TreeQNet: a webserver for Treatment evaluation with Quantified Network

Zhenlei Li, Ya Huang, Qingrun Li, Yidi Sun, Chen Li, Jiarui Wu, Haoran Zheng, Rong ZengBMC Bioinformatics 2022 · DOI

Introduction

Personalized medicine aims to match the right drug to the right patient, but cancer's inherent heterogeneity often complicates this goal. Traditional biomarker discovery typically focuses on "nodes"—the expression levels of individual genes or proteins. However, biological functions are emergent properties of complex interaction networks. Single-molecule biomarkers can be inconsistent across patients due to genetic backgrounds or stochastic cellular processes. TreeQNet addresses this by focusing on "edges"—the functional relationships and signaling "chatter" between molecules, such as kinases and their substrates. These interactions are often more stable and informative indicators of therapeutic response.

Methods and Architecture

TreeQNet is a web-based platform developed using a Model-View-ViewModel (MVVM) architecture. This structure ensures a responsive user experience while handling intensive backend computations.

TreeQNet System Architecture: The platform integrates a Data Layer for reference and uploaded omics data, a Model Layer for algorithmic processing, and a View Layer for interactive visualization of results.

The core innovation of TreeQNet lies in its multi-step algorithmic pipeline:

  1. Edge Transformation: Instead of raw expression, the system uses the EdgeBiomarker algorithm to calculate "edge strength" between kinases and substrates, integrating both joint expression and phosphorylation states.
  2. Feature Selection: The model identifies differentially correlated gene pairs (DCPs) using Pearson correlation, narrowing thousands of potential interactions down to the most predictive features.
  3. Elastic Net Regression: To process high-dimensional omics data (e.g., 1,696 edge features), the server employs Elastic Net regression. This statistical method combines LASSO and Ridge regression penalties to select critical features while managing highly correlated variables.
Elastic Net Prediction Model: The model utilizes kinase-substrate edge features as inputs to an Elastic Net regression framework to predict drug response, measured as Tumor Cell Growth Inhibition (TCGI%).

Results and Validation

The researchers validated TreeQNet using a reference database of 31 colorectal cancer (CRC) samples and a validation set of 13 CRC miniPDX (patient-derived xenograft) models. The platform's ability to predict Tumor Cell Growth Inhibition (TCGI) was tested for three targeted therapies: Afatinib, Gefitinib, and Regorafenib.

The results demonstrated a strong correlation between the model's predictions and actual observed responses in the miniPDX models:

  • Afatinib: $R^2 = 0.75$ ($p = 0.002$)
  • Gefitinib: $R^2 = 0.61$ ($p = 0.026$)
  • Regorafenib: $R^2 = 0.75$ ($p = 0.003$)

While experimental drug responses often show high variance, the TreeQNet model successfully "smoothed" the influence of outliers, providing a more reliable and narrower scale of predicted inhibition that closely tracked the ground truth.

Discussion

TreeQNet shifts the paradigm from abundance-based biomarkers to interaction-based network analysis. By leveraging phosphoproteomics, the tool provides a functional lens into the active signaling "wiring" of a tumor. This approach democratizes complex bioinformatics; clinical researchers can upload standard protein data files and receive visual predictions of drug efficacy without specialized coding knowledge.

Current limitations include a relatively small sample size and a focus on colorectal cancer with a limited drug library. Future updates aim to expand the platform to include more tumor types and a broader range of therapeutic models, further enhancing its utility in precision oncology.

Summary updated

BibTeX
@article{li2022treeqnet,
  author = {Zhenlei Li and Ya Huang and Qingrun Li and Yidi Sun and Chen Li and Jiarui Wu and Haoran Zheng and Rong Zeng},
  title = {{TreeQNet: a webserver for Treatment evaluation with Quantified Network}},
  journal = {BMC Bioinformatics},
  year = {2022},
  doi = {10.1186/s12859-022-05024-y},
}
Figure