Pathway-extended gene expression signatures integrate novel biomarkers that improve predictions of patient responses to kinase inhibitors

Abstract

Cancer chemotherapy responses have been related to multiple pharmacogenetic biomarkers, often for the same drug. This study utilizes machine learning to derive multi-gene expression signatures that predict individual patient responses to specific tyrosine kinase inhibitors, including erlotinib, gefitinib, sorafenib, sunitinib, lapatinib and imatinib. Support vector machine (SVM) learning was used to train mathematical models that distinguished sensitivity from resistance to these drugs using a novel systems biology-based approach. This began with expression of genes previously implicated in specific drug responses, then expanded to evaluate genes whose products were related through biochemical pathways and interactions. Optimal pathway-extended SVMs predicted responses in patients at accuracies of 70% (imatinib), 71% (lapatinib), 83% (sunitinib), 83% (erlotinib), 88% (sorafenib) and 91% (gefitinib). These best performing pathway-extended models demonstrated improved balance predicting both sensitive and resistant patient categories, with many of these genes having a known role in cancer aetiology. Ensemble machine learning-based averaging of multiple pathway-extended models derived for an individual drug increased accuracy to >70% for erlotinib, gefitinib, lapatinib and sorafenib. Through incorporation of novel cancer biomarkers, machine learning-based pathway-extended signatures display strong efficacy predicting both sensitive and resistant patient responses to chemotherapy.

Keywords: biochemical pathways; gene signatures; machine learning; systems biology; tyrosine kinase inhibitors.

FIGURE 1

Procedure for pathway gene selection. An initial set of genes with known associations to a particular TKI (here, we show a subset of sorafenib‐related genes) are selected and then evaluated by MFA, which was used to find a correlation between cell line drug sensitivity (GI₅₀) and the GE or CN of these genes in those cell lines (left). MFA correlation circles visualize these relationships (bottom). The gene list is extended, using pathway and interaction databases (i.e. PathwayCommons) to find genes related to curated genes which showed MFA correlation to GI₅₀ (one‐node distant genes; middle‐left). The list is extended again from the MFA‐correlating one‐node distant genes (two‐node distant genes; middle‐right). All curated and extended genes which showed an MFA correlation were then used as features to generate a final predictive SVM gene signature for the evaluated TKI (right). Genes within the best performing sorafenib signature are indicated in thick borders (black for curated genes; purple for pathway‐extended genes)

FIGURE 2

Accuracy of curated and pathway‐extended SVMs on TKI sensitive and resistant patients. The predictive accuracy of the best‐performing curated (C‐) and pathway‐extended (PE‐) models for each TKI was arranged based on their accuracy in classification of drug‐sensitive and drug‐resistant tumour patients. This illustrates how curated models are often only accurate towards one patient class (sensitive or resistant) but not both (red), which is an issue as the patient data were often imbalanced (number of sensitive | resistant patients in each study: lapatinib [‘Lap’; n = 8 | 23], imatinib [‘Ima’; n = 17 | 6], sunitinib [‘Sun’; n = 6 | 12], erlotinib [‘Erl’; n = 9 | 34], gefitinib [‘Gef’; n = 10 | 2] and sorafenib [‘Sor’; n = 21 | 46]). Conversely, predictions by PE SVMs were often more balanced (blue), possessing moderate to high accuracy for both sensitive and resistant patients, and consequently greater accuracy as a whole

FIGURE 3

Connection network for pathway‐extended TKI SVMs. Schematic relationships outlining the pathway connections for the best‐performing PE model for each drug in panels (A) sorafenib, (B) lapatinib, (C) gefitinib, (D) sunitinib, (E) imatinib and (F) erlotinib. All symbols indicated are gene names. The erlotinib model was highly interconnected and is represented as a table. Genes in red are features selected for the final PE‐Sor gene signature, whereas genes coloured green were chosen in a separate PE gene signature with comparable performance. Genes in black were not part of the final signature themselves but correlated with efficacy to the TKI of interest by MFA and expanded the gene pool through biochemical connections they possessed to one‐node or two‐node distant genes

FIGURE 4

Effect of removal of individual genes from signature on overall accuracy using patient tumour data. The patient classification accuracy and MCC of the strongest performing PE models are altered upon the removal of each component gene listed. These PE TKI gene signatures are (A) sorafenib [PE‐Sor], (B) lapatinib [PE‐Lap], (C) gefitinib [PE‐Gef], (D) sunitinib [PE‐Sun], (E) imatinib [PE‐Ima] and (F) erlotinib [PE‐Erl]. Blue and red bars denote the overall accuracy and MCC of the model after gene removal, respectively

FIGURE 5

Schematic of the pathway‐extended genes in the sorafenib model PE‐Sor. The best performing sorafenib model PE‐Sor is a nine‐gene model consists of a single curated gene (TGFB1) and eight genes selected by pathway extension (ELF5, RBP5, GC, PRKD2, SCNN1A, COL25A1, DACT1 and LHX8). This cell schematic provides context of the cellular mechanisms of action and/or known relationships between genes with a documented impact on sorafenib activity (‘curated’ genes; black borders) and those genes selected by pathway extension (purple borders). Genes with grey borders are neither curated nor pathway‐extended genes and are simply present to give context between genes and their known cellular functions. Thicker borders specify those genes in the PE‐Sor model, whereas gene colour coding indicates how GE and/or CN correlated to sorafenib GI₅₀ by MFA