Patient-derived organoids across cancers reveal conserved tumor heterogeneity and actionable therapeutic vulnerabilities

Abstract

We developed a pan-cancer patient-derived organoid (PDO) platform comprising 220 PDOs from 191 patients across 15 cancer types to advance functional precision oncology. Comprehensive characterization demonstrated high fidelity to parent tumors, with 93% histopathology concordance, 80% median genomic concordance for driver mutations, and a 0.85 median gene expression correlation. Expression profiles remained largely stable over 10 passages, ensuring reproducibility for long-term screening. Clonality analysis showed that 85% of dominant tumor clones were preserved, with genomic concordance directly reflecting clonal similarity. Even PDOs with lower concordance retained key oncogenic drivers, validating their utility as disease models. Functional assays revealed that 58% of PDOs from patients ineligible for US Food and Drug Administration–approved poly(adenosine 5′-diphosphate–ribose) polymerase inhibitors were sensitive to talazoparib, linked to DNA damage repair alterations. Furthermore, combination screens identified agents that overcome resistance, particularly in TP53-mutant models. Our platform enables the investigation of targeted therapies and molecular drivers of drug sensitivity, providing translational insights for personalized treatment beyond current biomarker guidelines.

Pan-cancer PDOs model tumor biology and clonality, enabling drug screening and uncovering responses beyond guidelines.

INTRODUCTION

Functional precision oncology platforms enable direct testing of cancer therapies in patient-derived tumor tissues, identifying vulnerabilities that may not be detectable through molecular profiling alone (1, 2). This approach, crucial for advancing precision medicine, addresses the complexity and heterogeneity of tumors, offering treatment options to patients without actionable biomarkers or with acquired drug resistance. Within functional precision oncology, patient-derived organoids (PDOs) have emerged as promising preclinical models. PDOs are increasingly integrated into functional precision oncology, alongside advanced technologies such as next-generation sequencing and clinical data, to uncover patient-specific molecular features and guide the development of more effective, targeted therapies.

PDOs have been generated from a wide variety of cancer types, and pan-cancer PDO cohorts have been reported (3, 4). These renewable resources are invaluable to study disease progression, test drug efficacy, conduct toxicity studies, and may even serve as patient avatars for treatment and coclinical trials (5–7). To ensure reliable results from these models, it is important to standardize and validate the PDO platforms used for these applications. However, inconsistencies among existing PDO pipelines, particularly in model characterization and validation—meaning the extent to which PDOs recapitulate key molecular, histological, and functional features of the parent tumor—remain a substantial challenge. For example, criteria to determine successful generation of PDOs are not universally defined, strategies to evaluate their fidelity to parent tumors at molecular level are not routinely applied, and biomarker assessment for cancer subtype classification is not consistently performed. In addition, concerted efforts to collect clinical data for screened PDOs are lacking, although this information is crucial to correlate drug responses with patient characteristics. As PDO pipelines continue to evolve and integrate into functional precision oncology, robust standardization efforts are necessary to ensure their effectiveness for personalized drug testing and reliability as models to predict patient responses to therapies.

To support this goal, we present a comprehensively characterized pan-cancer PDO platform comprising 220 PDOs from 191 patients, validated against parent tumors for histopathology, genomic features, and expression profiles. This work advances our PDO platform introduced in 2017, broadening its characterization and application (3). Here, we report high rates of concordance between PDOs and parent tumors in tissue and cytomorphology (8, 9), somatic mutations, somatic copy number alterations (CNAs), and expression profiles, confirming the platform’s reliability to generate preclinical models that faithfully recapitulate their matching tumor biology. As a proof of concept for its application in functional precision oncology, we tested the sensitivity of a subset of PDOs to talazoparib, a poly(adenosine 5′-diphosphate–ribose) polymerase inhibitor (PARPi). These PDOs were derived from patients with cancer who, based on current US Food and Drug Administration (FDA) guidelines, were not eligible for PARPi-based therapies (10–14). This application highlighted the capability of our platform to investigate targeted therapies and identify genomic and transcriptomic trends associated with drug sensitivity, while proposing potential combination therapy strategies across a wide range of cancers. Our well-characterized PDO platform offers promising preclinical models to advance precision oncology research with a goal to improve patient care.

RESULTS

Development and characterization of successful PDO models

We developed a robust PDO platform to establish, characterize, and leverage PDOs from tissues across 15 different cancer types. Between 2014 and 2022, we received 1472 tissue samples and initiated cultures for 1408. We observed successful PDO formation in 81% (n = 1140 of 1408) of samples (Fig. 1A). Cultures that failed to form PDOs at passage 0 (p0) were either contaminated with bacteria, particularly those from research autopsies (15), or did not have enough viable cells. PDOs that did not grow beyond five passages (p5) or failed to expand sufficiently for biobanking were classified as short-term cultures (53.5%, 610 of 1140) (Fig. 1A). We successfully biobanked 128 (11.2%) cultures that achieved expansion to 1 × 10⁶ cells by p5 before showing limited proliferative capacity and therefore could not be sustained further. An additional 402 cultures (35.3%) continued to expand and were biobanked at p5 (Fig. 1A). At the time of data freeze (April 2023), 235 PDOs had been established, which expanded consistently to p5. Fifteen of the established PDOs failed the single-nucleotide polymorphism (SNP) panel identification assay (SPIA) quality assessment, indicating a mismatch between the tumor or PDO and the matched germline sample, and were therefore excluded from further analyses. The remaining PDOs (n = 220) were characterized on the basis of their histopathology and/or molecular features and compared with their parent tissue (Fig. 1A). The final cohort described in this manuscript comprises 220 established and characterized pan-cancer PDOs derived from 191 patients (Fig. 1, A and B, and Table 1).

Fig. 1. Overview of the PDO pan-cancer cohort.

(A) PDO development and characterization workflow. All numbers were calculated at time of data freeze (April 2023). (B) Pie chart showing the number and fraction of established PDOs by primary tumor type. (C) Hematoxylin and eosin (H&E)–stained slides of three representative pairs of PDOs and parent tumors, which have highly concordant tissue morphologies. (D) H&E-stained slides of four representative PDO and parent tumor pairs, illustrating minor morphological differences. Left: A breast adenocarcinoma PDO exhibiting tubular and cribriform structures, in contrast to the solid growth pattern of the parent tumor. Middle left: A gastrointestinal (GI) adenocarcinoma PDO with papillary/micropapillary architecture and a high nuclear-to-cytoplasmic ratio, compared to the rounded tubular configuration with columnar nuclei in the parent tumor. Middle right: A colon adenocarcinoma PDO with rounder nuclei, contrasting with the distinctly columnar nuclei of the parent tumor. Right: An endometrial adenocarcinoma PDO showing a relatively higher nuclear-to-cytoplasmic ratio than the parent tumor. Scale bars, 150 μm.

Table 1. Patient and sample characteristics for established PDOs.

Characteristics	PDOs, total N (%)
Cohort
Patients	191
Samples	220
Age
Median (range) years	63 (23–94)
Sex assigned at birth
Female	80 (42%)
Male	74 (39%)
Not available	37 (19%)
Self-reported race
Asian or Pacific Islander	9 (5%)
Black	23 (12%)
White	98 (51%)
Not listed	10 (5%)
Declined	12 (6%)
Not available	39 (21%)
Self-reported ethnicity
Hispanic or Latino	8 (4%)
Non-Hispanic	113 (59%)
Declined	20 (11%)
Not available	50 (26%)
Prior treatment
No pretreatment	18 (9%)
1–3 pretreatment	79 (42%)
3+ pretreatment	43 (23%)
No information	51 (27%)

The success rate of PDO formation, determined as the proportion of initiated cultures that showed organoid formation out of the total cultures initiated, improved notably with our workflow from 57.1 to 81.8% over time and varied by sample source, with higher formation in fluid samples (92.5%) compared to autopsy samples (57.6%), as well as by cancer type, with the highest success in endometrial cancers (94.2%) and the lowest rates in prostate cancer (69.4%) (fig. S1, A to C). PDOs for some tumor types required longer time to successfully establish than others, probably due to suboptimal growth conditions or inherently slower tumor proliferation. For example, the median time to establish breast PDOs was 86.5 days compared to 36 days for ovarian tumors (table S1).

In our cohort, bladder, breast, endometrium, and lung cancer PDOs were mostly derived from primary lesions (over 80%), while colorectal, pancreas, skin, and prostate cancer PDOs were predominantly metastatic (over 65%) (table S1). Colorectal (28.6%), lung (12.3%), endometrium (10.5%), breast (10.9%), and pancreatic (10%) cancers represented the most common tumor types in the established cohort (Fig. 1B). Our PDO collection also included a small subset of melanomas (0.9%), cancers of unknown primary (1.4%) and normal organoids, i.e., PDOs from benign tissues (3.6%) (Fig. 1B).

In summary, our cohort of 220 PDOs, established from 191 patients from multiple race groups across 15 different cancer types and benign tissues with well-characterized molecular features, offers a diverse and robust set of models for functional precision oncology research. Table 1 summarizes the cohort demographics, including biological sex, self-reported race, and ethnicity, as well as total number of treatments received before tissue collections.

