Abstract
Graphical models are widely used to study complex multivariate biological systems. Network inference algorithms aim to reverse-engineer such models from noisy experimental data. It is common to assess such algorithms using techniques from classifier analysis. These metrics, based on ability to correctly infer individual edges, possess a number of appealing features including invariance to rank-preserving transformation. However, regulation in biological systems occurs on multiple scales and existing metrics do not take into account the correctness of higher-order network structure. In this paper novel performance scores are presented that share the appealing properties of existing scores, whilst capturing ability to uncover regulation on multiple scales. Theoretical results confirm that performance of a network inference algorithm depends crucially on the scale at which inferences are to be made; in particular strong local performance does not guarantee accurate reconstruction of higher-order topology. Applying these scores to a large corpus of data from the DREAM5 challenge, we undertake a data-driven assessment of estimator performance. We find that the “wisdom of crowds” network, that demonstrated superior local performance in the DREAM5 challenge, is also among the best performing methodologies for inference of regulation on multiple length scales.
Acknowledgments
The authors are grateful to the Associate Editor and three anonymous reviewers whose comments greatly improved this manuscript. In addition the authors thank Gustavo Stolovitzky and Sach Mukherjee for their feedback at an earlier stage of this research. This work was supported by the Centre for Research in Statistical Methodology grant (EPSRC EP/D002060/1). CJO was involved in designing the DREAM8 Challenge. CJO, RA, SEFS were not involved in the DREAM5 Challenge.
Appendix A: Proofs
Proof of Theorem 1.S-optimality of the benchmark network G follows immediately from the definition of each score. To prove nonuniqueness we begin by considering MSS1. For this, take G to be the network A→B→C→A and take G′ to be the network A←B←C←A; then both networks entail the same descendancy relationships. It follows from (D3; consistency) that both G and G′ maximise
Proof of Theorem 2. (a) Given ε>0. We proceed by constructing a sequence of pairs (G, Ĝ) indexed by p, the number of vertices, such that the scores associated with Ĝ end simultaneously to the required limits as p→∞. Define a data-generating network G on p vertices by the edge set
and consider an unweighted network Ĝ with topology E(Ĝ)=E(G)∪{(p, 1)}. i.e., Ĝ differs to G in the addition of a single edge (p, 1). The ROC curve corresponding to Ĝ is defined by the three points
so that by linear interpolation the area under the curve is
Here T is the number of true examples; for local scores this is E(G), for MSS1 this is #{(i, j):G(i→j)≠∅, i≠j} and for MSS2 this is ∑i,jei, j. To compute the area under this curve we must use nonlinear interpolation (Goadrich et al., 2004). Specifically, interpolation between the two points A, B with true/false positive values (TPA, FPA), (TPB, FPB) respectively we create new points for each of TPA+1, TPA+2, …, TPB–1, increasing the false positives for each new point by (FPB–FPA)/(TPB–TPA). By direct calculation we have that the area under this interpolated curve is
Fix ε>0. For (i), since Ĝ differs to G by just a single edge we have TP=2(p–2), FP=1, T=2(p–2), TPR=1,
Thus
For (ii-iii) and MSS1 note that Ĝ(i→j)≠∅ for all i≠j, so that Ĝ predicts every possible descendancy. Thus TP=2 p–3, FP=p2–3p+3, T=2p–3, TPR=1, FPR=1,
Thus
leading to TP =2p, FP =p(p–2), T=2p, TPR =1, FPR =1,
Thus
Taking p>max{P1, P2, P3}, we have shown that (G, Ĝ) satisfy the conclusions of the theorem.
(b) The converse result is proved similarly, taking G to be the cycle 1→2→…→p→1 and Ĝ to be the reverse 1←2←…←p←1.□
Proof of Theorem 3. Let P=(p0, …, pm) be a path from p0 to pm. Define the intermediate vertices in the path to be the vertices p1, …, pm–1. For each pair of vertices i and j we wish to calculate τi, j, the largest threshold τ for which i and j are connected by a path in HΓ.
