Limits...
Gene network inference by fusing data from diverse distributions.

Žitnik M, Zupan B - Bioinformatics (2015)

Bottom Line: In a simulation study, we demonstrate good predictive performance of FuseNet in comparison to several popular graphical models.Fusion of datasets offers substantial gains relative to inference of separate networks for each dataset.Our results demonstrate that network inference methods for non-Gaussian data can help in accurate modeling of the data generated by emergent high-throughput technologies.

View Article: PubMed Central - PubMed

Affiliation: Faculty of Computer and Information Science, University of Ljubljana, Ljubljana, Slovenia and Department of Molecular and Human Genetics, Baylor College of Medicine, Houston, TX, USA.

Show MeSH

Related in: MedlinePlus

Application of gene network inference algorithms to Poisson-distributed simulated data. Simulation studies on four network types were performed: random (see Supplementary Fig. S4), hub, scale-free and small world. These graph structures appear in many real biological networks. For each graph type, we generated data with n = 200 observations with P = 100 variables (nodes) at a low (first row) and high (second row) signal-to-noise ratio (SNR). Receiver operating curves and boxplots are shown for Poisson FuseNet (proposed here), the Local Poisson Graphical Model (LPGM) (Allen and Liu, 2013), the Graphical Lasso (GLASSO) (Friedman et al., 2007), the GLASSO on log-transformed data (Log-GLASSO) (e.g. cf. Gallopin et al., 2013) and the GLASSO on data transformed through nonparanormal Gaussian copula (NPN-Copula) (Liu et al., 2009)
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4542780&req=5

btv258-F3: Application of gene network inference algorithms to Poisson-distributed simulated data. Simulation studies on four network types were performed: random (see Supplementary Fig. S4), hub, scale-free and small world. These graph structures appear in many real biological networks. For each graph type, we generated data with n = 200 observations with P = 100 variables (nodes) at a low (first row) and high (second row) signal-to-noise ratio (SNR). Receiver operating curves and boxplots are shown for Poisson FuseNet (proposed here), the Local Poisson Graphical Model (LPGM) (Allen and Liu, 2013), the Graphical Lasso (GLASSO) (Friedman et al., 2007), the GLASSO on log-transformed data (Log-GLASSO) (e.g. cf. Gallopin et al., 2013) and the GLASSO on data transformed through nonparanormal Gaussian copula (NPN-Copula) (Liu et al., 2009)

Mentions: We simulated four network types, which are known to resemble the structure of real biological networks (Allen and Liu, 2013; Costanzo et al., 2010). We report receiver operator curves computed by varying the regularization parameter λ in Figure 3 and Supplementary Figure S4, boxplots of true and false positive rates for fixed λ as determined by stability selection in Figure 3, Supplementary Figures S2 and S4. Further, we evaluated precision and recall of the networks estimated from different data distributions in Supplementary Figures S2–S5.


Gene network inference by fusing data from diverse distributions.

Žitnik M, Zupan B - Bioinformatics (2015)

Application of gene network inference algorithms to Poisson-distributed simulated data. Simulation studies on four network types were performed: random (see Supplementary Fig. S4), hub, scale-free and small world. These graph structures appear in many real biological networks. For each graph type, we generated data with n = 200 observations with P = 100 variables (nodes) at a low (first row) and high (second row) signal-to-noise ratio (SNR). Receiver operating curves and boxplots are shown for Poisson FuseNet (proposed here), the Local Poisson Graphical Model (LPGM) (Allen and Liu, 2013), the Graphical Lasso (GLASSO) (Friedman et al., 2007), the GLASSO on log-transformed data (Log-GLASSO) (e.g. cf. Gallopin et al., 2013) and the GLASSO on data transformed through nonparanormal Gaussian copula (NPN-Copula) (Liu et al., 2009)
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4542780&req=5

btv258-F3: Application of gene network inference algorithms to Poisson-distributed simulated data. Simulation studies on four network types were performed: random (see Supplementary Fig. S4), hub, scale-free and small world. These graph structures appear in many real biological networks. For each graph type, we generated data with n = 200 observations with P = 100 variables (nodes) at a low (first row) and high (second row) signal-to-noise ratio (SNR). Receiver operating curves and boxplots are shown for Poisson FuseNet (proposed here), the Local Poisson Graphical Model (LPGM) (Allen and Liu, 2013), the Graphical Lasso (GLASSO) (Friedman et al., 2007), the GLASSO on log-transformed data (Log-GLASSO) (e.g. cf. Gallopin et al., 2013) and the GLASSO on data transformed through nonparanormal Gaussian copula (NPN-Copula) (Liu et al., 2009)
Mentions: We simulated four network types, which are known to resemble the structure of real biological networks (Allen and Liu, 2013; Costanzo et al., 2010). We report receiver operator curves computed by varying the regularization parameter λ in Figure 3 and Supplementary Figure S4, boxplots of true and false positive rates for fixed λ as determined by stability selection in Figure 3, Supplementary Figures S2 and S4. Further, we evaluated precision and recall of the networks estimated from different data distributions in Supplementary Figures S2–S5.

Bottom Line: In a simulation study, we demonstrate good predictive performance of FuseNet in comparison to several popular graphical models.Fusion of datasets offers substantial gains relative to inference of separate networks for each dataset.Our results demonstrate that network inference methods for non-Gaussian data can help in accurate modeling of the data generated by emergent high-throughput technologies.

View Article: PubMed Central - PubMed

Affiliation: Faculty of Computer and Information Science, University of Ljubljana, Ljubljana, Slovenia and Department of Molecular and Human Genetics, Baylor College of Medicine, Houston, TX, USA.

Show MeSH
Related in: MedlinePlus