NetCIS: a network-based common insertion site pipeline for case–control Sleeping Beauty screens

Abstract

Forward genetic screens are powerful tools for novel biological discoveries because they do not require a priori knowledge of genes controlling a phenotype of interest. The Sleeping Beauty (SB) transposon system is one of many commonly employed genetic screen tools used to induce both gain and loss of gene function mutations via insertional mutagenesis. In the past, SB screens have been successfully utilized to discover several cancer driver genes that could serve as potential therapeutic targets. More recently, SB technologies have been extended to other applications, including the discovery of mechanisms contributing to drug resistance and novel targets for immunotherapy. Importantly, existing tools for SB screen data analysis do not support data-driven case–control comparisons capable of analyzing screens with this design. To this end, we developed a network-based common insertion site (NetCIS) analysis tool to robustly identify common insertion sites (CISes) in a case–control phenotype selection screen. NetCIS uses an efficient graph-based algorithm for discovering statistically significant CISes that differ between cases and controls. We benchmark NetCIS against other insertional mutagenesis analysis tools using a previously published SB dataset and show that NetCIS is able to prioritize previous biologically validated genes similarly or better than competing methods, is the only method that effectively utilizes a case–control experimental setup, and can identify CISes involving unannotated regions of the genome that are overlooked by existing methods. Code for NetCIS and an accompanying tutorial can be found at https://github.com/RogersLabGroup/NetCIS.

Introduction

Genetic screens are essential for understanding the complex relationships between genes and phenotypes, enabling unbiased discovery and yielding functional insights in a high-throughput manner. Many complementary screening technologies exist, including chemical and radiation mutagenesis, targeted gene editing with Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR), and insertional mutagenesis using virus or transposon vectors. Sleeping Beauty (SB) is one of such transposon-based insertional mutation systems ideal for whole genome mutagenesis across vertebrate species [1]. SB screens are driven by two elements, the DNA transposon and the transposase enzyme, allowing for simple delivery in vitro. Likewise, in vivo screens can be easily implemented as transgenic SB mouse lines carrying transposon and inducible transposases that are widely available and have been successfully used for over 20 years [2–4]. SB transposons are unbiased and insert randomly at thymine/adenine (TA) dinucleotides throughout the genome. The transposon was engineered to include regulatory elements that enable both disruption and overexpression of the full-length or truncated coding sequence, depending on where the insertion lands within a gene, the orientation of the transposon, as well as the skew and kurtosis of the distribution of insertions [5]. In addition, the transposon itself can be mapped to individual insertion sites (ISes) using high-throughput sequencing approaches targeting the transposon sequence, facilitating easy identification of causal genes [6].

The SB system was engineered for insertional mutagenesis in vertebrate cells and initially employed as an in vivo forward genetic screen tool for cancer driver gene discovery [7, 8]. In this context, SB mutations can induce both hematopoietic and epithelial malignancies that accurately model the evolutionary pressures and genetic heterogeneity of human tumors [9]. To identify which mutations are potential drivers of malignancy, SB insertions are typically compared across independent samples, and sites where insertions are more frequent than expected by chance—either by an assumed statistical distribution [10], a simulated null distribution [11, 12], or a rule-based mapping derived from insertion orientation [13]—are deemed common insertion sites (CISes).

Many computational tools have been developed for CIS identification after insertional mutagenesis, including tools specifically designed for SB (IAS [5], TAPDANCE [10]) and those with more generalized insertional mutagenesis methods in mind (KC-RBM [13], TRANSIT [11]). Importantly, each of these existing methods has significant limitations including the reliance on annotation quality, one-to-one mappings of insertions to genes, insertion location distribution assumptions, and an inability to directly compare cases that underwent genetic selection with control or unselected cases [5, 10–16]. IAS [5] assumes a uniform null distribution over the set of all TA sites and calculates significance using a Fisher’s Exact method comparing insertion events at reference TA sites against the total number of reference TA sites, which may inflate statistical results. In addition, IAS only evaluates ISes that can be annotated to TA sites in the GRCm38 reference, and a one-to-one mapping of ISes to genes is enforced. TAPDANCE uses multiple window sizes to identify CISes and the Poisson distribution to calculate three P-values for: the insertions, the library, and the region, while applying multiple test corrections. Insertions are only annotated to genes if they are within a 20 kilobases (KB) distance and the genes are tested for significance using Fisher’s exact test. Similarly, KC-RBM identifies CISes using four window sizes based on the transcription start site of a gene for an insertion that is upstream sense, upstream antisense, downstream sense, or downstream antisense. Insertions are associated with a gene in a one-by-one mapping. Wilcoxon’s test is then used to determine optimal window sizes for the SB transposon using gene expression values with and without a given insertion with a 5% significance threshold determined by permuting gene expression values across samples 10 000 times. Notably, the output from KC-RBM does not test for statistical significance. TRANSIT defines CISes as gene regions that can be comparatively tested through a permutation resampling method or a Wilcoxon’s rank-sum test but only counts ISes if they are mapped to a reference TA site.