For k=0, …, p, let
To see this, note that
The conclusion of the Theorem follows by iterative application of Equation 20.□
Remarks on the proof of Theorem 3: To make the algorithm more memory efficient, we note that it is not necessary to store more than
For Theorem 4 we require a series of Lemmas:
Lemma 1The effects
are well-defined and satisfy 0≤eij≤1.
Proof. We show the sum in Equation 21 converges absolutely. Since all terms are non-negative, it is sufficient to prove that, for each i≠j such that H(i→j)≠∅, the sum is bounded above. Consider the partial sums
If Equation 21 diverges, so must
Suppose
since a path P of length m can be decomposed into a path from p0=i to pm–1=k and an edge from pm–1=k to pm=j. Therefore
from which we again conclude
Lemma 2Effects, as defined in Equation 21, satisfy (i) eij=δijfor all i, j s.t. H(i→j)=∅, (ii)
Proof of Lemma 2. The first statement follows immediately from Equation 21. For the second statement, consider two cases: Firstly, for i≠j such that i∉H(•, j) we have
Note that absolute convergence (Lemma 1) ensures the manipulation of infinite sums is valid. The proof for i∈H(•, j) is similar but requires slightly more care:
□
Lemma 3Define a matrixM(i) by (i)
Proof. Equivalence may be verified directly.□
Lemma 4SupposevsatisfiesM(i)v=v. Then for all k we must have∣vk∣≤vi∣.
Proof. Suppose not. Then there exists k such that (i) ∣vk∣>vi∣ and (ii) ∣vk∣ is maximal. Without loss of generality H(i→k)≠∅ since otherwise
Lemma 5There exists a unique solution toM(i)v=vwith vi=1.
Proof.Existence: The ith row of M(i) contains only zeros, except for the ith entry which is equal to 1. Hence the characteristic polynomial contains a root (eigenvalue) λ=1 with corresponding eigenvector
Uniqueness: Suppose v and v′ both satisfy the criteria. Then v″=v–v′ is also an eigenvector of M(i) with unit eigenvalue. From Lemma 4 we have
Proof of Theorem 4. Together Lemmas 1–5 prove that the effect matrix e={eij} exists and is unique. Moreover, the ith row ei of e may be computed as the unique eigenvector of M(i) corresponding to unit eigenvalue.
For moderate dimensional matrices M(i), eigenvectors may be calculated using any valid algorithm. One such algorithm, suitable for high dimensional problems, is the “power iteration” method. Specifically we compute the leading eigenvector v of M(i) as
Efficient computation in this setting is reviewed in Berkhin (2005).□
References
Banerjee, A. and J. Jost (2009): “Graph spectra as a systematic tool in computational biology,” Discrete Appl. Math., 157, 2425–2431.Search in Google Scholar
Berkhin, P. (2005): “A survey on pagerank computing,” Internet Mathematics, 2, 73–120.10.1080/15427951.2005.10129098Search in Google Scholar
Breitkreutz, D., L. Hlatky, E. Rietmanc and J. A. Tuszynski (2012): “Molecular signaling network complexity is correlated with cancer patient survivability,” Proc. Natl. Acad. Sci. USA 109, 9209–9212.10.1073/pnas.1201416109Search in Google Scholar PubMed PubMed Central
Cantone, I., L. Marucci, F. Iorio, M. A. Ricci, V. Belcastro, M. Bansal, S. Santini, M. di Bernardo, D. di Bernardo and M. P. Cosma (2009): “A yeast synthetic network for in vivo assessment of reverse-engineering and modeling approaches,” Cell 137, 172–181.10.1016/j.cell.2009.01.055Search in Google Scholar PubMed
Chuang, H. Y., E. Lee and T. Ideker (2007): “Network-based classification of breast cancer metastasis,” Mol. Syst. Biol. 3, 140.Search in Google Scholar
