A forward genetic screen identifies Sirtuin1 as a driver of neuroendocrine prostate cancer

Prostate cancer is the most common cancer in men, affecting one in six men in their lifetime. The vast majority present with regionally confined prostate adenocarcinoma that is often managed by active surveillance or local therapy, whereas for men with recurrent or advanced prostate cancer, the standard of care is androgen deprivation therapy (ADT) (Attard et al., 2015; Gelmann, 2002; Shen and Abate-Shen, 2010; Watson et al., 2015). Indeed, the androgen receptor (AR) is the most critical regulator of normal prostate differentiation as well as of all stages of prostate cancer progression (Abate-Shen and Shen, 2000; Gelmann, 2002; Shen and Abate-Shen, 2010). Consequently, prostate cancer treatments have been dominated by approaches to dampen AR signaling (Watson et al., 2015). However, while ADT initially leads to tumor regression, eventually tumors recur as castration-resistant prostate cancer (CRPC), so called because of its continued reliance on AR even in the absence of androgens (Scher and Sawyers, 2005). While further treatment of CRPC with second-generation anti-androgen therapies improves survival, many patients ultimately develop resistance and progress to aggressive disease variants that may no longer be dependent on AR (Watson et al., 2015). It is now well established that aggressive prostate cancer variants, including neuroendocrine prostate cancer (NEPC), arise through lineage plasticity (Beltran et al., 2016; Ku et al., 2017; Mu et al., 2017; Zou et al., 2017), defined as the transition from one differentiated cell state to another (Le Magnen et al., 2018). Thus, elucidating the mechanisms governing this process may improve treatment by overcoming plasticity-associated drug resistance.

We therefore sought to identify causal drivers of aggressive prostate cancer using a Sleeping Beauty (SB) forward genetics mutagenesis screen (Collier et al., 2005; Dupuy et al., 2005; Starr et al., 2009). SB is a two-component system composed of a mobile DNA transposon and a transposase enzyme that mediates transposon excision and random genomic insertion. Herein, biallelic mice carrying both the transposon and transposase are referred to as “active SB” mice (SB⁺), whereas mono-allelic mice carrying the transposase but lacking the transposon are designated “inactive SB” mice (SB^–). When both the transposase and transposon are present—as in the SB⁺ but not in the SB^– mice—the transposon can insert randomly into the genome and induce mutations. Insertions in or near cancer-relevant genes that aberrantly activate or inactivate these genes may promote or modulate tumorigenesis.

Unlike CRISPR or RNAi-based screens, which target known genes and usually in a defined direction (i.e., either gain- or loss-of-function), SB-mediated transposition is an unbiased approach in which mutations can occur anywhere in the genome in both coding and noncoding regions, and may result in either gain- or loss-of-function of nearby genes (Beckmann and Largaespada, 2020; Copeland and Jenkins, 2010). Furthermore, because mutagenesis occurs in autochthonous mice, mutations arise somatically during the natural course of disease progression and within the native tumor environment, analogous to the accumulation of mutations as occur during human cancer progression. These features make SB-mediated mutational screens well suited for uncovering novel mechanisms of cancer progression (Dupuy et al., 2009), as demonstrated by studies in colorectal cancer (Iida et al., 2024; March et al., 2011; Shimomura et al., 2023; Starr et al., 2009; Starr et al., 2011) as well as tumors of the central nervous system (Beckmann et al., 2019), pancreas (Mann et al., 2012), melanoma (Mann et al., 2015), and prostate (Ahmad et al., 2016; Rahrmann et al., 2009). However, SB mutagenesis has been less widely adopted than RNAi or CRISPR approaches, in part because the random nature of transposon insertions complicates identification of the genes driving the resulting phenotypes.

In the current study, we performed a SB transposon–based screen in a well-characterized autochthonous genetically engineered mouse model (GEMM) of prostate cancer based on haploinsufficient loss of Nkx3.1 and homozygous loss of Pten and Trp53 (NPp53 mice; Zou et al., 2017), modeling key genetic events frequently observed in advanced human prostate cancer (Abida et al., 2019; Cancer Genome Atlas Research, 2015). Compared with control SB-inactive NPp53-SB(−) mice, experimental SB-active NPp53-SB(+) mice develop more aggressive prostate cancer phenotypes with a high incidence of NEPC. Using an integrative systems biology approach—combining phenotypic, transcriptomic, and genomic analyses with state-of-the-art network-based algorithms—we have identified and validated mechanistic determinants of NEPC. Among these, we uncovered the nicotinamide adenine nucleotide (NAD)-dependent deacetylase sirtuin 1 (Sirt1). We demonstrate that expression of Sirt1 promotes NEPC while its silencing or pharmacological inhibition suppresses the NEPC phenotype. Overall, our study establishes a generalizable computational and experimental framework that integrates SB mutagenesis with phenotypic, genomic, and transcriptomic analyses to identify novel cancer drivers.

To identify mechanistic drivers of advanced prostate cancer, we used a SB mutagenesis screen, followed by integration of phenotypic, transcriptomic, and genomic analyses with systems biology algorithms. This integrative approach was combined with cross-species comparisons of mouse tumors and human cohorts to prioritize and functionally validate candidate genes (Fig. 1).

Step 1: We implemented the SB screen by crossing biallelic SB mice (Collier et al., 2005; Dupuy et al., 2005; Starr et al., 2009) with NPp53 mice (Zou et al., 2017) to generate the experimental NPp53-SB(+) and the control NPp53-SB(−) mice (Fig. 1, step 1). Phenotypic analyses revealed that the NPp53-SB(+) mice had more aggressive prostate cancer phenotypes than the control NPp53-SB(−) mice with higher metastatic incidence and a greater prevalence of NEPC compared with controls.

Step 2: To evaluate their conservation with human NEPC, we performed transcriptomic analyses comparing histologically defined NEPC and non-NEPC tumors from SB mice (Fig. 1, step 2). Differential gene expression analyses of these RNA sequencing (RNA-seq) profiles confirmed that mouse tumors classified as NEPC were molecularly distinct from those classified as non-NEPC; further, these mouse NEPC tumors recapitulated established gene expression signatures of human NEPC. We applied the VIPER algorithm to convert these RNA-seq profiles into protein activity signatures (Alvarez et al., 2016) and to infer master regulators (MRs) that are candidate drivers of NEPC. Comparison of MR signatures from the mouse tumors with corresponding signatures of human NEPC using the OncoMatch algorithm (Alvarez et al., 2018; Mundi et al., 2023; Vasciaveo et al., 2023) revealed significant conservation. Together, these analyses demonstrate that the mouse SB NEPC tumors engage molecular programs that are highly conserved with human NEPC.

Step 3: To identify SB transposon integration sites, we performed genomic sequencing of the mouse SB tumors and applied Transposon Annotation Poisson Distribution Association Network Connectivity Environment (TAPDANCE) analysis (Sarver et al., 2012) to identify common insertion sites (CIS)—i.e., those with statistically significant enrichment across multiple independent tumors (Fig. 1, step 3). We then mapped associated genes (“CIS genes”) and the location of the CIS within the genes. Differential gene expression analysis revealed a subset of CIS genes enriched in NEPC versus non-NEPC SB tumors.

Step 4: To identify CIS genes that represent candidate mechanistic determinants of NEPC, we implemented a cross-species integrative framework combining genomic and transcriptomic data from mouse and human NEPC tumors (Fig. 1, step 4). Specifically, integrating CIS mapping with MR transcriptomic analyses from mouse SB tumors, we applied the CINDy algorithm (Giorgi et al., 2014; Paull et al., 2021) to nominate CIS genes predicted to function as upstream modulators of NEPC MRs. Because such upstream modulators are more likely to exert causal control over transcriptional programs (Chen et al., 2014; Paull et al., 2021), this approach prioritized CIS genes with a high likelihood of driving the NEPC phenotype.

Step 5: Functional studies in human prostate cancer cell models and mouse organoid models using gain- and loss-of-function approaches in vitro and in vivo, as well as pharmacological inhibition, demonstrated that one of the top-ranked candidates, Sirt1, promotes NEPC (Fig. 1, step 5). Each of these steps is described in detail below.

We implemented the SB mutagenesis screen (Fig. 1, step 1) using the NPp53 mice (Nkx3.1^CreERT2/+; Pten^flox/flox; Trp53^flox/flox) (Fig. 2 A), a well-characterized autochthonous GEMM with prostate-specific loss of Pten and Trp53 (Zou et al., 2017) that models prevalent genetic events in advanced human prostate cancer (Abida et al., 2019). Tumor induction is spatially and temporally controlled in the prostatic epithelium via the Nkx3.1^CreERT2 allele (Wang et al., 2009b), which also confers haploinsufficiency for Nkx3.1, another common event in human prostate cancer (Cancer Genome Atlas Research, 2015). NPp53 mice develop poorly differentiated prostate cancer that can progress to CRPC following androgen deprivation, while further treatment using second-generation anti-androgens promotes more aggressive phenotypes, including NEPC (Zou et al., 2017). Given this latent potential toward aggressive prostate cancer, we reasoned that NPp53 mice would provide an ideal model to identify drivers of aggressive disease, including NEPC. As a control, we used the NP mice (Nkx3.1^CreERT2/+; Pten^flox/flox; Floc’h et al., 2012), which develop well-differentiated prostate adenocarcinoma that progresses to CRPC but does not advance to more aggressive phenotypes following androgen deprivation or anti-androgen treatment (Zou et al., 2017).