In this paper, we present a new method, called network-based common insertion site (NetCIS) Pipeline, which is explicitly designed for case–control analysis of SB screens and includes features to address several other limitations of existing tools. We tested the utility of this method through reanalysis of a previously published case–control SB screen designed to identify molecular drivers of T-cell infiltration into solid tumors [17]. We demonstrate that NetCIS can effectively discover differential CISes using a rigorous statistical criterion, including both previously validated genes known to enhance T-cell infiltration as well as ISes in unannotated regions.

Materials and methods

Screen design and sample collection

A recent SB screen was performed to identify T-cell intrinsic mutations that would improve tumor infiltration of adoptively transferred T-cell therapies [17]. This screen utilized a case–control design whereby cases were anatomically selected for location inside a solid tumor (infiltrated) and controls were located in the spleen (not infiltrated). Thus, mutations in intratumoral T-cells (case) could be directly compared to mutations in T-cells that did not infiltrate the tumor (control) to identify mutations that may be therapeutic targets to improve T-cell trafficking. Each mouse with SB-mutagenized T-cells was bilaterally challenged with tumors on the left (LT) and the right flanks (RT). At the endpoint, the LT, RT, and spleen (S) were harvested, and transposon-gDNA junctions were amplified using ligation-mediated polymerase chain reaction (LM-PCR) [5, 17, 18]. This study included 132 total mice with three different tumor models: B16 (n = 81), EL4 (n = 43), and LLC (n = 8). Not all tissues were harvestable, and thus, 312 out of 396 possible tissue samples were obtained. Of the 132 total mice, 17 were treated with anti-PD1. The breakdown of the individual counts for the tumor model, treatment group, anti-PD1 treatment, and the number of samples that passed the preprocessing quality control are further detailed in Supplemental Table 1. Our preprocessing quality control methods did not remove any tissue samples.

To address the limitations of prior SB screen tools, NetCIS analyzes case–control insertion data through a four-step pipeline (Fig. 1). The first step is to preprocess the fasta files into aggregated insertion count data that are normalized. Next, we generate pseudo-common insertion sites (pCISes), which are regions of insertions sites that are grouped together but have yet to be evaluated in a statistical case–control manner. Then, for each pCIS, we find other, overlapping pCISes in the corresponding case/control set, which collectively define each CIS. Finally, we analyze each CIS for statistical significance and run gene set enrichment on the genes adjacent to significant CIS events in cases and in controls.

Figure 1.

High-level overview of the NetCIS pipeline. Created with BioRender.com.

Step 1: preprocessing reads

The first step is to preprocess the fastq files into aggregated insertion count data that are normalized. LM-PCR was used to create an amplicon library for high throughput sequencing [5, 17, 18]. LM-PCR generates paired-end fastq files for the genomic junctions with the inverted repeat left (IRL) and inverted repeat right (IRR) regions of the transposon. Cutadapt (v. 4.7) [19] is used to trim either the IRL or IRR transposon sequence and the ligated adapter/primer sequence. Remaining genomic sequences are mapped to the reference genome using Bowtie2 (v. 2.5) [20] using the Mus musculus GRCm39 (v. June 2020, available here: https://www.ncbi.nlm.nih.gov/grc/mouse) genome assembly of chromosomes 1–19, X, Y, and M. The reads are then filtered with a mapping quality threshold, which is the −10 log 10 probability that the mapping position is incorrect (e.g. a value of 13 indicates a .05 probability of a mapping error). Also, only reads that are properly paired are retained. This results in IRL and IRR bam files that are then processed into individual insertions. For each insertion, the orientation of the promoter region is determined based on the genomic strand and the sequencing library (IRL/IRR) to interpret whether a CIS is disrupting or overexpressing a coding sequence. In general, disrupting insertions for a gene does not show an orientation bias, as either orientation of the transposon can lead to protein disruption, while overexpressing insertions will nearly all be in the 5′ to 3′ orientation. We run the preprocessing procedure twice, one for each transposon orientation, on the IRL and IRR files. If the same read successfully maps to both the positive (same orientation of the genome) and negative (opposite orientation of the genome) genomic orientation, then this read is considered ambiguous, and it is removed. How we trim the corresponding adapter and primers based on the sequence library, transposon orientation, and paired read in Cutadapt is depicted in Supplemental Fig. 1. Read counts are normalized to counts per million (CPM) per chromosome, and IRL and IRR insertions are combined into one TSV file. After preprocessing reads into ISes, Step 2 takes these insertions and generates the first CIS networks.

