Revealing disease-associated pathways by network integration of untargeted metabolomics

Journal name:
Nature Methods
Volume:
13,
Pages:
770–776
Year published:
DOI:
doi:10.1038/nmeth.3940
Received
Accepted
Published online

Abstract

Uncovering the molecular context of dysregulated metabolites is crucial to understand pathogenic pathways. However, their system-level analysis has been limited owing to challenges in global metabolite identification. Most metabolite features detected by untargeted metabolomics carried out by liquid-chromatography-mass spectrometry cannot be uniquely identified without additional, time-consuming experiments. We report a network-based approach, prize-collecting Steiner forest algorithm for integrative analysis of untargeted metabolomics (PIUMet), that infers molecular pathways and components via integrative analysis of metabolite features, without requiring their identification. We demonstrated PIUMet by analyzing changes in metabolism of sphingolipids, fatty acids and steroids in a Huntington's disease model. Additionally, PIUMet enabled us to elucidate putative identities of altered metabolite features in diseased cells, and infer experimentally undetected, disease-associated metabolites and dysregulated proteins. Finally, we established PIUMet's ability for integrative analysis of untargeted metabolomics data with proteomics data, demonstrating that this approach elicits disease-associated metabolites and proteins that cannot be inferred by individual analysis of these data.

At a glance

Figures

  1. PIUMet identifies disease-associated pathways and hidden components from untargeted metabolomic data.
    Figure 1: PIUMet identifies disease-associated pathways and hidden components from untargeted metabolomic data.

    The input to PIUMet are metabolomic peaks that differ between disease and control samples (peaks 2, 3 and 5 in the shown example). PIUMet then searches an underlying database. The PPMI nodes are proteins (circle nodes) or metabolites (square nodes). These nodes are connected via edges representing physical interactions among proteins, as well as substrate-enzyme and product-enzyme associations of metabolic reactions. PIUMet output is an optimum subnetwork of PPMI that connects disease features. This network represents dysregulated metabolic pathways in diseased cells, and its components display hidden proteins and metabolites that had not been detected in experiments. Hidden metabolites directly connected to disease features represent the putative identity of these features. Additionally, the resulting nodes and edges are scored based on their robustness to uncertainty in the underlying database. PIUMet also accept other omic data such as proteomics as an optional input.

  2. PIUMet.
    Figure 2: PIUMet.

    (a) PIUMet embraces the ambiguous identity of disease features. It first identifies putative metabolites matching a feature based on mass. It then represents each feature as a node (m/z), which is connected to the matched metabolites (M1−5). PIUMet reduces the ambiguity in the assignment and scores each of these metabolites. (b) In an optimum subnetwork of PPMI that links disease features, blue squares connected to triangles represent the inferred metabolites corresponding to the features. These metabolites are connected by high-confidence protein-protein and protein-metabolites interactions. (c) An example of an undesirable result that is biased toward highly connected nodes. (d) A comparison of a subnetwork in a tree structure with a subnetwork that captures the complex topology of metabolic reactions. (e) An example of a network generated from randomly chosen mock data sets, in which the majority of input nodes remain separated, and a few are connected via a long path of protein-protein and protein-metabolite interactions.

  3. The PIUMet subnetwork showing altered sphingolipid metabolism in the STHdh cell line model of HD.
    Figure 3: The PIUMet subnetwork showing altered sphingolipid metabolism in the STHdh cell line model of HD.

    PIUMet connects disease features via high-probability protein-protein and protein-metabolite interactions. Metabolites connected to disease features represent their putative identities. These metabolites along with the rest of the nodes are ranked based on the robustness scores, and are shown by different sizes. We experimentally verified the dysregulation of sphingolipids (upregulated and downregulated) using a targeted metabolomic platform. Also shown are hidden proteins that play a role in dysregulation of sphingolipids in diseased cells.

  4. Altered fatty acid and steroid metabolic processes identified by PIUMet.
    Figure 4: Altered fatty acid and steroid metabolic processes identified by PIUMet.

    Shown is a part of resulting network associated with these processes. Disease features are directly connected to metabolites representing their putative identities. The remaining nodes display hidden or experimentally undetected metabolites and hidden proteins of these pathways, with their sizes associated with robustness scores. We experimentally verified that the proteins and metabolites highlighted by red and orange boxes were altered in diseased cells.

