Source-linked AI summary
Parenclitic network analysis of methylation data for cancer identification
Alexander Karsakov, Thomas Bartlett, Iosif Meyerov, Alexey Zaikin, Mikhail Ivanchenko
TL;DR
The large number of CpG methylation measurements makes aggregate approaches useful for studying cancer-related methylation signatures. This paper represents methylation data as subject-specific parenclitic networks and classifies cancer status using network measures. Classifiers achieve 93–99% accuracy for almost all cancer types, while cancer-positive networks generally deviate from power-law degree distributions.
Problem
DNA methylation changes accompany early carcinogenesis, but the large number of CpG sites motivates aggregate measures that reduce dimensionality while retaining cancer-related genomic effects.
Source: Introduction · p. 1Method
The approach builds gene-node networks from methylation data using deviations from healthy-group relationships and classifies cancer status with 12 topological indices.
Source: Construction of networks · p. 2, Results and discussion · p. 6Results
Classifiers using network measures achieved 93–99% accuracy for almost all cancer types, and cancer-positive networks generally deviated from power-law degree scaling.
Source: Results and discussion · pp. 7–8Takeaways & Limitations
Reducing the classifier to 12 topological features did not substantially reduce accuracy relative to using 15,295 gene methylation levels and can help avoid overfitting.
Source: Conclusions · p. 8The limited number of healthy samples meant AUC could be calculated only in rare cases.
Source: Results and discussion · p. 7
Abstract
from arXiv · showhide
We make use of ideas from the theory of complex networks to implement a machine learning classification of human DNA methylation data, that carry signatures of cancer development. The data were obtained from patients with various kinds of cancers and represented as parenclictic networks, wherein nodes correspond to genes, and edges are weighted according to pairwise variation from control group subjects. We demonstrate that for the $10$ types of cancer under study, it is possible to obtain a high performance of binary classification between cancer-positive and negative samples based on network measures. Remarkably, an accuracy as high as $93-99\%$ is achieved with only $12$ network topology indices, in a dramatic reduction of complexity from the original $15295$ gene methylation levels. Moreover, it was found that the parenclictic networks are scale-free in cancer-negative subjects, and deviate from the power-law node degree distribution in cancer. The node centrality ranking and arising modular structure could provide insights into the systems biology of cancer.
Introduction
The paper uses parenclitic networks to classify cancer from DNA methylation while reducing the complexity of measurements. Its results indicate high classification accuracy and cancer-associated changes in network degree distributions.
Introduction
- Epigenetic modifications, including DNA methylation, can be environmentally modulated and persist through subsequent mitoses.The passage also notes that epigenetics mediates environmental effects on the genome and leaves a backdoor for environmental risk factors.Source: Introduction · p. 1
- DNA methylation changes accompany early human carcinogenesis and may provide quantitative signatures of illness or its risk.The large number of assessed CpG sites motivates aggregate measures that reduce dimensionality while retaining key cancer-related genomic effects.Source: Introduction · p. 1
- Earlier work linked gene-specific methylation measures to cancer-related changes and proposed gene-node networks whose edges quantify covariation with healthy profiles.Some prognostic edges were associated with patient survival, and their modules corresponded to biological functions.Source: Introduction · p. 1
- Parenclitic analysis connects each pair of measurement nodes when a subject deviates significantly from the healthy group's linear regression model.The resulting subject-specific network is expected to have different properties in health and disease.Source: Introduction · p. 2
- Approximately 93–99% accuracy was achieved using routine measures of parenclitic networks built from DNA methylation data.Healthy-tissue networks had power-law degree-distribution tails, while cancer networks deviated from power-law scaling; one exception had about 90% classifier performance.Source: Introduction · p. 2
Methods
The study aggregates gene-level methylation measurements into parenclitic networks, using healthy controls to define pairwise relationships and weighting deviations in cancer-positive and negative samples. It then characterizes the weighted graphs with network topology measures.
Data
- Methylation data from ten healthy and tumour tissue datasets were processed to retain 15,295 genes, with gene-wide averages used to represent methylation levels.Probes with non-unique mappings, SNP mappings, sex-chromosome mappings, or insufficient coverage were removed; missing values were imputed by KNN.Source: Data · p. 2
Construction of networks
- For each gene pair, the method fits a control-group linear regression and builds a complete weighted graph for each non-control sample, weighting edges by deviation from the model.The regression-error standard deviation scales the mismatch; for the ZFP106–TRIM9 example, Mahalanobis distance addresses cases where ordinary mismatch weights are similar despite visibly distinct clusters.Source: Construction of networks · p. 3, Figure 2 · p. 5
- For their data, the approach better captures abnormal methylation modifications in gene pairs and improves sample-classification accuracy by 1–3%.Source: Construction of networks · p. 4
Metrics of network topology
- The analysis characterizes weighted graphs with measures including node degree, shortest-path distance, diameter, degree centrality, graph efficiency, and betweenness centrality.Degree is the sum of incident edge weights; distance sums edge weights along the shortest path, and diameter is the maximum distance between vertices.Source: Metrics of network topology · pp. 4–5
Results and discussion
Network-based classifiers achieved 93–99% accuracy across almost all cancer types, while network degree distributions distinguished cancer-negative samples from most cancer-positive samples. The results also show that the compact network-feature approach generally retained performance relative to gene-level methylation data, with PRAD as a notable exception.
Results and discussion
- For BRCA examples, cancer networks typically contained star-like subgraphs around abnormally methylated genes, while healthy-subject networks were more homogeneous.Figure 3 displays only the 1,000 strongest edges and their incident nodes; its caption notes pronounced modular structure in the cancer network.Source: Results and discussion · p. 6, Figure 3 · p. 6
- The binary classifiers used 12 topological indices with SVM and RF, and evaluation used a two-step procedure followed by 10-fold cross-validation.AUC could be calculated only rarely because the available healthy-subject data were limited.Source: Results and discussion · pp. 6–7
- 93–99% accuracy was achieved for classifiers trained on network measures across almost all cancer types, with no considerable dependence on algorithm in most cases.The reported RF dissimilarity was noticeable for BRCA, PRAD, and THCA.Source: Table 2 · p. 7, Results and discussion · p. 7, Table 3 · p. 8
- Network measures showed no systematic weakness against the 15,295-gene average methylation baseline, and their smaller feature set prevented overfitting.In some cases, network measures enhanced classification results.Source: Results and discussion · p. 7
- Cancer-negative networks were scale-free, whereas cancer-positive networks deviated from power-law scaling except in PRAD, where classification accuracy was about 87–90%.The authors suggest prioritizing changes to network construction over adding other network measures.Source: Results and discussion · pp. 7–8
Conclusions
The parenclitic-network approach supports cancer classification with 12 global measures, with Random Forest accuracy comparable to using 15,295 genes’ average methylation levels. Network topology and centrality rankings may also offer insight into cancer biology.
Conclusions
- Twelve global network measures provide a basis for binary cancer-positive versus cancer-negative classification using Support Vector Machine and Random Forest algorithms.For almost all cancer types, the two algorithms perform much the same.Source: Conclusions · p. 8
- 90−99% accuracy was reached across cancer types with Random Forest on network measures, with no substantial difference from using average methylation levels of 15,295 genes.The authors note that reducing the features to 12 topological indices avoids overfitting that could arise with the relatively small sample count.Source: Conclusions · p. 8
- High-centrality nodes are candidates for significant roles in cancer development, while the observed modular structure’s correspondence to biological functions remains an open question.The authors also state that strong deviations from scale-free degree distributions in almost all cancer types mark global topology changes that remain to be understood.Source: Conclusions · p. 8