The SB system is comprised of two alleles: a conditionally activatable transposase in the Rosa26 locus (Rosa26^LSL-SB11) and a transgenic transposon allele (T2/Onc2) (Collier et al., 2005; Dupuy et al., 2005; Starr et al., 2009). Cre activation via the tamoxifen-inducible Nkx3.1^CreERT2/+ excises the stop cassette blocking SB11 expression, enabling transposase expression specifically in prostatic epithelium. The transposase then recognizes the terminal repeats of T2/Onc2, promoting its excision and random genomic integration (Dupuy et al., 2005; Starr et al., 2009). Therefore to generate the relevant SB mice, we crossed the Rosa26^LSL-SB11 andT2/Onc2 alleles with NPp53 mice to produce experimental NPp53-SB(+) mice (Nkx3.1^CreERT2/+; Pten^flox/flox; Trp53^flox/flox; Rosa26^{LSL-SB11/LSL-SB11}; T2/Onc2), which carry both transposase and the transposon, and control NPp53-SB(−) mice (Nkx3.1^CreERT2/+; Pten^flox/flox; Trp53^flox/flox; Rosa26^{LSL-SB11/LSL-SB11}), which lack the transposon (Fig. 2 A). Tamoxifen induction of NPp53-SB(+) mice results in Cre-mediated activation of the transposase, resulting in T2/Onc2 transposition, as confirmed by PCR (Fig. S1 A), whereas NPp53-SB(−) mice have no transposition (Fig. S1 A). For comparative analyses, corresponding NP-SB(−) mice (for Nkx3.1^CreERT2/+; Pten^flox/flox; Rosa26^{LSL-SB11/LSL-SB11}) and the NP-SB(+) mice (for Nkx3.1^CreERT2/+; Pten^flox/flox; Rosa26^{LSL-SB11/LSL-SB11}; T2/Onc2) were also generated and analyzed (Fig. S1, B and C).

We performed tamoxifen-mediated tumor induction of SB mice at 2–3 mo of age and monitored them for up to 12 mo or until they succumbed to prostate cancer, after which prostate tumors and other tissues were collected for phenotypic and molecular analyses (Fig. 2 A). Compared with the NPp53-SB(−) mice, the NPp53-SB(+) mice developed more aggressive prostate cancer, as evidenced by their histology, reduced survival, and increased metastasis (Fig. 2, B–I; Fig S1, D–K;,Table 1, and Table S1). Specifically, the NPp53-SB(+) mice displayed a significant reduction in survival relative to the NPp53-SB(−) mice (log-rank P value <0.0001; Fig. 2 B). In contrast, the NP-SB(+) mice showed no significant decrease in survival or acceleration of their prostate cancer phenotype compared with the NP-SB(−) mice (log-rank P value = 0.87; Fig. S1 B, Table 1, and Table S1), consistent with prior observations that they do not progress to advanced disease (Zou et al., 2017). NPp53-SB(+) mice also showed markedly higher metastatic incidence to lymph nodes, lung, and liver, predominantly in mice with NEPC tumors (P < 0.0001; Fig. 2, C–E; Fig. S1, F and G; and Table 1). These findings indicate that SB mutagenesis of NPp53 tumors promotes highly aggressive, metastatic prostate cancer with frequent NEPC, arising de novo in hormonally intact, untreated mice.

As is characteristic of NPp53 mice (Zou et al., 2017), histopathological analyses of NPp53-SB(−) and NPp53-SB(+) prostate tumors revealed heterogeneous phenotypes (Fig. 2, D, E, G, and H; Fig. S1, H–K;,Table 1, and Table S1). However, while NPp53-SB(−) tumors were predominantly squamous or sarcomatoid, similar to the baseline NPp53 tumors (Zou et al., 2017), a significant fraction of NPp53-SB(+) tumors displayed NEPC features, which are more rarely seen in baseline NPp53 tumors (P < 0.0001; Fig. 2 F, Table 1, and Table S1). Based on previously published histopathological criteria (Ittmann et al., 2013; Shappell et al., 2004), NEPC in mouse prostate tumors was classified based on characteristic histopathological features, including (1) nesting or trabecular growth patterns, (2) small cells with scant cytoplasm, and (3) high mitotic activity (Table S1). Further, tumors classified as NEPC in NPp53-SB(+) mice expressed canonical NEPC markers—Sox2, synaptophysin (Syp), chromogranin A (ChgA), and Sox2—alongside reduced AR expression (Fig. 2, G–I). Despite comparable tumor weights (Fig. S1 D), NEPC tumors were significantly more proliferative than non-NEPC tumors, as shown by Ki67 staining (P < 0.0001; Fig. 2 G and Fig. S1 E).

To characterize their transcriptomic phenotype (Fig. 1, step 2), we performed RNA-seq on NEPC or non-NEPC tumors from NPp53-SB(+) or NPp53-SB(−) mice (n = 27 and n = 8, respectively; Fig. 3, A–C; and Table S2 A). Differential gene expression and principal component analyses (PCAs) revealed that NEPC NPp53-SB(+) tumors clustered separately from both non-NEPC NPp53-SB(+) tumors and NPp53-SB(−) tumors (Fig. 3, A and B), indicating their distinct transcriptional state. Notably, genes that are differentially expressed in NEPC SB tumors included those linked to cancer progression or neuroendocrine differentiation in humans, such as RASSF1, which is down-regulated in the NEPC SB tumors and is frequently silenced by promoter hypermethylation across multiple malignancies, including prostate cancer (Liu et al., 2002), and CBX2, which is up-regulated in the NEPC SB tumors and in human prostate cancer has been shown to be associated with metastasis and poor clinical outcome (Clermont et al., 2016).

To directly assess their similarity with human NEPC, we compared RNA-seq profiles from mouse SB tumors with well-characterized gene signatures of human AR-driven and neuroendocrine CRPC (Labrecque et al., 2019). Unsupervised clustering revealed that mouse NEPC SB tumors were enriched for signatures of known NEPC subtypes, namely NEURO I and NEURO II, and have correspondingly reduced AR-related gene expression, whereas non-NEPC SB tumors displayed the opposite pattern (P < 0.05; Fig. 3 C and Table S2 B).

To extend these findings to protein activity-based analysis, we applied the VIPER algorithm (Alvarez et al., 2016) to infer differential protein activity from gene expression signatures of the NEPC and non-NEPC NPp53-SB(−) and NPp53-SB(+) tumors (Fig. 4 A and Table S3 A), using a mouse prostate cancer interactome previously constructed from murine tumor RNA-seq data (Vasciaveo et al., 2023). Protein activity– based clustering identified three molecularly distinct subtypes (C₁–C₃) composed predominantly of NEPC or non-NEPC tumors (Fig. 4 B and Table S3 B). Subtypes C₁ and C₂ consisted primarily of NPp53-SB(−) tumors and non-NEPC NPp53-SB(+) tumors, whereas subtype C₃ was composed primarily of NEPC NPp53-SB(+) tumors. Notably, C₃ tumors exhibited robust expression of canonical NEPC markers, including Syp and ChgA, as confirmed by quantitative PCR (Fig. 2 I and Fig. 4 B). Further, proteins differentially active in human NEPC—derived from a signature of metastatic CRPC patients that have both CRPC-adenocarcinoma (CRPC-adeno) and NEPC (CRPC-NE) (Beltran et al., 2016)—were significantly enriched among proteins differentially active in C₃ tumors relative to C₁ and C₂ (Fig. 4 B).

The proteins inferred to be most differentially active in C₃ tumors, which represent candidate MRs of NEPC, include those previously implicated in NEPC, such as Myt1 (Fig. 4 B) (Guo et al., 2019). Notably, 13 of the top 15 MRs in C₃, including Lmx1b, Dner, Snca, and Rab3c, have functions associated with neural development or nervous system function (Ding et al., 2003; Fischer von Mollard et al., 1994; Han et al., 2021; Hsieh et al., 2013; Siddiqui et al., 2016). In addition, the C₃ tumors exhibited high activity of MRs linked to aggressive prostate cancer, including Foxm1 and Cenpf, as well as NEPC regulators, including Nsd2, Foxa2, and Foxa1 (Fig. 4 B) (Han et al., 2022; Li et al., 2026; Vasciaveo et al., 2023).

To directly assess conservation between mouse C₃ NEPC tumors and human NEPC patients, we leveraged the OncoMatch algorithm (Vasciaveo et al., 2023) (Fig. 4, A and C; Fig. S2, and Table S4 A–D). For this analysis, we queried two well-characterized human prostate cancer cohorts that include NEPC patients: the Beltran cohort described above (n = 34 CRPC-adeno and n = 15 CRPC-NE, see above; Beltran et al., 2016) and the Stand Up to Cancer–Prostate Cancer Foundation study (SU2C) cohort consisting of posttreatment metastatic biopsies from mCRPC patients (n = 210 non-NEPC, n = 22 NEPC, n = 34 unclassified; Abida et al., 2019). Protein activity profiles for each human tumor were inferred using VIPER (Table S4 A–D) with a previously generated human prostate cancer interactome (Vasciaveo et al., 2023).