  5. Comparison of disease-associated components identified in separate analyses of lipidomics and phosphoproteomics, and in an integrative analysis of those data.
    Figure 5: Comparison of disease-associated components identified in separate analyses of lipidomics and phosphoproteomics, and in an integrative analysis of those data.
  6. Regions of the resulting network obtained from integrative analysis of lipidomics and phosphoproteomics, displaying the dysregulation of high-scoring, hidden components.
    Figure 6: Regions of the resulting network obtained from integrative analysis of lipidomics and phosphoproteomics, displaying the dysregulation of high-scoring, hidden components.

    (ac) High-scoring proteins belonging to three subsets of nodes. The first subset contains nodes that increase in robustness when lipidomics and phosphoproteomics are considered together compared to lipidomics alone. DHCR7 (a) and FASN (c) are high-scoring members of this subset, which were significantly altered in diseased cells (two-tailed Student's t-test). The second subset includes nodes that increase in robustness when lipidomics and phosphoproteomics are considered together compared to phosphoproteomics alone. RASA1 (b) is the highest-scoring node in this network, whose encoded protein is significantly upregulated in diseased cells. Finally, the third subset contains proteins that are only identified by multi-omic analysis of lipidomics and phosphoproteomics. We confirmed that ERCC6-encoded protein (c), a high-scoring node in this subset, is significantly upregulated in diseased cells.

  7. A flow chart describing different steps of the PIUMet algorithm.
    Supplementary Fig. 1: A flow chart describing different steps of the PIUMet algorithm.
  8. The comparison of the disease-specific score distributions from real and random results.
    Supplementary Fig. 2: The comparison of the disease-specific score distributions from real and random results.

    The figure shows that the scores for the real results are significantly higher than random results (P value = 4.238E-37).

  9. The frequency of resulting hidden components from real data in the networks inferred from randomly selected metabolite features.
    Supplementary Fig. 3: The frequency of resulting hidden components from real data in the networks inferred from randomly selected metabolite features.
  10. A Venn diagram comparing the potential metabolites matching disease features based on mass compared to the putative ones inferred by PIUMet.
    Supplementary Fig. 4: A Venn diagram comparing the potential metabolites matching disease features based on mass compared to the putative ones inferred by PIUMet.

    The metabolites inferred by PIUMet are significantly enriched for the metabolites detected by a targeted metabolomic platform (hypergeometric test P value= 6.00×10− 4). All of the metabolites identified by the targeted platform were dysregulated in diseased cells.

  11. Altered sphingolipids in STHdh Q111 cells compared to STHdh Q7 cells. These sphingolipids were measured in nine biological replicates.
    Supplementary Fig. 5: Altered sphingolipids in STHdh Q111 cells compared to STHdh Q7 cells. These sphingolipids were measured in nine biological replicates.

    Two-sided student t-test analysis showed significant changes in C24:0 Ceramide (d18:1) with P value =1.2×10−11 (a), C16:1 Ceramide (d18:1) with P value =7.74×10−5 (b), C24:1 Ceramide (d18:1) with P value =0.02 (c), and C24:1 sphingomyelin (SM) with P value =2.07×10−8 (d). The height of the bar plot shows the average of each sphingolipid levels, while the error bar shows their standard deviation.

  12. PIUMet inferred Sphingosine-1-phosphate (S1P) as a disease-modifying hidden component of dysregulated sphingolipid pathway.
    Supplementary Fig. 6: PIUMet inferred Sphingosine-1-phosphate (S1P) as a disease-modifying hidden component of dysregulated sphingolipid pathway.

    (a) S1P was significantly downregulated (P value = 1.99×10−4, two-sided student t-test) in STHdh Q111 cells compared to STHdh Q7 cells. The bar plot shows the average levels of S1P that were measured in nine biological replicates. The error bars show the standard deviation of S1P levels (b). The treatment of diseased cells with an analogue of S1P (FTY720-P) significantly decreased apoptosis (P value = 7.98×10−5, two-sided student t-test). The bar plot shows the average percentage of cell death, and the error bars show the standard deviation from two independent experiments with twenty replicates each. (c) Calcein (green) detects viable cells and monitors changes in cell shape and morphology, while propidium iodide (PI – red) stains late apoptotic cells with damaged membranes.

