Online GESS: prediction of miRNA-like off-target effects in large-scale RNAi screen data by seed region analysis

RNA interference (RNAi) is a post-transcriptional gene regulatory mechanism [1] that has been widely used for functional genomics studies both in cell lines and organisms. The synthetic duplexes referred to as small interfering RNAs (siRNAs) or short hairpin RNAs (shRNAs) used for RNAi partner with the RNA Induced Silencing Complex (RISC) to target messenger RNAs for degradation in a sequence-specific manner [2, 3]. It has been well established that synthetic duplexes have both on-target activities (reducing expression of intended gene) as well as off-target activities (leading to reduced expression of unintended genes) [4]. A significant fraction of false-positives from RNAi screens is due to off-target effects (OTEs) [5, 6]. Many efforts have been made to reduce the number of false positive results due to sequence-specific OTEs, including improved algorithms for RNAi reagent design and incorporation of chemical modifications into siRNAs [7]. Another approach has been to develop bioinformatics strategies to identify possible OTEs in screening results.

OTEs have been linked to the mechanism of action of miRNAs, in which a short sequence of bases 2–8 on the 5′ end of a strand of the RNAi duplex (usually the antisense strand), also called the ‘seed region’, is complementary to the 3′ untranslated regions (UTRs) of multiple mRNAs, causing degradation of their associated transcripts [8, 9]. To improve the interpretation of RNAi datasets and to help minimize follow-up experimental efforts, it is important to identify transcripts that are likely to have scored as a result of seed sequence-based targeting. A number of off-target analysis algorithms have been developed and made available to the scientific community. For example, the Haystack algorithm analyzes RNAi off-target effects based on a predictive model trained with published datasets [10]. The model takes into account four types of seed matches and the length of 3’ UTR regions. It requires seed sequence diversity across the dataset and is optimized for large RNAi datasets of sufficient scale with normally distributed scores. Genome-wide Enrichment of Seed Sequence matches (GESS) is another bioinformatics method developed for OTE analysis of RNAi screen datasets [11]. It identifies candidate off-targeted transcripts by investigating the association between matches of the seed regions of RNAi reagents in 3’UTRs with phenotypes observed in large-scale screens. In GESS, RNAi reagents are grouped into two categories: siRNA/shRNAs that score in a screen as “with phenotype” or “active”, and reagents considered “without phenotype” or “inactive”. The algorithm calculates a seed match frequency (SMF) for active and inactive siRNAs/shRNAs for each tested sequence (transcript of a target gene). It is expected that transcripts (and in particular, 3’UTR regions) that are significantly over-represented for seed region matches among active RNAi reagents are more likely to be off-targets. The GESS algorithm has been used successfully to identify off-targeted transcripts in several medium- to large-scale datasets. For example, MAD2 was identified as an off-targeted transcript in a spindle assembly checkpoint components screen and TGFβ-R2 in a screen for novel components of the TGFβ pathway [11]. GESS was also used in analyzing data from a screen for genes required for homologous recombination and predicted RAD51 as a candidate off-targeted gene; RAD51 OTEs were later confirmed experimentally to be responsible for the activity of many siRNAs identified as hits in the primary screen [12].

Previously, MATLAB was used to program and run the GESS algorithm. Standalone versions of the GESS MATLAB code were provided for several operating systems, including Windows, Linux and Mac. Although it is possible to install and use these stand-alone MATLAB versions, it is not easy for biologists lacking programming and informatics expertise to implement GESS in its original form, as the user has to: 1. download and install the program; 2. provide files for the reference sequences; and 3. prepare separate files for siRNAs/shRNAs, phenotype and reagent sequence information. In addition, the run time of the program is not optimal for large files. For example, it can take up to 30 hours to analyze 10,000 siRNAs against 27,500 3’UTRs using the standalone MATLAB version of GESS. Hence, an open, user-friendly online tool with improved performance would be of interest to the scientific community.

Online GESS was developed as a Java web application. Twitter Bootstrap 3 front-end framework and jQuery JavaScript library were used to develop the web pages. At the back end, Online GESS contains reference sequences corresponding to 3’UTRs (the region thought to be the most sensitive to miRNA-off-target effects), 5’UTRs, coding sequences (CDS) or full-length transcripts (including non-coding RNAs) in the human, mouse and Drosophila genomes. The human and mouse sequences are obtained from the NCBI RefSeq database. Although these sequences are derived from GenBank records, RefSeq records are non-redundant and have gone through additional levels of validation, annotation, and manual curation. Transcript sequences, as well as CDS and UTR annotations, are retrieved. The Drosophila transcript sequences are obtained from FlyBase (flybase.org) [13], a comprehensive database of Drosophila information that is curated by experts to ensure quality and includes sequences, gene annotation, mutant alleles and publications. Because curation and annotation of reference sequences is an ongoing effort, we have implemented a mechanism for synchronizing reference sequences with each new RefSeq and FlyBase release [14, 15].

After a user uploads their annotated screen results (i.e. sequences of active and, if available, inactive RNAi reagents) in Excel, comma-separated values or tab-delimited text format, the online GESS tool extracts the seed sequences from active and inactive RNAi reagent sequences, then searches the transcript sequences for perfect matches. If a set of inactive RNAi reagents is not provided, the program creates a theoretically inactive set by replacing the first nucleotide of each seed region with the complementary nucleotide. The program then calculates the frequencies of matches among active and inactive RNAi reagents, and identifies transcripts that are significantly enriched among active RNAi reagents using the Fisher exact test and Yates chi-square test. When the sample size is small, the p value from the Fisher’s Exact Test is selected; otherwise, the p value from the Yates Chi Square test is used. Transcripts are then ranked based on the selected p-value. Ranks are later used for calculating multiple hypothesis correction. Three multiple hypothesis correction methods are used in the analysis, the Bonferroni, Bonferroni step-down and Benjamini & Hochberg algorithms, listed in order from most to least stringent correction. Detailed information about the GESS algorithm and analysis methods can be found in the original publication [11].

In Drosophila, short hairpin RNAs have been used for large-scale in vivo screens [26]. We analyzed results from two screens for embryonic phenotypes associated with maternally loaded shRNAs [24, 25]. We did not find any potential off-target genes with either dataset, possibly due to the small size of these studies (1000 or 2300 shRNAs, respectively). As more transgenic shRNA screens are done, we anticipate that GESS will prove useful to detect potential off-targeted transcripts that might be associated with in vivo Drosophila RNAi screen data.

RNAi is a powerful tool for systematic study of gene functions but results must be analyzed carefully, as screens are associated with false positive and false negative results. Further validation of results, such as screening with multiple independent RNAi reagents, performing qPCR to verify correlation between knockdown efficiency and phenotypic strength, or RNAi-resistant “rescue” experiments, is time and labor intensive. Detection of potential off-targeted transcripts via automated pre-processing based on our current knowledge of the sources of off-targets, including miRNA-like effects, provides one way to focus limited resources on the most promising candidates. To help support automated detection of off-targeted transcripts in RNAi data, we have implemented a web-based application of seed region analysis for identification of potential off-target transcripts, based on the GESS algorithm. This tool allows users to run off-target analysis with ease. Users can analyze human, mouse or Drosophila datasets directly using built-in reference sequence database. In addition, screen data can be analyzed based on a custom reference database, making it possible to analyze RNAi screen results from any organism and at any scale.