Castrate-resistant prostate cancer response to taxane is determined by an HNF1-dependent apoptosis resistance circuit Authors: Ilya S Senatorov, Joel Bowman, Keith H Jansson, Aian Neil Alilin, Brian J Capaldo, Ross Lake, Morgan Riba, Yasmine C Abbey, Crystal Mcknight, Xiaohu Zhang, Sonam Raj, Michael L Beshiri, Paul Shinn, Holly Nguyen, Craig J Thomas, Eva Corey, Kathleen Kelly Published: 2024 ABSTRACT SummaryMetastatic castrate-resistant prostate cancer (mCRPC) is a genetically and phenotypically heterogeneous cancer where advancements are needed in biomarker discovery and targeted therapy. A critical and often effective component of treatment includes taxanes. We perform a high-throughput screen across a cohort of 30 diverse patient-derived castrate-resistant prostate cancer (CRPC) organoids to a library of 78 drugs. Combining quantitative response measures with transcriptomic analyses demonstrates that HNF1 homeobox A (HNF1A) drives a transcriptional program of taxane resistance, commonly dependent upon cellular inhibitor of apoptosis protein 2 (cIAP2). Monotherapy with cIAP2 inhibitor LCL161 is sufficient to treat HNF1A+ models of mCRPC previously resistant to docetaxel. These data may be useful in future clinical trial designs.Keywords: prostate cancer, organoid, xenograft, pharmacology, drug resistance, HNF1A, BIRC3, apoptosis, docetaxel Introduction Prostate cancer new diagnoses in the United States have been estimated at 288,300, accounting for the largest number of new cancer cases and second highest projected cancer-related cause of mortality in men.1 Current treatment options vary based on disease symptoms, localization, aggressiveness, and risk of recurrence, while taking into account patient comorbidity and age. Localized prostate cancer is treated with surgery or radiation, while androgen deprivation therapy (ADT) and administration of androgen receptor (AR) antagonists remain the initial approach for metastatic prostate cancer.2 Although ADT has been demonstrated to yield an approximately 85% response rate when used as a first-line treatment in metastatic prostate cancer,3 resistance in the form of castrate-resistant prostate cancer (CRPC) eventually occurs in nearly all patients. Mechanisms of CRPC include either reactivating AR-mediated signaling pathways through receptor mutation or bypass signaling, or less frequently, by losing AR signaling pathway dependence observed as lineage switching.4 The use of microtubule stabilizers such as docetaxel and cabazitaxel to treat metastatic CRPC (mCRPC) extends overall survival and is recommended as a second line of therapy.5 However, treatment is toxic and can affect quality of life, and ineffective treatment consumes time that could be used for other therapies. Currently, the molecular determinants of clinical response and intrinsic/acquired resistance remain poorly understood, and as a result, molecular biomarkers to predict taxane responsiveness have not been established.6 A major challenge for characterizing oncology drugs with translatable clinical efficacy and predictive biomarkers is to recapitulate the heterogeneity of tumors in contemporary patient populations. High-throughput drug screening against cancer cell lines has historically served as a starting point for determining drug candidates, but these models suffer from artificial selection for in vitro growth as well as nonrepresentative establishment of tumor subtypes. An alternative model source, patient-derived xenografts (PDXs), generally better reflect drug sensitivity, recapitulating the human source better than cell lines.7,8 While PDXs have the advantage of capturing some aspects of in vivo tumor biology, scaling PDX projects to high-throughput levels is not practical, and these systems also reflect adaptation to a mouse-specific environment.9,10 Patient-derived organoids (PDOs) allow a substantial representation of tumor heterogeneity while being scalable to high-throughput assays, allowing for both drug efficacy and biomarker discoveries in models representative of patient cancers.11 For prostate cancer, a well-characterized cohort of roughly 30 PDXs has been established,12 from which our lab previously generated a biobank of organoids that maintain molecular and phenotypic characteristics of their parent PDX and half of which can be passed for at least 10 generations.13 Organoid models from various cancers have been demonstrated to recapitulate their parent PDX drug sensitivities when tested in vitro to nearly 90% efficacy.14 Importantly, organoid models recapitulate both sensitivities and resistances to specific therapies, including chemotherapies, observed through retrospective analysis of response in the source patients.15,16 In this work, we use first-generation organoids derived from propagating PDX cohorts as well as a small number of PDOs to facilitate high-throughput screening of therapeutic compounds, many of which have well-studied and/or established targets. Our aim has been to determine drug sensitivity profiles among a mCRPC patient representative cohort, leading to a description of therapeutic vulnerabilities in advanced prostate cancer and the molecular properties associated with response or resistance. While various drugs demonstrated validated activity in limited models, docetaxel showed a robust range of responses across the majority of models, the latter being consistent with clinical experience. Computation and experimental analyses revealed HNF1 homeobox A (HNF1A) as a previously unappreciated determinant of taxane responses. Results The PDX-PDO biobank recapitulates heterogeneity found in human PCa A biobank of 30 individual first-generation CRPC-PDX-PDO models was screened generating response data to 78 compounds (Table S1). Organoids were obtained from 25 models within the LuCaP PDX cohort,12 from four organoid cultures previously established from NIH Clinical Center patient biopsies,13 and from cell lines including LNCaP, RWPE, and normal prostate basal cells. Overall, the subtypes analyzed consisted of 15 adenocarcinomas, 6 experimentally castrated adenocarcinomas, 5 neuroendocrine prostate cancers (NEPCs), an amphicrine prostate cancer (expressing both adenocarcinoma and neuroendocrine characteristics in single cells), 1 double-negative prostate cancer (expressing neither adenocarcinoma nor neuroendocrine characteristics), and normal prostate basal stem cells. These encompass tumors obtained from the prostate as well as various common sites of metastasis and were screened against a compound library spanning multiple mechanistic classes (Table S2). To minimize differences in long-term in vitro growth potential of the PDX models, the screen was conducted on first-generation organoids, defined as those formed immediately post isolation from parental PDX tumors. Transcriptomics of first-generation organoids showed that the model cohort contained a spread of classically used lineage, AR-driven, and neuroendocrine markers; in addition, there was a distribution of recently described transcription-based