  13. Western blot results showed that the DHCR7-encoded protein was significantly downregulated (P value = 0.025, two-sided student t-test) in the STHdh Q111 cells compared to STHdh Q7 cells.
    Supplementary Fig. 7: Western blot results showed that the DHCR7-encoded protein was significantly downregulated (P value = 0.025, two-sided student t-test) in the STHdh Q111 cells compared to STHdh Q7 cells.

    The boxplot shows the measured protein levels in six biological replicates (black dots).

  14. Altered fatty acids in STHdh Q111 cells compared to STHdh Q7 cells.
    Supplementary Fig. 8: Altered fatty acids in STHdh Q111 cells compared to STHdh Q7 cells.

    These fatty acids were measured in nine biological replicates. Two-sided student t-test analysis showed significant changes in eicosapentaenoic acid (EPA, P value=7.8×10−6, a) and dihomo-gamma-linolenic acid (DHGLA, P value=0.01, b). The height of the bar plot shows the average of each fatty acid levels, while the error bars show their standard deviation.

  15. Western blot results showed that fatty acid synthase enzyme, encoded by the FASN gene, was significantly upregulated (P value = 0.025, two-sided student t-test) in the STHdh Q111 cells compared to STHdh Q7 cells.
    Supplementary Fig. 9: Western blot results showed that fatty acid synthase enzyme, encoded by the FASN gene, was significantly upregulated (P value = 0.025, two-sided student t-test) in the STHdh Q111 cells compared to STHdh Q7 cells.

    The boxplot shows the measured protein levels in six biological replicates (black dots).

  16. Western blot results of measuring RASA1 and ERCC6 protein levels.
    Supplementary Fig. 10: Western blot results of measuring RASA1 and ERCC6 protein levels.

    (a) Western blot results showed a significant increase (P value = 0.008, two-sided student t-test) in the RASA1 protein levels. The boxplot shows the measured protein levels in three biological replicates (black dots). (b) The plot shows Western blot results indicating a significant increase (P value =0.028, two-sided student t-test) in the ERCC6 protein levels. The boxplot shows the measured protein levels in six biological replicates (black dots).

  17. The comparison of degree distributions of features from terminal and detectable metabolite feature (DMF) sets.
    Supplementary Fig. 11: The comparison of degree distributions of features from terminal and detectable metabolite feature (DMF) sets.

    The degree of a feature shows the number of metabolites corresponding to the feature.

