Biomarker discovery by integrated joint non-negative matrix factorization and pathway signature analyses

Naoya Fujita,Katsuhiko Murakami,Kenta Nakai,Shinji Mizuarai

doi:10.1038/s41598-018-28066-w

Naoya Fujita, Katsuhiko Murakami + Show 2 more

Open Access

PDF Available

https://doi.org/10.1038/s41598-018-28066-w

Copy DOI

Export

Save

Cite

Abstract
Highlights/Summary
Full-Text PDF
Similar Papers

Abstract

Listen

Predictive biomarkers are important for selecting appropriate patients for particular treatments. Comprehensive genomic, transcriptomic, and pharmacological data provide clues for understanding relationships between biomarkers and drugs. However, it is still difficult to mine biologically meaningful biomarkers from multi-omics data. Here, we developed an approach for mining multi-omics cell line data by integrating joint non-negative matrix factorization (JNMF) and pathway signature analyses to identify candidate biomarkers. The JNMF detected known associations between biomarkers and drugs such as BRAF mutation with PLX4720 and HER2 amplification with lapatinib. Furthermore, we observed that tumours with both BRAF mutation and MITF activation were more sensitive to BRAF inhibitors compared to tumours with BRAF mutation without MITF activation. Therefore, activation of the BRAF/MITF axis seems to be a more appropriate biomarker for predicting the efficacy of a BRAF inhibitor than the conventional biomarker of BRAF mutation alone. Our biomarker discovery scheme represents an integration of JNMF multi-omics clustering and multi-layer interpretation based on pathway gene signature analyses. This approach is also expected to be useful for establishing drug development strategies, identifying pharmacodynamic biomarkers, in mode of action analysis, as well as for mining drug response data in a clinical setting.

Highlights

Precision medicine for cancer patients with molecular targeted drugs and predictive biomarkers is expected to lead to a paradigm shift from one-size-fits-all medicine to patient-specific medicine[1]
Comprehensive genomic and pharmacological data of large collections of cancer cell lines have been published as the Cancer Cell Line Encyclopedia (CCLE)[3,4]
To illustrate the robustness of our multi-omics clustering method against missing values, Joint non-negative matrix factorization (JNMF) was first applied to simulated data

Summary

Introduction

Precision medicine for cancer patients with molecular targeted drugs and predictive biomarkers is expected to lead to a paradigm shift from one-size-fits-all medicine to patient-specific medicine[1]. There is some discordance between databases, especially in terms of the compound sensitivity profiles, these databases generally show reasonable consistency[5,6] These multi-dimensional genomic and pharmacological datasets have been used to perform multi-omics analyses with the goal of understanding the relationships between cancer genomes and drug responders. The top-performing method was found to be a kernel method with multiview and multitask learning, which uses all of the genetic profiles provided[7] This challenge is focused on providing a benchmarked set of algorithms, it is difficult to translate the results obtained from the predictors for clinical application. We sought to resolve these issues to facilitate the use of multi-omics analyses for understanding relationships between the cancer genome and drug responders through development of comprehensive prediction models with multi-genetic features. Among the many techniques available to handle multiple inputs[14], JNMF is theoretically and practically equivalent to a standard NMF method with concatenated inputs

Methods

Results

Conclusion