Novel machine learning methods for gene-network profiling and dynamic cancer micro-outcome prediction.
This project will use novel machine learning methods our group developed for network-based gene signature profiling and cellular-level dynamic cancer micro-outcome prediction, such as cell fraction, cell-cell interaction, and local brain networks. We propose to integrate multi-omics, neuroimaging, and flow cytometry data in identifying novel gene-network signatures. Different from the traditional single-gene differential expression (DE) studies, our approaches incorporate the connection structures of the genome (e.g., gene or CpG co-regulations and protein-protein interactions), which are usually ignored in the marginal DE analyses, into outcome prediction. Integrating inter-feature connections could help identify novel gene signatures with marginally weak DE effects but have strong co-regulation effects on other disease-triggering hub genes or pathways. Such findings may help better understand the mechanisms of cancer development and progression. Models, algorithms, and software developed from this project are directly applicable to pan-cancer studies. New gene network signatures identified will also enable more accurate prognostics and shed light on future gene therapy development.
We propose to implement the following specific Aims.
Aim-1. Develop machine/deep learning methods for high-dimensional network-based signature detections for different forms of outcomes.
Aim-2. Apply the methods to UK Biobank data for clinically meaningful scientific discoveries.
Aim-3. Implement and release software, and analytic pipelines.
The duration of this project is 3 years.