References

  1. DeBerardinis, R.J. & Thompson, C.B. Cellular metabolism and disease: what do metabolic outliers teach us? Cell 148, 11321144 (2012).
  2. Patti, G.J., Yanes, O. & Siuzdak, G. Innovation: Metabolomics: the apogee of the omics trilogy. Nat. Rev. Mol. Cell Biol. 13, 263269 (2012).
  3. Baker, M. Metabolomics: from small molecules to big ideas. Nat. Methods 8, 117121 (2011).
  4. Dunn, W.B. et al. Mass appeal: metabolite identification in mass spectrometry-focused untargeted metabolomics. Metabolomics 9, 4466 (2013).
  5. Johnson, C.H., Ivanisevic, J., Benton, H.P. & Siuzdak, G. Bioinformatics: the next frontier of metabolomics. Anal. Chem. 87, 147156 (2015).
  6. Cho, K., Mahieu, N.G., Johnson, S.L. & Patti, G.J. After the feature presentation: technologies bridging untargeted metabolomics and biology. Curr. Opin. Biotechnol. 28, 143148 (2014).
  7. Grapov, D., Wanichthanarak, K. & Fiehn, O. MetaMapR: pathway independent metabolomic network analysis incorporating unknowns. Bioinformatics 31, 27572760 (2015).
  8. Kuo, T.-C., Tian, T.-F. & Tseng, Y.J. 3Omics: a web-based systems biology tool for analysis, integration and visualization of human transcriptomic, proteomic and metabolomic data. BMC Syst. Biol. 7, 64 (2013).
  9. Karnovsky, A. et al. Metscape 2 bioinformatics tool for the analysis and visualization of metabolomics and gene expression data. Bioinformatics 28, 373380 (2012).
  10. Krumsiek, J. et al. Mining the unknown: a systems approach to metabolite identification combining genetic and metabolic information. PLoS Genet. 8, e1003005 (2012).
  11. Li, S. et al. Predicting network activity from high throughput metabolomics. PLoS Comput. Biol. 9, e1003123 (2013).
  12. Yeger-Lotem, E. et al. Bridging high-throughput genetic and transcriptional data reveals cellular responses to alpha-synuclein toxicity. Nat. Genet. 41, 316323 (2009).
  13. Tuncbag, N. et al. Simultaneous reconstruction of multiple signaling pathways via the prize-collecting steiner forest problem. J. Comput. Biol. 20, 124136 (2013).
  14. Huang, S.-S.C. & Fraenkel, E. Integrating proteomic, transcriptional, and interactome data reveals hidden components of signaling and regulatory networks. Sci. Signal. 2, ra40 (2009).
  15. Trettel, F. et al. Dominant phenotypes produced by the HD mutation in STHdh(Q111) striatal cells. Hum. Mol. Genet. 9, 27992809 (2000).
  16. Maceyka, M., Harikumar, K.B., Milstien, S. & Spiegel, S. Sphingosine-1-phosphate signaling and its role in disease. Trends Cell Biol. 22, 5060 (2012).
  17. Di Pardo, A. et al. FTY720 (fingolimod) is a neuroprotective and disease-modifying agent in cellular and mouse models of Huntington disease. Hum. Mol. Genet. 23, 22512265 (2014).
  18. Di Menna, L. et al. Fingolimod protects cultured cortical neurons against excitotoxic death. Pharmacol. Res. 67, 19 (2013).
  19. Deogracias, R. et al. Fingolimod, a sphingosine-1 phosphate receptor modulator, increases BDNF levels and improves symptoms of a mouse model of Rett syndrome. Proc. Natl. Acad. Sci. USA 109, 1423014235 (2012).
  20. Valenza, M. & Cattaneo, E. Emerging roles for cholesterol in Huntington's disease. Trends Neurosci. 34, 474486 (2011).
  21. Kreilaus, F., Spiro, A.S., Hannan, A.J., Garner, B. & Jenner, A.M. Brain cholesterol synthesis and metabolism is progressively disturbed in the R6/1 mouse model of Huntington's disease: a targeted GC-MS/MS sterol analysis. J. Huntingtons Dis. 4, 305318 (2015).
  22. Yehuda, S., Rabinovitz, S. & Mostofsky, D.I. Essential fatty acids and the brain: from infancy to aging. Neurobiol. Aging 26 (Suppl. 1), 98102 (2005).
  23. Block, R.C., Dorsey, E.R., Beck, C.A., Brenna, J.T. & Shoulson, I. Altered cholesterol and fatty acid metabolism in Huntington disease. J. Clin. Lipidol. 4, 1723 (2010).
  24. Puri, B.K. et al. Ethyl-EPA in Huntington disease: a double-blind, randomized, placebo-controlled trial. Neurology 65, 286292 (2005).
  25. Puri, B.K. et al. Reduction in cerebral atrophy associated with ethyl-eicosapentaenoic acid treatment in patients with Huntington's disease. J. Int. Med. Res. 36, 896905 (2008).
  26. López, M. & Vidal-Puig, A. Brain lipogenesis and regulation of energy metabolism. Curr. Opin. Clin. Nutr. Metab. Care 11, 483490 (2008).
  27. Li, S.-H. & Li, X.-J. Huntingtin-protein interactions and the pathogenesis of Huntington's disease. Trends Genet. 20, 146154 (2004).
  28. Stevnsner, T., Muftuoglu, M., Aamann, M.D. & Bohr, V.A. The role of Cockayne Syndrome group B (CSB) protein in base excision repair and aging. Mech. Ageing Dev. 129, 441448 (2008).
  29. Subba Rao, K. Mechanisms of disease: DNA repair defects and neurological disease. Nat. Clin. Pract. Neurol. 3, 162172 (2007).
  30. Razick, S., Magklaras, G. & Donaldson, I.M. iRefIndex: a consolidated protein interaction database with provenance. BMC Bioinformatics 9, 405 (2008).
  31. Wishart, D.S. et al. HMDB 3.0—The Human Metabolome Database in 2013. Nucleic Acids Res. 41, D801D807 (2013).
  32. Thiele, I. et al. A community-driven global reconstruction of human metabolism. Nat. Biotechnol. 31, 419425 (2013).
  33. Aranda, B. et al. PSICQUIC and PSISCORE: accessing and scoring molecular interactions. Nat. Methods 8, 528529 (2011).
  34. Frolkis, A. et al. SMPDB: The Small Molecule Pathway Database. Nucleic Acids Res. 38, D480D487 (2010).
  35. Hucka, M. et al. SBML Forum. The systems biology markup language (SBML): a medium for representation and exchange of biochemical network models. Bioinformatics 19, 524531 (2003).
  36. Bornstein, B.J., Keating, S.M., Jouraku, A. & Hucka, M. LibSBML: an API library for SBML. Bioinformatics 24, 880881 (2008).
  37. Huang, S.S. et al. Linking proteomic and transcriptional data through the interactome and epigenome reveals a map of oncogene-induced signaling. PLoS Comput. Biol. 9, e1002887 (2013).
  38. Bailly-Bechet, M., Braunstein, A., Pagnani, A., Weigt, M. & Zecchina, R. Inference of sparse combinatorial-control networks from gene-expression data: a message passing approach. BMC Bioinformatics 11, 355 (2010).
  39. Huang, W., Sherman, B.T. & Lempicki, R.A. Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources. Nat. Protoc. 4, 4457 (2009).
  40. Saghatelian, A. et al. Assignment of endogenous substrates to enzymes by global metabolite profiling. Biochemistry 43, 1433214339 (2004).
  41. Tautenhahn, R., Patti, G.J., Rinehart, D. & Siuzdak, G. XCMS Online: a web-based platform to process untargeted metabolomic data. Anal. Chem. 84, 50355039 (2012).
  42. NCBI Resource Coordinators. Database resources of the National Center for Biotechnology Information. Nucleic Acids Res. 41, D8D20 (2013).
  43. Gnad, F., Gunawardena, J. & Mann, M. PHOSIDA 2011: the posttranslational modification database. Nucleic Acids Res. 39, D253D260 (2011).
  44. Schreiber, E., Matthias, P., Müller, M.M. & Schaffner, W. Rapid detection of octamer binding proteins with 'mini-extracts', prepared from a small number of cells. Nucleic Acids Res. 17, 6419 (1989).
  45. Ng, C.W. et al. Extensive changes in DNA methylation are associated with expression of mutant huntingtin. Proc. Natl. Acad. Sci. USA 110, 23542359 (2013).

