Individual-specific edge-network analysis for disease prediction

Abstract

Predicting pre-disease state or tipping point just before irreversible deterioration of health is a difficult task. Edge-network analysis (ENA) with dynamic network biomarker (DNB) theory opens a new way to study this problem by exploring rich dynamical and high-dimensional information of omics data. Although theoretically ENA has the ability to identify the pre-disease state during the disease progression, it requires multiple samples for such prediction on each individual, which are generally not available in clinical practice, thus limiting its applications in personalized medicine. In this work to overcome this problem, we propose the individual-specific ENA (iENA) with DNB to identify the pre-disease state of each individual in a single-sample manner. In particular, iENA can identify individual-specific biomarkers for the disease prediction, in addition to the traditional disease diagnosis. To demonstrate the effectiveness, iENA was applied to the analysis on omics data of H3N2 cohorts and successfully detected early-warning signals of the influenza infection for each individual both on the occurred time and event in an accurate manner, which actually achieves the AUC larger than 0.9. iENA not only found the new individual-specific biomarkers but also recovered the common biomarkers of influenza infection reported from previous works. In addition, iENA also detected the critical stages of multiple cancers with significant edge-biomarkers, which were further validated by survival analysis on both TCGA data and other independent data.

Figure 1.

Concept of edge-network and iENA. (A) The comparison between new edge-network and conventional node-network. (B) The computational framework of iENA.

Figure 2.

The sample organization of dataset about influenza infection. The subjects are divided into two groups according to the clinical symptom chart based on the standardized symptom scoring: symptomatic (Sx) group with nine subjects (subjects S1,S5,S6,S7,S8,S10,S12,S13,S15) and asymptomatic (Asx) group with eight subjects (subjects S2,S3,S4,S9,S11,S14,S16,S17). The non-symptom samples (in grey) which have no significant clinical symptom, may have obvious changes in a network level; and we will identify edge-biomarkers for detecting early-warning signals before the time point of symptom samples (in red). The samples labeled in yellow colour indicates the time points predicted by our iENA based on dataset GSE52428, which is clearly earlier than the clinical symptom except one subject (S5). Actually, for the subject S5, the tipping point predicted by iENA coincides with the first symptom point at 45 h, but in this paper, we count it as an incorrect prediction.

Figure 3.

The consistence between our findings and previously reported markers in the study of influenza infection from dataset GSE52428. (A) The overlap between marker genes identified from Sx individuals and Asx individuals. (B) The overlap among iENA identified disease marker genes (i.e. Sx genes) and previously reported 22 marker genes and 50 marker genes. (C) The overlap ratio as the percentage of individual edge-biomarkers recovered from prior-known genes in each sample. (D) The overlap ratio as the percentage of individual node-biomarkers recovered from prior-known genes in each sample.

Figure 4.

The edge-biomarkers identified by iENA for dataset GSE52428. (A) Protein–protein associations on STRING network. (B) Regulatory associations on IPA network annotated as ‘Antimicrobial Response, Inflammatory Response’.

Figure 5.

Prediction performance of iENA for dataset GSE52428. (A) The predicted curve of early-warning signals or tipping points for all subjects by our edge-biomarkers. (B) The prediction accuracy as ROC curve. (C) The individual prediction curves. The red circle points the clinically diagnosed infection-time (symptom) for the corresponding subject, and the green star mark indicates the predicted infection-time (tipping point) by our prediction cut-off. Clearly, we correctly predicted all of the symptom cases except S5 before their clinical symptom. Actually, for the subject S5, the tipping point predicted by iENA coincides with the first clinical symptom point at 45 h, but in this paper, we count S5 as an incorrect prediction.

Figure 6.

Prediction results on another dataset of influenza symptom from GSE30550 for validation. (A) The detected marker genes (recovered genes) from GSE30550 analysis, and the Common Sx gene from GSE52428 analysis, and the 22-gene marker from our previous study. (B) The ROC of prediction. (C) The prediction curves of individuals. The red circle points the clinically diagnosed infection-time for the corresponding subject, and the green star mark indicates the predicted infection-time by our prediction cut-off. Clearly, we correctly predicted all of the symptom cases except S5 before the clinical symptom, and the accuracy of the prediction is about 90%. Although the tipping point of subject S5 predicted by iENA coincides with the first clinical symptom point at 45 h, we count S5 as an incorrect prediction in this paper.

Figure 7.

The results on TCGA cancer data. (A) The iENA analysis on breast cancer (BRCA), where stage ii is identified as the tipping point. (B) The survival-day comparison for different breast cancer stages. (C) The survival analysis by coding-genes of edge-biomarkers on TCGA BRCA data. (D) The survival analysis by coding-genes of edge-biomarkers on GEO independent data. (E) The iENA analysis on liver cancer (LIHC), where stage ii is identified as the tipping point. (F) The survival-day comparison for different liver cancer stages. (G) The survival analysis by coding-genes of edge-biomarkers on TCGA LIHC data. (H) The survival analysis by coding-genes of edge-biomarkers on GEO independent data. Here, the edge-biomarkers are the DNB members.