CRPC subclasses of PDO models for the WNT pathway and for stem, basal, and luminal cells.17 We observed a demonstrable range of RB1 and AR signaling signatures (Figure 1A). Robust AR signatures are characteristic of adenocarcinoma CRPC, while RB1 loss signatures are characteristic of NEPC and a small subclass of aggressive adenocacinomaCRPC.18 Principal component analysis demonstrated predictable separation of model clusters based on lineage phenotype (Figure 1B). To determine the extent of patient sample representation found in the CRPC cohort used here, gene set variance analysis (GSVA) on PCa-relevant transcriptomic signatures was compared between the organoid cohort and patient data samples from the Stand Up to Cancer (SU2C 2019) cohort (Figure 1C). Importantly, AR signaling and neuroendocrine scores,19 as well as transcriptomic signatures modeling common oncodrivers such as RB1 loss20 and Myc activity,21 showed that the organoid cohort distributed throughout the observable range derived from patient samples. The PDX-PDO biobank recapitulates heterogeneity found in human prostate cancer (A) Composite heatmap and barplots denoting transcriptomic signatures and genes modeling key signaling and lineage characterizations. Models are organized by lineage. ARPC, adenocarcinoma; Exp Cas, selected in vivo to be castrate resistant; DN: double negative; AMP, amphicrine; NE, neuroendocrine prostate cancer. RB loss and RB loss-CRPC signatures were scored as described in STAR Methods. (B) Principal component analysis plot for transcriptomic clustering of organoid models. Abbreviations are as in (A). (C) GSVA barplots showing the organoid cohort and SU2C patient transcriptomic signature distributions among the range of observed values. See also Table S1. First-generation PDO screening yields reproducible compound sensitivities To maximize resolution in categorizing responsive and nonresponsive models, each compound in the screen was tested using a 10-point dose curve with a 1:3 serial dilution, resulting in robust drug-response curves for each compound against each model. Dose responses were summarized into single area under the curve (AUC) values normalized by the dose range tested (nAUC), which are illustrated as a heatmap in Figure 2. Validation of screen results was conducted by comparison to in-house biological replicates, experiments generated by hand, of multiple model-drug pairs. Normalized “in-house” AUC (nAUC) was plotted against nAUC derived from the screen for 140 unique organoid-compound pair response curves, yielding a linear regression with a Pearson R of 0.81 (Figure S1A). These data showed robust reproducibility of curves generated via the high-throughput screen as compared to individual experimental manipulation of models. To determine whether organoid proliferation rate influenced a determination of drug responses, we performed a Spearman correlation across the compounds with highest nAUC ranges versus the first-generation proliferation rate for 11 organoid models. However, we did not observe any statistical significance, adjusted for multiple comparisons, to suggest proliferation-governed responses (Table S3). PDO drug screening identifies patterns of therapeutic compound sensitivity and resistance Normalized area under curve (nAUC) heatmap showing summarized dose-response curve values for screened compounds across the organoid cohort. See also Figure S1 and Tables S2 and S3. Docetaxel contained the largest range of responses in the library, closely followed by MCL-1 inhibitor S63845 and farnesyltransferase inhibitor tipifarnib (Figure S1B). Supporting the validity of the high-throughput screen, we also noted that the screen results recapitulated observations previously made in experimental NEPC models,22 such as the vulnerability to BCL-2 family targeting agents of NEPC (3/5 models) compared to adenocarcinoma (0/15 models). In addition, the double-negative LuCaP 173.2 model was vulnerable uniquely to the fibroblast growth factor receptor inhibitor erdafitinib as previously shown.22,23 Transcriptomic analysis reveals molecular commonalities within docetaxel-responsive and -nonresponsive mCRPC models Despite docetaxel being a clinical staple in CRPC treatment, the clinical importance of experimentally selected response and resistance mechanisms is mostly unclear.24,25,26 Response to docetaxel was used as the contrast for determining gene enrichment that may predict molecular features of responsive and nonresponsive CRPC phenotypes and/or useful biomarkers. Differential gene expression (DGE) was run on full transcriptome sequencing of screened organoid models. Notable among the most differentially expressed genes between docetaxel-responsive and -nonresponsive models were two genes, HNF1A and SPINK1, which annotate the presence of a gastrointestinal transcriptional circuit, previously demonstrated to be associated with more aggressive prostate cancer27 (Figure 3A). To concentrate gene enrichment interpretations on adenocarcinoma and minimize the selection of lineage markers, pathway enrichment of DGE was conducted on the model cohort separately including and excluding neuroendocrine models. Gene sets enriched in both analyses were kept as likely to be representative of biological difference between responders and nonresponders. These results showed that nonresponsive models were enriched in gene sets involving fatty and amino acid metabolism, glucuronidation, and HNF1A transcriptional circuit activity, while responsive models contained enrichments in gene sets characterizing signaling pathways related to cell cycle and DNA maintenance and damage response pathways28 (Figure 3B). Transcriptomic analysis reveals molecular commonalities associated with docetaxel-responsive vs. nonresponsive mCRPC models (A) Volcano plot showing differentially expressed genes between RB1/TP53 intact docetaxel responders and nonresponders. (B) Gene set enrichment analysis showing gene sets enriched in RB1/TP53 intact organoid docetaxel responders and nonresponders. (C) Linear regression plot of nAUC for docetaxel in the complete organoid models cohort vs. Docetaxel response scores. See also Figure S2. To elucidate biological characteristics influencing docetaxel response in mCRPC, we derived a set of genes predictive for taxane responsiveness using the “nearest shrunken centroid” method29 applied to adenocarcinoma transcriptomic data for those models defined as either docetaxel responders or nonresponders (nAUC ≤ 0.5 standard deviations below or ≥0.5 standard deviations above the mean, respectively). Using as little as 19 genes (GLYATL1, CEACAM5, ALDH3A1, SLC17A4, UGT2B15, KLK1, DPP4, KLF12, NYNRIN, GLYATL2, NUDT10, TDO2, SERTAD4, A2ML1, RAMP1, SPINK1, NPFFR1, SFRP1, and HNF1A), the derived docetaxel score signature linearly modeled the docetaxel response range across organoid models, including those responding in an intermediate range that were not originally included in the signature derivation (Figure 3C). To determine how broadly the organoid