Download references

Author information

Affiliations

  1. Department of Biological Engineering, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA.

    • Leila Pirhaji,
    • Pamela Milani,
    • Timothy Curran,
    • Forest M White &
    • Ernest Fraenkel
  2. Department of Chemistry and Chemical Biology, Harvard University, Cambridge, Massachusetts, USA.

    • Mathias Leidl &
    • Alan Saghatelian
  3. Koch Institute for Integrative Cancer Research, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA.

    • Timothy Curran &
    • Forest M White
  4. Broad Institute, Cambridge, Massachusetts, USA.

    • Julian Avila-Pacheco,
    • Clary B Clish &
    • Ernest Fraenkel
  5. Salk Institute for Biological Studies, La Jolla, California, USA.

    • Alan Saghatelian

Contributions

L.P. and E.F. designed the approach. L.P. implemented the algorithm and performed the computational analyses. P.M. prepared cell cultures and performed western blot and cell viability experiments. M.L. and A.S. performed the untargeted lipidomic experiments. T.C. and F.M.W. measured global levels of phosphoproteins. J.A.-P. and C.B.C. performed targeted lipidomic experiments. L.P. and E.F. wrote the manuscript, and all the authors approved the final version.

Competing financial interests

L.P. and E.F. are co-founders of ReviveMed, Inc., and have filed a provisional patent based on the work described here (application number US 62/203,292).

Corresponding author

Correspondence to:

Author details

Supplementary information

Supplementary Figures

  1. Supplementary Figure 1: A flow chart describing different steps of the PIUMet algorithm. (347 KB)
  2. Supplementary Figure 2: The comparison of the disease-specific score distributions from real and random results. (70 KB)

    The figure shows that the scores for the real results are significantly higher than random results (P value = 4.238E-37).