Step 2: generating pseudo-common insertion site networks

The second step generates pCISes, which are regions of ISes that are grouped together based on their proximity to each other (based on genomic distance) but have yet to be evaluated for significance based on a case–control comparison. The insertions are separated by chromosome, and a graph structure is created per chromosome, where nodes are genomic ISes, and edges connect nodes based on a given genomic distance. Nodes intrinsically have unique IDs such that no two nodes with the same ID can be added to a graph. Therefore, individual IS are grouped by genomic position and contain the sum of read counts from all IS. Each IS is added as a node to the graph with additional information, including read counts and unique sample IDs of the insertions (Fig. 2A). After all nodes have been added, the genomic distance between each node and all other nodes is calculated. Edges are added between pairs of nodes with distances below a given threshold (Fig. 2B). We empirically tested the ranking of previously biologically validated genes (Aak1, Ehhadh, and Sprr1b, see Discussion for details) [21] for 5, 10, 20, and 50 KB thresholds and found that 20 KB provided the best overall ranking of these genes (Supplemental Table 2). Aak1 and Sprr1b were stably ranked highly across different thresholds, while Ehhadh dramatically improved from ranks of 838 and 881 for 5 KB and 10 KB to rank 6 for 20 KB and 50 KB. Previous work also shows that the fraction of significant genes found within SB insertional elements falls to baseline levels around 20 KB from the insertion event [13]. Thus, we recommend 20 KB as a starting point for users of NetCIS but do encourage exploration of other threshold values. The connectivity between nodes is then used to find subgraphs within the larger chromosomal graph (Fig. 2C). Each of these subgraphs is called a pCIS and includes the set of IS that spans a certain genomic distance. Each pCIS can be associated with either a case or control sample and is compared to other pCISes to discover which overlapping pCISes will be used to evaluate as a cohesive CIS.

Figure 2.

Visualization of how a pCIS is determined through finding subgraphs within the broader graph and how two pCIS are compared to each other to determine the overall CIS. (A) An example of eight ISes with corresponding read count values. (B) Using an edge threshold of 10, any IS can have an edge connecting it to any other IS that is within 10 bp. (C) These edges are used to determine subgraphs from the larger graph. In this case, three subgraphs mean three pCIS. (D) An example of overlapping case (red) and control (purple) pCIS. (E) ISes that are present in one pCIS but not the other are added with a read count of 0 and are denoted as an empty circle. (F) The sum of all samples’ read counts to each IS is used for the set of values for each case and control pCIS for statistical testing, in this instance, the rank-sum test. (G) if a case pCIS does not overlap with a control pCIS (or vice versa) then the set of read counts is compared against the same number of insertions with all zero read counts as shown.

Step 3: discovering common insertion sites

The third step uses the collection of all pCISes identified in Step 2 and finds overlapping pCISes in the opposite case or control set. A CIS is found by testing for any overlap between the genomic regions of a case pCIS with control pCISes (Fig. 2D). This results in a CIS that contains pCISes from both groups (one case and one control, one case and multiple controls, multiple cases and one control), or if no overlapping insertions were observed, a CIS containing a pCIS from only one group (case or control). If an overlap between the two pCISes was observed, each set of IS within the pCISes is expanded with the union of the IS between the pCISes (Fig. 2E). A read count of zero is recorded for an IS that is observed in one group only, which enables modeling of IS present in one pCIS but not observed in the other, while also preventing an inflated number of zero read count IS. This also yields a defined genomic region as a candidate CIS where the number of insertions in cases and controls can then be tested for statistical significance and ranked.

Step 4: common insertion site analysis

The fourth step analyzes each CIS for statistical significance and runs gene set enrichment on the genes associated with CISes in cases or controls. For each CIS, the genomic range, number of nonzero ISes, unique samples, and total read count are recorded for each treatment pCIS. A Wilcoxon’s rank-sum test with a two-sided alternative hypothesis is used to test if the read counts, or CPM, from cases versus controls for the pCISes are drawn from the same distribution (Fig. 2F). If there is not a corresponding pCIS to be compared against, e.g. a case pCIS does not have an overlapping control pCIS, then the rank-sum test uses an equal number of zeros to compare against the ISes from the case pCIS (Fig. 2G). A Fisher’s exact test with a two-sided alternative hypothesis is used to determine if, within a single CIS region, there is an association between the occurrence of a pCIS with nonzero read counts and case/control status. P-values are corrected for multiple hypothesis testing by controlling the false discovery rate using the Benjamini–Yekutieli method. All statistical tests used are implemented in Scipy (v. 1.12.0) [22]. Finally, CISes are ranked using a rank aggregation of three metrics: rank-sum significance before correction, total read count, and total number of samples. To do the rank aggregation, we used the python package ranky, using the Borda count median method [23]. The Fisher’s exact test was not used for aggregation because it would fail to identify a CIS as significant if all possible samples in the case and control groups were present in that CIS. We hypothesize that this scenario would be one that is biologically interesting, and we did not want a metric to rank this poorly. However, we keep the Fisher’s exact test in our analysis as it may still provide valuable information regarding differences in the case/control status of the mice involved in a CIS.