cohort replicates patient samples, we assigned response scores to samples contained within the publicly available SU2C/PCF Dream Team clinical data transcriptomic database (see STAR Methods). The spread of scores observed in our organoid cohort overlapped the majority range of scores observable in the SU2C cohort, suggesting useful coverage (Figure S2A). To determine optimal docetaxel response score cutoffs applicable in other cohorts, a bootstrapping algorithm was used to simulate a range of projected variabilities. Thresholds were subsequently derived via receiver operating characteristic curve analysis, yielding an AUC of 0.95 for docetaxel, allowing for high specificity (95%) and selectivity (94%) in predicting responders, as well as high specificity and sensitivity (100%) in predicting nonresponders (Figure S2B). Using derived cutoffs for declaring predicted responders, we assigned patient samples predicted to be sensitive and resistant to docetaxel (Figure S2A). Predicted responders and nonresponders then were analyzed for whole transcriptome-determined characteristics. Using gene set enrichment analysis (GSEA), we observed that the biological characteristics differentiating responders and nonresponders in this clinical cohort had striking similarities to our organoid cohort (Figure S2C). Notably, for the 25 gene sets describing the biology separating responders and nonresponders in the organoid cohort, 23 were in agreement between the organoid and SU2C cohorts, and the two divergent gene sets can be explained by high false discovery rate values in the SU2C dataset. These data demonstrate that the 19-gene subset describing taxane responses in the organoid cohort was reflected in the global transcriptomic characteristics of patient samples, supporting the suggestion that the docetaxel response score can serve as a surrogate marker for predictive characteristics found to differentiate responsive and nonresponsive mCPRC tumors. Docetaxel response score outperforms histological phenotype, chromosomal instability, and replication stress in modeling docetaxel response in mCRPC We analyzed several clinically relevant molecular characteristics in tandem with docetaxel responses. Strikingly, the expression of HNF1A mRNA and protein seemed to be nearly exclusive to docetaxel-resistant organoid models (Figures 4A and S3A). RB1 and TP53 genotype status or transcriptionally defined chromosomal instability or replication stress scores did not correlate with response as well as the derived docetaxel response score. Of particular interest are two models derived from independent metastases within an individual patient. LuCaP 170.3 expresses HNF1A and was resistant to docetaxel treatment in vitro, while LuCaP 170.2 does not express HNF1A and was responsive. Docetaxel response phenotypes were validated in PDX trials, confirming the docetaxel response score prediction for a variety of models including LuCaPs 170.2 and 170.3 as well as additional adenocarcinoma (LuCaPs 136 and 167) and NEPC (LuCaP 93) models (Figure 4B). Docetaxel response score outperforms histological phenotype, chromosomal instability, and replication stress in modeling docetaxel response in mCRPC (A) Composite summary for 26 organoid models summarizing relative quantification of HNF1 RNA, select transcriptomic signatures, and Z-scored nAUC responses to selected drugs organized relative to docetaxel response signature scores. (B) In vivo docetaxel tumor responses. Assays were terminated when tumors reached 2,000 mm3 or at 5 weeks, whichever came first. A minimum of 5 mice were used per group. Error bars ± SEM. p values were calculated using the Student’s t test, two-tailed, unpaired. Correlations reported as R-squared between docetaxel nAUC and the remaining compounds in the figure: 0.4631 (topotecan), 0.4461 (volasertib), 0.2034 (navitoclax), 0.07768 (S63845), 0.2444 (WEHI.539), 0.267 (berzosertib), and 0.3402 (danusertib). See also Table S4. In addition, we observed compounds that shared response patterns with docetaxel, including the BCL-XL targeting agents, navitoclax and WEHI 539 (see Figure 4A). By contrast, the MCL-1 inhibitor S63845 did not share this pattern and even demonstrated promising activity in some of the most taxane-resistant models. In further analysis, the phosphatidylinositol 3-kinase inhibitor (Pi3K) AMG-511 was the one compound besides docetaxel in the screen with a statistically significant association with HNF1A expression (Table S4). However, the other four screened compounds targeting Pi3K did not produce the observed association, suggesting a potential off-target effect driving the AMG-511 association. HNF1A confers functional docetaxel resistance without preventing drug-target interaction To test whether HNF1A is a surrogate marker or has functional consequence in conferring docetaxel resistance, we ectopically expressed HNF1A or empty vector in LuCaP 23.1 organoids, a model with one of the strongest predictive scores within the cohort for docetaxel response. Lentivirus transduction achieved efficiencies of ∼90% within the organoid model (Figure S3B). Docetaxel responses in genetically modified HNF1A+ LuCaP 23.1 organoids revealed a shift in half-maximal docetaxel response curves of ∼10-fold, demonstrating a direct functional role for HNF1A in conferring docetaxel resistance that could be overcome at sufficiently high doses of drug (Figure 5A). As a specificity control, the replication stress-selective ATR inhibitor, berzosertib, was also tested. We observed no increased or decreased sensitivity associated with ectopic HNF1A expression (Figure 5A). Strikingly, transcriptomic analysis of ectopic HNF1A+ LuCaP 23.1 compared to empty vector revealed that nearly every gene in the docetaxel response score changed directionality upon the introduction of HNF1A, suggesting that the derived docetaxel score strongly reflects HNF1A-driven activity in mCRPC (Figure S3C). HNF1A confers functional docetaxel resistance without preventing drug-target interaction (A) Drug response curve showing docetaxel and berzosertib responses in 2 biological replicates with induced HNF1A expression. (B) Schematic showing confocal microscopy method for measuring mitotic arrest, defined as overlap of H3S10Phos stain in DAPI-positive regions in organoids treated with vehicle or docetaxel. Stains were DAPI (blue) and H3S10Phos (red). Scale bar, 50 μm. (C) Barplot summary showing percentage of cells across increasing time points and docetaxel doses (1) detected in mitotic arrest and (2) viably surviving. ∗∗∗∗ denotes p < 0.001. See also Figure S3. To determine whether drug-target engagement is preserved in HNF1A+ models, mitotic arrest was tracked following dose-dependent exposure to docetaxel at four time points over 48 h. A dose and time-dependent accumulation of cells positive for the mitotic marker, serine 10 phosphorylated histone H3 (H3S10phos+) peaked at 24 h for empty vector and HNF1A+ LuCaP 23.1 (Figure 5B). Higher doses were required for a response in the HNF1A+ variant, which even at the