PDOs are morphologically similar to parent tumors

We performed a thorough histopathology review of hematoxylin and eosin (H&E)–stained sections of 189 PDOs and their parental tissues to determine whether tumor cytomorphology and cellular characteristics were maintained in the organoids (8, 9). Additional seven samples were embedded and confirmed to be tumor, but their corresponding parental tissues were not available for concordance assessment. A select subset of immunohistochemistry (IHC) markers was used to confirm biomarker expression. The remaining 24 PDOs had not been embedded because of logistical issues: Their establishment predated the systematic application of histopathological review, and by the time of data freeze, prioritization of limited material was given for molecular characterization.

A high concordance was observed in 93% (176 of 189) of PDOs closely matching their parent tissues. Most of them (77%, 145 of 189) mirrored the cellular architectures of their parental tumors and were consistent with the histopathology of the given tumor type. For example, PDOs derived from colon and endometrial adenocarcinomas preserved their characteristic tubular (glandular) structures. Similarly, PDOs from high-grade urothelial carcinomas retained features of the parent tumors, including evidence of divergent differentiation, such as squamous morphology (Fig. 1C). The remaining subset of PDOs, 31 of 189 (16.4%), showed minor discrepancies where the PDO was morphologically slightly different from the parental tumors with different cytomorphologies (21 of 31), histology grades (7 of 31), or biomarker expressions (3 of 31) (Fig. 1D and table S2). Only 7% (13 of 198) of the samples failed pathological review potentially due to contaminations of the tissue samples with neighboring cells, resulting in normal cell overrepresentation from cancerous tissue (8 of 13), benign-appearing tissues that yielded tumor cell growth (4 of 13), or a potential mix up of two PDO samples (1 of 13) (fig. S1D and table S2). In some cases, small clones—either normal or cancerous—expanded to become the dominant population in organoid culture, further emphasizing the need for rigorous pathological review. To address these issues, two histopathologists independently assessed the neighboring sections to determine the degree of contamination in our samples and, if needed, recharacterized the PDOs as either cancerous or normal. These measures to refine the classification of our PDOs ensured their reliability and accuracy as disease models for further research.

Genomic features are conserved between PDOs and parent tumors

The next step in the pipeline was to assess how closely PDOs represented their parent tumors by comparing their genomic features. We analyzed mutational profiles of 147 unique pairs of PDOs and parent tumors using whole-exome sequencing (WES; n = 114 pairs) and targeted gene panels (n = 33 pairs) (table S3). We identified a strong positive correlation among tumor-PDO pairs for tumor mutation burden (TMB) (r = 0.79, P < 2.2 × 10⁻¹⁶) and microsatellite instability (MSI) scores (r = 0.94, P < 2.2 × 10⁻¹⁶) (Fig. 2A). The mutational burden was not significantly different between tumors and matched PDOs (paired t test, P = 0.77; fig. S2A). The ploidy correlation was modest (r = 0.34, P < 0.00025); however, after excluding 21 outlier pairs (18% of the dataset), the correlation significantly improved to 0.81 (fig. S2, B and C).

Fig. 2. Comparative analysis of genomic characteristics and mutation profiles between PDOs and parent tumors.

(A) Pearson correlation among PDO and parent tumors for aggregate measures of genomic characteristics. Asterisk (*) denotes a significant correlation (P < 0.05). (B) Box plots showing the distribution of genomic concordance between PDOs and parent tumors across 12 cancer types, calculated on the basis of the proportion of shared nonsilent mutations. Bar plot at the bottom shows the percentage of samples that belong to the low (<40%), mid (40 to 80%), and high (≥80%) concordance groups. (C) Heatmap showing the most frequent COSMIC SBS mutational signatures, selected on the basis of contributions of >0 in a minimum number of samples in the cohort. The density distribution plot on the right shows the distribution of the cosine similarity scores of COSMIC SBS signature profiles between paired PDOs and parent tumors. (D) Oncoprint displaying the top 21 most frequent cancer driver genes in the cohort, selected on the basis of a mutation frequency of 9% or higher in either the tumor or the PDO sample groups. Bar plots on the right indicate the frequency of each mutated gene within the tumor and PDO sample groups.

To assess genomic fidelity at the mutation level between tumors and their matched PDOs, we calculated a metric termed somatic nucleotide variant (SNV) concordance, defined here to include single-nucleotide variants and small insertions/deletions. This metric represented the proportion of tumor mutations also detected in the matched PDOs. Across the cohort, the median SNV concordance was 60%, with approximately one-third of PDO-tumor pairs showing high SNV concordance (≥80%) (Fig. 2B). SNV concordance varied by cancer type, being highest for prostate tumors (84%, n = 5) and lowest for the only esophageal tumor in the cohort (15%, n = 1) (Fig. 2B) (16, 17). The low concordance in this single esophageal pair could be attributed to the low inferred cancer cell purity of 27% for the PDO. However, unexpectedly, in general, the SNV concordance was independent of tumor purity in our cohort (Pearson r = 0.13, P = 0.18), as, for the pairs with SNV concordance of ≤40%, the average purity was 92.5% in PDOs and 66% in tumors (table S3 and fig. S2D).

When we restricted our WES SNV concordance analysis to 200 pan-cancer driver genes (18), our SNV concordance increased to 80% (fig. S2E and table S3). These results are in alignment with the previously reported somatic concordance of 77.6% for a pan-cancer cohort (4). For targeted panels, our SNV concordance (median = 88%) was even more strongly conserved, with 15 of 33 tumor-PDO pairs showing a 100% concordance (fig. S2, F and G, and table S3). Consistent with previous reports, these findings confirm a robust SNV concordance between PDOs and parent tumors in our cohort, especially for known cancer driver genes.

In addition to SNV concordance, we tested whether the characteristic patterns of somatic mutations were conserved between PDOs and parent tumors. We quantified the contributions of 65 COSMIC mutation signatures for each sample and identified strongly conserved mutational signature profiles between the PDO and parent tumor pairs (cosine similarity = 0.83) (Fig. 2C) (19, 20). Single base substitution 1 (SBS1), the clock-like aging signature, was the most common signature among the samples in our cohort, consistent with reports of SBS1 being the most common COSMIC signature found across many cancer types (21). We also identified conserved signatures that have known prognostic associations with survival in certain cancer types. For example, the apolipoprotein B mRNA editing enzyme, catalytic subunit 3 (APOBEC3)-induced mutational signatures SBS2 and SBS13 were highly conserved between the bladder cancer PDO-tumor pairs (cosine similarity = 0.97) (22), suggesting that PDOs faithfully recapitulate this important mutational process.

PDOs recapitulate driver mutations of parent tumors

Cancer driver mutations are critical for selecting models in drug screening, especially for targeted therapies and precision medicine. To evaluate our models, we assessed whether PDOs exhibited any bias toward specific driver mutations. Our analysis revealed no significant differences in driver mutation rates between PDOs and their corresponding parent tumors (Fig. 2D and table S4). This finding remained consistent when the analysis was extended to include all mutated genes across the cohort [false discovery rate (FDR) = 1] (table S5).

To further examine how well PDOs recapitulated individual driver mutations, we compared driver mutation frequencies in PDOs to those observed in parent tumors and to expected frequencies in corresponding cancer types from The Cancer Genome Atlas (TCGA) (Fig. 2D and fig. S2F). We identified tumor protein p53 (TP53) as the most frequently mutated gene in our cohort (tumor = 61% and PDO = 60%; n = 114), consistent with its status as the most frequently mutated gene in TCGA pan-cancer cohorts (44.6%, n = 6408). Besides TP53, KRAS (tumor = 34% and PDO = 34%) and APC (tumor = 28% and PDO = 28%) were the top mutated genes, a pattern representative of the higher proportion of colorectal tumors in our cohort (∼31%) (Fig. 2D and table S4).

To test whether PDOs served as good disease-specific models, we analyzed the frequency of cancer-specific mutations within each corresponding cancer type (fig. S3, A to H, and table S6). For most PDOs, driver mutations overlapped well with their corresponding parent tumors and occurred at expected frequencies. For example, colorectal tumors (n = 34) were enriched in TP53 (tumor = 74.3% and PDO = 73.5%), KRAS (tumor = 65.7% and PDO = 67.6%), and APC (tumor = 57.1% and PDO = 55.9%) mutations (fig. S3A). These genes were also the top three most frequently mutated drivers in the TCGA–colon adenocarcinoma and TCGA–rectum adenocarcinoma cohorts (TP53 = 59.5%, KRAS = 42.2%, and APC = 74.2%) (23, 24). Similarly, lung tumors (n = 15) were enriched in KRAS [tumor = 46.7%, PDO = 46.7%, TCGA–lung adenocarcinoma (LUAD) = 46.1%], TP53 (tumor = 40%, PDO = 46.7%, and TCGA-LUAD = 50.8%), EGFR (tumor = 20%, PDO = 26.7%, and TCGA-LUAD = 14.3%), and KEAP1 (tumor = 20%, PDO = 20%, and TCGA-LUAD = 17.4%) mutations (fig. S3B) (23, 24). Bladder tumors (n = 14) were enriched in mutations for TP53 [tumor = 64.3%, PDO = 64.3%, and TCGA–bladder urothelial carcinoma (BLCA) = 49.2%], KMT2D (tumor = 50%, PDO = 50.0%, and TCGA-BLCA = 27.7%), PIK3CA (tumor = 35.7%, PDO = 28.6%, and TCGA-BLCA = 20%), and FGFR3 (tumor = 28.6%, PDO = 21.4%, and TCGA-BLCA = 12.3%) (fig. S3C). These findings show that PDOs faithfully recapitulate the driver mutation frequencies of their parent tumors and closely mirror TCGA cohort data, supporting their utility as robust models for disease-specific drug screening and precision medicine.