Dash, D. (2003): “Caveats for causal reasoning with equilibrium models,” PhD thesis, Intelligent Systems Program, University of Pittsburgh.Search in Google Scholar
Davis, J. and M. Goadrich (2006): “The relationship between Precision-Recall and ROC curves.” Proceedings of the 23rd International Conference on Machine Learning. ACM Press, New York, pp. 233–240.10.1145/1143844.1143874Search in Google Scholar
De Jongh, M. and M. J. Druzdzel (2009): “A comparison of structural distance measures for causal Bayesian network models.” In: Klopotek, M., Przepiorkowski, A., Wierzchon, S.T., Trojanowski, K. (Eds.), Recent Advances in Intelligent Information Systems, Academic Publishing House EXIT, Warsaw, pp. 443–456.Search in Google Scholar
Drummond, C. and R. C. Holte (2004): “What ROC curves can’t do (and cost curves can).” Proc. 1st Workshop ROC Analysis in AI, pp. 19–26.Search in Google Scholar
Fawcett, T. (2006): “An introduction to ROC analysis,” Pattern Recogn, 27, 861–874.10.1016/j.patrec.2005.10.010Search in Google Scholar
Feiglin, A., A. Hacohen, A. Sarusi, J. Fisher, R. Unger and Y. Ofran (2012): “Static network structure can be used to model the phenotypic effects of perturbations in regulatory networks,” Bioinformatics, 28, 2811–2818.10.1093/bioinformatics/bts517Search in Google Scholar PubMed
Feizi, S., D. Marbach, M. Médard and M. Kellis (2013): “Network deconvolution as a general method to distinguish direct dependencies in networks,” Nat. Biotechnol., 31, 726–733.Search in Google Scholar
Filosi, M., R. Visintainer, S. Riccadonna, G. Jurman and C. Furlanello (2014): “Stability indicators in network reconstruction,” PloS One, 9, e89815.10.1371/journal.pone.0089815Search in Google Scholar PubMed PubMed Central
Goadrich, M., L. Oliphant and J. Shavlik (2004): “Learning ensembles of first-order clauses for recall-precision curves: A case study in biomedical information extraction.” Proceedings of the 14th International Conference on Inductive Logic Programming (ILP).10.1007/978-3-540-30109-7_11Search in Google Scholar
Heiser, L. M., A. Sadanandam, W. L. Kuo, S. C. Benz, T. C. Goldstein, S. Ng, W. J. Gibb, N. J. Wang, S. Ziyad, F. Tong, N. Bayani, Z, Hu, J. I. Billig, A. Dueregger, S. Lewis, L. Jakkula, J. E. Korkola, S. Durinck, F. Pepin, Y. Guan, E. Purdom, P. Neuvial, H. Bengtsson, K. W. Wood, P. G. Smith, L. T. Vassilev, B. T. Hennessy, J. Greshock, K. E. Bachman, M. A. Hardwicke, J. W. Park, L. J. Marton, D. M. Wolf, E. A. Collisson, R. M. Neve, G. B. Mills, T. P. Speed, H. S. Feiler, R. F. Wooster, D. Haussler, J. M. Stuart, J. W. Gray and P. T. Spellman (2012): “Subtype and pathway specific responses to anticancer compounds in breast cancer,” Proc. Natl. Acad. Sci. USA, 109, 2724–2729.10.1073/pnas.1018854108Search in Google Scholar PubMed PubMed Central
Hill, S. M., Y. Lu, J. Molina, L. M. Heiser, P. T. Spellman, T. P. Speed, J. W. Gray, G. B. Mills and S. Mukherjee (2012): “Bayesian inference of signaling network topology in a cancer cell line,” Bioinformatics, 28, 2804–2810.10.1093/bioinformatics/bts514Search in Google Scholar PubMed PubMed Central
Iwasaki, Y. and H. A. Simon (1994): “Causality and model abstraction,” Artif. Intell., 67, 143–194.Search in Google Scholar
Jeong, H., B. Tombor, R. Albert, Z. N. Oltvai and A. -L. Barabási (2000): “The large-scale organization of metabolic networks,” Nature, 407, 651–654.10.1038/35036627Search in Google Scholar PubMed