maximal dose tested remained at approximately 50% the level of the control cells (Figure 5C). Decreased survival was first observed around 48 h, 24 h after mitotic arrest peaks in both models. Docetaxel AUC determined at various time points over 7 days demonstrated kinetically similar responses (Figure S3D). Taken together, these data suggest that the mechanism of HNF1A+-induced docetaxel resistance is decreased sensitivity, which at pharmacologically usable doses (e.g., 1 nM) is translated to responsive and nonresponsive phenotypes (Figure 5). We note that untreated HNF1A+ models demonstrated similar proliferation rates to the parental cells (Figure S3D). HNF1A confers docetaxel resistance via apoptotic blunting Since apoptosis modulators correlated highly with docetaxel response in the screened cohort, the capacity for apoptosis induction by docetaxel was analyzed. Importantly, cleaved caspase-3 induction following docetaxel treatment was severely blunted in treated LuCaP 23.1 HNF1A+ variant cells compared to the parental control (Figures 6A and S4A). This implies that a consequence of HNF1A expression is inhibited apoptosis signaling following taxane-dependent mitotic arrest. By comparison, no change was observed in the response to navitoclax or S63845 (MCL1 inhibitor) targeting of intrinsic baseline apoptotic activity, suggesting that HNF1-mediated apoptotic blunting acted upon docetaxel-initiated apoptosis components (Figure S4B). HNF1A confers docetaxel resistance via apoptotic blunting (A) Confocal microscopy showing cleaved caspase-3 (green), DAPI (blue), and microtubules (red) in response to docetaxel treatment over 24 h. DAPI-stained nuclei were used for cell detection. Scale bar, 20 μm. (B) Volcano plot of genes differentially expressed between empty vector and HNF1A+ LuCaP 23.1 biological replicates. Apoptosis-associated genes are shown in red. (C) IGV screenshot of HNF1A and H3K27Ac ChIP-seq fragments relative to BIRC3 transcription start site and first exon. BigWig tracks were group autoscaled based on antibody, red tracks are from HNF1A high models, and blue tracks are from HNF1A low models. HNF1A motifs were derived using RGT HNT algorithm. See also Figure S4 and Table S5. We analyzed control vs. HNF1A+ differential expression for genes related to apoptosis to gain insight into a mechanistic rationale for docetaxel response differences, and we observed enriched expression of BIRC3 (encoding cellular inhibitor of apoptosis protein 2 [cIAP2]) in the majority of HNF1A+ models (Figure 6B). A potential mechanistic role for BIRC3 was further supported by the finding that its expression was statistically correlated with HNF1A status in both the screened organoid cohort (p = 0.0203) and the SU2C Dream Team 2019 patient cohort (p = 0.0172) (Figure S4C). To investigate a direct vs. an indirect effect of HNF1A on BIRC3 transcription, we separately performed chromatin immunoprecipitation sequencing (ChIP-seq) using antibodies targeting HNF1 and acetylated H3K27 (H3K27Ac) in the LuCaP 23.1 and LuCaP 170 models. As shown in Figure 6C, we observed HNF1A binding at a predicted HNF1A consensus nucleotide site at the transcription start site of BIRC3 in the LuCaP23.1-HNF1A transfectant but not the parental line. Differential binding analysis of the genetically modified LuCaP 23.1 model additionally revealed a set of sites targeted by HNF1A. Consistent with the HNF1A activity at the BIRC3 locus, we observed depletion of H3K27Ac surrounding the BIRC3 HNF1A promoter-binding site in parental LuCaP 23.1 but H3K27Ac presence in HNF1A+ LuCaP 23.1. In addition, H3K27/Ac signal was present surrounding BIRC3 in LuCaP 170.3 but not 170.2. A global analysis of H3K27Ac data was then used to identify binding motifs present in all models and was combined with the differentially bound sites from the HNF1A data to create a transcription factor motif binding database. GSEA on the set of differentially bound H3K27Ac sites for all models of high HNF1A activity against low HNF1A activity was found to be significantly enriched (adjusted p value = 2.366E−08) in the experimentally derived HNF1A-binding sites, including the BIRC3 locus (Figure S4D). To further characterize cIAP2 involvement, we confirmed the predicted differential cIAP2 protein expression and its maintenance under docetaxel treatment (Figure S4E), implying a potential target of resistance. cIAP2 can be pharmacologically inhibited with LCL161 (Figure 7A), which we found to be therapeutically active in HNF1A-transduced LuCaP 23.1 versus the non-expressing counterpart. Similarly, natively expressing HNF1A LuCaP 170.3 was more sensitive than LuCaP 170.2 (HNF1A negative), although both models demonstrated growth inhibition at sufficiently high drug concentrations (Figure 7B). Importantly, monotherapy with the cIAP2 inhibitor LCL161 proved sufficient to extinguish in vivo xenograft growth in the naturally occurring HNF1+ LuCaP 170.3 model but not the co-occurring HNF1-negative LuCaP 170.2 metastasis, demonstrating the potential of this target as a possible alternative therapy for docetaxel-resistant mCRPC (Figure 6E). As predicted, combined therapy in vitro of LCL161 and DTX demonstrated a significantly greater shift in responsiveness at lower docetaxel concentrations for HNF1A-positive compared to negative models (Figure 7C). Targeting cIAP2 is efficacious in HNF1A+ BIRC3-expressing mCRPC (A) Drug response curves to LC161 (cIAP2 inhibitor) alone in LuCaP 23.1 with and without exogenous HNF1A expression (left) and in LuCaPs 170.2 and 170.3. (B) In vivo responses to LCL161 in LuCaPs 170.2 (HNF1A−) and 170.3 (HNF1A+). A minimum of 5 mice were used per group. Error bars ± SEM. p values were calculated using the Student’s t test, two-tailed, unpaired. (C) Drug responses showing combination of docetaxel and LCL161 (cIAP2 inhibitor) in LuCaP 23.1 with and without exogenous HNF1A expression (left) and in LuCaPs 170.2 and 170.3. See also Figure S4. Discussion Despite the widespread use of taxanes in the treatment of mCRPC, the determinants of therapeutic response and resistance and biomarkers capturing the probability of response have not been established. Although initial efficacy is observed, mechanisms of resistance, as well as unnecessary treatment exposure and toxicity, are outstanding issues as objective responses to docetaxel therapy exist in just 12%–17% of patients,30 and the progression-free survival at 3 years on docetaxel approaches 0%.31 As one of the few extensive PDO high-throughput drug screens for mCRPC, we tested a variety of approved therapeutics and tool compounds against a heterogeneous cohort of PDO models spanning most PCa phenotypes encountered in the clinic, identifying a robust distribution of responses to docetaxel. The analysis of ∼30 models and the associated molecular markers allowed us to observe the role of HNF1A in mediating docetaxel resistance in mCRPC. As evidenced in naturally occurring and genetically modified HNF1A-expressing models versus non-expressing counterparts, there is