Clonal architecture partly explains differences in genomic concordance

As noted above, most PDOs retained cancer-specific driver mutations, including those with low SNV concordance (<40%), supporting their relevance as good experimental models (fig. S3, A to H, and table S7). In breast cancer, however, low-concordance pairs (n = 13) lacked key driver mutations (Fig. 3A and table S7), raising the possibility that perhaps clonal selection in PDO cultures may have contributed to this mutational discordance. To investigate this, we inferred the clonal composition in breast cancer tumor-organoid (TO) pairs to visualize how mutation-level concordance relates to clonal structure (Fig. 3B and fig. S4). In a representative low-concordance pair (WCM2108, concordance = 6%), the tumor and PDO retained a shared truncal ancestor (cluster 0) while displaying divergent subclonal trajectories (distinct branches). This indicates that the low mutation overlap reflects sampling of different subclones rather than a lack of biological relatedness. Conversely, in a high-concordance pair (WCM2137, concordance = 98%), the clonal hierarchy was largely mirrored between the tumor and PDO (Fig. 3B). Clonal trees of the remaining breast TO pairs display a more heterogeneous relationship, where both clonal presence and prevalence vary between the tumor and PDO (fig. S4). Although clonal selection of distinct subclones frequently drives divergence, substantial variability in mutation burden across clones and the absence of dominant shared mutation clusters in some pairs (e.g., n = 62 versus n = 114 in WCM2124) indicates that clonal architecture alone does not fully explain mutation-level concordance.

Fig. 3. Cellular prevalence of mutations in PDOs compared to matched tumors.

(A) Oncoprint showing the frequency of breast cancer–specific driver genes (18) in tumors and matched PDOs, ordered by their genomic concordance levels. (B) Graphical representation of shared clonal architecture between tumor and matched PDOs. Lowest concordance (left) and highest concordance (right) examples from the breast cancer cohort sore are shown. Genes along the edges indicate known driver mutations. Sphere of cells represent proportion of clonal subpopulations in a sample. (C) Scatter plot comparing cellular prevalence of mutation clusters inferred using PyClone-VI. Each point represents a distinct cluster, with its estimated prevalence in the tumor (x axis) and the corresponding PDO (y axis). Points are colored by degree of concordance. The diagonal line represents equal prevalence in both PDO and matched tumor. (D) Box plots for fraction of clones per TO pair within each clonal divergence category (highly concordant, moderately concordant, and divergent), stratified by SNV concordance status (<40, 40 to 80, and ≥80%). Each dot represents one TO pair in that divergence category. Median values of each box plot are marked by dark red dot. P values are from pairwise Wilcoxon tests, and only significant P < 0.05 is shown. Clone classification was based on absolute differences in cellular prevalence inferred using PyClone-VI. (E) Bar plots showing the median percentage [±interquartile range (IQR)] of dominant tumor clones across different primary tumor sites. Clone status was defined on the basis of changes in cellular prevalence between tumor and PDO: expanded (>1.25× increase), contracted (<0.75× decrease), unchanged (within ±25%), or lost (detected in tumor, absent in PDO). Only clones with >10% cellular prevalence in the tumor were considered dominant and used for this analysis. Tumor types with fewer than two matched pairs were excluded.

To better understand global patterns of clonal preservation and divergence across tumor types, we extended our analysis to the full cohort by applying PyClone-VI to all available TO pairs (n = 132) (25). Since PyClone-VI clusters mutations with similar cellular prevalence under the assumption that they arise from the same cellular lineage, we interpreted each mutation cluster as representing a distinct clonal population. We then assessed clonal divergence by comparing the cellular prevalence of each cluster between tumor and PDO. Clusters were categorized as highly concordant (<10% absolute difference), moderately concordant (10 to 29%), or divergent (≥ 30%) based on their absolute difference in cellular prevalence. On the basis of the absolute difference in cellular prevalence between tumor and PDO, we found that the majority of clones were either highly concordant or moderately concordant (62.2%), indicating substantial preservation of clonal structure (Fig. 3C). When overlaid with SNV concordance levels, TO pairs with high SNV concordance (≥80%) showed significantly higher fraction of highly concordant clones (P = 0.01) and a lower fraction of highly divergent clones (P = 0.01) compared to those with low SNV concordance (<40%) (Fig. 3D). These findings suggested that genomic similarity is associated with clonal stability during organoid derivation.

In addition, clone size was significantly associated with concordance category (χ² = 46.6, df = 4, P = 1.8 × 10⁻⁹). Logistic regression further revealed that smaller clones exhibited significantly higher odds of divergence compared with large clones [medium clones: odds ratio (OR) = 1.50, P = 0.0098; small clones: OR = 2.26, P = 0.00027]. This pattern may reflect greater susceptibility of low mutation subclones to clonal shifts during organoid establishment, including potential loss of minor tumor subclones or expansion of low-prevalence populations that were near the detection limit in the parental tumor.

The patterns observed also pointed to possible expansion, contraction, or loss of clones during PDO establishment and growth. To quantify these dynamics, we calculated the fold change in cellular prevalence between tumor and PDO for dominant tumor clones (defined as those with >10% prevalence in the tumor). At the cohort level, 85% of dominant tumor clones were retained in the matched PDO, while 15% (n = 96) were lost. Among retained clones, 24% (n = 153) showed contraction (fold change < 0.75) and 18% (n = 116) showed expansion (fold change > 1.25). We further stratified these patterns by tumor type and observed that, across most cancer types, a substantial fraction of dominant clones remained stable (unchanged) or increased in prevalence (expanded) in PDOs, supporting the overall preservation of key clonal populations (Fig. 3E). Notably, colorectal and bladder tumors displayed the highest median proportion of unchanged clones (∼67%), whereas stomach tumors showed broader variability in clonal outcomes, including elevated rates of clonal loss and expansion. These findings point toward tissue-specific variability in how clonal populations are reshaped in PDO cultures.

Divergence in CNA profiles between PDOs and parent tumors

To determine the genomic fidelity of established PDOs for CNAs, we performed a comparative analysis similar to the one conducted for mutations. We calculated CNA concordance based on the number of CNA events shared between PDOs and parent tumors relative to the total number of events in the parent tumor (Fig. 4A). Our pan-cancer cohort had a median CNA concordance rate of 74%, and nearly half of these pairs had a high concordance of ≥80% (Fig. 4A). These concordance rates accurately captured the shared patterns of genomic CNAs between the PDO-tumor pairs (Fig. 4B). CNA concordance was marginally correlated with SNV concordance per PDO-tumor pair (r = 0.26, P = 0.0076, n = 105) (Fig. 4C). Similar to mutations, CNA concordance did not correlate with tumor purity, ploidy, or the fraction of the genome altered (Fig. 4D and fig. S5, A and B).

Fig. 4. Comparative analysis of CNA between PDOs and parent tumors.

(A) Box plots showing the distribution of CNA concordance between PDOs and parent tumors across 12 cancer types, calculated on the basis of the proportion of shared amplified, deleted, or neutral regions in the genome. Bar plot at the bottom shows the percentage of samples that belonged to the low (<40%), mid (40 to 80%), and high (≥80%) CNA concordance groups. (B) Heatmap showing the log₂ ratios for the copy number segments across the human exome [excluding chromosomes (chr) X and Y] in a paired PDO-tumor analysis. The heatmap shows five representative samples each from the low (<40%), mid (40 to 80%) and high (≥80%) CNA concordance groups. (C) Scatter plot with a linear regression line (and a gray region for 95% confidence intervals) between the CNA (x axis) and mutation/SNV concordance (y axis) for the cohort. The points are colored by cancer type. (D) Pearson correlation among the two quantified types of concordance and the parent tumor purity, MSI, and TMB. Asterisk (*) denotes a significant correlation (P < 0.05). (E) The significantly deleted copy number peaks (blue) identified in the GISTIC analysis of the tumor samples (right) or in the PDOs (left). The number of genes in the peak are indicated as n in the parentheses under the cytoband descriptor. The networks denote the enriched biological processes associated with the genes in the cytobands (as marked by the green arrows and circles). Note that only genes appearing in the enriched Gene Ontology (GO) terms within the cytoband-associated networks are shown; for the complete gene list, please refer to table S8. STAT, signal transducers and activators of transcription; JAK, Janus kinase.

In our CNA concordance analysis, we identified 28 PDO-tumor pairs with a low concordance rate of <40%. To better understand this low concordance, we further investigated these pairs for known driver CNAs, similar to our approach with mutations. We observed that PDOs with <40% concordance had poor overlap in driver CNAs with their parent tumors. However, driver CNAs were still present in the PDOs, although they did not exactly match those in the parent tumors (fig. S5C). This finding motivated us to collect PDOs with low concordance (<40% with parent tumors) for both CNA and SNV (n = 12) and characterize them for driver events independent of their parent tumors. In this analysis, we identified key cancer drivers (such as TP53 and KRAS) in PDOs (fig. S5D), indicating that despite their low overall concordance with parent tumors, these PDO still harbor essential cancer-driving alterations and could be good standalone cancer models.