In a key difference from prior gene-centric methods like IAS, a CIS does not have to be annotated to a genomic marker for the CIS to be included in the analysis. However, we have provided an option in NetCIS when running analysis.py to remove CISes with no genomic annotation with the --remove_empty_cis flag. This does not result in a difference in the relative ranking of annotated CISes, but with a reduced total number of possible statistical tests to run, results in less stringent adjusted p-values and a better (smaller) rank for the annotated CISes compared to instances where the unannotated CISes were kept in. Annotation was performed using the Mouse Genome Informatics (MGI) genetic markers list file named MRK_List2.rpt [24]. The distance an IS can be annotated to a gene body varies, with previous work ranging from 3–301 KB [10, 13, 15, 25, 26]. Because there is no consensus for how near a gene body an insertion must be to confidently attribute it to the gene, NetCIS allows user choice of edge threshold, which not only is used to annotate a CIS to a genetic marker but is also used as a metric to connect IS together into a pCIS. An edge threshold of 20 KB was chosen in this study, as this empirically recapitulated experimentally validated results and agrees with previous literature. The annotation file gives two options (marker and feature type) to be annotated. Depending on a user's selection, there may be very few or many gene hits for a given CIS. Selecting ‘Gene’ marker type and leaving the feature type empty is recommended.

Gene set enrichment

Annotated CISes are selected for gene set enrichment if they are either significant in rank-sum or Fisher’s exact tests with a P-value less than 0.05 before correction. The CISes are then split into two groups based on their enrichment in either LT or S treatment groups based on the log fold change of the read count difference between the case and control pCIS that makes up the CIS. Pathway enrichment is run on the top 100 genes enriched either in cases or controls using Enrichr [27–29] implemented in Python by GSEApy [30]. The gene sets used for this analysis are the mouse ontology gene sets [31–33] as a file named m5.all.v2024.1.Mm.symbols.gmt from the Molecular Signatures Database [34, 35]. Gene sets with less than 10 or more than 300 genes, and redundant gene sets with a Jaccard dissimilarity less than 0.5, were removed since these sets tend to be less meaningful and can add unnecessary multiple hypothesis tests [36]. A background list of genes is created from all the genes annotated to all CISes.

Addressing local hopping bias

NetCIS directly compares a case sample set to a control sample set, such that insertions present in both groups are evaluated for differences in quantity and distribution. This proves effective in handling a known SB bias called local hopping [37]. Local hopping events are insertions that occur close to the original transposon location, as a nearby insertion location is more favorable than a distant location. Local hopping is expected to occur in both the case and control sample sets. Thus, direct comparison will not result in calling local hopping events as significant CISes.

Sequence depth quality control thresholds

There are no rigid criteria to differentiate between a ‘bad’ and a ‘good’ quality SB dataset. However, the nonzero mean for the read counts should be a minimum of 10 CPM, and ideally, greater than 50 CPM [11]. All samples we use to demonstrate our method in this study meet the 10 CPM threshold, and the B16 tumor model was found to meet the 50 CPM threshold. These averages and other basic statistics have been split into tumor model and treatment group and can be found in Supplemental Table 3.

Benchmarking network-based common insertion site

NetCIS was benchmarked against three existing insertional mutagenesis analysis tools (IAS, KC-RBM, and TRANSIT). TAPDANCE, though conceptually relevant, has several dependencies on deprecated code and was not implemented for benchmarking. Mouse reference genome version GRCm39 was used for all tools except IAS, which utilizes a custom file structure based on GRCm38. The result file for IAS can be found at the original publication as Supplemental Table 2 [17]. Result files for KC-RBM and TRANSIT are found in Supplemental Files 2 and 3.

Results

To address limitations of existing tools for case–control SB forward genetic screen analysis, we developed a NetCIS method for analysis of SB case–control screens by adapting previous work that leveraged graphs to represent retroviral transposon insertion events [38]. NetCIS uses graphs to identify pCISes in case and control samples. A pCIS is a set of ISes that are grouped together by their genome proximity. These pCISes are used to determine a CIS, where a case pCIS overlaps with a control pCIS as determined by their respective genomic regions defined by their IS, or in a situation where there is no overlap in a pCIS. CISes are tested for significance by comparing their read counts per ISes with the Wilcoxon’s rank-sum test as well as the number of sample cases and controls that are found in the CIS using Fisher’s exact test. The final list of candidates is ordered by implementing a rank aggregation approach on the statistical test results.

NetCIS does not require that insertions are annotated to a reference TA site nor that a CIS is annotated to a gene or other genomic feature. NetCIS also minimizes the impact of annotation quality by tracking ISes by location, independent of statistical significance, annotation status, or mapping to a reference TA site. As such, NetCIS allows one-to-many mapping of ISes to genes that may more accurately reflect biological systems, given that protein-coding genes are frequently overlapping in many species, including mammals [39]. Finally, NetCIS utilizes a case–control experimental design whereby observed, unselected insertion patterns derived from a control screen serve as the baseline for comparison with insertions observed under selection conditions to control for SB biases like local hopping. NetCIS is implemented in Python 3.11 and has been optimized to be computationally efficient. More details on the NetCIS method are described in the Methods section.