a right shift in organoid docetaxel dose-response curves and a parallel loss of in vivo efficacy. The HNF1A-mediated docetaxel resistance we describe in this work demonstrated preservation of the ability for drug-target interaction based on analyses of microtubule staining and the observation of blocked metaphase plates at various concentrations and times after docetaxel treatment. Importantly, a role for decreased apoptosis secondary to HNF1A expression was apparent from the notable and selectively reduced levels of cleaved caspase-3 observed in mitotically arrested (i.e., target-engaged) cells expressing HNF1A. Consistent with a decreased apoptotic response, we noted that the expression of cIAP2, encoded by BIRC3, was statistically more likely to be expressed in HNF1A+ mCRPC, in both the organoid cohort and the SU2C Dream Team PCa patient cohort. The efficacy of targeting cIAP2 pharmacologically both in vitro and in vivo was selective to models expressing HNF1A. The role of inhibitor of apoptosis proteins (IAPs) across many solid tumor types has been characterized to show strong prognostic value for patient survival, tumor stage, and drug sensitivity to modulators of cell death targeting intrinsic and extrinsic apoptosis and necrosis.32 Data from LNCaP cells suggest that BIRC3 may have a pro-oncogenic role in prostate cancer.33 The use of IAP inhibitors in the clinic has been demonstrated to be well tolerated,34,35,36 though clinical efficacy of IAP inhibitors as monotherapy has been lacking and most ongoing clinical trials focus on combining IAP inhibitors with chemotherapy and/or radiation.37 While we have observed inhibition of cIAP2 as showing monotherapy efficacy, it is most likely that any clinical application would require combined use with docetaxel, with potential for increasing efficacy due to tumor heterogeneity and possibly preventing one form of anticipated resistance. The elucidation of a conserved cIAP2-targetable vulnerability in HNF1A-driven taxane-resistant mCRPC potentially provides a therapeutic opportunity in patient cohorts otherwise predicted to fail standard taxane treatment. The absence of validated predictive biomarkers for mCRPC continues to be a key limitation in treatment, leading to a lack in efficacious personalized medicine. The observation that HNF1A was exclusively expressed in the most nonresponsive models to docetaxel is intriguing, and importantly, the gene signature derived to model docetaxel response consists of genes under the influence of the HNF1A transcriptional circuit. Future work to validate key antibody-based markers encoded in the derived taxane response score is one direction toward establishing practical markers that prospectively predict a potential lack of taxane efficacy. HNF1A and HNF4G previously were shown to activate a gastrointestinal (GI) transcriptional circuit in prostate cancer, mediating castrate resistance as a result of lineage plasticity.27 The GI lineage has been defined by a signature based upon correlates of SPINK1 expression in mainly primary prostate cancer, and HNF1A was one gene within the top 100 most correlated genes. Although SPINK1 is contained within the docetaxel response signature, no other genes in the two signatures overlap. The lack of overlap is likely due to biological differences between the samples used to derive the signatures as well as technical differences in sample purity and preparation (i.e., surgical specimens of primary prostate cancer vs. organoid samples of CRPC). Nevertheless, HNF1A and SPINK1 expression are major defining characteristics that overlap, suggesting that docetaxel responsiveness in patients with GI lineage adenocarcinoma be evaluated. An interesting possibility is that inhibition of AR pathway signaling may artificially select for HNF1A+ mCRPC that will go on to drive docetaxel resistance in patients subsequently receiving taxanes as therapy for mCRPC. A variety of mechanisms underlying taxane resistance have been proposed, mainly based upon the selection and analysis of prostate cell line models. Two categories of resistance mechanisms have been described: (1) lack of drug-target engagement and (2) decreased sensitivity downstream of microtubule dysregulation. Taxanes act as agents that disrupt microtubule dynamics, resulting in inefficient spindle separation and ultimately mitotic apoptosis.38 In addition, alterations to interphase microtubules that disrupt intracellular transport lead to cell death in interphase.39 Related to the latter, taxane sensitivity due to decreased AR-dependent transcription following cytoplasmic retention of AR and resistance secondary to microtubule-independent transport of ARv7 have been demonstrated.24,40,41 However, there was no significant correlation between AR-dependent transcription and taxane AUC in the cohort used here (r = 0.27, with a p value of 0.1093), and there was no evidence that ectopic HNF1 modulated AR-dependent genes. Drug-target engagement mechanisms of resistance include alterations via structural mutation or post-translational modification to microtubule subunits or microtubule-associated proteins,42,43 although mutations to microtubules in clinical samples appear to be rare.44 Increased expression of multi-drug resistance (MDR) efflux pumps has been described as a resistance mechanism45,46; however, we observed no association between MDR efflux pump expression patterns and HNF1A status across the screen (Table S5). In support of this observation, a phase 3 clinical trial comparing docetaxel to MDR-resistant cabazitaxel as a first-line therapy for mCRPC demonstrated no difference in overall survival between cabazitaxel versus docetaxel regimens.47 Although we have identified HNF1A expression as one major source of taxane-initiated apoptotic resistance that is commonly expressed in several models, we anticipate that there is heterogeneity and contributions from other mechanisms across the cohort. Finally, the high-throughput drug screen data did not reveal any compounds to be preferentially efficacious for the models nonresponsive to docetaxel but point to broadly active drugs of interest for future investigation. For example, we noted the MCL-1 inhibitor S3845 and farnesyltransferase inhibitor tipifarnib as two compounds with similarly robust responses when compared to docetaxel, suggesting a similar approach could be used to investigate mechanisms of sensitivity and resistance. Additionally, both of these compounds contained docetaxel nonresponders among their responsive cohorts, suggesting the biology governing mCRPC vulnerability to these compounds is separate from that governing docetaxel response. Expression of MCL-1 has been shown to be higher in prostate cancers with advanced Gleason grade,48,49 protect prostate cancer cells from chemotherapy-induced cell death,50 and be a promising target to deal with senescence in prostate cancer.51 Tipifarnib has been less characterized mechanistically in prostate cancer but has been shown to have efficacy against prostate cancer cell