476 | VOL.6 NO.7 | JULY 2009 | NATURE METHODS CORRESPONDENCE profiling and qPCR recovered a nonuniform distribution of microRNAs (Fig. 2a); we observed up to four orders of magnitude dif- ference between the most and least frequently detected microRNAs. Only 61% (SREK-SOLiD) and 52% (modban-Solexa) of the microRNAs varied within a single order of magnitude (Fig. 2a). These results showed the inherent quantification biases of both DGE profiling and qPCR based on microRNA sequence, complicating comparison of microRNA amounts in a sample. Correction of the biological dataset with the frequency bias obtained using the synthetic RNA pool did not improve the correlation between the library-preparation methods (data not shown). We therefore used the synthetic small RNA dataset to explore the potential basis of systematic biases. Although we found clear effects of certain terminal mono- and dinucleotides ( Supplementary Fig. 4), we could not identify a satisfactory correction model based on primary (RNA sequence) and secondary (for example, folding characteristics) parameters (Supplementary Fig. 5 and Supplementary Note). This might be explained by our observation that even single nucleotide differences influenced the read frequencies (Supplementary Fig. 6). RNA ligase preferences 7 may contribute to the observed different terminal nucleotides over the read frequency spectrum. In addition, the reverse-transcriptase reaction as well as the PCR could be a contributor to the bias 8 . To determine whether DGE profiling allows for differential expression analysis, we sequenced small RNA libraries from rat spleen and liver (SREK-SOLiD). In parallel, we analyzed the input RNA by qPCR. Similar to our previous results, qPCR data differed substantially from the read frequencies within a sample (Fig. 2b). However, differential expression results between samples obtained by qPCR and DGE profiling were strongly correlated (Fig. 2b), showing that the systematic biases do not prohibit the comparison of relative microRNA amounts between samples. Despite the limitations described here, small RNA profiling by DGE is the method of choice for studying small RNA expres- sion. In contrast to most other existing methods, DGE profiling is hybridization-independent, accurate in discriminating microRNA family members that differ by only a single nucleotide, capable of detecting 5′ and 3′ end variability (for example, isoMirs), and as the approach does not require a priori information, it can be used to simultaneously detect known and discover new biomolecules. Note: Supplementary information is available on the Nature Methods website. ACKNOWLEDGMENTS This work was supported by the Cancer Genomics Center and Netherlands Bioinformatic Center through Netherlands Genomics Initiative (to E.C.), Fred Hutchinson Cancer Research Center and Canary Foundation funding (New Development funds to M.T.), Core Center of Excellence in Hematology Pilot Grant P30 DK56465 (to M.T.), Pacific Northwest Prostate Cancer Specialized Program of Research Excellence Grant P50 CA97186 (to M.T.) and a Rosetta Inpharmatics Fellowship in Molecular Profiling (to S.K.W.). We thank P. Toonen for animal care. COMPETING INTERESTS STATEMENT The authors declare competing financial interests: details accompany the full- text HTML version of the paper at http://www.nature.com/naturemethods/. Sam E V Linsen 1,7 , Elzo de Wit 1,7 , Georges Janssens 1 , Sheila Heater 2 , Laura Chapman 2 , Rachael K Parkin 3 , Brian Fritz 3,6 , Stacia K Wyman 3 , Ewart de Bruijn 1 , Emile E Voest 4 , Scott Kuersten 2 , Muneesh Tewari 3,5 & Edwin Cuppen 1 1 Hubrecht Institute and University Medical Center Utrecht, Cancer Genomics Center, Utrecht, The Netherlands. 2 Life Technologies, Austin, Texas, USA. 3 Human Biology Division, Fred Hutchinson Cancer Research Center, Seattle, Washington, USA. 4 Department of Medical Oncology, University Medical Center Utrecht, Utrecht, The Netherlands. 5 Clinical Research Division, Fred Hutchinson Cancer Research Center, Seattle, Washington, USA. 6 Present address: Illumina, Inc., San Diego, California, USA. 7 These authors contributed equally to this work. e-mail: e.cuppen@niob.knaw.nl 1. Berezikov, E., et al. Genome Res. 16, 1289–1298 (2006). 2. Ruby, J.G. et al. Cell 127, 1193–1207 (2006). 3. Kuchenbauer, F. et al. Genome Res. 18, 1787–1797 (2008). 4. Berezikov, E. et al. Nat. Genet. 38, 1375–1377 (2006). 5. Lau, N.C., Lim, L.P., Weinstein, E.G. & Bartel, D.P. Science 294, 858–862 (2001). 6. Griffiths-Jones, S., Grocock, R.J., van Dongen, S., Bateman, A. & Enright, A.J. Nucleic Acids Res. 34, D140–D144 (2006). 7. Romaniuk, E., McLaughlin, L.W., Neilson, T. & Romaniuk, P.J. Eur. J. Biochem. 125, 639–643 (1982). 8. Taube, R., Loya, S., Avidan, O., Perach, M. & Hizi, A. Biochem. J. 329, 579–587 (1998). RNAiCut: automated detection of significant genes from functional genomic screens To the Editor: RNA interference (RNAi) is a popular functional genomic technology for identifying genes involved in a biological process. Although higher scores for genes in an RNAi screen suggest more central roles in the pathway, estimating the score threshold separating pathway- or process-relevant hits from noise remains dif- ficult (Supplementary Table 1) and is typically done manually. To overcome this subjective approach, we built a fully auto- mated system, RNAiCut, that objectively and robustly identifies score thresholds from functional genomic data by introducing the use of the connectivity of subgraphs of protein-protein interac- tion (PPI) networks 1,2 . Unlike some previous work 3 , our method does not overlap RNAi and PPI data to find interacting regulators. Instead, its guiding hypothesis is that true positive hits in an RNAi experiment are densely interconnected in the PPI network. For the k highest-scoring genes (k = 1, 2, 3…), RNAiCut computes the edge count of the induced subgraph and estimates the P-value of finding a PPI subgraph of at least this size that is induced by k randomly cho- sen nodes that have the same degrees as these genes (Supplementary Methods and Supplementary Results). The plot of these P-values as a function of k is typically V-shaped, and we take the global min- imum as the score threshold (Fig. 1). We used RNAiCut to com- pute thresholds for several Drosophila melanogaster RNAi screens 4 (Supplementary Figs. 1–10 and Supplementary Tables 2–3). RNAiCut chose successful thresholds, as measured by Gene Ontology (GO) 5 enrichment: the gene lists with above-threshold scores were enriched for functions relevant to the screen, compared to the rest (Supplementary Table 4). When the manual screener’s threshold was later in the ranked list of hits than the RNAiCut threshold, choosing RNAiCut’s threshold may reduce the poten- tially high number of false positives. When RNAiCut’s threshold was later, the GO enrichment for RNAiCut’s cutoff was at least as good as for the manually determined cutoff, revealing additional pathway-relevant genes (Supplementary Results). Although some of the additional hits identified by RNAiCut may be false positives, analyzing them may be useful given their apparent connectivity to © 2009 Nature America, Inc. All rights reserved.