After deriving independent protein activity signatures for each mouse and human tumor, we assessed cross-species conservation by comparing the relative enrichment of their respective MR programs. Specifically, we used the aREA algorithm (Alvarez et al., 2016) to assess the 25 most activated (25↑) and 25 most inactivated (25↓) candidate MRs from each human tumor within the protein activity profiles of each mouse tumor (Table S4 A–D). A fixed MR set (25↑ + 25↓) was used to ensure comparability across cohorts because prior studies have shown that ∼50 MRs are sufficient to capture canalized, functionally relevant genetic alterations (Paull et al., 2021; Vasciaveo et al., 2023).

Following Vasciaveo et al. (2023), mouse tumors were classified as high-fidelity (i.e., conserved) models of human NEPC tumors if they met a stringent OncoMatch significance threshold (P ≤ 10⁻⁵, by Bonferroni corrected one-tail aREA test). Using this criterion, each mouse C₃ NEPC tumor represented a high-fidelity model of at least one human NEPC tumor (Fig. 4 C and Fig. S2). Conversely, high-fidelity mouse NEPC models were identified for 66.7% and 45.5% of NEPC samples in the Beltran (Fig. 4 C) and SU2C cohorts (Fig. S2), respectively. Together, these analyses define the transcriptional programs and MRs associated with NEPC in mouse SB tumors and demonstrate their conservation in human NEPC.

CIS refers to transposon insertion loci that are recurrent across multiple independent tumors and, therefore, are more likely to dysregulate genes causally associated with the accompanying tumor phenotypes (Ranzani et al., 2013; Sarver et al., 2012). Because transposon insertions occur randomly and at high frequency throughout the genome, statistical approaches—such as Poisson distribution–based analyses—are required to distinguish statistically significant CIS events from random background. Accordingly, to identify transposon-dysregulated genes associated with NEPC in mouse SB tumors, we performed genomic sequencing to map transposon insertion sites (>1 million total sites), followed by identification of the CIS (122 total CIS), and the assignment of CIS to nearby genes (330 genes, hereafter CIS genes) (Fig. 1, step 3).

Specifically, we carried out multiplexed Illumina sequencing of NPp53-SB(+) tumors (n = 74), using unique barcodes for each tumor sample (Table S5; see Materials and methods). Statistically significant CIS were identified with the TAPDANCE software (Sarver et al., 2012) and visualized using a Circos plot (Fig. 5 A and Table S6). This analysis identified 122 CIS distributed across multiple chromosomes on both DNA strands (Fig. 5 A and Table S6). We further used TAPDANCE to identify genes within 20 kb of each CIS. One CIS lacked a nearby gene, while the remaining 121 CIS mapped to 330 CIS genes, with most CIS associated with multiple genes; these relationships are illustrated in the outer ring of the Circos plot (Fig. 5 A and Table S6).

Because TAPDANCE provides precise CIS coordinates and insertion directionality (Sarver et al., 2012), we were able to analyze CIS strandedness and map the insertion sites relative to associated genes (Fig. 5, A–D; Fig. S3, A–C; Fig. S4, and Table S6). CIS insertion patterns at individual genes were visualized using lollipop plots, which depict both strandedness and location relative to transcription start sites (TSS), exons, introns, and putative upstream or downstream regulatory regions (Fig. 5 C; and Fig. S3, A and C).

As expected, given the random nature of transposition (Collier et al., 2005; Dupuy et al., 2005; Dupuy et al., 2009), CIS insertions were distributed across both the positive and negative strands and throughout the gene bodies at positions predicted to affect gene activity and RNA expression (Fig. 5, C and D; Fig. S3, A–C; Fig. S4, and Table S6). These included insertions at or near TSS (e.g., Evi2a, Tgfbr2, and Nisch), within or adjacent to exons (e.g., Sirt1, Arid1a, and Pp2r2a), within introns (e.g., Arid1a, Ezh2, and Msi2), and within putative 5′ or 3′ regulatory regions (e.g., Aifm3) (Fig. 5 C; and Fig. S3, A and C).

Additionally, in a subset of tumors for which both the genomic and RNA-seq data were available (e.g., Sirt1 and Msi2), the locations of CIS insertions were consistent with observed differences in their RNA transcript (Fig. S4), consistent with an impact on gene expression. Interesting, some insertions (e.g., Smurf2 and Tcf12) were predicted to generate gene fusions or alter isoform usage, potentially affecting gene function (Fig. S3 C and Table S6). Notably, CIS fusion events were more frequent in NEPC than in non-NEPC tumors (P = 0.045; Fig. 5 D and Fig. S3 B), suggesting selective enrichment of CIS in NEPC. Consistent with this, RNA-seq analysis comparing NEPC and non-NEPC SB tumors (Fig. 3) identified 75 CIS genes significantly enriched among differentially expressed genes (Fig. 5 B and Table S7). These NEPC-associated CIS genes were, therefore, prioritized for further analyses.

Among the 75 NEPC-associated CIS genes, we sought to identify those most likely to represent mechanistic determinants of NEPC and to be conserved in human prostate cancer (Fig. 1, step 4). Based on prior work showing that the causal events shaping tumor cells often act as upstream modulators of MR activity (Chen et al., 2014; Paull et al., 2021), we reasoned that CIS genes with causal roles in NEPC would preferentially function as upstream modulators of NEPC MRs (Fig. 5, E and F). We further hypothesized that conserved drivers could be identified through cross-species integration of genomic and transcriptomic data from mouse SB tumors with transcriptomic data from human prostate cancer cohorts (Fig. 5 E).

To identify CIS genes likely to function as upstream modulators of NEPC regulatory programs, we applied the CINDy algorithm (Giorgi et al., 2014), an updated implementation of Modulator Inference by Network Dynamics (Wang et al., 2009a). CINDy assesses the statistical significance of the conditional mutual information, $I[MR;T|M]$⁠, where MR denotes a candidate MR, T represents its transcriptional targets (regulon), and M is a candidate modulator protein (Fig. 5 E). CINDy analysis was performed on two complementary human patient cohorts, namely (1) the Cancer Genome Atlas (TCGA) cohort, comprising primary, treatment naïve prostate tumors (N = 498; Cancer Genome Atlas Research, 2015), and (2) the SU2C mCRPC cohort described above (N = 266; Abida et al., 2019). For cross-species integration, we interpreted these results in the context of the MR programs inferred from mouse NEPC and non-NEPC SB tumors (Fig. 4 B). Specifically, we filtered CINDy modulator–MR relationships to those in which (1) the candidate modulator (M) corresponded to one of the 75 NEPC-enriched CIS genes (human orthologs), (2) the candidate MR was a transcription factor with significant differential VIPER protein activity between mouse NEPC and non-NEPC SB tumors (Fig. 4 B), and (3) the CINDy P value was <0.05. CIS genes meeting all three criteria were classified as candidate upstream modulators of the NEPC master regulatory program (Table S7).

To prioritize these candidates, we applied Fisher’s method to integrate three independent lines of evidence: (1) CIS significance as assessed by TAPDANCE (Table S6), (2) differential expression between NEPC and non-NEPC tumors (Fig. 5 B and Table S7), and (3) CINDy-inferred significance as upstream modulators of NEPC MRs (Table 2 and Table S7). Candidates were ranked by their VIPER-inferred differential protein activity between NEPC and non-NEPC tumors (Fig. 5, E and F; and, Table 2, and Table S7), reflecting their predicted functional relevance to NEPC.

Among the prioritized candidates, we focused on the NAD-dependent deacetylase Sirt1 (Fig. 1, step 5). Sirt1 exhibited strong VIPER-inferred differential protein activity in NEPC versus non-NEPC tumors (VIPER score, 8.59, VIPER FDR, 4.2 X 10⁻¹⁸; Table 2), and a high percentage of CIS events in NEPC tumors (67%; Table 2). RNA-seq coverage analysis also suggested differences in the Sirt1 transcript between NEPC and non-NEPC tumors, although interpretation was limited by the small sample size (Fig. S4). Further, the predicted regulon of SIRT1 includes several established NEPC regulators, including MYT1, NEUROD1, INSM1, and ASCL1 (Fig. 6 A). Consistent with these findings, SIRT1 was highly expressed (>70% positive cells) in patient samples with neuroendocrine pathology but showed minimal expression (<5%) in non-NEPC tumors (Fig. 6 B). Based on this convergent evidence, we evaluated the functional relevance of SIRT1 using gain- and loss-of-function approaches in human prostate cancer cell lines induced toward NEPC and in a mouse NEPC organoid model (Figs. 6, 7, 8, and S5).

First, we examined the effects of SIRT1 loss-of-function in Lymph Node Carcinoma of the Prostate (LNCaP) cells (Horoszewicz et al., 1980), an AR-dependent prostate cancer cell line that can be induced to differentiate to a neuroendocrine-like phenotype when deprived of androgens and treated with agents that increase the intracellular levels of cyclic AMP (hereafter denoted +cAMP; Fig. 6 C) (Bang et al., 1994; Burchardt et al., 1999; Shen et al., 1997). SIRT1 is expressed at low levels prior to NEPC induction (day 0) but was significantly increased following induction (day 7). This induction-associated increase in SIRT1 expression was effectively suppressed by two independent SIRT1-directed shRNAs (Fig. S5 A).