lines.52,53 Its role in prostate cancer pharmacotherapy is yet to be tested in clinical trials. In terms of phenotypic associations with compound selectivity, we noted agents generally found to be active against hematological malignancies such as BCL-2 family acting agents to be preferentially efficacious against NEPC organoids, where NEPC organoids displayed a lower average nAUC compared to adenocarcinoma (0.539568 vs. 0.75574, respectively), an observation which has been previously described.22 An important consideration regarding the sensitivities and resistances observed in this screen is the 7-day duration of drugging. This time point was chosen to maintain robust in vitro first-generation growth in order to limit the influence of non-drug-induced cell death. Importantly, there are a few compounds that demonstrated a lack of efficacy in the screen, which likely could have proven efficacious had the drugging been conducted for a longer period of time. For example, agents targeting AR signaling such as enzalutamide showed very poor activity across the entire screen at the 7-day time point reported, whereas we previously have shown that a 14-day time point creates a range of responses across several of these PDO models.13 In summary, we have quantified an extensive inventory of drug responses using a broadly representative cohort of mCRPC models and identified a previously unappreciated association between HNF1A expression and taxane sensitivity. It will be important to design prospective clinical trials to investigate the range of consequences following docetaxel/cabazitaxel treatment that may be associated with HNF1 expression. Limitations of the study Although organoid and PDX models often are accurate predictors of initial responses to therapeutics, several physiological factors such as drug pharmacology and effects of the immune system in patients may influence response outcomes. In addition, there may be unknown factors related to organoid culture media and/or methods that influence the results. The conclusions presented here were not analyzed in appropriate taxane-treated patient populations due to a lack of available applicable data. Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Kathleen Kelly (kellyka@mail.nih.gov). Materials availability This study did not generate new materials. Data and code availability •Organoid RNA-seq data used in this article were previously uploaded to the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo) under accession ID GEO: GSE236808.54 The ChIP-seq and RNA-seq data generated in this article have been deposited into GEO and the Database for Genotypes and Phenotypes (https://www.ncbi.nlm.nih.gov/gap/). The processed data can be retrieved using the accession IDs GEO: GSE252203 (ChIP-seq) and GEO: GSE252216 (RNA-seq), while the raw data are accessible to authorized users under accession ID dbGAP: phs001587.v3.p1. All remaining data are accessible in the main text or supplementary materials. Organoid RNA-seq data used in this article were previously uploaded to the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo) under accession ID GEO: GSE236808.54 The ChIP-seq and RNA-seq data generated in this article have been deposited into GEO and the Database for Genotypes and Phenotypes (https://www.ncbi.nlm.nih.gov/gap/). The processed data can be retrieved using the accession IDs GEO: GSE252203 (ChIP-seq) and GEO: GSE252216 (RNA-seq), while the raw data are accessible to authorized users under accession ID dbGAP: phs001587.v3.p1. All remaining data are accessible in the main text or supplementary materials. •This paper does not report original code. This paper does not report original code. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This research was supported by the Intramural Research Program of the NIH, National Cancer Institute, Center for Cancer Research; Intramural Research Program of the NIH, National Center for Advancing Translational Sciences; Prostate Cancer Foundation (Young Investigator Awards to S.R. and M.L.B.); and Department of Defense Prostate Cancer Research Program (W81XWH-18-1-0767 to K.K.). Support from the CCR Single Cell Analysis Facility was provided by FNLCR Contract HHSN261200800001E. Palantir Foundry was used in the integration, harmonization, and analysis of drug screen data inside the secure NIH Integrated Data Analysis Platform (NIDAP). Author contributions Conceptualization, I.S.S., K.K., K.H.J., and C.J.T.; methodology, I.S.S., K.K., K.H.J., E.C., and H.N.; software, I.S.S., J.B., B.J.C., and R.L.; validation, I.S.S., K.H.J., and A.N.A.; formal analysis, I.S.S., J.B., B.J.C., and R.L.; investigation, I.S.S., K.H.J., J.B., A.N.A., M.R., Y.C.A., C.M., X.Z., P.S., M.L.B., and S.R.; writing – original draft, I.S.S. and K.K.; writing – review and editing, I.S.S., K.K., K.H.J., R.L., E.C., and C.J.T.; visualization, I.S.S., J.B., R.L., and B.J.C.; funding acquisition, K.K. Declaration of interests E.C. received sponsored research funding from Sanofi, Gilead, AbbVie, Genentech, Janssen Research, AstraZeneca, GSK, Bayer Pharmaceuticals, Forma Pharmaceuticals, Foghorn, Kronos, and MarcoGenics. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesH3K27ac AntibodyDiagenodeC15410196; RRID:AB_2637079HNF1A (D7Z2Q) AntibodyCell Signaling89670; RRID:AB_2728751Caspase-3 AntibodyCell Signaling9662; RRID:AB_331439c-IAP2 (58C7) AntibodyCell Signaling3130; RRID:AB_10693298Bcl-xL (54H6) AntibodyCell Signaling2764; RRID:AB_2228008Cleaved Caspase-3 (Asp175) (5A1E)Cell Signaling9664; RRID:AB_2070042Mcl-1 (D2W9E) AntibodyCell Signaling94296; RRID:AB_2722740Bim (C34C5) AntibodyCell Signaling2933; RRID:AB_1030947GAPDH (6C5) AntibodyAbcamAb8245; RRID:AB_2107448α Tubulin AntibodyMillipore SigmaT6074; RRID:AB_477582Phospho-Histone H3 (Ser10) (6G3)Cell Signaling9706; RRID:AB_331748Goat anti-Mouse AF647 AntibodyThermo Fisher ScientificA28181; RRID:AB_2536165Goat anti-Rabbit AF568 AntibodyThermo Fisher ScientificA11011; RRID:AB_143157Biological samplesPatient-derived xenograftsUniversity of WashingtonLuCapPatient-derived organoidsNational Cancer InstituteN/AChemicals, peptides, and recombinant proteinsBortezomibSelleckchemS2619PrENR mediaDrost et al.55N/ADocetaxelNCI DCTDNSC 759850LCL 161NCI DCTDNSC 772868Therapeutics LibraryThis paperSee Table S2Critical commercial assaysMACS Human Tumor Dissociation KitMiltenyi Biotec130-095-929Matrigel Matrix Growth Factor ReducedCorning354234ChIP-IT High SensitivityActive Motif53040Accel-NGS-2S DNA Library KitSwift Biosciences21096Protein G DynabeadsThermo Fisher Scientific10004DLipofectamine 2000Thermo Fisher Scientific11668019LentiBOOSTSiron BiotechSB-P-LV-101-12CellTiter-Glo 3DPromegaG9681RNeasy KitQiagen74104Halt Protease and Phosphatase InhibitorThermo Fisher Scientific784404–20% Criterion TGX Precast Midi Protein GelBio-Rad5671094PHEM BufferElectron Microscopy Sciences11163μ-Slide 8 WellIbidi80826Deposited dataOrganoid RNA-seq dataAgarwal et al.54GEO:GSE236808ChIP-seq dataThis paperGEO: GSE252203RNA-seq dataThis paperGEO: GSE252216CHiP-seq and