We examined whether certain CNA events were enriched in PDOs relative to tumors in general. We analyzed CNA profiles using Genomic Identification of Significant Targets in Cancer (GISTIC) to identify significant CNAs separately for PDOs and tumors (fig. S6, A and B, and table S8). As GISTIC identifies recurrent CNAs at the cohort level, the observed significant peaks reflect aggregate recurrence patterns and are not intended as formal paired comparisons between matched tumors and PDOs. Among the large cluster of genes, we identified loci 8p23.1 and 9p21.3, which were significantly deleted in tumors but not in PDOs (Fig. 4E). The deleted regions of loci 8p23.1 enriched in defensins genes and 9p21.3 enriched in interferon-γ (IFN-γ) genes. Deletion in locus 9p21.3 has previously been shown to be associated with an immune cold tumor microenvironment (TME) and a poor response to immune checkpoint immunotherapy (26). Tumors experience heightened selection pressure in the TME from immune surveillance; therefore, loss of IFN-γ loci might be one of the mechanisms of immune escape for the tumor, whereas PDOs are not subject to the same selective pressure in vitro. In summary, our analysis suggests that while PDOs generally maintain a high degree of CNA concordance with their parent tumors, significant divergences in specific CNA events, particularly those associated with immune selection, may occur.

Transcriptomic drift and differences among biological processes and pathways

To quantify the conservation of key biological processes and pathways between the tumors and the early (passage number ≤ 7) and late passage PDOs, we compared the expression profiles collected for these samples. Consensus clustering of the 5000 most variable genes identified six robust transcriptomic subgroups (k = 6) (Fig. 5A). Several clusters exhibited strong tissue-of-origin specificity, including colorectal-dominated cluster 3, bladder-enriched cluster 5, and breast-predominant cluster 6. In contrast, cluster 2 reflected heterogeneous lineage composition, primarily enriched for upper gastrointestinal (GI) and pancreatic tumors (fig. S7A). The expression profiles for early PDOs and matched tumor had a good overall concordance [median r = 0.85 (0.72 to 0.95), n = 72 pairs] for the top 5000 most variable genes. This concordance was significantly higher than that observed for randomly paired tumor-PDO backgrounds (P < 0.001) (Fig. 5B). Similarly, expression profiles for matched early and late passage PDOs had significantly higher concordance (median r = 0.92) compared to randomly paired PDOs (median r = 0.44), indicating preserved gene expression over multiple passages (Fig. 5C).

Fig. 5. Comparative analysis of gene expression profiles between PDOs and parent tumors.

(A) Uniform Manifold Approximation and Projection (UMAP) plot to show the clustering of early passage PDOs based on their top 5000 most variable genes. Samples are colored by primary site (upper GI aggregates esophagus, stomach, and GI samples) and shaped by sample localization (circle, metastatic; triangle, primary). Dashed ellipses delineate the six consensus clusters (k = 6). Plot shows cancer types with ≥2 PDOs. (B) Violin plots showing the distribution of Spearman correlation coefficients for top 5000 most variable genes between expression profiles of tumors and PDOs from early passage (passage number ≤ 7) shown in green and between randomly matched pairs shown in gray. (C) Histograms showing the distribution of correlations between matched early and late passage PDOs against a randomly matched pair background. (D) Clustered heatmap of differentially regulated 50 cancer hallmark pathways between PDOs and parent tumors. The heatmap shows the ORs of differences in pathway expression levels, where color red denotes higher, and color blue denotes lower pathway expression in PDOs compared to the tumors. Only significant OR (P < 0.01) are shown in the plot. TNFA, tumor necrosis factor A; PI3K, phosphatidylinositol 3-kinase; MTOR, mechanistic target of rapamycin; TGF-β, transforming growth factor–β. (E) Bar plot showing the enriched pathways in the genes differentially down-regulated in at least five different cancer types. x axis denoted the adjusted (adj) P values (−log₁₀ scale) from hypergeometric enrichment tests. The bar plot is annotated by the number of down-regulated genes found in a pathway versus total number of genes in that pathway. PDGF, platelet-derived growth factor. (F) Volcano plot for paired DGE analysis between early (passage number ≤ 7) and late passage PDOs (passage number > 7) (n = 84 pairs). All significant differentially regulated genes between the two groups [adjusted P < 0.1 and absolute log₂ fold change (log₂FC) of >1] are labeled in the plot. NS, not significant.

Next, we aimed to understand whether the cancer hallmark pathways differed between the PDOs and parent tumors. We calculated single-sample gene set enrichment analysis (ssGSEA) scores for each pathway in the cancer hallmark pathway collection (n = 50) from Molecular Signatures Database (MSigDB) (27). In cancer-specific models, we applied a linear regression analysis to test the association of parent tumors and early passage PDOs with each cancer hallmark pathway ssGSEA scores and reported the exponent of coefficients as OR (Fig. 5D). We adjusted these models for tumor purity to account for purity associated difference among the PDOs and tumors. Most of the hallmark pathways were not differentially enriched in either the tumors or the PDOs. However, we did identify some differences among the colorectal tumors (n = 26 pairs) where the tumors showed enrichment in pathways associated with the TME such as immune responses and angiogenesis, which are not retained in the PDOs (Fig. 5D) (28). We repeated the same analysis for the early versus late passage PDOs and found no significant differences in any of the hallmark pathways in any of the cancer types analyzed, further supporting the stability of the expression profiles.

To quantify the extent of transcriptomic drift between parent tumors and PDOs, we performed differential gene expression (DGE) analyses between early-passage PDOs and their matched tumors and between early- and late-passage PDOs, using paired, cancer-specific comparisons. In our analysis of early-stage PDOs versus tumors, we defined pan-cancer differentially expressed genes (DEGs) as those that were consistently up-regulated or down-regulated in at least five of the seven cancer types analyzed (fig. S7, B and C). With this selection, we identified 241 down-regulated and 0 up-regulated pan-cancer DEGs in early-stage PDOs relative to parent tumors. These down-regulated genes showed a significant enrichment in the β-integrin family of genes that are associated with the composition and organization of the extracellular matrix (ECM) (Fig. 5E and table S9). The down-regulated DEGs included genes involved in ECM organization, such as collagens, proteoglycans, and glycoproteins, genes related to cell adhesion and migration including cadherins and integrins, as well as adhesion/junctional proteins. We also found markers associated with angiogenesis and vascular development as well as surface markers for immune cells and cytokine/chemokine signaling (tables S10 and S11). Together, these findings showed that the early-stage PDOs lack critical features of the TME, resulting in a diminished ability to mimic the complex interactions between the tumor cells and the stromal/immune compartments of the TME (28).

To investigate transcriptomic drift across passages, we performed a paired DEGs analysis between early passage and late passage PDOs (n = 84 matched pairs). There were no significantly up-regulated genes and only 10 significantly down-regulated genes in late-passage relative to early-passage PDOs, suggesting that global expression profile remained largely stable across passages (Fig. 5F). All genes down-regulated in late-passage PDOs were predominantly associated with inflammatory, immune, and stress-response programs. These included genes involved in major histocompatibility complex class II antigen presentation [human leukocyte antigen (HLA) (HLA-DRA, HLA-DRB1, and HLA-DRB5), chemokine signaling (CXCL14), inflammatory responses (S100A8/A9), and tissue remodeling or inflammation-associated pathways (CHI3L1), as well as interferon- and stress-related genes such as interferon alpha-inducible protein 27 (IFI27) and peptidase inhibitor 3 (PI3). A possible explanation for these observations is that early-passage PDOs retain transient inflammatory or stress-associated transcriptional signals linked to tissue dissociation and initial culture establishment, which attenuate with subsequent passaging.

PDO as functional precision medicine models for PARPi sensitivity

Having established that our PDO platform reliably recapitulates parent tumor biology, we aimed to demonstrate the application of PDOs as functional precision oncology models using PARPi drug sensitivity screens. We selected PDOs across GI, bladder, ovary, and prostate cancers and optimized a 7-day treatment for capturing long-term cellular response and DNA damage effects, as confirmed by imaging and viability assays (fig. S8A and table S12). Using a threshold of 75% cell death at 10 μM, PDOs were classified as sensitive (n = 16) or resistance (n = 17) to talazoparib (Fig. 6A). We evaluated RADiation sensitive 51 (RAD51), phospho-histone H2AX (pγH2AX), and geminin staining as potential indicators for PARPi sensitivity but found no significant correlation with talazoparib response (Spearman r = 0.37, P = 0.07), primary tumor type, nor prior platinum response (Fisher’s exact test, P = 0.26 and 0.24, respectively) (fig. S8, B to D) (29, 30).

Fig. 6. PDOs as a functional precision oncology model to test their sensitivity against talazoparib.

