diff --git a/assays/2_9-StatisticalAnalysis-ProteomeDeconvolution/protocols/2_9-StatisticalAnalysis-ProteomeDeconvolution.md b/assays/2_9-StatisticalAnalysis-ProteomeDeconvolution/protocols/2_9-StatisticalAnalysis-ProteomeDeconvolution.md
index 1c9d118c8604fff4448f44f7d085183b8e8e02fd..5811d94640a067946296968595b8c3509bb6eedb 100644
--- a/assays/2_9-StatisticalAnalysis-ProteomeDeconvolution/protocols/2_9-StatisticalAnalysis-ProteomeDeconvolution.md
+++ b/assays/2_9-StatisticalAnalysis-ProteomeDeconvolution/protocols/2_9-StatisticalAnalysis-ProteomeDeconvolution.md
@@ -3,3 +3,5 @@
 Unless stated otherwise all data analyses were performed in R using the statistical package e1071 for support vector regression (http://www.R-project.org), ggplot2 for visualization (Wickham, 2009; Wickham, 2016; https://ggplot2.tidyverse.org) and NFM for Nonnegative Matrix factorization (Gaujoux and Seoighe, 2010; http://cran.r-project.org). To assign MS-identified proteins to a certain organellar fraction, several deconvolution approaches have been applied. Unsupervised deconvolution was performed by Nonnegative matrix factorization (NMF). We determined the rank of factorization using the Brunet- and Lee-Seung-algorithms for ranks from 2 to 10 with 50 repetitions (Gaujoux and Seoighe, 2010). Both algorithms were further constrained to sum to one in a sample, so that its value can be interpreted as the relative proportion of a fraction in a sample. We used random as well as double singular value decomposition for seeding (Boutsidis and Gallopoulos, 2008). The final factorization was performed 500 times with random seeding and a rank of 5. We determined the fractions specific for an organelle by scaling the coefficients per protein and clustered them via k-means with Euclidean distance. An enrichment of organelle-specific proteins in a cluster was determined via Fisher’s Exact Test and proteins specific for an organelle were used for classification. Proteins were assigned either to peroxisomes, plastids, mitochondria, and other organelles based on the fraction with the highest coefficient.
 For supervised deconvolution we used the distribution of marker proteins normalized to its maximum in all samples. Quadratic programming as well as Support vector regression were constrained to yield positive values. Nu-type SVR was performed with a linear kernel and a set of Nu values (0.25; 0.5; 0.75; 1.0). To test the quality of prediction, deconvolution was repeated 500 times splitting the dataset into a training set containing 70% of the dataset and a testing set using the remaining 30% of the data. Proteins were associated to the organelle with the highest coefficient and consensus classification was achieved by simple majority vote of all four algorithms.
 
+- Gaujoux, R., and Seoighe, C. (2010). A flexible R package for nonnegative matrix factorization. BMC Bioinf. 11, 367. doi: 10.1186/1471-2105-11-367
+- Boutsidis, C., and Gallopoulos, E. (2008). SVD based initialization: A head start for nonnegative matrix factorization. Pattern recognition 41, 1350–1362. doi: 10.1016/ j.patcog.2007.09.010
\ No newline at end of file