  3. Supplementary Figure 3: The frequency of resulting hidden components from real data in the networks inferred from randomly selected metabolite features. (28 KB)
  4. Supplementary Figure 4: A Venn diagram comparing the potential metabolites matching disease features based on mass compared to the putative ones inferred by PIUMet. (93 KB)

    The metabolites inferred by PIUMet are significantly enriched for the metabolites detected by a targeted metabolomic platform (hypergeometric test P value= 6.00×10− 4). All of the metabolites identified by the targeted platform were dysregulated in diseased cells.

  5. Supplementary Figure 5: Altered sphingolipids in STHdh Q111 cells compared to STHdh Q7 cells. These sphingolipids were measured in nine biological replicates. (111 KB)

    Two-sided student t-test analysis showed significant changes in C24:0 Ceramide (d18:1) with P value =1.2×10−11 (a), C16:1 Ceramide (d18:1) with P value =7.74×10−5 (b), C24:1 Ceramide (d18:1) with P value =0.02 (c), and C24:1 sphingomyelin (SM) with P value =2.07×10−8 (d). The height of the bar plot shows the average of each sphingolipid levels, while the error bar shows their standard deviation.

  6. Supplementary Figure 6: PIUMet inferred Sphingosine-1-phosphate (S1P) as a disease-modifying hidden component of dysregulated sphingolipid pathway. (221 KB)

    (a) S1P was significantly downregulated (P value = 1.99×10−4, two-sided student t-test) in STHdh Q111 cells compared to STHdh Q7 cells. The bar plot shows the average levels of S1P that were measured in nine biological replicates. The error bars show the standard deviation of S1P levels (b). The treatment of diseased cells with an analogue of S1P (FTY720-P) significantly decreased apoptosis (P value = 7.98×10−5, two-sided student t-test). The bar plot shows the average percentage of cell death, and the error bars show the standard deviation from two independent experiments with twenty replicates each. (c) Calcein (green) detects viable cells and monitors changes in cell shape and morphology, while propidium iodide (PI – red) stains late apoptotic cells with damaged membranes.

  7. Supplementary Figure 7: Western blot results showed that the DHCR7-encoded protein was significantly downregulated (P value = 0.025, two-sided student t-test) in the STHdh Q111 cells compared to STHdh Q7 cells. (74 KB)

    The boxplot shows the measured protein levels in six biological replicates (black dots).

  8. Supplementary Figure 8: Altered fatty acids in STHdh Q111 cells compared to STHdh Q7 cells. (57 KB)

    These fatty acids were measured in nine biological replicates. Two-sided student t-test analysis showed significant changes in eicosapentaenoic acid (EPA, P value=7.8×10−6, a) and dihomo-gamma-linolenic acid (DHGLA, P value=0.01, b). The height of the bar plot shows the average of each fatty acid levels, while the error bars show their standard deviation.

  9. Supplementary Figure 9: Western blot results showed that fatty acid synthase enzyme, encoded by the FASN gene, was significantly upregulated (P value = 0.025, two-sided student t-test) in the STHdh Q111 cells compared to STHdh Q7 cells. (109 KB)

    The boxplot shows the measured protein levels in six biological replicates (black dots).

  10. Supplementary Figure 10: Western blot results of measuring RASA1 and ERCC6 protein levels. (118 KB)

    (a) Western blot results showed a significant increase (P value = 0.008, two-sided student t-test) in the RASA1 protein levels. The boxplot shows the measured protein levels in three biological replicates (black dots). (b) The plot shows Western blot results indicating a significant increase (P value =0.028, two-sided student t-test) in the ERCC6 protein levels. The boxplot shows the measured protein levels in six biological replicates (black dots).

  11. Supplementary Figure 11: The comparison of degree distributions of features from terminal and detectable metabolite feature (DMF) sets. (57 KB)

    The degree of a feature shows the number of metabolites corresponding to the feature.

PDF files

  1. Supplementary Text and Figures (4650 KB)

    Supplementary Figures 1–11 and Supplementary Table 1

Excel files

  1. Supplementary Table 2 (26 KB)

    The summary of untargeted lipidomic data fromSTHdh cell lines.

  2. Supplementary Table 3 (11 KB)

    The summary of global phospho-proteomic data from STHdh cell lines.

Additional data