(A) Dose response curves for talazoparib sensitivity in PDOs. (B) Oncoprint of genes relevant to PARPi activity, with PDO samples grouped by sensitivity to talazoparib. Bar plots to the right show gene frequencies in the sensitive (s) versus the resistant (r) PDO groups. HRD stands for homologous recombination deficiency. Colors in the “BRCA2 mutant cases” annotation bar represent individual samples, consistent with (C). (C) Lollipop plot showing variants across PDOs that harbored BRCA2 gene mutations (n = 5). Each color represents a different PDO sample, consistent with the BRCA2 mutant cases annotation bar in the oncoprint shown in (B). Asterisks denote germline variant, while an “R” marks the PDO resistant to talazoparib. (D) GSEA enrichment plots showing up-regulation of mismatch repair and homologous recombination pathways in PDOs sensitive to talazoparib compared to those resistant to it. Preranked gene list used for this analysis was obtained from the DGE analysis between talazoparib-sensitive versus talazoparib-resistant (reference) groups, adjusted for tumor type. NES, normalized enrichment score. (E) Volcano plot showing the relationship between expression of curated HRD genes and talazoparib response (measured as lowest percentage of cell viability) across the PDO cohort (n = 24). The x axis displays the Pearson correlation coefficient (r), and the y axis represents the significance level as −log₁₀(FDR). Red points indicate genes with a statistically significant correlation (FDR < 0.05). A negative correlation coefficient indicates that higher gene expression is associated with lower cell viability (i.e., increased drug sensitivity). Significant genes associated with sensitivity are labeled in the plot. The red dashed line marks the threshold for statistical significance at FDR = 0.05. TLZ, talazoparib. (F) Oncoprint showing genetic alterations and expression profiles of selected genes in PDOs treated with talazoparib, along with their synergy responses to combination treatments with selected inhibitors.

To identify molecular characteristics associated with sensitivity to talazoparib, we analyzed the mutation profiles of PDOs classified as either sensitive (n = 11) or resistant (n = 13) to the drug based on sequencing data (Fig. 6B). Among genes relevant to PARPi sensitivity, talazoparib-sensitive PDOs had a higher frequency of mutations in breast cancer gene 2 [BRCA2 (n = 4)], breast cancer gene 1 [BRCA1 (n = 2)], ATM (n = 2), and NBN (n = 1) genes compared to resistant PDOs. This finding aligns with expectations, as mutations in BRCA1 and BRCA2 genes impair DNA repair pathways, rendering cancer cells more sensitive to PARPis. In contrast, mutations in the base excision repair pathway were more common in resistant PDOs (46%) than in sensitive PDOs (18%), which could indicate an adaptive resistance mechanism. Among the BRCA2-mutated PDOs, those sensitive to talazoparib had higher TMB (median TMB: mutated = 12, wild-type = 1.9). In addition, BRCA2-mutated PDOs sensitive to talazoparib harbored multiple mutations in the BRCA2 gene, including the pathogenic variant Asn³¹²⁴ → Ile (N3124I) (31, 32). Notably, all talazoparib-sensitive PDOs carried BRCA2 Ser¹⁹⁷⁹ → Arg (S1979R), a variant of uncertain significance (VUS), in either germline or somatic alterations (Fig. 6C).

Next, we analyzed all nonsynonymous mutations, including those outside DNA repair pathways, to identify those differentially associated with talazoparib sensitivity. We noted that talazoparib-sensitive PDOs had mutations in key cellular processes that intersect with DNA repair, including chromatin remodeling (CREBBP and PHF12), oxidative (SLC7A11) or cellular stress responses (FMNL1 and CRISP3), DNA repair gene activation (ALPK2), and epigenetic regulation (UHRF1BP1L and ZC3H4) (fig. S8E).

To investigate differences in expression profiles, we analyzed DEGs between talazoparib-sensitive (n = 13) and talazoparib-resistant (n = 10) PDOs adjusting for tumor type–specific effects. Talazoparib-sensitive PDOs showed up-regulation of pathways associated with homologous recombination deficiency (HRD) signatures (homologous recombination and DNA repair pathways) and cell cycle dependencies [early region 2 binding factor (E2F) targets and G2M checkpoint] (Fig. 6D and fig. S8F). To assess whether baseline DNA repair gene expression is associated with talazoparib response, we correlated expression of curated homologous recombination repair (HRR) pathway genes with cell viability in the PDO cohort (Fig. 6E). Higher expression of several HRR genes was associated with increased talazoparib sensitivity, reflected by negative Pearson correlation coefficients. After correction for multiple testing, CHEK1, FANCI, and FANCB remained significantly associated with sensitivity (FDR < 0.05). These results indicate that elevated expression of checkpoint and Fanconi anemia pathway genes is linked to greater talazoparib sensitivity in PDOs.

To externally validate our PDO findings, we analyzed genomic and transcriptomic data from the Cancer Dependency Map (DepMap) (33, 34), comprising 562 cancer cell lines treated with talazoparib. Using the same analytical framework applied to PDOs, GSEA comparing sensitive and resistant cell lines revealed strong concordance at the pathway level. In both datasets, talazoparib-sensitive models showed significant enrichment of cell cycle and DNA repair programs, including E2F targets, G2M checkpoint, homologous recombination, and DNA repair pathways (fig. S8, F and G). DepMap additionally identified pathways such as MYC targets and mismatch repair, likely reflecting its larger sample size and broader lineage diversity relative to the clinical PDO cohort. In addition, higher expression of homologous recombination and replication checkpoint genes was consistently associated with increased talazoparib sensitivity in DepMap cell lines, similar to our observations with PDOs. CHEK1 was a significant correlate of sensitivity in both our cohort and the DepMap datasets (FDR < 0.05), while Fanconi anemia pathway genes were associated with sensitivity in each cohort, albeit with different family members reaching significance (PDOs, FANCI/FANCB; DepMap, FANCE/FANCC), consistent with functional redundancy within this pathway (Fig. 6E and fig. S8H).

In contrast to pathway-level concordance, overlap at the level of individual somatic mutations was limited. Linear regression analyses adjusting for tumor lineage, restricted to the set of differentially mutated genes identified in the PDO cohort (fig. S8E), did not identify any mutations significantly associated with sensitivity after FDR correction (fig. S8I). BRCA2 and base excision repair pathway mutation frequencies in DepMap did not recapitulate trends observed in the PDO cohort (table S13), and mutational burden failed to stratify talazoparib response in BRCA2-mutant models (P = 0.68), unlike the pattern identified in our cohort. Collectively, the lack of genomic correlates underscores that mutational features alone were insufficient to capture the resistance phenotype in the broader cell line compendium. However, the transcriptional pathways (i.e., dependence on DNA repair and cell cycle checkpoint pathways) were highly conserved, suggesting that gene expression profiles provide a more generalizable predictor of talazoparib response across diverse model systems than genomic alterations.

Last, we tested combination treatments to sensitize seven pancreas and GI cancer PDOs with moderate resistance to talazoparib. We selected six drugs targeting key components of the DNA damage repair or previously proposed to have synergistic effect with PARPis: adavosertib (Wee1 inhibitor), AZD7648 [DNA dependent protein kinase (DNA-PK) inhibitor], cerlasertib [ataxia telangiectasia and rad3-related kinase (ATR) inhibitor], gemcitabine (nucleoside analog), prexasertib [checkpoint kinase 1 (CHK1) inhibitor], and temozolomide (alkylating agent), with the latter three being FDA-approved oncology drugs. All six drugs enhanced talazoparib sensitivity to varying degrees (Fig. 6E and fig. S9). The strongest synergy was observed with temozolomide (synergy scores > 20) in three of seven PDOs, with two additional lines approaching the threshold (8.8 and 9.4). Consistent with our previous findings, all of these lines harbor TP53 alterations (35). Prexasertib showed synergy in three lines, while gemcitabine, cerlasertib, and AZD7648 exhibited synergy in two lines each. PDOs with moderate baseline sensitivity to talazoparib responded to more synergistic combinations than resistant PDOs, particularly those with additional DNA repair pathway alterations beyond TP53. Notably, synergy profiles varied across PDOs, highlighting the treatment response heterogeneity and the importance of using PDOs as functional models for personalized combination strategies.

Overall, these findings highlight the complexity of talazoparib sensitivity, showing that it is not limited to mutations in BRCA genes or prior response to chemotherapy. Instead, multiple factors may influence sensitivity to PARPi. In addition, several DNA damage repair inhibitors exhibited synergy with talazoparib, although the degree of synergy varied across PDOs based on their genomic composition. These results emphasize the importance of a functional precision oncology approach to identifying optimal therapeutic combination for individual patients.

DISCUSSION

We successfully established and comprehensively characterized a cohort of 220 PDOs from 191 patients and demonstrated their application in functional precision oncology by conducting a PARPi drug screens on a subset. Our cellular prevalence and clonality analysis, spanning multiple tumor types, revealed a level of resolution that provides insights into the extent and variability of clonal preservation during PDO establishment—an aspect not previously characterized at this scale. This work further validates the robustness of our pan-cancer PDO platform, initially introduced in 2017 (3).

Standardizing PDO generation is key to advancing functional precision oncology. An important consideration in this process is the success rate of PDO cultures, which varies widely depending on tumor type and culture conditions. In this study, the PDO formation rate was ∼81%, with 46% successfully biobanked, which was an improvement over the 38.6% success rate reported in our initial work (3). While some studies do not clearly specify their success criteria (36–43), a few define success based on the ability to expand PDOs (44–46), whereas others focus on the number of passages or the ability to biobank cultures (3, 4, 17, 43, 47–50). Reported success rates in these cases generally range from 20 to 70%. Conversely, studies with more stringent criteria that require more than 10 passages, molecular characterization, or long-term cultures exceeding 6 months, tend to report lower success rates of around 10% (17, 50), with an exceptional study achieving a remarkable 70% success rate in establishing neuroendocrine carcinoma organoids (50). Platforms seeking to establish PDOs as a resource for both functional precision oncology and drug discovery and validation should prioritize models that can expand beyond p5 and can be banked at a minimum of 1 × 10⁶ cells per vial. On the basis of these criteria, our platform’s success rates aligned well with other large-scale PDO studies (4), confirming the robustness of our approach in establishing long-term cultures.