Johannes, M., J. C. Brase, H. Fröhlich, S. Gade, M. Gehrmann, M. Fälth, H. Sültmann and T. Beissbarth (2010): “Integration of pathway knowledge into a reweighted recursive feature elimination approach for risk stratification of cancer patients,” Bioinformatics, 26, 2136–2144.10.1093/bioinformatics/btq345Search in Google Scholar PubMed
Jurman, G., S. Riccadonna, R. Visintainer and C. Furlanello (2011a): “Biological network comparison via Ipsen-Mikhailov distance,” arXiv, 1109.0220.Search in Google Scholar
Jurman, G., R. Visintainer and C. Furlanello (2011b): “An introduction to spectral distances in networks,” Fr. Art. In. 226, 227–234.Search in Google Scholar
Lakhina, A., K. Papagiannaki, M. Crovella, C. Diot, E. D. Kolaczyk and N. Taft (2004): “Structural analysis of network traffic flows.” Proceedings of International Conference on Measurement and Modeling of Computer Systems, ACM, New York, USA, pp. 61–72.10.1145/1005686.1005697Search in Google Scholar
Maathuis, M. H., D. Colombo, M. Kalisch and P. Bühlmann (2010): “Predicting causal effects in large-scale systems from observational data,” Nat. Methods, 7, 247–248.Search in Google Scholar
Marbach, D., T. Schaffter, C. Mattiussi and D. Floreano (2009): “Generating realistic in silico gene networks for performance assessment of reverse engineering methods,” J. Comp. Biol., 16, 229–239.Search in Google Scholar
Marbach, D., R. J. Prill, T. Schaffter, C. Mattiussi, D. Floreano and G. Stolovitzky (2010): “Revealing strengths and weaknesses of methods for gene network inference,” Proc. Natl. Acad. Sci. USA, 107, 6286–6291.10.1073/pnas.0913357107Search in Google Scholar PubMed PubMed Central
Marbach, D., J. C. Costello, R. Küffner, N. M. Vega, R. J. Prill, D. M. Camacho, K. R. Allison, M. Kellis, J. J. Collins and G. Stolovitzky (2012): “Wisdom of crowds for robust gene network inference,” Nat. Methods, 9, 796–804.Search in Google Scholar
Milenković, T., F. Ioannis, L. Michael and N. Pržulj (2009): “Optimized null models for protein structure networks,” PLoS One, 4, e5967.10.1371/journal.pone.0005967Search in Google Scholar PubMed PubMed Central
Morrison, J. L., R. Breitling, D. J. Higham and D. R. Gilbert (2005): “GeneRank: using search engine technology for the analysis of microarray experiments,” BMC Bioinformatics, 6, 233.10.1186/1471-2105-6-233Search in Google Scholar PubMed PubMed Central
Nelander, S., W. Wang, B. Nilsson, Q. B. She, C. Pratilas, N. Rosen, P. Gennemark and C. Sander (2008): “Models from experiments: combinatorial drug perturbations of cancer cells,” Mol. Syst. Biol., 4, 216.Search in Google Scholar
Oates, C. and S. Mukherjee (2012): “Network inference and biological dynamics,” Ann. Appl. Stat., 6, 1209–1235.Search in Google Scholar
Oates, C.J., J. Korkola, J. W. Gray and S. Mukherjee (2014a): “Joint estimation of multiple related biological networks,” Ann. Appl. Stat., to appear.10.1214/14-AOAS761Search in Google Scholar
Oates, C. J., F. Dondelinger, N. Bayani, J. Korola, J. W. Gray and S. Mukherjee (2014b): “Causal network inference using biochemical kinetics,” Bioinformatics, to appear.10.1093/bioinformatics/btu452Search in Google Scholar PubMed PubMed Central
Page, L., S. Brin, R. Motwani and T. Winograd (1999): “The pagerank citation ranking: bringing order to the web,” Technical Report, Stanford InfoLab.Search in Google Scholar
Pearl, J. (2000): Causality: models, reasoning, and inference. Cambridge University Press, Cambridge, UK.Search in Google Scholar