In the control LNCaP cells, NEPC induction resulted in increased expression of genes comprising the SIRT1 regulon that are predicted to be activated by SIRT1 (ONECUT2, MYT1, NEUROD1, INSM1, and ASCL1), and a corresponding decrease in expression of those predicted to be down-regulated by SIRT1 (TWIST1, SNAI1, TP53, RB1, and TWIST2) (Fig. 6 D). However, this change in gene expression was abrogated by silencing SIRT1 (Fig. 6 D). Furthermore, NEPC induction of the LNCaP cells (day 7) resulted in increased expression of NEPC markers (SOX2, CHGA, NEUROD1, ASCL1, and INSM1), which was abrogated by depletion of SIRT1 (Fig. 6 D), supporting the idea that SIRT1 supports induction of NEPC differentiation of LNCaP cells.

To extend these findings, we performed gain-of-function studies using a human prostate cancer cell line, 22RV1 (Sramkoski et al., 1999), which exhibits low SIRT1 expression in both untreated cells (day 0) and following NEPC induction with +cAMP (day 7) (Fig. 6, E and F; and Fig. S5 B). We, therefore, investigated whether enforced SIRT1 expression promotes NEPC differentiation in vitro and in vivo following orthotopic implantation (Fig. 6, E–J; and Fig. S5, B–F). SIRT1 was activated in 22RV1 cells using a previously described CRISPR activation (CRISPRa) system (Arriaga et al., 2024). Cells expressing either a control sgRNA or two independent SIRT1-targeting sgRNAs were generated, both of which produced robust increases in SIRT1 mRNA and protein levels (Fig. 6, F and G; and Fig. S5 B). Although SIRT1 activation had modest or negligible effects on cell proliferation and colony formation (Fig. S5, C, E, and F), it induced a marked up-regulation of key NEPC markers, including SOX2, CHGA, and NEUROD1 (Fig. 6 G).

Consistent with these in vitro findings, orthotopic implantation of 22Rv1 CRISPRa cells into the prostates of male host mice revealed only modest effects on tumor growth (Fig. 6, I and J), but a striking increase in expression of NEPC markers, including SYP and INSM1, with minimal effects on AR expression (Fig. 6 H and Fig. S5 D). Together, these results demonstrate that gain-of-function of SIRT1 promotes NEPC differentiation in human prostate cancer models.

To more directly assess the role of Sirt1 in NEPC, we developed a mouse organoid model with an overt NEPC phenotype (Fig. 7). This organoid was derived from a baseline (i.e., non-SB) NPp53 mouse that developed a rare spontaneous NEPC tumor, as defined by characteristic histopathological features, low expression of AR, pan-CK, and vimentin, and high expression of the neuroendocrine markers Insm1 and Syp (Fig. 7 A). The derived organoid faithfully recapitulated the parental tumor phenotype, retaining low AR, pan-CK, and vimentin expression and high Insm1 and Syp expression (Fig. 7 A). Notably, Sirt1 expression was high in this NEPC organoid model (Fig. 7 B).

To directly assess the functional role of Sirt1 in maintaining the NEPC phenotype, we silenced its expression in the NEPC organoid model using an inducible shRNA and evaluated the effects on NEPC differentiation and tumor growth both in vitro and in vivo (Fig. 7, B and C; and Fig. 8, A–G). Sirt1 silencing resulted in a striking reversal of neuroendocrine features, evident by increased expression of AR, pan-CK, and vimentin, accompanied by reduced expression of the neuroendocrine markers Insm1 and Syp, both in vitro and following subcutaneous growth in vivo (Fig. 8, C and F). This phenotypic reversion was associated with a profound reduction in tumor growth in vivo (Fig. 8, D and E). These findings demonstrate that Sirt1 is required for maintenance of the NEPC state and that its suppression is sufficient to revert the NEPC phenotype and inhibit NEPC tumor growth.

Given that Sirt1 depletion both prevented NEPC induction and reversed established NEPC, we next asked whether pharmacological inhibition of SIRT1 could phenocopy these effects in human prostate cancer cells and mouse NEPC organoids (Fig. 8, H and I; and Fig. S5, G–I). For these studies, we used the SIRT1-specific inhibitor selisistat, which has been shown to inhibit SIRT1 lysine deacetylase activity in multiple cellular contexts (Adams et al., 2024; Napper et al., 2005; Westerberg et al., 2015). In our models, selisistat effectively inhibited lysine deacetylase activity at 5–10 μM without compromising overall cell viability (Fig. S5 H), consistent with doses used previously (Adams et al., 2024).

Pharmacological inhibition of SIRT1 abrogated NEPC induction in LNCaP cells and reversed the NEPC phenotype in the mouse organoid model (Fig. 8, H and I; and Fig. S5, G–I). This was reflected by altered expression of SIRT1-regulated targets, including ONECUT2 and MYT1 (Fig. S 5 I), as well as decreased expression of neuroendocrine markers such as SOX2,INSM1, and ASCL1 (Fig. 8 I and Fig. S5 I). Collectively, these findings demonstrate that SIRT1 promotes NEPC differentiation, while its genetic or pharmacological inhibition suppresses or reverses the NEPC phenotype.

Forward genetic screens, particularly those based on SB mutagenesis, have proven powerful for unbiased identification of cancer drivers and therapeutic targets (Beckmann and Largaespada, 2020; Copeland and Jenkins, 2010). Because SB mutagenesis is implemented in autochthonous mouse models, it enables the discovery of genetic events as they arise somatically during disease progression within the native tumor microenvironment, closely mirroring the evolutionary processes that occur in human cancers. In this study, we employed a SB mutagenesis screen to identify molecular drivers of NEPC, an aggressive prostate cancer variant that is increasing in incidence as a consequence of therapy resistance (Abida et al., 2019; Watson et al., 2015). By integrating tumor phenotypic analyses with transcriptomic profiling and genomic mapping of transposon insertion sites in a cross-species framework, we identified the NAD-dependent deacetylase SIRT1 as a targetable driver of NEPC.

Unlike prior SB mutagenesis screens in prostate cancer mouse models, which were based on loss of the Pten tumor suppressor gene alone (Ahmad et al., 2016; Rahrmann et al., 2009), our study employed a model with concurrent loss of Pten and Trp53 (NPp53 mice), which is predisposed to develop aggressive disease (Zou et al., 2017). Using the NPp53 model as the basis for the screen enabled identification of drivers of advanced prostate cancer and, in particular, facilitated interrogation of the molecular determinants of NEPC, a highly lethal subtype that has emerged in the context of treatment resistance (Beltran et al., 2019; Watson et al., 2015). Notably, NEPC phenotypes in NPp53-SB(+) mice arise in the setting of intact androgen signaling and in the absence of therapeutic intervention, providing a unique opportunity to investigate the de novo cellular events that give rise to lethal prostate cancer.

A major limitation of SB and other forward genetic screens is the challenge of prioritizing functional drivers from large candidate gene sets. Our integrative framework addresses this limitation by combining SB–derived genomic and transcriptomic data with regulatory network–based analyses across both mouse and human prostate cancer, enabling systematic identification of CIS genes most likely to control disease-relevant transcriptional programs. Specifically, we integrate phenotypic, genomic, and transcriptomic analyses of mouse SB tumors with human prostate cancer cohorts to identify and prioritize CIS genes with mechanistic roles in driving NEPC. This work builds on our prior studies demonstrating that cross-species reverse engineering of transcriptomic data can uncover gene regulatory networks and nominate modulators that influence transcription factor activity, thereby revealing key determinants of cancer (Aytes et al., 2014; Giorgi et al., 2014; Paull et al., 2021; Vasciaveo et al., 2023).

Although this study focuses on SIRT1, our screen identified several additional candidate CIS genes of interest. Some have been previously implicated in NEPC, particularly the lysine methyltransferase Ezh2 (Enhancer of Zeste 2, PRC2 subunit) (Dardenne et al., 2016; Ku et al., 2017). Others, such as the TGFβ type II receptor (Tgfbr2), have not been directly linked to NEPC but are components of signaling pathways—here, TGFβ—known to regulate lineage plasticity and prostate cancer progression (Hao et al., 2018). Additional candidates, including the RNA-binding protein Msi2, have not been previously associated with NEPC but play established roles in cancer and stem cell maintenance (Kharas et al., 2010). Notably, although Evi2a resides together with Evi2b and Omg within the Nf1 tumor suppressor locus (O'Connell et al., 1990) —a frequent target of SB mutagenesis—Evi2a was the only gene in this region to pass all prioritization criteria.

SIRT1 is a member of the sirtuin family, originally identified in non-mammalian systems for their roles in lifespan regulation (Haigis and Sinclair, 2010), and encodes an NAD-dependent lysine deacetylase that promotes heterochromatin formation (Vaquero et al., 2004). Beyond its canonical epigenetic functions, SIRT1 plays key roles in the nervous system and has been implicated in neurodegenerative and neuroendocrine diseases (Fujita and Yamashita, 2018). SIRT1 also functions in cancer, in part through interactions with p53 (Lin and Fang, 2013). Notably, SIRT1 was first identified in yeast as a regulator of the SWI/SNF chromatin remodeling complex (Mittal and Roberts, 2020), raising the possibility that its role in NEPC may be mediated through modulation of SWI/SNF activity—particularly given the established importance of this complex in prostate cancer and lineage plasticity (Cyrta et al., 2020; Thienger et al., 2026).