Beyond success rates, validating PDO models against their parent tumors is also essential to ensure accurate representation of the parental tissue. Most studies perform histopathology and genomic profiling to characterize PDOs, but not all report the degree of concordance with the parent tumors (38, 49). Similar to our study, while most studies report good morphological and marker staining concordance, by showing representative images, some have reported contamination with normal cells within their cohorts (36). Genomic analysis typically involves reporting SNVs and/or CNAs using different sequencing methods, including whole-genome sequencing (WGS) (38, 39, 44, 50), WES (3, 17, 37, 45, 48, 49, 51, 52), or targeted panels (4, 41–43, 46, 47). Because of the differences between sequencing methods in depth and coverage, direct comparisons of genomic concordance among different studies are challenging. Targeted panels (100 to 500 genes) tend to yield higher concordance (4, 42, 43, 47) than when broader coverage of ∼20,000 genes, such as WES, is used (3, 38, 45, 49, 50). Studies using autologous pairs of PDOs and tissues generally show good correlations for SNVs and CNAs (60 to 90%) (3, 4, 37–45, 49, 50). In cases where parental tissues are unavailable, studies often compare top mutated cancer genes with those in publicly available cohorts (4). Some studies also investigate genomic stability across passages, typically demonstrating consistent concordance (4, 17, 39, 43, 47–49). In our cohort, we report a median SNV concordance of 60% (80% when only considering cancer driver genes) and a median CNA concordance of 74% from 114 pan-cancer matched pairs using WES, which is comparable to our own previously reported concordance from 14 tumor-PDO pairs (3). In addition, our targeted panels demonstrated 88% concordance, which aligns with another pan-cancer platform of 32 tumor-PDO pairs (76.9%) (4). Regarding transcriptomic analysis, most studies examine variations in gene expression under different medium conditions (4), across passages (39, 43), and in relation to IHC or subtype classifications (47–49). Overall, our PDOs maintained stable expression profiles between early (p5) and late (p10) passages.

A small subset of our PDOs (∼20%) showed variable degrees of histopathological discrepancies. In cases of minor discrepancies, most PDOs maintained over 50% SNVs concordance, while those with major histological discordance showed low genomic concordance (0 to 25%). Despite these discrepancies, PDOs with minor histological discrepancies still showed sufficient genomic concordance to be useful models (table S2). In addition, 12 PDOs with low concordance (<40%) for both SNVs and CNAs still retained key cancer driver mutations (table S3). While these deviations from parent tumors might make some PDOs unsuitable for functional precision oncology, their expandability and ability to preserve key cancer-related features highlight their value as cancer disease models to study cancer etiology and facilitate drug discovery. Ultimately, it remains to be determined how much genomic divergence between a PDO and its parent tumor is compatible with effective precision oncology applications.

A key challenge associated with PDO models is their inability to retain the TME and capture complex interactions among cancer cells and the TME. In our cohort, TME-related genes were down-regulated in PDOs compared to tumors, as expected when culturing PDOs independent of their microenvironment (28). To address this limitation, we are biobanking autologous cancer-associated fibroblasts, tumor-infiltrating immune cells, and peripheral blood mononuclear cells separately, with the aim to develop coculture systems that integrate distinct components of the TME (28, 53). The ECM also influences gene expression and drug sensitivity, yet current culture matrices such as Matrigel are undefined and lack physiological relevance. Using synthetic hydrogel-based matrices could provide a more controlled environment and reveal how matrix composition alters inhibitor responses (54). We believe that these cocultures and hydrogel-based cultures will enable the study of cancer cell interactions with the TME and provide insights into tumor self-seeding and resistance mechanisms. Once challenges in scaling coculture systems are overcome, these more complex and physiologically relevant PDO models could have substantial impact in precision oncology.

Another key challenge for PDOs is the time required to generate stable models. While we prioritized establishing PDOs as long-term models, we acknowledge that tumors evolve over time, and PDOs—especially those requiring longer establishment times—may not fully capture the dynamic landscape of parent tumor. Some platforms address this concern by conducting experiments immediately after dissociating specimens into single cells. This temporarily preserves the diversity of cell types from the parent tissue in short-term cultures (2, 55) but may limit the number of drugs that can be tested.

Despite these challenges, PDOs are valuable models for functional precision oncology, as demonstrated by our assessment of PARPi sensitivity in selected PDOs. Our drug screening confirmed the feasibility of using PDOs as functional precision oncology models for assessing talazoparib responsiveness. While BRCA2 mutations were linked to sensitivity, many sensitive PDOs lacked known BRCA risk mutations, suggesting that additional mechanism contribute to talazoparib sensitivity, potentially expanding PARPi indications beyond current biomarkers. We identified alternative molecular signatures, including chromatin remodeling and stress response genes, associated with sensitivity. Notably, distinct BRCA2 variants, such as the pathogenic N3124I and the recurrent S1979R variants (currently classified as VUS), were linked to talazoparib response. Pathway analysis in sensitive PDOs revealed up-regulation of mismatch repair and ECM-related pathways, along with down-regulation of antigen-presenting pathways, indicating a complex interplay of DNA repair reliance and immune evasion mechanisms. In addition, combination drug screening identified several inhibitors that enhanced talazoparib sensitivity, with the strongest synergy observed with temozolomide (35). Synergy often cooccurred with TP53 mutations and an alteration in DNA damage repair pathways. These results highlight the importance for comprehensive molecular profiling and direct drug sensitivity testing with PDO models to enhance precision oncology strategies and expand therapeutic applications.

In summary, we have developed a well-characterized pan-cancer PDO platform that includes histopathology and genomic and transcriptomic comparison with their parent tumors. This resource serves as a valuable tool to advance cancer modeling and support research in functional precision oncology, illustrated by the inclusion of some of our models in the global Human Cancer Model Initiative (56).

MATERIALS AND METHODS

Sample collection

The organoid platform at the Englander Institute for Precision Medicine develops PDOs through biopsies, surgical resections, or rapid autopsy procedures. Fresh tissue samples were collected with written informed patient consent with the approval from Weill Cornell Medicine (WCM) institutional review boards (IRBs): #0908010582, #1807019405, #1008011221, #1011011386, #19-05020075, and #1305013903; and Memorial Sloan Kettering Cancer Center IRBs: #15-318 and #06-107.

PDO establishment and culture

PDOs were established using modified standard protocols (3). Fresh tissue specimens were collected with portions frozen in Tissue-Tek optimal cutting temperature (OCT) compound for genomic or histopathological analysis. Medium formulations are detailed in table S14, and material catalog number are in table S15. Remaining tissue was transported to the laboratory, washed three times in transport medium, mechanically dissected into ∼2-mm pieces, and enzymatically digested in a 20-fold volume of collagenase medium on a shaker at 200 rpm and 37°C until the solution turned cloudy (30 to 45 min). Resulting cell pellet was centrifuged at 1300 rpm for 3 min, washed with washing medium, and resuspended in a cell medium/Matrigel mixture (1:2, v/v) with tissue-specific culture medium. Up to five 100-μl drops were plated into a six-well cell suspension culture plate, polymerized at 37°C for 30 min, and overlaid with 3 ml of tissue-specific culture medium. Medium was refreshed every 3 to 4 days. PDO formation is characterized by the observation of a three-dimensional (3D) structure, with evidence of cell proliferation and cluster formation. PDOs were passaged in Matrigel using TrypLE Express for 7 to 10 min or Accutase for 5 to 7 min, depending on tumor type, and replated as described above. Monthly mycoplasma testing was performed using the polymerase chain reaction (PCR) mycoplasma detection kit on cell lysates instead of medium. Biobanking was initiated when cell pellets exceeded 1 million cells, using recovery cell culture freezing medium for overnight storage in −80°C, followed by transfer to liquid nitrogen. The PDO success rate was calculated as the proportion of initiated cultures that exhibited organoid formation out of all initiated cultures. PDO samples that were directly cryopreserved after processing were excluded from this calculation.

PDO medium

Medium formulations are detailed in table S14. For homemade Noggin and R-spondin medium, human embryonic kidney 293T cells with respective cassettes were thawed into a T175 flask with Dulbecco’s modified Eagle’s medium (DMEM), 10% fetal bovine serum, and penicillin/streptomycin. Once cells reach 90 to 100% confluence, they were split with TrypLE Express and replate into twenty 150-mm plates with 25 ml of DMEM. After 3 days, when cells should have reached at least 70% confluence, cells were starved by aspirating the medium, washing with 5 ml of phosphate-buffered saline (PBS), and adding 25 ml of washing medium. After 24 hours, the conditioned medium was collected, filtered through 0.22-μm filters, tested for mycoplasma, and stored at −20°C until use.

PDO pathological characterization

PDO histopathology was verified by comparing fixed p5 organoid sections to parent tumor sections using our cytology and histology platforms (8, 9). PDOs were released from Matrigel droplets using Cell Recovery Solution, suspended in a fibrinogen/thrombin (10:1, v/v) pellet, and fixed with 4% paraformaldehyde in PBS. Samples were then sent to WCM Center for Translational Pathology for formalin fixation, paraffin embedding, and staining. H&E, Ki67, and additional cancer-specific staining are listed in table S14. A comprehensive cytomorphologic comparison of PDO and primary lung tumor (histopathology H&E section) was performed by evaluating cytological features, including nuclear and cytoplasmic characteristics, as well as morphological features. In addition, IHC was conducted and compared between the tumor and the organoid where necessary.

