A Controllability Reinforcement Learning Method for Pancreatic Cancer Biomarker Identification

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Pancreatic cancer is highly malignant with poor prognosis.
• Transcriptional data offers potential for identifying novel pancreatic cancer biomarkers.
• Existing network-based biomarker discovery methods have limitations, such as not analyzing RNA or relying on mutation data.

Purpose of the Study:

• To propose a novel method, RDDriver, for identifying pancreatic cancer biomarkers.
• To leverage multi-layer heterogeneous transcriptional regulation networks for biomarker discovery.
• To overcome limitations of existing methods by incorporating RNA data and network controllability.

Main Methods:

• Constructed a regulation network including long non-coding RNA, microRNA, and messenger RNA.
• Employed Relational Graph Convolutional Network (RGCN) for node representation learning.
• Utilized Deep Q-Network (DQN) principles and the Popov-Belevitch-Hautus criterion for RNA scoring and prioritization.

Main Results:

• RDDriver was trained on simulated networks and applied to regulation networks.
• The method demonstrated effectiveness in identifying potential cancer driver RNAs.
• Comparative experiments showed RDDriver's performance against eight other methods.

Conclusions:

• RDDriver presents an effective approach for pancreatic cancer biomarker discovery.
• The method's integration of RGCN and DQN advances network-based biomarker identification.
• This study highlights the potential of multi-layer transcriptional networks in cancer research.