The role of SIRT1 in cancer is highly context dependent, with evidence supporting both oncogenic or tumor-suppressive functions depending on cellular state and disease stage (Lin and Fang, 2013). Prior studies have reported tumor-suppressive effects based on preinvasive phenotypes following SIRT1 loss, as well as tumor-promoting roles in advanced disease (Byles et al., 2012; Di Sante et al., 2015; Huang et al., 2021; Huang et al., 2022; Powell et al., 2011). This functional duality underscores the need to define the stage- and context-specific roles of SIRT1 during tumor progression.

Although SIRT1 is up-regulated in prostate cancer, including in NEPC (Huang et al., 2021; Huffman et al., 2007; Natani et al., 2021; Ruan et al., 2018), its functional significance in this setting has remained unclear. Our findings reconcile these observations by demonstrating that SIRT1 plays a pivotal role in promoting NEPC, revealing a context-dependent function that extends beyond general tumor growth to the regulation of lineage plasticity and neuroendocrine differentiation. Notably, while our studies establish a role for SIRT1 in NEPC, the SB data analysis does not prove that the insertion events lead to gain of function of Sirt1.

Importantly, our data suggest that targeting SIRT1 may suppress or reverse progression toward NEPC. Pharmacologic SIRT1 inhibitors, such as selisistat, are Food and Drug Administration–approved and have been evaluated in Phase I trials (Westerberg et al., 2015), highlighting SIRT1 as an attractive and clinically actionable target for lethal prostate cancer that warrants further investigation in future clinical studies.

All experiments using animals were performed according to protocols approved by the Institutional Animal Care and Use Committee at Columbia University Irving Medical Center. The mice used in this study are on a mixed strain (C57BL/6J;129S2) background and were housed in pathogen-free barrier conditions under 12-h light/dark cycles and with temperature and humidity at 20–25°C and 30–70%, respectively. Since the focus of our study is on prostate cancer, only male mice were used. The SB mice have a Cre-inducible transposase allele (Rosa26-LSL-SB11;Geurts et al., 2003; Starr et al., 2009) with or without the T2/Onc2 transposon allele (National Cancer Institute [NCI] Mouse Repository, Frederick National Laboratory, strain 01XGA; Dupuy et al., 2005) such that the SB(−) mice have only the Rosa26-LSL-SB11 allele, while the SB(+) mice have both the Rosa26-LSL-SB11 and the T2/Onc2 alleles. SB(−) and SB(+) mice were crossed with: (1) NPp53 mice (Zou et al., 2017) to generate NPp53-SB(−) mice (Nkx3.1^CreERT2/+; Pten^flox/flox; Trp53^flox/flox; R26^{LSL-SB11/LSL-SB11}) and NPp53-SB(+) mice (Nkx3.1^CreERT2; Pten^flox/flox; Trp53^flox/flox; R26-^{LSL-SB11/LSL-SB11}; T2/Onc2); or (2) NP mice (Floc’h et al., 2012) to generate NP-SB(−) mice (Nkx3.1^CreERT2; Pten^flox/flox; R26^{LSL-SB11/LSL-SB11}) and the NP-SB(+) mice (Nkx3.1^CreERT2; Pten^flox/flox; R26^{LSL-SB11/LSL-SB11}; T2/Onc2). Since the T2/Onc2 transposon is located on chromosome 1, insertions on this chromosome were disregarded in subsequent analysis to overcome complications related to local hopping (Liang et al., 2009). A summary of all mice used in this study is provided in Table S1.

Tumor induction was achieved using the Nkx3.1^CreERT2 allele, which has an inducible Cre recombinase that is specifically activated in luminal prostatic epithelial cells following delivery of tamoxifen (Wang et al., 2009b). Mice were genotyped prior to tumor induction. For tumor induction, mice were administered vehicle (corn oil) or tamoxifen (100 mg/kg in corn oil) (T5648; Sigma-Aldrich) via oral gavage at 2–3 mo of age once daily for 4 consecutive days. Mice were monitored three times weekly, and euthanized when their body condition score was <1.5, or when they experienced body weight loss ≥20% or signs of distress, such as difficulty breathing or bladder obstruction. At their terminal endpoint, mice were euthanized via carbon dioxide inhalation followed by cervical dislocation and were necropsied. Harvested tissues were visualized using an Olympus SZX16 microscope. Tissues were snap-frozen in liquid nitrogen for DNA or RNA isolation, or were fixed in 10% formalin (Thermo Fisher Scientific) for H&E, immunohistochemical and immunofluorescence staining.

Histopathological scoring of the prostate tumors was assessed independently by two pathologists (S. de Brot and A. Rodriguez-Calero) based on evaluation of H&E-stained tissues, some with confirmation by immunohistochemistry for Syp and Ki67 as summarized in Table S1. As described in Ittmann et al. (2013) and Shappell et al. (2004), histomorphologic features indicative of neuroendocrine differentiation included: (1) organoid nesting trabecular growth pattern, with peripheral palisading, and rosette formation; (2) tumor cells smaller than the diameter of three small lymphocytes, having scant cytoplasm, with oval or spindled hyperchromatic nuclei; and (3) high mitotic count; and (4) frequently presence of extensive necrosis.

Immunostaining was done using 3 μm formalin-fixed sections as described (Zou et al., 2017). Briefly, sections were deparaffinized in xylene followed by heat-induced antigen retrieval using Citrate-Based Antigen Unmasking Solution at pH 7.0 (Vector Labs). Sections were blocked in 10% normal goat serum, incubated with primary antibodies overnight at 4°C, and with secondary antibodies for 1 h at room temperature. For immunohistochemistry, signal detection and visualization were performed using the Vectastain ABC system followed by NovaRed Substrate Kit (Vector Labs), counterstained with hematoxylin and mounted with Clearmount (American Master*Tech Scientific). Images were captured using an Olympus VS120 whole-slide scanning microscope. For immunofluorescence, tissues were incubated with primary and secondary antibodies as above and then stained with DAPI and mounted with Vectashield antifade mounting medium (Vector Labs). Images were captured using a Leica TCS SP5 confocal microscope. Quantification of Ki67 staining was performed manually on a minimum of 8,000 cells per group based on three independent tumors with a minimum of five sections per tumor, as described in Zou et al. (2017). All antibodies and secondary antibodies are described in Table S8.

Validation of mRNA expression levels was done by quantitative real-time PCR using the QuantiTect SYBR Green PCR kit (Qiagen) using mouse Gapdh as the control (Zou et al., 2017). Primer sequences are provided in Table S8.

RNA-seq was performed on 35 RNA samples from 32 independent mouse prostate tumors, including 8 NPp53-SB(−) tumor samples and 27 NPp53-SB(+) tumor samples that included 13 histologically defined as NEPC and 22 histologically characterized as non-NEPC (Table S1). We also analyzed four NP-SB(−) and four NP-SB(+) tumors, all of which were histologically defined as non-NEPC (Table S1). RNA was prepared from snap-frozen tissues that were homogenized in TRIzol (Thermo Fisher Scientific) and extracted using the MagMAX-96 total RNA isolation kit (Thermo Fisher Scientific). Total RNA was enriched for mRNA using poly-A pull-down; only samples having between 200 ng and 1 µg and with an RNA integrity number >8 were used.

Libraries were made using an Illumina TruSeq RNA prep-kit version 2 or TruSeq Stranded mRNA library prep kit and sequenced using an Illumina HiSeq2500/4000 or NovaSeq6000. RNA-seq profiles were generated by mapping mRNA reads to the mouse reference genome (version GRCm38 mm10), using kallisto version 0.44.0. RNAseq raw counts were normalized and variance stabilized using DESeq2 (version 1.36.0) package (Bioconductor) in R-studio 2023.03.0+385, R version 4.2.3 (R Foundation for Statistical Computing).

For the generation of gene expression heatmaps, variance stabilized counts were z-scaled by row and visualized using the ComplexHeatmap (version 2.14.0) package (Bioconductor) in R-studio 2023.03.0+385, R version 4.2.3 (R Foundation for Statistical Computing). PCA was performed using the plotPCA function from the DESeq2 (version 1.42.1) package, using the top 500 highly variable genes computed from variance stabilized count data.

To identify genes differentially expressed in histologically defined NEPC versus non-NEPC SB mouse tumors, we used edgeR (version 4.0.16). Read counts were modeled using Genewise Negative Binomial Generalized Linear Models with Quasi-Dispersion Estimation (glmQLFit) after having estimated gene dispersion using Empirical Bayes Tagwise Dispersions for Negative Binomial GLMs (estimateGLMTagwiseDisp). The contrast matrix was generated by comparing SB(+) tumors that displayed NEPC with those that did not display NEPC. The list of statistically significantly (FDR <0.1, Benjamini-Hochberg corrected) differentially expressed genes is provided in Table S2 A; the dataset is provided in GSE271053. To compare the mouse RNA-seq data with human NEPC gene signatures, mouse genes were mapped to their human orthologs and compared with the AR, NEURO I, and NEURO II gene expression sets reported in Labrecque et al. (2019). Data were visualized as a heatmap and provided in Table S2 B.