DNA and RNA extraction

Genomic DNA and RNA were extracted from various samples, including formalin-fixed paraffin-embedded (FFPE) tissues, OCT-embedded frozen tissue, peripheral blood mononuclear cells, saliva, and organoid cell pellets for profiling. Material catalog numbers are in table S15. DNA from FFPE tissue was extracted using the Maxwell 16 FFPE Plus DNA Kit on the Maxwell 16 instrument. Tissue sections (10 μm) were macrodissected on the basis of the annotated H&E slides. DNA from OCT-embedded tissue was extracted using the Maxwell 16 Tissue DNA Purification Kit or Omega Bio-Tek Mag-Bind Blood & Tissue DNA HDQ Kit. For saliva, DNA was extracted using Oragene collection kit. DNA from buffy coat fractions was extracted using the Maxwell 16 LEV Blood DNA Kit or Omega Bio-Tek Mag-Bind Blood & Tissue DNA HDQ Kit. Organoid cell DNA was extracted using the QIAGEN DNeasy Blood & Tissue Kit or Omega-Bio-Tek Mag-Bind Blood & Tissue DNA HDQ Kit. RNA from organoid cells was extracted with the Maxwell 16 LEV simplyRNA Cells Kit. DNA quantity and quality were assessed using a Qubit 4 Fluorometer and an Agilent 4200 TapeStation machine. RNA was similarly quantified using the Qubit 4 Fluorometer, and quality was evaluated using the Agilent 2100 Bioanalyzer.

Whole-exome sequencing

WES library preparation and sequencing

The Exome Cancer Test v1.0 (EXaCT-1) library preparation was done using FFPE DNA as templates for PCR amplification of two independent glyceraldehyde-3-phosphate dehydrogenase amplicons (52). Amplicon yields from FFPE DNA were measured with Agilent Bioanalyzer and compared to those from intact reference DNA template. The sample-to-reference yield ratio served as a quantitative indicator of FFPE DNA integrity, predicting sample performance in HaloPlex target enrichment. FFPE DNA samples were categorized on the basis of yield ratio: A (>0.2), B (0.05 to 0.2), and C (<0.05), with input DNA amounts adjusted accordingly. Acceptable DNA requirements included >225 ng for fresh-frozen high–molecular weight DNA, >500 ng for high-quality FFPE DNA, and >1000 ng for low-quality FFPE DNA, with Qubit 4 Fluorometer absorbance ratios of A₂₆₀/A₂₈₀ (1.8 to 2.0) and A₂₆₀/A₂₃₀ (>2.0).

For the samples processed after the end of the year 2020, a newer version of the Exome Cancer Test, called EXaCT-2, was used. EXaCT-2 is based on SureSelect Human All Exon V6 (57), with augmented coverage of approximately 1400 cancer-related genes, enabling the detection of mutations at very low allelic frequencies. The capture kit also included chromosome tiling probes in intergenic regions to provide better analysis of CNAs (58, 59). Less than 200 ng of FFPE DNA sample was enzymatically fragmented using SureSelectXT HS and XT Low Input enzymatic fragmentation kit. Material catalog numbers are in table S15. The fragmented DNA underwent end repair, A-tailing, and adapter ligation with SureSelectXT Low Input Reagent Kit. Each library was made with a unique dual index with SureSelectXT Low Input Dual Index P5 Indexed Adaptors. Then, the indexed libraries were captured using the custom-designed EXaCT-2 Target Enrichment Human All Exon probes and amplified, and their quality and quantities were assessed by Agilent 4200 TapeStation System and Invitrogen Qubit Fluorometer. Final libraries were pooled and sequenced on an Illumina NovaSeq 6000 sequencer at paired-end sequencing (2 × 150 cycles).

WES data alignment and processing

Raw sequencing reads in BCL format were processed through bcl2fastq 2.19 (Illumina) for FASTQ conversion and demultiplexing. Quality control was performed with fastp (60). The sequencing reads were aligned to the human reference genome hs37d5 using the Burrows-Wheeler Alignment tool (bwa v0.7.17, using “mem” option). After alignment, PCR duplicates were marked using Sambamba (v0.6.7) for samples sequenced with SureSelect. Samples sequenced with HaloPlex did not require this step since HaloPlex is amplicon based. This was followed by realignment around the indels, fixing mates and base quality score recalibration using Genome Analysis Toolkit (v4.2.0.0), and all unmapped reads [Mapping quality (MAPQ) = 0] being removed. Accurate pairing between each tumor (or PDO) and its corresponding germline was determined using SPIA (61), which calculates genetic distance between samples based on a selected set of SNPs. Sample pairs with a SPIA score of >0.4 were deemed mismatched and excluded from further genomic analysis. CLONET (62) was applied to estimate tumor purity and ploidy. MSI for each sample was calculated using MSIsensor-pro (63).

Somatic mutation detection

For EXaCT-1, somatic mutations were identified in matched tumor (or PDO) and normal pairs using SNVseeqer (64). Mutations were filtered on the basis of a minimum read depth of 10 in the tumor or PDO with at least two reads supporting the alternate allele. Variant allele frequency (VAF) filters were applied as follows: ≥25% for variants of unknown significance and ≥10% for known cancer genes based on COSMIC Cancer Census (20). A list of clinically relevant mutations based on the EIPM Knowledge Base was always reported. The final set of mutations were annotated using Oncotator (65).

For EXaCT-2, somatic mutations were identified using a method adapted from the Multi-Center Mutation Calling in Multiple Cancers (MC3) pipeline (66). The adapted pipeline included six SNV callers [VarScan (67), SomaticSniper (68), MuTect2 (69), Strelka2 (70), RADIA (71), and MuSE (72)] and four indel callers [VarScan (67), MuTect2 (69), Strelka2 (70), and Pindel (73)]. Each tool was run independently, and all mutations identified were merged. When combining overlapping mutations, the read depth and VAFs from different tools were averaged. This combined set of variants was then filtered on the basis of the following criteria: within-sample allele strand bias frequency (0.1), a total read depth of 30× in tumor or PDO, and a population frequency greater than 0.01 in either TOPmed [Trans-Omics for Precision Medicine (74)] or gnomAD [Genome Aggregation Database (75)]. Frequently mutated but biologically unrelated genes such as TTN and the MUC were also excluded (58). The final set of mutations from EXaCT-2 were annotated using Nirvana (v3.3.103.11) (76). TMB was calculated for each sample by counting the total number of nonsynonymous mutations (including truncating mutations), normalized to the number of bases in the targeted exonic regions per million. Mutations that correlated differently with one group versus the other were identified using Fisher’s exact test and a P value threshold of <0.05.

Copy number analysis

CNAs were identified from aligned reads using CNVkit v0.9.10 (77). Segments with a log₂ ratio of ≥0.2 were classified as copy number gains, while those with a log₂ ratio of ≤−0.2 were classified as copy number losses. These copy number calls were then adjusted for tumor purity and ploidy using CLONET v2 (62). For gene-level analysis, copy number segments were annotated for genes based on GENCODE v19 (78). Log₂ ratios for multiple segments spanning a gene were combined by calculating a weighted average based on the overlap of segment coordinates, resulting in a gene-specific log₂ ratio. The fraction of genome altered (FGA) was calculated by dividing the total number of bases with copy number gains or losses by the size of the human genome (3,054,815,472 bases).

Driver CNAs were identified using GISTIC2 (v2.0.23) (79) applied to the segmented log₂ ratios. GISTIC was run with the reference file hg19.UCSC.add_miR.140312.refgene.mat and a confidence threshold of 0.99. This analysis was conducted separately for tumors and PDOs in the cohort. CNA peaks that were differentially deleted or amplified in either tumor or PDO, but not in both, were identified on the basis of a P value of <0.01. Network analysis of genes within these significant CNA peaks was performed using the enrichGO function from the clusterProfiler package (v4.6.2) visualized with the cnetplot function from the enrichplot package (v1.22.0) in R.

Identification of mutational signatures

Mutational signature analysis was performed using deconstructSigs (v 1.9.0) (80). First, the mut.to.sigs.input function was used to transform the mutation data into a matrix of 96 possible trinucleotide contexts. Then, the whichSignatures function was applied to infer the contributions of each sample to the 60 COSMIC signatures from the reference signatures.exome.cosmic.v3.may2019, with the tri.counts.method parameter set to “exome.” To evaluate the similarity between mutational signature profiles of tumor and PDO pairs, cosine similarity values were computed using the lsa package (v0.73.3) in R. The contributions of the most frequent signatures in the cohort were visualized using the ComplexHeatmap package (v2.13.0) in R.