Peters, J. and P. Bühlmann (2013): “Structural Intervention Distance (SID) for evaluating causal graphs,” arXiv, 1306.1043.Search in Google Scholar
Prill, R. J., D. Marbach, J. Saez-Rodriguez, P. K. Sorger, L. G. Alexopoulos, X. Xue, N. D. Clarke, G. Altan-Bonnet and G. Stolovitzky (2010): “Towards a rigorous assessment of systems biology models: The DREAM3 challenges,” PLoS One, 5, e9202.10.1371/journal.pone.0009202Search in Google Scholar PubMed PubMed Central
Prill, R. J., J. Saez-Rodriguez, L. G. Alexopoulos, P. K. Sorger and G. Stolovitzky (2011): “Crowdsourcing network inference: the DREAM predictive signaling network challenge,” Sci. Signal, 4(189), mr7.10.1126/scisignal.2002212Search in Google Scholar PubMed PubMed Central
Scutari, M. and R. Nagarajan (2013): “Identifying significant edges in graphical models of molecular networks,” Artif. Intell. Med., 57, 207–217.Search in Google Scholar
Shrivastava, A. and P. Li (2014): “A new space for comparing graphs,” arXiv, 1404.4644.Search in Google Scholar
Simon, N. and R. Tibshirani (2012): Comment on “Detecting novel associations in large datasets” by Reshef et al. Technical Report, Stanford, CA, USA.Search in Google Scholar
Städler, N. and S. Mukherjee (2014): “Multivariate gene-set testing based on graphical models,” Biostatistics, to appear.10.1093/biostatistics/kxu027Search in Google Scholar PubMed
Tong, A. H. Y., G. Lesage, G. D. Bade, H. Ding, H. Xu, X. Xin, J. Young, G. F. Berriz, R. L. Brost, M. Chang, Y. Q. Chen, X. Cheng, G. Chua, H. Friesen, D. S. Goldberg, J. Haynes, C. Humphries, G. He, S. Hussein, L. Ke, N. Krogan, Z. Li, J. N. Levinson, H. Lu, P. Ménard, C. Munyana, A. B. Parsons, O. Ryan, R. Tonikian, T. Roberts, A. M. Sdicu, J. Shapiro, B. Sheikh, B. Suter, S. L. Wong, L. V. Zhang, H. Zhu, C. G. Burd, S. Munro, C. Sander, J. Rine, J. Greenblatt, M. Peter, A. Bretscher, G. Bell, F. P. Roth, G. W. Brown, B. Andrews, H. Bussey and C. Boone (2004): “Global mapping of the yeast genetic interaction network,” Science, 303, 808–813.10.1126/science.1091317Search in Google Scholar PubMed
Wang, C., J. Xuan, L. Chen, P. Zhao, Y. Wang, R. Clarke and E. Hoffman (2008): “Motif-directed network component analysis for regulatory network inference,” BMC Bioinformatics, 9(Suppl. 1), S21.Search in Google Scholar
Warshall, S. (1962): “A theorem on Boolean matrices,” J. ACM., 9, 11–12.Search in Google Scholar
Weile, J., K. James, J. Hallinan, S. J. Cockell, P. Lord, A. Wipat and D. Wilkinson (2012): “Bayesian integration of networks without gold standards,” Bioinformatics, 28, 1495–1500.10.1093/bioinformatics/bts154Search in Google Scholar PubMed PubMed Central
Winter, C., G. Kristiansen, S. Kersting, J. Roy, D. Aust, T. Knösel, P. Rümmele, B. Jahnke, V. Hentrich, F. Rückert, M. Niedergethmann, W. Weichert, M. Bahra, H.J. Schlitt, U. Settmacher, H. Friess, M. Büchler, H.-D. Saeger, M. Schroeder, C. Pilarsky, R. Grützmann (2012): “Google goes cancer: improving outcome prediction for cancer patients by network-based ranking of marker genes,” PLoS Comput. Biol., 8, e1002511.Search in Google Scholar
Yates, P. D. and N. D. Mukhopadhyay (2013): “An inferential framework for biological network hypothesis tests,” BMC Bioinformatics, 14, 94.10.1186/1471-2105-14-94Search in Google Scholar PubMed PubMed Central
Supplemental Material
The online version of this article (DOI: 10.1515/sagmb-2014-0012) offers supplementary material, available to authorized users.
©2014 by De Gruyter