Insertion sites require nonparametric tests

The IS read counts are normalized to CPM for each tumor model, tissue source, and chromosome. The number of ISes per library differed substantially between IRL and IRR, with up to 10 times more ISes found in the IRL library (Table 1). While the reason for this is unclear, we speculate that this could reflect that the IRL library primer is more efficient than the IRR primer. All downstream analysis was based on the joint set of insertions (IRL IRR). Kernel density plots were created for all tumor models (Fig. 3A), B16 (Fig. 3B), EL4 (Fig. 3C), and LLC (Fig. 3D) and are not normally distributed. Overall, the distributions are relatively similar between the LT and RT treatment groups, with slight differences from the S group, which reflects samples derived from the spleen. Read counts per sample are normalized using a CPM measure (see Methods for details), and thus, any discoverable CISes will not be due to overall differences in the read count distribution between the mutagenized T-cells that can traffic intratumorally versus circulating T-cells that can be found in the spleen. The observed distributions tend to exhibit multiple modes, confirming the need for nonparametric tests such as Wilcoxon’s rank-sum test.

Table 1.

Number of ISes per library (IRL, IRR) and per treatment (LT, RT, S).

Library	Treatment	Number of ISes
IRL	LT	51 095
	RT	49 523
	S	34 435
IRR	LT	5000
	RT	4522
	S	11 295

Figure 3.

Kernel density plots of the read count or count per million (CPM) for tumor models (A) all, (B) B16, (C) EL4, and (D) LLC. Treatment groups LT, RT, and S.

Network-based common insertion site finds statistically significant common insertion site

The insertions for each treatment group were processed into a graph per chromosome, and pCISes were identified as distinct subgraphs within the larger graph. This resulted in 9867, 9358, and 11 844 total pCISes for LT, RT, and S, respectively (Table 2). To explore potential CISes, we analyzed the overlap of pCISes between treatment groups based on their genomic coordinates.

Table 2.

Number of pCISes per treatment group (LT, RT, S).

	LT	RT	S
Total pCIS	9867	9358	11 844
pCIS with CPM > 10	4816	4882	5889
pCIS with CPM > 50	3400	3470	4019

Also shown are the number of pCISes with an average CPM greater than 10 and greater than 50.

A possible CIS is determined by finding where a pCIS from one treatment group overlaps with another pCIS in the other treatment group based on the physical genomic range of their constituent IS (Fig. 2D and E). After determining overlapping pCIS, the resulting combination of IS are all considered part of the same candidate CIS. For the samples we analyzed, this process resulted in 18 080, 17 495, and 15 327 CISes for LT-S, RT-S, and LT-RT respectively (Table 3). For the sake of brevity, subsequent results focus only on the LT treatment group compared to the S control group. Results for LT-S, RT-S, and LT-RT comparisons can be found in Supplemental File 1 and are accessible from the Data Availability section.

Table 3.

Number of CISes per case and control group (LT-S, RT-S, LT-RT); also shown are the number of CISes with an average CPM greater than 10 and greater than 50.

	LT-S	RT-S	LT-RT
Total CIS	18 080	17 495	15 327
CIS with CPM > 10	9006	8921	7736
CIS with CPM > 50	6122	6109	5503

Of the 18 080 CISes resulting from the LT-S contrast, 8053 are enriched in the left tumor and 10 027 in the spleen (Table 4). Statistical significance of a CIS was determined using Wilcoxon’s rank-sum test on the normalized read counts of the left tumor and the spleen, which tests if the observed read counts for any IS mapping to the candidate CIS region for the two samples being contrasted (in this case, the left tumor and spleen) are drawn from the same distribution. Further, NetCIS tests if the proportion of samples in the CIS that are from the left tumor and the spleen is significant using Fisher’s exact test. This is completed by constructing a contingency table of the total case and control samples and the number of cases and control samples that have IS included in the candidate CIS of interest. These results are then visualized through volcano plots with the rank-sum before multiple test correction (Fig. 4A), rank-sum after correction (Fig. 4B), Fisher’s exact before correction (Fig. 4C), and Fisher’s exact after correction (Fig. 4D). Applying the rank-sum test results in 2021 significant CISes (P < .05) before correction and 24 significant CISes (FDR < 0.05) after correction, while the Fisher’s exact test results in 213 significant CISes (P < .05) before correction and 4 significant CISes (FDR < 0.05) after correction (Table 4). The Fisher’s exact test produces fewer significant results than the Rank-sum test, although there was an overlap of 152 CISes between the two tests before correction and an overlap of 3 CISes after correction. For this example, the rank-sum test identifies ~75% of what is discoverable using Fisher’s exact test. The Fisher’s exact test only uses the presence/absence information of a CIS event with nonzero read counts (and its association with the case/control status), whereas the rank-sum test directly uses read counts, which results in increased sensitivity. The strongest candidate CIS events would be identified by both approaches. For increased sensitivity, we used the rank-sum test results in the aggregated ranking of CIS (see Methods for details).

Table 4.