RNA-seq raw dataThis paperdbGap: phs001587.v3.p1Experimental models: Cell linesLenti-X 293TTakara Bio632180Experimental models: Organisms/strainsNSG MiceCharles RiverStrain Code 614Recombinant DNAEF1a_mCherry_P2ADr. P. MaliAddgene plasmid # 135003EF1a_HNF1A_P2A_Hygro_BarcodeDr. P. MaliAddgene plasmid # 120448VSV-GDr. T. ReyaAddgene plasmid # 14888psPAX2Dr. D. TronoAddgene plasmid # 12260Software and algorithmsNextflow Core analysis pipeline10.5281/zenodo.3240506https://nf-co.re/chipseq/2.0.0/R version 4.1R Core Teamhttps://www.r-project.orgGraphPad Prism 8GraphPadhttps://www.graphpad.comPalantirPalantirhttps://www.palantir.com/HALO Image Analysis PlatformIndica Labshttps://indicalab.com Experimental model and study participant details PDX and organoid samples Patient-derived organoid tumors (PDOs) were established from prostate tumor biopsies (NCI.PC155, NCI.PC44, NCI.PC35.1, and NCI.PC35.2) at the National Cancer Institute (NCI, Bethesda, MD) and descriptions of the originating tumor characteristics as well as available demographic data have been published.13 LuCaP patient-derived xenografts (PDXs) were provided by Dr. E. Corey and Dr. R.L. Vessella, University of Washington; Seattle, WA, and descriptions of the originating tumor characteristics as well as available demographic data have been published.12 Tumors had their identities periodically confirmed using STR analysis by Laragen Inc. Prostate tumors are male in gender. PDX tumors were passaged in male NOD scid gamma (NSG) mice at the NCI and propagated in vitro as 3-dimensional organoid cultures as previously described.13 There was no selection based on phenotype, genotype, or growth characteristics for organoid models analyzed in the high-throuput drug screen. Animal studies All animal procedures including LuCaP PDX tumor propagation and in vivo drug studies were conducted at the NCI under an NCI Animal Care and Use Committee approved protocol and were carried out in accordance with ACUC recommendations. Six week old male NOD SCID gamma (NSG) mice and test compounds (Docetaxol; LCL161) were obtained from the NCI Division of Cancer Treatment and Diagnosis (DCTD) Frederick, MD. Mice were randomly assigned to treatment groups and subcutaneously injected on their flanks with 2 million processed LuCaP PDX tumor cells or implanted with ∼20mm3 tumor chunks. When tumors exceeded approximately 150 cubic millimeters, mice were randomly assigned into treatment and vehicle control arms. Docetaxel was administered 12 mg/kg IP biweekly. LCL161 was administered 75 mg/kg PO biweekly. Tumor measurements were taken biweekly using digital calipers until control or treatment arms reached 2000 cubic millimeters. A minimum of five mice per treatment arm were used. Method details In vitro PDX-Organoid drug response high-throughput screen (HTS) Prior to screening procedures, all PDO and PDX models had their identity validated using single tandem repeat (STR) analyses, and two screened models were removed from data analysis due to failed STR validation. For drug response assays, PDXs were enzymatically degraded using the MACS human Tumor Dissociation Kit (Miltenyi Biotec, 130-095-929) and seeded at 2000 cells/well in 384-well format in 3D-culture composed of Matrigel (Corning, 354234) mixed with Prostate-specific ENR growth media (PrENR)55 in a ratio of 80/20%, respectively, with a final volume of 20 μL/well, and incubated at 37 C for 1 h prior to adding an additional 30 μL media to each well. The following day plates were treated with a 10-point dose curve and a 1:3 serial dilution per compound and allowed to grow for one week, with one media/drug change 72 h after the initial dosing. Bortezomib (2.3 μM) was used as positive control for cytotoxicity. At the end of the week, viability in each well was quantified using CellTiter-Glo 3D (Promega, G9681) according to manufacturer’s protocol. Individual drug validation screens were conducted in-house using the same protocol performed manually. Proliferation assays Organoids were seeded at 1500 cells/well in 384 well plates. CellTiter-Glo 3D (Promega, G9681) measurements were taken at 1 h and 7 days post seeding. Proliferation was reported as percent change in luminescence. RNA library preparation, sequencing, and differential gene analysis RNA libraries were prepared from first generation organoids using RNeasy kit (QIAGEN, 74104) according to manufacturer’s instructions and profiled for quality using a Bioanalyzer (Agilent). Poly-A–enriched and Illumina–barcoded libraries were prepared and sequenced on a NextSeq 500 (Illumina), to a depth of 75 million base pairs (150 bp paired end) at the Illumina Sequencing Facility, Center for Cancer Research. RNA sequencing data was aligned to grch37 using the nextflow core RNAseq pipeline (https://github.com/nf-core/rnaseq). Transcripts were quantified using featureCounts from the subread package (http://bioconductor.org/packages/release/bioc/html/Rsubread.html), and the R bioconductor package edgeR (https://bioconductor.org/packages/release/bioc/html/edgeR.html) was used for performing differential gene expression analysis. Data was normalized using TMM, and differential expression was assayed using the quasi likelihood framework F test. The bioconductor package clusterProfiler was used to perform gene set enrichment analysis (GSEA) and overrepresentation analysis (https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html). Immunoblotting PDO Sample lysates were prepared in RIPA buffer supplemented with HALT protease and phosphatase inhibitor (Thermo Fisher, 78440). Protein concentrations were quantified using a BCA assay (Thermo Fisher, 23225) and standardized to equivalent loading into a 4–20% CriterionTMTGXTM Precast Midi Protein Gel (Bio-Rad, 5671094). Gels were transferred to polyvinylidene difluoride membranes and blocked in TBS with 0.1% Tween 20. Primary antibodies include HNF1A (Cell Signaling 89670), Caspase-3 (Cell Signaling 9662), Cleaved Caspase-3 (Cell Signaling 9661), cIAP2 (Cell Signaling 3130), MCL-1 (Cell Signaling 94296), BCL-XL (Cell Signaling 2764), BIM (Cell Signaling 2933), and GAPDH (Abcam ab8245). Confocal microscopy PDO models were seeded into 8 well chamber slides (Ibidi, 80826) coated with Poly-D-Lysine at 105 cells per well and allowed to adhere overnight at 37C. If drugging was intended, drugs were added to wells after 1 h adhesion. For images requiring microtubule visualization, microtubule stabilization steps were added prior to fixation including permeabilizations with PHEM Buffer (Electron Microscopy Sciences, 11163) supplemented with 0.2% Triton X-100. Otherwise, slides were processed at the given time point to fixation under 4% formaldehyde. Fixed slides were washed three times at room temperature in TBS +0.05% Tween 20, blocked for 1 h at room temperature with PHEM +2% BSA, and incubated with primary antibodies including HNF1A (Cell Signaling 89670), Alpha Tubulin (Sigma T6074), H3S10phos (Cell Signaling 9706), or Cleaved Caspase-3 (Cell Signaling 9664). Secondary antibodies used were