Protein activity-based cluster analysis was done following VIPER-based analysis of RNA-seq profiles from the mouse SB tumors including the 13 histologically defined NEPC tumors and the 22 histologically defined non-NEPC tumors (see above). First, a differential gene expression signature (DGES) was computed for each sample as above (Table S2 A). Then the VIPER algorithm (Alvarez et al., 2016) was used to transform each DGES into a differential protein activity signature. A summary of VIPER protein activity is provided in Table S3 A.

Clusters were identified by the hierarchical clustering, as implemented by hclust (version 4.3.2), with Ward’s agglomeration method (Fionn and Legendre, 2014). Spearman’s correlation of the respective protein activity profiles was used as a sample-to-sample distance metric. This yielded a list of MR proteins representing candidate drivers of the cell states associated with each cluster. For visualization purposes, only the 15 most statistically significant differentially MRs are shown, as assessed by P value integration across the samples of each cluster (n = 3), using Stouffer’s method. A summary of protein activity for each cluster is provided in Table S3 B.

For analyses of the mouse tumors, we used 35 SB tumor samples, comprising 13 histologically defined NEPC tumors and 22 histologically defined non-NEPC tumors (Table S1). The human tumor profiles were collected from two independent cohorts, namely, (1) the Beltran cohort (n = 49 patients), comprising 34 adenocarcinoma and 15 NEPC patients, as downloaded on November 12, 2020, from cBioPortal as FPKM measurements (Beltran et al., 2016) and (2) the SU2C cohort (n = 266 patients), comprising 210 non-NEPC, 22 NEPC patients and 34 unclassified tumors, as downloaded on October 4, 2019, from cBioPortal as FPKM measurements based on the 2019 SU2C-mCRPC dataset (Abida et al., 2019). For these analyses, we first generated an independent DGES for each mouse and patient-derived tumor sample by computing Z-scores obtained by dividing the difference between the expression of each gene in the sample and the mean expression in the associated cohort by the standard deviation.

Differentially active proteins were identified by analyzing each DGES using VIPER (Alvarez et al., 2016) for the human samples and metaVIPER (Ding et al., 2018) for the mouse samples, with the species-appropriate regulatory networks, see Vasciaveo et al. (2023). The mouse signatures were processed integrating both the human SU2C network and mouse prostate cancer network as described in Vasciaveo et al. (2023). Protein activity signatures for the mouse tumors are provided in Table S3 A and for the human tumors in Table S4 C and D.

To match the mouse SB prostate tumors to human prostate cancer patients, we used the OncoMatch algorithm (Vasciaveo et al., 2023). Specifically, we used the aREA algorithm (Alvarez et al., 2016) to assess enrichment of patient tumor-specific MRs—i.e., the top 50 most differentially active proteins in each human tumor, including the 25 most activated (25↑) and 25 most differentially inactivated proteins (25↓)—in proteins differentially active in the mouse SB tumors. For this purpose, mouse proteins were humanized (i.e., mapped to their human protein orthologs) before running aREA (Vasciaveo et al., 2023). Normalized enrichment scores (NES) were converted to P values and the Bonferroni method was used to correct for multiple hypothesis testing. The value of $OMScore=−Log10 (p−value)>5$ was used as a conservative statistical threshold to assess tumor/model fidelity. The OncoMatch data are provided in Table S4.

Targeted genomic sequencing was performed to identify transposon insertion sites in NPp53-SB(+) prostate tumors following the protocol published in Janik and Starr (2013). We analyzed 74 tumors for which the NPp53-SB(+) T2/Onc2 transposition was confirmed by PCR (Fig. S1 A). Genomic DNA was extracted and digested with restriction enzymes flanking the left (BfaI) or right (NlaIII) side of the transposon, as in Janik and Starr (2013). The resulting 148 samples were subjected to linker-mediated PCR (LM-PCR) as published in Janik and Starr (2013). Samples were prepared using the primers in Table S8, and the barcodes described in Table S5. Samples were sequenced using the Illumina HiSeq with 100 bp read (Azenta).

CIS, which are defined as regions of the genome having transposon insertions (i.e., integration sites) across multiple independent tumors at a significantly greater frequency than that expected by chance (Dupuy et al., 2005), were identified using the TAPDANCE software (version 3.1), which both maps insertion reads and computes statistically significant CIS (Sarver et al., 2012). Reads were mapped using the University California Santa Cruz indexed bowtie 1.2.3 compatible mm10 reference genome. All default parameters were used, and the $library percent, which discards insertion regions that are represented by a low number of reads, was set to 0.001 (i.e., insertion regions represented by lower than 0.001% of total reads assigned to the tumor are discarded). As noted above, since the T2/Onc2 transposon is located on chromosome 1, insertions on this chromosome were disregarded in subsequent analysis to overcome complications related to local hopping (Liang et al., 2009).

This analysis identified 122 unique CIS (Table S6 A). For 121 of the 122 CIS, we identified 330 CIS-associated genes based on annotated genomic location within 20 kbp of the CIS. One CIS lacked any annotated gene feature within 20 kbp of the insertion. A list of the 330 CIS-associated genes together with the standardized TAPDANCE output is provided in Table S6 B. For each CIS-associated gene we also report the proportion of NEPC tumors calculated based on two-sided binomial test (stats) using p = proportion of neuroendocrine samples (27/74) for neuroendocrine versus adenocarcinoma.

The 330 CIS-associated genes were represented in a Circos plot, using the circlize (0.4.16) package. Strandedness was assessed from the raw TAPDANCE mapping output file, with the height of the dots representing each transposon’s read count. CIS were marked based on TAPDANCE output file (Table S6 A) and gene positions were acquired by bioMart (2.60.1) and plotted in the outer side of the Circos plot. Co-occurrence analysis of pairs of CIS-associated genes was performed using the somaticInteractions function using the maftools (2.20.0) package and is represented in the center of the Circos plot.

Lollipop plots for selected CIS-associated genes were generated using trackViewer (1.41.5) and GenomicRanges (1.57.1) packages. Gene structures were obtained from the mm10 assembly, using the TxDb.Mmusculus.UCSC.mm10.knownGene (3.10.0) package. Transposon insertion data from TAPDANCE output were plotted, with insertion heights representing read counts. In Figs. 5 and S3 and insertions in the same direction as gene transcription are shown with red arrows, opposite in green.

To identify genes where transposon insertions correlate with altered transcript structure, we performed two complementary analyses. First, genome-wide isoform switching analysis was conducted using IsoformSwitchAnalyzeR (version 1.16.0) (Vitting-Seerup and Sandelin, 2019). Salmon (version 1.9.0)-quantified transcript abundances were imported with GTF annotation (GENCODE version M25) and filtered for lowly expressed transcripts. Differential isoform usage between NEPC and non-NEPC conditions was tested using DEXSeq (version 1.56.0), with significance thresholds of q <0.05 and |ΔIF| >0.1. Protein domains were annotated using HMMER against the Pfam database. Results were intersected with CIS genes to identify candidates with both insertions and isoform switching. Second, fine-grained coverage analysis was performed for prioritized CIS genes with intronic insertions. RNA-seq reads were extracted from BAM files across gene bodies using Rsamtools (version 2.26.0), and coverage was compared between NEPC samples harboring insertions, SB-negative controls (n = 12), and SB-positive-non-NEPC controls (n = 10). Coverage was normalized within each group to enable comparison of relative patterns across the gene body.

To detect transposon-mRNA fusion transcripts, paired end RNA-seq reads were aligned to a composite reference genome consisting of mm10 (with the En2 locus masked) and the T2/Onc2 transposon sequence (2,163 bp) appended as chromosome “chrSB” as described in Temiz et al. (2016). Alignments were performed using HISAT2 (version 2.2.1) with splice-site annotations from GENCODE version M25. Transposon-genome fusion junctions were identified using the FUSION FINDER approach with default parameters. Detected fusions were cross-referenced with DNA-based CIS insertion coordinates to identify events where both DNA insertion and RNA fusion evidence were present. Data for all TAPDANCE analysis are provided in Table S6.

Candidate MRs regulating genes differentially expressed in NEPC versus non-NEPC mouse SB tumors were identified as described above. CINDy was then used to prioritize candidate modulators of MR activity (Giorgi et al., 2014), based on the analysis of RNA-seq profiles from human primary prostate tumors from the TCGA patient cohort (N = 498; Cancer Genome Atlas Research, 2015) and metastatic CRPC biopsies from the SU2C patient cohort (N = 266; Abida et al., 2019). Briefly, for each cMR (as above), candidate modulator gene [M], and MR transcriptional target [T]—as inferred by the ARACNe algorithm (Basso et al., 2005)—we computed the statistical significance of the conditional mutual information $I[cMR;T|M]$⁠, using RNA-seq profiles from both human cohorts, independently. The statistical significance (P value) of each candidate modulator gene was assessed by integrating the P values across all of its targets, as described in Giorgi et al. (2014), and across cohorts. Statistical significance was assessed at P ≤ 0.05 (FDR corrected).

To prioritize CIS genes, we assessed each gene by integrating three independent statistics, including (1) P value as assessed by TAPDANCE analysis, (2) differential expression in NEPC versus non-NEPC samples, and (3) CINDy-predicted activity as NEPC MR modulators. The P values generated for each of these three analyses were integrated using Fisher’s method. We then ranked the list by their VIPER scores. The data are presented in Table 2 and Table S7.

