**5. Conclusions**

In conclusion, this work proposed a short list of testable hypotheses involving specific biological signatures for NF1-deficient nerve sheath tumors. Verification of these mechanisms in in-vitro and in-vivo models of nerve sheath tumors as well as human NF1 nerve sheath tumor tissue needs active and extensive experimental work. Although we analyzed tumor datasets from four different studies, the addition of other neurofibromatosis-driven tumor datasets will greatly aid identification of commonalities or critical differences to inform therapeutic decisions across the family of neurofibromatoses. This study, together with future work, may guide the repositioning of clinically approved drugs in the context of NF1.

**Supplementary Materials:** The following are available online at http://www.mdpi.com/2073-4425/11/2/226/s1. Figure S1: (a) A density plot indicating the pairwise Pearson correlation of LVs (not including self-self-pairs). A threshold of 0.5 was selected to eliminate non-orthogonal latent variables. Figure S2: Schematic showing the generation of the ensemble of random forests. (A) Stratified split of original dataset into "model" set and "independent test" set. (B) Generation of first ensemble of random forests using training set and test set generated from the "model" set for all 962 latent variables. (i) Example density plot showing distribution of model accuracy (F1) scores for each class. (ii) Example plot showing distribution of feature importance scores from the 500 random forests. (C) Generation of the second ensemble of random forests using restricted feature set of only 98 latent variables. The newly generated models were tested on the "independent test" set. (i) Example plot showing the distribution of model accuracy scores from the second ensemble of random forests. Figure S3: A boxplot representing the decrease in pairwise distances between all tumor samples of the same tumor type from Figure 1A (green, based on gene clustering) to 1C (orange, based on LV clustering). Figure S4: Small molecule-target networks showed enrichment of LV-dependent target classes. (A) Cluster 1 of LV-correlated VIPER proteins was enriched for HDAC inhibitors. (B) Cluster 3 of LV-correlated VIPER proteins was enriched for kinase inhibitors, particularly inhibitors of kinases responsible for cell cycle progression (CDKs). (C) Cluster 5 of LV-correlated VIPER proteins was enriched for HDAC inhibitors. Table S1: Pan-NF1 MultiPLIER results across RNA-seq data for 77 nerve sheath tumors. Table S2: Significance of latent variable status being correlated with gene mutation status. Table S3: Summary of latent variables selected by random forest model and their correlated genes and signatures. Table S4: Tumor deconvolution scores across NF1 patient samples. Table S5: metaVIPER protein scores across all NF1 patient samples. Table S6: Spearman correlation scores of RF-selected latent variables with metaVIPER scores. Table S7: Drug set enrichment analysis of latent variables.

**Author Contributions:** Conceptualization, R.J.A., J.B., S.J.C.G.; methodology, R.J.A., J.B., J.N.T., C.S.G., A.B., J.G., S.J.C.G.; software, R.J.A., J.B., J.N.T., C.S.G., A.B., J.G., S.J.C.G.; formal analysis, R.J.A., J.B., A.B., S.J.C.G.; investigation, R.J.A., J.B., S.J.C.G.; resources, X.Z., C.I.M., A.H., C.A.P.; writing—original draft preparation, R.J.A., J.B., S.J.C.G.; writing—review and editing, J.B., R.J.A., C.S.G., A.B., J.G., and S.J.C.G.; supervision, S.J.C.G., C.S.G., J.O.B., J.G.; project administration, S.J.C.G.; funding acquisition, S.J.C.G., J.G. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported with funds from the Neurofibromatosis Therapeutic Acceleration Program.

**Acknowledgments:** We would like to acknowledge our colleagues at Sage Bionetworks, Matthew Wall and Michael Mason, for many helpful conversations.

**Conflicts of Interest:** The authors have no conflicts of interest to declare, with the exception of Angela Hirbe and Christine A. Pratilas, who are serving as a co-editor for this issue of Genes. Both Angela Hirbe and Christine A. Pratilas have abstained from any editorial duties relating to this manuscript.
