Detection for disease tipping points by landscape dynamic network biomarkers

Xiaoping Liu, Xiao Chang, Siyang Leng, Hui Tang, Kazuyuki Aihara, Luonan ChenNational 2018 · DOI

TL;DR

This paper introduces a novel, model-free method called landscape dynamic network biomarkers (l-DNB) that can identify "tipping points" in disease progression using only a single patient's omics data. This allows for early detection of imminent disease deterioration and the identification of critical genes (biomarkers) that drive these transitions, with applications in predicting influenza severity and cancer prognosis.

What Problem Does This Paper Address?

Diseases often progress through gradual changes that eventually lead to a drastic, qualitative shift from a healthy state to a diseased one. This transition point, known as a "tipping point" or "critical state," is crucial because it's the last stage where intervention might reverse the disease. However, identifying this critical state is challenging because the molecular profiles of normal and critical states are often very similar, while the diseased state is markedly different. Traditional biomarkers, which rely on differences between healthy and diseased states, fail to capture these subtle critical transitions. Existing dynamic network biomarker (DNB) methods, while theoretically sound, often require multiple samples, which are not always available for individual patients in clinical settings. This paper aims to bridge this gap by developing a method that can reliably detect these critical states and provide early warnings using single-sample omics data.

Methods

The core innovation is the landscape dynamic network biomarker (l-DNB) methodology. It builds upon the concept of DNBs, which are groups of molecules that signal an impending transition. The l-DNB approach works as follows:

  1. Single-Sample Network (SSN) Construction: For a given patient's omics data, a network is built where nodes are genes and edges represent their correlated expression. This network is "single-sample" because it's specific to that individual's data.
  2. Local DNB Score Calculation: For each gene, a "local module" is defined (the gene itself and its immediate neighbors in the SSN). A score, Is(x), is calculated for this module based on three criteria derived from DNB theory:
  3. Single-sample Expression Deviation (sED): How much a gene's expression deviates from the average in reference samples.
  4. Single-sample Pearson Correlation Coefficient (sPCCin): The average correlation among genes within the local module.
  5. Single-sample Pearson Correlation Coefficient (sPCCout): The average correlation between genes within the module and genes outside the module. These scores quantify the "criticality" of each gene's local network.

  6. Global DNB Score and Biomarker Identification: The local DNB scores are used to create a "landscape." The top-k genes with the highest local DNB scores are identified as the individual DNB members for that sample. The global DNB score (IDNB) is an average of these top-k local scores. A higher IDNB suggests the patient is closer to a critical state.

This approach is "model-free" and doesn't require pre-defined models or extensive learning from large datasets, making it adaptable to individual patients.

Datasets

The l-DNB method was validated using the following datasets:

  • Influenza Virus Infection: A time-series dataset from 17 healthy human volunteers inoculated with H3N2 influenza virus, with gene expression profiles measured at 16 time points. This dataset was used to test early-warning signals for severe influenza symptoms.
  • Tumor Datasets (TCGA): Three cancer datasets from The Cancer Genome Atlas (TCGA) were used:
  • Lung Adenocarcinoma (LUAD)
    • Kidney Renal Clear Cell Carcinoma (KIRC)
    • Thyroid Carcinoma (THCA)

These datasets were used to identify critical stages of tumor deterioration and to evaluate the prognostic power of identified biomarkers.

Key Insights and Ideas

  • Single-Sample Biomarkers: The l-DNB method can identify individual biomarkers specific to a single patient, moving beyond population-level biomarkers. This is crucial for personalized medicine.
  • "Dark Matter" Genes: The method can identify genes that are critical for disease progression but do not show differential expression between healthy and diseased states. These "dark matter" genes are often missed by traditional methods.
  • Predicting Prognosis: Beyond detecting critical states, the identified DNB members can be categorized into "pessimistic" (indicating poor prognosis) and "optimistic" (indicating good prognosis) biomarkers, offering valuable insights into patient outcomes.
  • Model-Free Advantage: The method's model-free nature makes it robust and adaptable, avoiding the need for large, labeled datasets for training, which is a significant advantage in clinical settings.

Key Results

  • Influenza Early Warning: The l-DNB method accurately predicted severe influenza symptoms in symptomatic subjects at least 8 hours before the appearance of symptoms, while asymptomatic subjects showed stable, low scores.
  • Cancer Critical Stages: The method identified critical stages for LUAD (stage IIB), THCA (stage III), and KIRC (stage II), preceding significant tumor deterioration.
  • Prognostic Biomarkers: For LUAD, THCA, and KIRC, the identified pessimistic and optimistic biomarkers significantly correlated with patient survival times, demonstrating their predictive power. For instance, LUAD patients with pessimistic biomarkers had significantly shorter survival times.
  • Robustness: The method showed robustness across different cancer types and was validated using reference samples, indicating its general applicability.

Limitations and Future Work

  • Reference Sample Size: While the method is robust, the paper notes that small reference sample sizes (e.g., <10) can significantly impact rankings.
  • Common Biomarkers: Identifying common biomarkers or a universal criticality threshold across an entire disease population might require larger datasets.
  • Future Directions: The authors suggest that the l-DNB methodology has great potential for direct application in preventive and personalized medicine, and for analyzing molecular mechanisms underlying disease progression and cellular processes like differentiation and proliferation. They also highlight its applicability to other non-linear biological processes.

Summary updated

Abstract

A new model-free method has been developed and termed the landscape dynamic network biomarker (<i>l</i>-DNB) methodology. The method is based on bifurcation theory, which can identify tipping points prior to serious disease deterioration using only single-sample omics data. Here, we show that <i>l</i>-DNB provides early-warning signals of disease deterioration on a single-sample basis and also detects critical genes or network biomarkers (i.e. DNB members) that promote the transition from normal to disease states. As a case study, <i>l</i>-DNB was used to predict severe influenza symptoms prior to the actual symptomatic appearance in influenza virus infections. The <i>l</i>-DNB approach was then also applied to three tumor disease datasets from the TCGA and was used to detect critical stages prior to tumor deterioration using an individual DNB for each patient. The individual DNBs were further used as individual biomarkers in the analysis of physiological data, which led to the identification of two biomarker types that were surprisingly effective in predicting the prognosis of tumors. The biomarkers can be considered as common biomarkers for cancer, wherein one indicates a poor prognosis and the other indicates a good prognosis.
BibTeX
@inproceedings{liu2018detection,
  author = {Xiaoping Liu and Xiao Chang and Siyang Leng and Hui Tang and Kazuyuki Aihara and Luonan Chen},
  title = {{Detection for disease tipping points by landscape dynamic network biomarkers}},
  booktitle = {National},
  year = {2018},
  doi = {10.1093/nsr/nwy162},
}
Figure