Total number of CISes that have passed either an uncorrected or corrected significant P-value of .05 for both the rank-sum test and Fisher’s exact test.

	All CIS	LT enriched	S enriched
Total CIS	18 080	8053	10 027
Significant rank-sum before correction	2021	904	1117
Significant rank-sum after correction	24	20	4
Significant Fisher’s exact before correction	213	117	96
Significant Fisher’s exact after correction	4	3	1

The number of CISes is further broken down into those that are enriched in LT ISes or S ISes.

Figure 4.

Volcano plots for rank-sum (A, B) and Fisher’s exact (C, D) tests, both with significance before correction (P-value <.05) (A, C) and significance after multiple hypotheses correction (FDR < 0.05) (B, D). Each dot corresponds to a CIS. Dots on the left represent ISes where total read count of the LT is less than the ISes in S, and vice versa for dots on the right.

The FDR correction excludes many CISes, some of which could be true positives; however, correction substantially reduces the number of false positives if present. This was deemed preferable to prioritize gene targets for labor-intensive wet-bench validation [21]. It is important to note that a complete gold standard list of genes that increase tumor infiltration of T-cells does not exist, and so we are unable to calculate the accuracy of FDR correction in identifying true or false positives. Additionally, increasing both sample size and sequencing depth will enable enhanced recovery of informative CISes.

Left tumor common insertion sites are enriched in adhesion pathways

Gene hits were generated as described and analyzed for pathway enrichment (Fig. 5). The top 10 significant pathways before correction are shown for the top 100 LT enriched genes (Fig. 5A). There is a cluster of cell–cell adhesion GO terms, two of which passed multiple hypothesis testing (FDR < 0.05): ‘GO BP Synaptic Membrane Adhesion’ and ‘GO BP Cell Cell Adhesion Via Plasma Membrane Adhesion Molecules’ at an adjusted P-value of .002 and .003 and with the genes Cdh19, Mdga2, Magi2, Cdh7, Cadm1, Lrrc4c, Cdh13, Atp2c1, Ptprd, Kirrel3 and Mdga2, Magi2, Lrrc4c, Cadm1, Ptprd, respectively. The enrichment in these two processes suggests that insertion events in genes involved in cell–cell adhesion may influence T-cell migration intratumorally and lead to enrichment of the T-cell population in tumors. The top 10 significant pathways before correction for the top 100S enriched genes are shown (Fig. 5B), but there were no pathways that passed a standard significance threshold (FDR < 0.05) after correction for multiple hypothesis testing.

Figure 5.

Gene-set enrichment network plots for the top 100 LT enriched genes (A) and the top 100 S enriched genes (B). Each dot represents a gene-set pathway, with the color corresponding to the hypergeometric p-value of the pathway, and the size of each dot corresponding to the hits ratio, the fraction of genes that overlap with the genes in each pathway. All these pathways are significant before correction, except for two pathways from the LT enrichment that passed a multiple hypothesis testing threshold: GOBP Synaptic Membrane Adhesion and GOBP Cell Cell Adhesion Via Plasma Membrane Adhesion Molecules.

Validation of network-based common insertion site through a contrast analysis

We hypothesized that left and right tumors would have highly similar, but not identical, CISes. If true, this would result in the identification of few or no significant CISes after multiple hypothesis correction. To verify that NetCIS can find real differences between cases and controls, a contrast analysis consisting of LT as the case group and RT as the control group was performed. T-cell clones that respond to tumors in an antigen-specific manner by trafficking to tumors would be expected to be observed in both left and right tumors. In addition, T-cell clones that infiltrate tumors via nonantigen-specific mechanisms (vascular leakage, chemokine gradients, etc.) are expected to be observed in one or both tumors. Our analysis resulted in many fewer significant CISes from the LT-RT contrast as expected (Fig. 6). Overall, this contrast analysis yielded only one significant (FDR < 0.05) CIS for rank-sum and one for Fisher’s exact test after multiple hypothesis testing correction. This was substantially fewer than the significant CISes produced when LT was compared to S. These results support the utility of the NetCIS pipeline for discovering true effects for sample contrasts where differences are expected.

Figure 6.

Volcano plots for rank-sum (A, B) and Fisher’s exact (C, D) tests, both with significance before correction (A, C) and significance after multiple hypotheses correction (B, D). Each dot is a CIS. Dots on the left represent ISes where total read count of the LT is less than the S, and vice versa for dots on the right.

Network-based common insertion site ranks validated genes similarly or better than other methods

The top 100 genes identified by each method were selected according to the following: NetCIS results were sorted by the aggregated rank and CISes that were not annotated to any genes were removed and re-ranked, IAS results were sorted primarily on FDR and secondarily on the average enrichment score favoring smaller FDR and larger average enrichment, KC-RBM results were sorted by descending order of number of hits, and TRANSIT results were sorted primarily by the adjusted P-value and secondarily on the absolute log-fold change favoring smaller adjusted P-values and larger absolute log-fold change. Concordance between methods was visualized by creating a 4-way Venn diagram of each method’s top 100 genes (Supplemental Fig. 2). NetCIS and KC-RBM had high concordance with other methods, with only 24 and 27 genes not shared with any other method, respectively. IAS and TRANSIT had less concordance, with 80 and 64 genes not shared with other methods, respectively. NetCIS and KC-RBM results were most similar, sharing 69 genes in the top 100. IAS had minimal overlap with other methods, which may be caused by reference gene annotation differences between GRCm38 (IAS) and GRCm39 (all other methods).