All studies involving human tissue specimens were performed according to protocols approved by the Human Research Protection Office and Institutional Review Board at the University of Bern (Project ID: 2019-00328 and 2022-00978). Only anonymized tissues were used, and patient consent was obtained. SIRT1 expression was evaluated in retrospectively collected tissue from patients with advanced prostate cancer (N = 7). Five patients harbored prostate cancer with neuroendocrine differentiation (5/7). From those, three patients (3/7) presented a small cell neuroendocrine carcinoma with known prostate cancer and two patients (2/7) harbored a prostate cancer with neuroendocrine differentiation. In the remaining two patients (2/7), an acinar adenocarcinoma of the prostate was found. SIRT1 (10E4) from #04-1557; Sigma-Aldrich was stained on the Leica BOND RX Automated Stainer. The dilution for SIRT1 was 1:200 in AB diluent from Biosystems, and it was incubated for 30 min at room temperature. The pretreatment was conducted with citrate epitope retrieval buffer for 30 min at 100°C. The detection was done with BOND Polymer Refine DAB detection kit #DS9800, including counterstain with hematoxylin. The percentage of tumor cells presenting a SIRT1 nuclear expression was reported.

LNCaP cells (CRL-1740; ATCC) and 22Rv1 cells (CRL-2505; ATCC) were maintained in Roswell Park Memorial Institute medium (RPMI-1640; ATCC) with 10% fetal bovine serum (FBS) (Thermo Fisher Scientific) and HEK-293FT cells (R700-07; Invitrogen) were cultured in Dulbecco’s Modified Eagle Medium (DMEM) with 10% FBS (Thermo Fisher Scientific). Cells were authenticated by STR profiling and tested negative for Mycoplasma using the Universal Mycoplasma Detection Kit (#30-1012 K; ATCC).

Analyses of loss or gain of function in LNCaP and 22RV1 cells, respectively, was performed as described in Arriaga et al. (2024) and Papachristodoulou et al. (2021). For loss-of-function studies, we analyzed the consequences of silencing SIRT1 in LNCaP cells using two independent shRNAs based on LT3GEPIR T3G-GFP-(miR-E)-PGK-Puro-IRES-rtTA3 (#111177; Addgene) (Fellmann et al., 2013). The sequences of the shRNA are provided in Table S8. For gain-of-function studies, we analyzed the consequences of overexpression of SIRT1 in CRISPRa-engineered 22RV1 cells, as described (Arriaga et al., 2024). A non-targeting sgRNA (sgNT) and two individual sgRNAs targeting SIRT1 (sgSIRT1 #1 and #2) were generated based on the protospacer sequences identified in the hCRISPRa-v2 library in Horlbeck et al. (2016), and subcloned into the sgRNA template pU6-sgRNA EF1Alpha-puro-T2A-BFP (#60955; Addgene). Primer sequences are provided in Table S8. Procedures for the generation and infection with lentivirus were done as in Arriaga et al. (2024) and Papachristodoulou et al. (2021).

For analyses in culture, cells were seeded at 2 × 10⁵ cells (22RV1) or 5 × 10⁵ cells (LNCaP) per 10 cm dish in RPMI + 10% FBS. The following day, cells were changed to RPMI phenol red-free + 5% charcoal-stripped serum (CSS) and grown in DMSO and ethanol or 1 mM dibutyryl cyclic-AMP (db-cAMP, STEMCELL Technologies) and 0.5 mM 3-isobutyl-1-methylxanthine (IBMX; Thermo Fisher Scientific).

Colony formation and proliferation assays were performed as described in Papachristodoulou et al. (2021). Briefly, for colony formation assays, 22RV1 cells (1 × 10³ cells) were seeded in triplicate in 6-well plates and grown for 12 days. Studies controlled for effects on cell proliferation following induction toward neuroendocrine differentiation. Colonies were visualized by crystal violet staining and quantified using ImageJ. To measure proliferation, we performed 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) proliferation assays. 1 × 10⁴ cells were seeded in triplicate in 96-well plates and grown up to 72 h. MTT-based proliferation was quantified by measuring OD at 560 nm in a Varioskan LUX multimode microplate reader (Thermo Fisher Scientific).

For analyses of tumor growth in vivo, 1 × 10⁶ 22RV1 cells in 30 μl of 50/50 PBS/Matrigel (Corning) were injected into the anterior prostate (AP) of immunodeficient Athymic Nude mice (Envigo) as described (Arriaga et al., 2024). Mice were euthanized at 40 days after injection, and tumors were collected and analyzed as described (Arriaga et al., 2024).

For all cell culture studies, validation of mRNA expression levels was performed by quantitative real-time PCR with QuantiTect SYBR Green PCR Kit (QIAGEN, Hilden, Germany), using GAPDH as an internal control. Relative expression levels were calculated using the 2^−ΔΔCT method, as described (Papachristodoulou et al., 2021). Sequences of the primers are provided in Table S8. Validation of protein levels was performed by western blot using total cell lysates extracted with radioimmunoprecipitation assay buffer, as described (Papachristodoulou et al., 2021). Antibodies described in Table S8.

A mouse NEPC organoid line was generated from a YFP-lineage marked NPp53 mouse having a rare NEPC tumor (Zou et al., 2017), using protocols developed to establish mouse prostate organoids (Chua et al., 2014; Giacobbe et al., 2025, Preprint). Briefly, YFP-marked cells were FACS sorted with the BV421/FITC channels. Sorted cells were resuspended in organoid culture media as described in Chua et al. (2014) (hepatocyte culture medium [Corning], 5% heat-inactivated charcoal-stripped fetal bovine serum [CS-FBS, Gibco], 1× GlutaMAX [Gibco, Thermo Fisher Scientific], 10 ng/ml epidermal growth factor [EGF, Corning], 10 µM Y-27632 ROCK inhibitor [STEMCELL Technologies], 100 nM dihydrotestosterone [DHT; Sigma-Aldrich], 5% Matrigel [Corning], and 1× antibiotic-antimycotic [Gibco]). Cells were seeded at a density of 5,000 cells per well onto ultralow-attachment 96-well plates (3474; Corning). The samples were monitored for organoid growth via fluorescence imaging for 7 days.

For passaging, mouse organoids were dissociated by digestion in 1 ml 0.25% Trypsin-EDTA (Gibco) for 10 min at 37°C; digestion was stopped by the addition of 2 ml modified Hank’s Balanced Salt Solution (HBSS; STEMCELL Technologies) supplemented with 2% FBS. The dissociated cells were collected by centrifugation at 350 g for 5 min at 4°C and resuspended in organoid culture medium, as above. Cells were counted using a Countess II Automated Cell Counter (Thermo Fisher Scientific) and plated at a seeding density of 5,000 cells/well onto ultralow-attachment 96-well plates (Corning). Organoid lines were established following five consecutive passages.

For histopathological characterization, organoids were harvested by centrifugation at 250 g × 5 min at 4°C, fixed in 1 ml 10% formalin at 4°C overnight (Thermo Fisher Scientific), and placed in 80 μl HistoGel (Epredia) before embedding. Paraffin-embedded organoids were sectioned (3 µm) and subjected to histopathological analysis and imaging as above. All antibodies and secondary antibodies are described in Table S8.

Analyses of loss-of-function in mouse organoids was performed as described in Giacobbe et al. (2025, Preprint). Briefly, we analyzed the consequences of silencing Sirt1 in mouse organoids using two independent shRNAs based on LT3GEPIR T3G-GFP-(miR-E)-PGK-Puro-IRES-rtTA3 (#111177; Addgene) (Fellmann et al., 2013). The sequences of the shRNA are provided in Table S8. Upon infection of organoids with the LT3GEPIR T3G-GFP-(miR-E)-PGK-Puro-IRES-rtTA3 based lentiviruses, puromycin was added in the organoid culture medium (5 µg/ml) and cells were selected for 4 days. Following removal of the puromycin, doxycycline (0.5 µg/ml) was added for 2 consecutive days to induce shRNA-mediated silencing and was maintained in culture for 10 days. Tumor organoids were subjected to histopathological analysis and validation of mRNA expression levels was performed as above.

Functional analyses were performed using organoid lines that were infected with the lentiviruses as described above. For tumor growth assays, organoids (not dissociated) (equivalent to 2 × 10⁵ cells) were injected into the flank of male immunodeficient Athymic Nude mice (Envigo). To maintain shRNA expression, doxycycline was continuously provided in the drinking water at a concentration of 2 mg/ml and refreshed three times weekly. Tumors were monitored by caliper measurement twice weekly and tumor volumes were calculated using the formula [volume = (width)² × length/2]. Mice were sacrificed when the tumor size reached 2,000 mm³ or if the body condition score of the host mice was <1.5 or if they exhibited signs of distress. At the time of sacrifice, YFP-positive allograft tumors were visualized by ex vivo fluorescence using an Olympus SZX16 microscope (Ex490–500/Em510–560 filter) and collected for histopathological analysis.