Anti-mouse 647 (Thermo Fischer A28181) and Anti-Rabbit 568 (Invitrogen A11011). Images were visualized on Carl Zeiss LSM 780 laser scanning confocal microscope using a 40x Plan-Apochromat (N.A. 1.4) DIC oil objective. Z stacks were collected using a 1 μm step size and converted to a maximum intensity projection using Zen software. Images were analyzed, including automated cell detection and signal intensity quantification, using the HALO Image analysis program. Mitotic arrest assay LuCaP 23.1 empty vector control and LuCap 23.1 HNF1A + models were seeded into 8 well chamber slides (Ibidi, 80826) coated with Poly-D-Lysine at 105 cells per well and allowed to adhere overnight at 37C. Docetaxel was added after 1 h adhesion. Slides were treated as described under confocal microscopy protocol and costained with DAPI + H3S10phos (Cell Signaling 9706). Quantification of percent cells in cell-cycle arrest was determined by using H3S10phos+/DAPI + cells, detected using the automated HALO Image analysis program. HNF1A ectopic expression model generation HNF1A and empty vector plasmids were a gift from Prashant Mali.56 LuCaP 23.1 organoids were dissociated to single cells for transduction. Lentivirus was made by transfecting Lenti-X 293T cells (Clontech 632180) with EF1a_HNF1A_P2A_Hygro_Barcode (Addgene plasmid # 120448) or empty vector EF1a_mCherry_P2A_ (Addgene plasmid # 135003) plus psPAX2 and VSV-G packaging plasmids using Lipofectamine 2000 (Invitrogen 11668019). LuCaP 23.1 cells were transduced with generated lentivirus via centrifugation and LentiBOOST(Sirion Biotech). Following appropriate selection, HNF1A expression was confirmed via RNAseq and western blotting. CHiP sequencing Chromatin sample preparation was obtained using the Active Motif ChIP-IT High Sensitivity kit per manufacturer instructions (Active Motif, Carlsbad, CA). Samples were immunoprecipitated overnight using ChIP-seq grade H3K27ac antibody (Diagenode, cat. No C15410196) or HNF1A (Cell Signaling 89670), bound to protein G dynabeads (Thermo Fisher, Waltham, MA). A fraction of sample was not exposed to antibody and used as input control. Samples underwent crosslink reversal with RNase and proteinase K, and DNA sequencing libraries were prepared using the Accel-NGS-2S DNA Library Kit (Swift Biosciences, Ann Arbor, MI). Libraries were sequenced using 100-bp paired end reads using NovaSeq 6000 on the Illumina platform at the CCR-Sequencing Facility (Frederick, MD). Sequence analysis was conducted using Nextflow Core analysis pipeline (https://doi.org/10.5281/zenodo.3240506). Quantification and statistical analysis HTS results were uploaded into the Palantir platform (Palantir, Denver, CO), where pipeline-generated compound response curves were normalized to DMSO negative controls and berzosertib positive controls for each data point. Normalized Area Under Curve (nAUC) was calculated by dividing an individual compound-model AUC by maximum possible AUC for the exposed dose range. Responder class was assigned on a drug-by-drug basis where “responders” were defined as models with nAUC at least 0.5 standard deviations below mean nAUC for that drug, and “nonresponders” at least 0.5 standard deviations above mean nAUC. Validation data linear regressions and pie charts were analyzed using spearman regression in R. Data visualization for linear regressions, principle components analysis, differential gene volcano plots, dot plots, and histograms were created using GGplot2 package (https://ggplot2.tidyverse.org) for R. Heatmaps were created using the ComplexHeatmap package for R (https://doi.org/10.1093/bioinformatics/btw313). Docetaxel-response signature Docetaxel-response signature was generated using the nearest shrunken centroid method29 implemented in the pamr package for R (https://www.bioconductor.org/packages//2.7/bioc/html/pamr.html) to learn a classifier for discrimination between organoid model strong and weak responders to docetaxel. The thresholds were selected for a minimal number of genes to reduce possible misclassification error and increase overall efficacy. These genes were selected from a matrix of zFPKM-standardized RNA sequencing data57 and crossed with the weights across all organoid models. A final signature consisting of 19 genes with associated weights allowed for their summation and generation of a single numerical classifier, which demonstrated a continuous relationship with docetaxel nAUC when tested via linear regression against all organoid docetaxel nAUC values. Signature value cutoffs for prediction of responders and nonresponders using the signature were determined via a bootstrapping Receiver Operating Characteristic (ROC) curve package, pROC package for R, to maximize sensitivity and specificity for the internal organoid cohort, which was subsequently applied to clinical samples for predicted docetaxel responders and nonresponders. Clinical data analysis SU2C Dream Team 2019 data were downloaded from cbioportal.18 Data was zFPKM-standardized57 prior to Docetaxel Score assignment or gene set enrichment analysis. Published: December 9, 2024 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.xcrm.2024.101868. Supplemental information Document S1. Figures S1–S4 and Tables S1 and S3–S5 Table S2. Excel datasheet of drugs used in the high-throughput screen, related to Figure 2 Document S2. Article plus supplemental information References Associated Data Supplementary Materials Document S1. Figures S1–S4 and Tables S1 and S3–S5 Table S2. Excel datasheet of drugs used in the high-throughput screen, related to Figure 2 Document S2. Article plus supplemental information Data Availability Statement •Organoid RNA-seq data used in this article were previously uploaded to the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo) under accession ID GEO: GSE236808.54 The ChIP-seq and RNA-seq data generated in this article have been deposited into GEO and the Database for Genotypes and Phenotypes (https://www.ncbi.nlm.nih.gov/gap/). The processed data can be retrieved using the accession IDs GEO: GSE252203 (ChIP-seq) and GEO: GSE252216 (RNA-seq), while the raw data are accessible to authorized users under accession ID dbGAP: phs001587.v3.p1. All remaining data are accessible in the main text or supplementary materials. Organoid RNA-seq data used in this article were previously uploaded to the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo) under accession ID GEO: GSE236808.54 The ChIP-seq and RNA-seq data generated in this article have been deposited into GEO and the Database for Genotypes and Phenotypes (https://www.ncbi.nlm.nih.gov/gap/). The processed data can be retrieved using the accession IDs GEO: GSE252203 (ChIP-seq) and GEO: GSE252216 (RNA-seq), while the raw data are accessible to authorized users under accession ID dbGAP: phs001587.v3.p1. All remaining data are accessible in the main text or supplementary materials. •This paper does not report original code. This paper does not report original code. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.