Using the same sorting criteria applied to each method’s results, the relative ranking of previously experimentally validated genes Aak1, Ehhadh, and Sprr1b were evaluated (Supplemental Table 4). Overall, NetCIS and IAS ranked these three genes highly in the top 10, with NetCIS ranking of Aak1, Ehhadh, and Sppr1b as 1, 2, 8, respectively, and IAS ranking them at 3, 1, and 6, respectively. Despite the lack of statistical significance criteria on which to rank KC-RBM results, KC-RBM still ranked Aak1, Ehhadh and Sprr1b as 1, 2, and 17, respectively. In contrast, NetCIS has a significant criterion that identifies 24 CISes of significance. Of these 24 hits, 15 are CISes that are unannotated and unable to be found by any method. Of the remaining 9 CISes, 10 potential genes are identified and KC-RBM can only reasonably find 5 genes in the top 26 genes based on the number of gene hits. TRANSIT was unable to effectively identify Aak1, Ehhadh, or Sprr1b, ranking them at 178, 1748, and 21 906, respectively.

Discussion

Data quality and preprocessing considerations

SB transposon mutagenesis is an effective tool for functional genomics, and case–control SB study designs are increasingly common. Existing computational methods to analyze screen results have not been designed to directly compare two observed insertion distributions in an unbiased way. NetCIS implements this capability, effectively handling SB system biases while simultaneously boosting signal and eliminating reliance on genome annotation.

Data quality is a function of biology underlying the screen, methodology used to generate sequencing libraries, and the sequencing method used. We wanted to discover insertions that promote increased intratumoral accumulation. These could be found in genes that support activation, proliferation, viability, trafficking, and we would expect to see selection for these in tumors relative to the spleen. In spleens, we expect to see selection for proliferation, viability, and things that would likely lead to T-cell malignancy, but not necessarily T-cell infiltration. Thus, there is no unselected sample, but the strength or degree of selection may vary, such as there being more stringent or obvious CISes found in the tumors. This selection would have an impact on the distribution of the normalized read counts.

We would also expect the sequence library generation to have an impact on the distribution of the normalized read counts. LM-PCR is an amplicon library that relies on PCR to amplify transposon and gDNA junctions. Thus, LM-PCR is also subject to PCR-induced biases, including primer efficiency (IRL versus IRR differences) and sequence composition that leads to difficulty in amplification (repetitive or GC rich regions). Interestingly, the IRL and IRR libraries generated using the same DNA template yielded different number of insertions, indicating that the IRL primer may be more efficient than the IRR primer. As such, the joint set (IRL IRR) of ISes was analyzed in the present study.

Functional insights of unannotated regions identified by NetCIS

NetCIS improves upon previous methods in a few ways. First, NetCIS does not discard insertions that cannot be annotated to a TA site within the genome but instead allows natural genetic variation within the insertional data. Secondly, NetCIS provides a flexible approach to analysis, allowing users to use different versions of reference data for CIS annotations. Lastly, NetCIS doesn’t limit CISes to a single gene, rather, NetCIS allows for potentially more complex CIS annotations by allowing CISes to be annotated to more than one gene, as well as not discarding CISes that could not be annotated.

Even though these changes differ from previous analysis tools, we showed that NetCIS is able to effectively analyze a case–control insertional mutagenesis screen in T-cells to identify genes that have been shown to support T-cell trafficking intratumorally (Aak1, Ehhadh, Sprr1b). Furthermore, including ISes that do not map to an annotated feature increases the discovery power of SB screens. In support of this argument, NetCIS also identified CISes in poorly annotated regions that could be transcription factors, super enhancers, or potentially an insertion changing the chromatin accessibility structure, that all could be missed due to the strict annotation requirements. As such, we found three highly ranking CISes that are associated with genes not found by IAS.

Indeed, the top ranked CIS was an unannotated region located upstream of cAMP responsive element binding protein 5 (Creb5), which was also annotated to another CIS at rank 26 (Fig. 7A). Creb5 functions in both T-cell development and migration [40, 41]. Initially, Creb5 was thought to be too distant from the top ranked CIS to be deemed a functional target; however, the CIS shares significant sequence identity with the first intron of the human CREB5 ortholog, suggesting the presence of a conserved regulatory sequence that is annotated in the human reference genome but unannotated in the mouse reference genome.

Figure 7.