Studies were performed using a selective SIRT1 inhibitor, Selisistat EX-527 (purity: 99.78% from Selleckchem) resuspended in DMSO (Adams et al., 2024; Napper et al., 2005). To determine the dose of Selisistat that inhibits SIRT1 activity, we examined the overall acetylation of proteins using anti-acetyl Lysine antibodies (Abcam; ab22550) (Table S8). For studies in LNCaP cells, we seeded 5 × 10⁵ cells per 60 mm dish in RPMI + 10% FBS. The following day, cells were changed to RPMI phenol red-free + 5% CSS and treated with either DMSO and ethanol, or 1 mM db-cAMP (from 500 mM stock in 50/50 in ethanol/PBS) (STEMCELL Technologies), 0.5 mM IBMX (from 1 M stock in DMSO) (Thermo Fisher Scientific), or Selisistat EX-527 (from 50 mM stock in DMSO) (Selleckchem). For studies in NEPC organoids, we seeded cells from dissociated organoids at a density of 5,000 cells per well onto ultralow-attachment 96-well plates (3474; Corning). The following day, organoids were treated with either DMSO or Selisistat EX-527 (Selleckchem) for 10 days as above, and samples were subjected to validation of mRNA expression levels as described above.

Statistical analysis was performed using Graphpad Prism software (version 9.4.1) and R Studio (version 1.3.1093), R (version 4.1.2). Survival curves for the SB mice were plotted based on Kaplan-Meier analysis and significance was calculated based on the log-rank test using the survival (3.5-5) and survminer (0.4.9) packages in R. Metastasis and NEPC frequencies were calculated and significance was evaluated based on the Fisher’s exact test in R. Tumor weight of SB tumors was analyzed and significance was calculated based on Welch’s two sample t-test in R. Analysis of Ki67 staining was performed using Graphpad Prism (version 9.4.1) and significance was calculated with one-way ANOVA with Tukey’s multiple comparisons test. Functional validation experiments were plotted using Graphpad Prism and significance was calculated based on one-way ANOVA with Dunnett’s multiple comparisons test for tumor weights, mRNA relative expression from cell lines, relative number of colonies from colony formation assay, and OD from MTT assay. Significance for the above-mentioned statistical tests was assumed for P < 0.05.

Fig. S1 shows additional analyses of the prostate phenotype of SB mice (related to Fig. 2). Fig. S2 shows additional OncoMatch analyses (related to Fig. 4). Fig. S3 shows lollipop representation of selected CIS genes (related to Fig. 5). Fig. S4 shows additional analyses of selected CIS genes (related to Fig. 5). Fig. S5 shows additional validation studies in human prostate cancer cells (related to Figs. 6 and 8) (source files for Figs. S1 and 6). Table S1 shows summary of SB mouse phenotypes. Table S2 shows differential gene expression analysis in SB mice. (A) Differential gene expression of NEPC versus non-NEPC mouse SB tumors. (B) Relative expression levels of AR and Neuro markers from Labrecque et al. (2019). Table S3 shows protein activity signatures and cluster analyses of of mouse SB tumors. (A) Protein activity signatures from mouse SB tumors. (B) Protein activity of cluster analyses of mouse SB tumors. Table S4 shows oncomatch analysis between SB mice and prostate cancer patients. (A) OncoMatch analysis between SB mice and patients from the SU2C cohort. (B) OncoMatch analysis between SB mice and patients from the Beltran cohort. (C) SU2C protein activity. (D) Beltran 2016 protein activity. Table S5 shows SB barcodes. Table S6 shows a summary of CIS analyses and CIS-associated genes. (A) List of CIS determined by TAPDANCE analysis. (B) List of CIS-associated genes determined by TAPDANCE analysis. (C) List of all fusion events overlapping with a CIS region. (D) List of all genes with fusion events. (E) Isoform switching analyses of CIS-associated genes. (F) Domain details from isoform switching analyses of CIS-associated genes. (G) Full data from isoform switching analyses of CIS-associated genes. Table S7 shows prioritization of CIS-associated genes. Table S8 shows the key resources table.

RNA-seq expression profiling data have been deposited in the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) with the following accession codes: GSE271066 (a description of all data from the study), GSE271054 (LM-PCR-based targeted genomic sequencing from NPp53-SB(+) mice), GSE271053 (mouse RNA-seq data from NP/NPp53-SB(−) and NP/NPp53-SB(+) mice). Code for all analyses is publicly available at https://github.com/alevax/sleeping-beauty-paper.

We are indebted to Timothy Starr, the primary architect of the TAPDANCE software, for his assistance and insights in applying it to our SB screen. We thank Peter Sims for valuable comments on the manuscript. Some parts of Figs. 1, 2, 4, 6, 7, and 8 were created with BioRender.com using an institutional license sponsored by Columbia University’s Vagelos College of Physicians and Surgeons Office for Research. These studies were supported by Flow Cytometry and Genomics and High Throughput Screening Core facilities, which are funded in part through Herbert Irving Comprehensive Cancer Center support (HICCC) grant P30 CA013696.

This work was supported by grants R01CA173481, R01CA283068, and R01CA183929 to C. Abate-Shen, R01CA251527 to M.M. Shen, P01CA265768-03 and a grant from the T.J. Martell Foundation for Cancer Research to M.M. Shen and C. Abate-Shen, the NCI’s Center for Cancer Systems Therapeutics (CaST) award U54CA274506 to A. Califano and C. Abate-Shen, the NCI Outstanding Investigator award R35 CA197745 to A. Califano, and the National Institutes of Health (NIH) Shared Instrumentation Grants S10 OD012351, S10 OD021764, and S10OD032433 to A. Califano, a Prostate Cancer Challenge Award to M.M. Shen and A. Califano, and a Prostate Cancer Challenge Award to M.A. Rubin. F. Nunes de Almeida was supported by a 2020 American Association for Cancer Research (AACR)-AstraZeneca Stimulating Therapeutics Advancements through Research Training (START) Grant (20-40-12-NUNE). A. Vasciaveo was supported by a US Department of Defense Early Investigator Research Award (W81XWH19-1-0337) and an Early Career Development Pilot Award NIH/NCI Cancer Center, funded through the HICCC Cancer Center Support grant P30CA013696. A. Giacobbe was supported by a Prostate Cancer Foundation (PCF) (19YOUNG09) Young Investigator Award; a US Department of Defense Prostate Cancer Research Program Award (W81XWH-18-1-0193); and an International Cancer Research Fellowship Outgoing iCARE funded by the Associazione Italiana per la Ricerca sul Cancro (AIRC) (17887) and the Marie Curie Actions COFUND. M. Zou was supported in part by the National Center for Advancing Translational Sciences, NIH, grant number UL1TR001873. S. Park was supported by a Bladder Cancer Advocacy Network Young Investigator Award (1068480), a HICCC postdoctoral Pilot Award, and a fellowship from the American Cancer Society (PF23-1142485 and RP-16-237-06-COUN). J.M. Arriaga was supported by the Dean’s Precision Medicine Research Fellowship from the Irving Institute for Clinical and Translational Research at Columbia University Irving Medical Center (UL1TR001873), and a Prostate Cancer Foundation Young Investigator Award. A.L.E. Wang was supported by the National Institute of Diabetes and Digestive and Kidney Diseases T35 grant (T35DK093430). J. Lu was supported by a Prostate Cancer Foundation (PCF) Young Investigator Award. C. Abate-Shen is an American Cancer Society Research Professor supported in part by a generous gift from the F.M. Kirby Foundation.

Author contributions: Francisca Nunes de Almeida: conceptualization, data curation, formal analysis, investigation, methodology, supervision, validation, visualization, and writing—original draft, review, and editing. Alessandro Vasciaveo: conceptualization, data curation, formal analysis, investigation, methodology, resources, software, supervision, visualization, and writing—original draft, review, and editing. Arianna Giacobbe: formal analysis, investigation, validation, and writing—review and editing, conceptualization, visualization, writing - original draft. Matteo Di Bernardo: conceptualization, data curation, formal analysis, investigation, software, and writing—original draft, review, and editing. Min Zou: conceptualization, data curation, formal analysis, investigation, methodology, and visualization. Ainsley Mike Antao: formal analysis, investigation, validation, and visualization. Simone de Brot: formal analysis and investigation. Antonio Rodriguez-Calero: validation and writing—review and editing, formal analysis. Alexander Chui: data curation, formal analysis, software, visualization, and investigation. Alexander L.E. Wang: data curation, formal analysis, investigation, software, and visualization. Nicolas Floc’h: methodology. Jaime Y. Kim: formal analysis, investigation, resources, visualization, and writing—review and editing. Soonbum Park: methodology and resources. Stephanie N. Afari and Jinqiu Lu: resources, methodology. Timur Mukhammadov: investigation, methodology, and resources. Nicholas Ortega and Jingqiang Wang: investigation, methodology, resources. Sai Sampath Josyula: formal analysis and software. Juan Martin Arriaga: methodology and writing—review and editing. Michael M. Shen: funding acquisition, supervision, and writing—review and editing. Mark A. Rubin: investigation and writing—review and editing, resources, visualization, supervision. Andrea Califano: conceptualization, data curation, formal analysis, funding acquisition, methodology, project administration, resources, software, supervision, and writing—original draft, review, and editing. Cory Abate-Shen: conceptualization, formal analysis, funding acquisition, investigation, methodology, project administration, resources, supervision, validation, visualization, and writing—original draft, review, and editing.