranking (CIGER) framework is proposed to target a more realistic but more challenging setting in which overall rankings in gene expression profiles induced by de novo chemicals are predicted. The experimental results show that CIGER significantly outperforms existing methods in both ranking and classification metrics. Furthermore, a drug screening pipeline based on CIGER is proposed to identify potential treatments of drug-resistant pancreatic cancer. Our predictions have been validated by experiments, thereby showing the effectiveness of CIGER for phenotypic compound screening of precision medicine.

Keywords: attention; cancer therapy; drug repurposing; gene expression; graph neural network; learning-to-rank; machine learning; pancreatic cancer; phenotype screening; precision medicine.

Graphical abstract

Figure 1

Overview architecture of chemical-induced gene expression ranking (CIGER) This model consists of the four main components: feature-mapping, attention, prediction, and learning-to-rank objective function. It takes input as a tuple of chemical structure, cell line, and L1000 genes and then predicts the ranking of genes in the corresponding gene expression profile. Note that the multi-head attention zoom-in is detailed architecture of the multi-head attention layer in CIGER and is separated from the main figure.

Figure 2

LINCS L1000 data statistical analysis (cell lines, dosages, and time exposures are shown in random order) (A) Gene expression profiles in chemical-cell line space (i.e., yellow denotes missing profiles for chemical-cell line pairs, and red and blue denote that the pairs with the corresponding correlation scores are smaller (unstable) and larger (stable) than 0.6, respectively). (B) Proportion of profiles by cell lines. (C) Proportion of profiles by time exposures. (D) Proportion of profiles by chemical concentrations

Figure 3

Drug screening pipeline using CIGER This model is trained with the LINCS L1000 dataset to learn the relation between gene expression profiles and molecular structures (i.e., SMILES). Then, molecular structures retrieved from the DrugBank database are put into CIGER to generate the corresponding gene expression profiles. Finally, these profiles are compared with treatment profiles calculated from treated and untreated samples to find the most potential treatments for that disease

Figure 4

In vitro experiments of dipyridamole, AZD-8055, linagliptin, and preladenant with the combination of metformin and vitamin C as a positive control (A) Quantification of TET2 levels in drug treatments. Dipyridamole and linagliptin can significantly increase TET2 level after 24-h treatments. (B) Quantifications of GATA6 expressions in drugs treatment. Linagliptin increased GATA6 expressions in PANC-1 after 24-h treatment. Data are presented as means ± SD (n = 3). (C) Linagliptin and metformin vitamin C increased 5hmc levels in PANC-1 cells after 24-h treatment. Quantifications of 5 hmc dot blots (n = 3), data are represented as means ± SD. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001 (unpaired two-tailed t test and one-way ANOVA). (D) The effect of drug treatments on clonogenic survival as a measure of growth rate. All data are presented as means ± SD (n = 3). ∗∗p < 0.005, ∗∗∗∗p < 0.0001 (analyzed using one-way ANOVA)