Introduction

Although metazoan whole-genome sequences continue to accumulate at a rapid pace, comprehensive annotation of these genomes lags significantly behind. Particularly lacking is identification of non-coding regulatory sequences, which comprise a substantial fraction of the genome yet typically are completely lacking in available annotations. One reason why regulatory annotation so often lags behind genome sequencing is that historically, finding regulatory elements in the genome has been difficult even in well-studied model organisms because of their distant positions from target genes, the absence of a clear universal biochemical regulatory sequence marker, and the cell type specificity of regulatory element activity [1–4]. Although both empirical and computational methods for regulatory element discovery have been developed (for review see [5–8]), a variety of limitations prevent them from being easily adopted for efficient annotation of newly-sequenced genomes. Empirical approaches are costly; can be difficult to validate depending on the availability of biological resources such as cell lines, antibodies, and tissue samples, or the existence of relevant technologies, such as transgenesis; and carry apparent false-positive and false-negative rates that can be surprisingly high (false-positive rates range as high as 40% for some ChIP-based methods [9–11] and from 10–20% for some ATAC-seq studies [12, 13]). Because regulatory elements may be functional only in certain cell types or under specific conditions, assays need to be applied to multiple tissues over many developmental stages and/or under varying environmental conditions in order to achieve comprehensive annotation. Although computational methods in theory avoid these issues, many approaches still rely on experimental data either for training or as input, which often negates their advantages. For non-model organisms, where only limited functional genomic data tend to be available, regulatory annotation is therefore often highly challenging.

We previously developed SCRMshaw, a supervised cis-regulatory module machine-learning method that shows strong performance in predicting transcriptional regulatory sequences (“enhancers”) [14–19]. When trained on known enhancers from the fruit fly Drosophila melanogaster, SCRMshaw successfully predicts regulatory sequences across a wide range of insect species [17, 19–21]. Importantly, SCRMshaw requires as input only a sequenced genome and a basic gene annotation, making it highly suitable for producing preliminary regulatory annotations rapidly following initial publication of new genome assemblies. While benchmarking enhancer discovery methods is complicated due to the fact that there are no established true positive/true negative data sets against which to compare competing approaches [14], success rates from SCRMshaw appear to be on a par with or better than those from other rigorously-evaluated methods. Importantly, while we expect that other sequence-based enhancer discovery methods will, similar to SCRMshaw, be capable of cross-species discovery, only SCRMshaw so far has a track record of success in the cross-species setting.

We present here a detailed updated SCRMshaw protocol that replaces the earlier protocol published in 2019 [16] and aligns with the procedures described in our recent publication presenting the regulatory annotation of over 30 insect genomes [19]. The current protocol incorporates changes to the original SCRMshaw protocol as described by Asma and Halfon [14] and Asma et al. [19]. It includes additional steps developed by Asma et al. [19] to implement an end-to-end pipeline that begins with a sequenced genome and ends with a fully-annotated regulatory genome formatted for easy comparison to other SCRMshaw-annotated genomes.

The SCRMshaw pipeline consists of four basic parts (Fig 1). A pre-processing step checks the input files for proper formatting and removes minor unmapped/unannotated scaffolds from the genome. The SCRMshaw algorithm itself, which consists of three underlying statistical methods, is then run to scan the genome for high-scoring sequences. This step is dependent on training data, which can be downloaded from our GitHub site. Custom training sets can also be developed by the user; for instructions on custom training set generation, we refer the reader to Kazemian and Halfon [16]. The basic protocol described here is for “SCRMshaw_HD” [14], which requires use of a high-performance computing cluster and a minimum of 25 available nodes. However, SCRMshaw can also be run in its original lighter-weight form on a desktop computer by following the relevant protocol notes and/or referring to [16]. The third step in the SCRMshaw pipeline is post-processing, which determines the final set of enhancer predictions based on the SCRMshaw scores. Details on this step can be found in [19] and are diagrammed in Fig 2. The final step in the SCRMshaw pipeline, which is optional, maps putative enhancer target genes to their Drosophila orthologs. We find that this step is helpful in that it provides recognizable gene names to what are otherwise frequently arbitrary designations, and enables comparison of locus-by-locus results between SCRMshaw outputs from multiple species. Note that target gene assignments are based solely on nearest-gene relationships, which is a simple but often non-accurate method of determining enhancer targets. Users may want to incorporate additional information into their analysis if accurate enhancer-target relationships are a primary goal of their study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The SCRMshaw pipeline.

The left side shows pre-processing steps, the right side, post-processing. Input to SCRMshaw consists of the genome sequence and gene annotation. A protein sequence annotation is supplied later for the ortholog mapping step. Adapted from [19].

https://doi.org/10.1371/journal.pone.0311752.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Post-processing.

Post-processing is described in [14], with modifications in [19]. (A) Scores from each of the 25 individual SCRMshaw instances are assessed and (B) any 500 bp window whose score is below the value of the 5000^th ranked score is eliminated by having its score reset to zero (B). (C) The “elbow” point of the SCRMshaw score curve of the 5000 top scores from each instance is then determined, and (D) any scores below the elbow point are reset to zero. (E, F) After these two rounds of score evaluation, windows are grouped together and (G) subjected to peak calling on 10 bp intervals. (H) Final predictions are chosen as peaks with an amplitude above the elbow point of the amplitude curve, represented by a red dot. Adapted from [19].

https://doi.org/10.1371/journal.pone.0311752.g002

Materials and methods

The protocol described in this peer-reviewed article is published on protocols.io (doi:10.17504/protocols.io.e6nvw1129lmk/v2) and is included for printing as S1 File with this article.

Supporting information

Acknowledgments

We thank members of the Halfon lab for helpful comments on the protocol and the manuscript.