Previously overlooked mutagenized regions identified by NetCIS re-analysis. (A) Murine Creb5 with its human ortholog is pictured with SB insertions shown in red and blue. The location of an identified CIS upstream of Creb5 is highlighted by a semi-transparent box. (B) Exon–intron schematic of murine Ptprd with SB insertions is shown in red and blue. (C) The genomic region preceding the microRNA Mir101c is shown with SB insertions in red and blue. Exon sites were obtained using IGV and SB insertions were plotted to the corresponding locations using a custom Python script.

NetCIS also identified Protein tyrosine phosphatase receptor type D (Ptprd) as a novel gene candidate (Fig. 7B) as it was found to be significant in both Rank-sum and Fisher’s exact tests (Fig. 4). SB mutations in Ptprd appear to be disruption of expression events, such that loss of Ptprd function is hypothesized to improve T-cell infiltration. Ptprd is a negative inhibitor of Stat3, and loss of Ptprd results in increased Jak–Stat signaling [42]. In accordance with this, increased Jak–Stat signaling has been shown to improve the persistence of therapeutic engineered T-cells [43], supporting a potential role for Ptprd loss in intratumoral T-cell accumulation.

Despite several advances made by NetCIS over existing methods, there are important limitations of the pipeline. First, NetCIS does not utilize insertion orientation to differentiate insertions at the same IS. This means that insertions on both the sense and anti-sense strands can be annotated to a mono-allelic gene. Second, NetCIS is inherently limited to datasets that have a case–control setup. Third, some regions of the reference genome exhibit bias in mapping [44]. Investigation of our top twenty-ranked CISes revealed three CIS regions located in reported problematic regions of the mm39 assembly of the M. musculus reference genome [45]. One of these was Mir101c (Fig. 7C), a noncoding microRNA significantly mutated more frequently in tumor (case) as opposed to the spleen (control) and exhibiting the lowest adjusted P-value of any CIS tested. While published literature suggests Mir101c may be biologically relevant to T-cell function [46, 47], further investigation is warranted due to its inclusion in a problematic region.

Future applications for network-based common insertion site

NetCIS is capable of identifying novel genomic regions influencing phenotypes that include unannotated loci. Excitingly, CISes in unannotated regions can be inferred to have some function, as the genetic selection implemented in the T-cell infiltration study favored transposon integration to these loci, and these data may be important for future study of these regions. While bulk or pooled sequencing in the dataset does limit direct observation of cis-regulatory networks (e.g. cooperative mutations within a single cell), future work can implement analysis of single-cell sequencing technologies to take advantage of precise insertional knowledge per cell. This would require an extension of NetCIS to process single-cell sequencing data, identify clonally expanded insertional profiles from the single cells, implement analyses such as comparing case and control clonal populations, and determine exact co-mutations relevant to the phenotype being tested. It is important to note that no matter what method is used, we are only able to speculate if the findings are novel and still require biological validation for confirmation.

Although NetCIS was created using a dataset biologically focused on T-cell infiltration, NetCIS is not limited to this specific use case. NetCIS is applicable to any case–control SB experimental design. A few examples of this could be in identifying cancer drivers in specific cancer contexts, genetic factors leading to cancer metastasis, or drug resistance mechanisms. To better enable future users of NetCIS, we provide an option for users to supply their own file of insertions. This means that users who already have insertional data or data that was generated with a different experimental design than what NetCIS expects for preprocessing can still use NetCIS for analysis. Details on this can be found at the end of the README.md file in the NetCIS code repository (see Data Availability).

Conclusion

In this paper, we present NetCIS, a NetCIS analysis method that is data-driven and can robustly analyze case–control SB screen sequencing results. NetCIS is able to successfully identify novel annotated CISes as well as CISes in unannotated regions in the context of SB genetic screens. However, unlike existing tools, NetCIS can effectively address biases of local hopping while reducing reliance on strict reference annotations. When tested functionally on data from a previously published SB screen, NetCIS performed strongly, recapitulating previously validated genes Aak1, Ehhadh, and Sprr1b, as well as identifying several novel gene candidates like Ptprd, Creb5 and Mir101c. While limitations exist for NetCIS, future adaptations will be made to resolve insertion orientation mapping for mono-allelic genes and utilize single-cell IS sequencing methods. Overall, NetCIS is an important step forward in the creation of a more comprehensive understanding of the relevant biological factors that underlie observed phenotypes in case–control SB screens and will be an important tool for disentangling the complex relationships between genes and phenotypes of interest.

Key Points

• NetCIS can analyze case–control Sleeping Beauty screens in an annotation-agnostic approach.

• Novel annotated and unannotated regions are discovered by NetCIS.

• We find previously validated genes are ranked highly by NetCIS.

Funding

This work was supported by the National Institutes of Health (CA276005) and the Minnesota Partnership for Biotechnology and Medical Genomics.

Data availability

NetCIS code and tutorial can be found at https://github.com/RogersLabGroup/NetCIS. Data used in the tutorial and Supplemental File 1, the full NetCIS results from testing 5, 10, 20, and 50 KB edge thresholds, can be found at https://doi.org/10.5281/zenodo.18024128.