Targeted gene panels

Library preparation and sequencing for targeted panels

Our targeted gene panels included the TruSight Oncology 500 (TSO500; Illumina) and the Oncomine Focus Assay (Thermo Fisher Scientific). Library preparation and sequencing were performed by WCM Clinical Genomics Laboratory. DNA was extracted from FFPE or cell pellets with a minimal concentration of 3.33 ng/μl using the Maxwell 16 FFPE Plus DNA Kit. Material catalog numbers are in table S15. Genomic DNA was fragmented into 90 to 250 base pairs (bp) using Covaris E220evolution Focused-ultrasonicator. For TSO500 sequencing, TSO500 DNA/RNA high-throughput kit was used for end repair, A-tailing, and adapter ligation. Libraries were amplified, indexed with Illumina DNA/RNA UD indexes, captured through dual hybridization, then enriched, and cleaned up. Libraries were normalized using the TSO500 bead-based normalization method, and quantity was checked using Qubit dsDNA HS Assay Kit. Sequencing was performed on NovaSeq 6000 SP Reagent Kit and NovaSeq6000 system (Illumina) with PhiX Control v3 as control. Analysis was performed using Illumina TruSight 500 Local App v2.2. For Oncomine, DNA targets were amplified with Ion AmpliSeq Library Kit with DNA primer pools. Primer sequences were partially digested, ligated to the Ion Xpress Barcode Adaptors, and purified with Agencourt AMPure XP Kit. The library was quantified by comparing it to Escherichia coli DH10B control library with duplicates showing SD values higher than 5.83 being repeated. Sample libraries were diluted to 50 pM for sequencing. Samples were loaded on Ion 540 Chip Kitand Ion 540 Kit-Chef.

Variant detection in targeted panels

For TSO500, the sequenced reads were analyzed using the TSO500 LocalApp 2.2 pipeline and Pindel (v0.2.5b9, 20160729) (73). Indels were filtered to include variants larger than 20 bp, with at least 10× depth, that did not appear in population databases at >1%. Pathologists manually reviewed clinically relevant hotspot variants that might be missed by the pipeline. The filtered variants were reviewed using Velsera’s Clinical Genomics Workspace for clinical annotation. For Oncomine, the sequenced reads were analyzed using the standard Oncomine Comprehensive Assay v2 (Thermo Fisher Scientific) Ion Reporter software (v5.0) with “OnCORseq Comprehensive v2.0–DNA–Single Sample” workflow (v5.6), previously validated at WCM (81). To better capture complex indels, an additional workflow was run with SNP realignment disabled and the “Allow_Complex” setting enabled. Identified variants were annotated using the Torrent Variant Annotator v2.0 plug-in. Synonymous variants occurring at a frequency of >1% in the 5000 Exomes or ExAC databases (within Ion Reporter) were excluded. The identified variants were manually reviewed by pathologists and summarized in clinical reports.

RNA sequencing

RNA sequencing library preparation and sequencing

RNA libraries preparation and sequencing was performed by WCM Genomics Core Laboratory. For polyadenylate (polyA) selection, samples were processed using Illumina TruSeq Stranded mRNA Sample Library Preparation kit and Agilent high-throughput sample preparation Bravo B system according to the manufacturer’s instructions. Material catalog numbers are in table S15. Normalized cDNA libraries were pooled and sequenced on Illumina NovaSeq 6000 sequencer with paired-end 100 cycles. Raw sequencing reads were converted from BCL to FASTQ format using bcl2fastq 2.19 (Illumina). For ribosomal RNA (rRNA) depletion method, rRNA was removed from total RNA using Illumina Ribo Zero Gold for human/mouse/rat kit. mRNA was fragmented, reverse-transcribed into cDNA using Illumina TruSeq Stranded Total RNA Sample Library Preparation kit. The cDNA fragments underwent end repair, adapter ligation, and PCR enrichment to create the final cDNA library. These cDNA libraries were pooled and sequenced on Illumina NovaSeq 6000 sequencer with paired-end 100 cycles.

RNA sequencing alignment and quantification

RNA sequencing analysis was conducted as previously described (82). Sequenced reads were mapped to the human reference genome hg19/GRCh37 using the STAR aligner (RRID:SCR_004463) v2.7.2b and STAR-Fusion version 1.9.1 (83). HTSeq v0.11.3 (84) (RRID:SCR_005514) was used to quantify gene counts, and cufflinks v2.2.1 (85) (RRID:SCR_014597) was used to quantify fragments per kilobase of transcript per million mapped reads (FPKMs) with GENCODE v19 (RRID:SCR_014966) (86). Tumor purity was inferred from FPKMs using “Estimation of STromal and Immune cells in MAlignant Tumours using Expression data” (ESTIMATE) (87).

Consensus clustering of expression profiles

Unsupervised clustering was performed on the top 5000 most variable genes using the ConsensusClusterPlus R package (v1.68.0). Partition Around Medoids clustering method with Spearman distance was assessed over 1000 iterations, with a subsampling of 90% of samples and 80% of features per iteration (seed = 455). The optimal cluster number (k = 6) was determined by evaluating silhouette scores, cumulative distribution function curves (k = 2–13), and consensus matrix stability. For visualization, esophagus, stomach, and GI samples were grouped as “upper GI,” and primary sites with fewer than three samples were excluded.

Cancer hallmark pathways analysis

Hallmark pathway enrichment scores (ESs) were calculated for each sample using ssGSEA (88), implemented with GSVA package (version 1.46.0) in R (version 4.2.2). The signature set used for this analysis was obtained from the Human MSigDB collection of 50 hallmark pathways (H) (27). Linear regression models were applied using the lm function in R (version 4.2.2) to identify differentially enriched hallmark pathways between PDOs and tumors within each cancer type. The linear model equation was as follows: ES ∼ TO + Pair + TP, where ES represents enrichment scores for a given hallmark pathway, TO is a binary variable (0 for tumor and 1 for PDO), Pair denotes factor covariates representing PDO-tumor pairs, and TP stands for tumor purity, inferred from ESTIMATE. β Coefficients from the linear model were converted to OR using the formula 10^β. P values from all models were adjusted using Bonferroni correction. Differentially enriched pathways were identified using cutoffs of an adjusted P < 0.1 and an absolute OR of >1.5. For visualization, the OR for nonsignificant pathways was set to 0. The ORs for each hallmark pathway across cancer types were clustered and visualized using the ComplexHeatmap package (v2.14.0) in R.

DGE analysis

DGE analysis was performed on counts data using DeSeq2 (v1.42.1) (89), with PDO–parent tumor pairing and tumor purity (inferred from ESTIMATE) included as covariates. DGE analysis between tumor and early-stage PDOs was conducted separately for each cancer type with at least three unique tumor-PDO pairs (n = 7 cancer types). DEGs were identified on the basis of an absolute log₂ fold change (log₂FC) > 2 and an adjusted P < 0.01. The results were visualized as volcano plots for each cancer type using the EnhancedVolcano package (v1.20.0). DEGs identified as up-regulated (log₂FC > 2) or down-regulated (log₂FC < −2) in ≥5 cancer types were selected for pathway enrichment analysis using EnrichR (90). To ensure accurate statistical inference, we applied a one-sided hypergeometric test using a custom background rather than the default genome-wide background. The test parameters were defined as follows: N represents the total number of genes in the background/assay (25,374 genes), K is the number of genes in the specific pathway, n is the total number of selected DEGs (243), and k is the number of DEGs overlapping with the pathway. P values were adjusted for multiple testing using the Benjamini-Hochberg procedure. Significantly enriched pathways and biological processes were plotted using the ggpubr package (v0.6.0) in R. To confirm that a DEG was truly specific to a specific cancer type, the DEG was cross-checked against the remaining cancer types using a more lenient threshold of absolute log₂FC > 1, adjusted P < 0.05, and a base mean > 10 to ensure that it was not present in other cancer types. DGE analysis of early versus late passage PDOs were analyzed in a paired DESeq2 model, using matched pair IDs as covariates in the model design. DGE analysis of cases sensitive to versus resistant to a drug was also analyzed in a paired DESeq2 model, adjusting for tumor type as a confounder in the model design. For drug sensitivity DEGs, GSEA was performed using the fgsea package in R, querying hallmark (all pathways) and Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, and Gene Ontology (GO) Biological Process gene sets subset to pathways matching the terms DNA_REPAIR, HOMOLOG, MISMATCH, MMR, or BER. Correlation between gene expression and talazoparib sensitivity (measured as lowest percentage of cell viability) was assessed for a curated panel of HRD/DNA damage response (DDR) genes using Pearson correlation across the PDOs screened for talazoparib (n = 24), with statistical significance defined as FDR < 0.05.

Genomic and transcriptomic concordance analysis

Somatic mutations concordance

Concordance for somatic mutations was assessed for nonsilent single-nucleotide variants and truncating mutations. SNV concordance was calculated by dividing the number of shared mutations between the PDO and the parent tumor by the total number of mutations in the parent tumor. To minimize the possibility of false nonoverlapping mutations, we gathered all mutations from each PDO-tumor pair and manually reviewed them for evidence in samples where the mutation was absent. This review was conducted using bcftools (v1.9) and/or Integrative Genomics Viewer (2.3.57). For driver gene SNV concordance, the mutation lists were filtered to include 200 pan-cancer driver genes (18), and concordance was calculated as described above.

Copy number concordance

To assess CNA concordance, the total number of bases covered by segments in the parent tumor was first calculated, denoted as x. bedtools (v2.28.0) was used to identify overlapping CNA segments between the tumor and the matched PDO. For regions with overlap, the number of bases where the PDO’s copy number status (gain, loss, or neutral) matched that of the tumor was counted, denoted as y. CNA concordance was then calculated as y divided by x. Gains were defined as log₂ ratios of >0.50, losses were defined as log₂ ratios of <−0.50, and regions with log₂ ratios between −0.50 and 0.50 were considered copy number neutral.

Expression profile concordance

Expression concordance between a parent tumor and its PDO was calculated using Spearman correlation across top 5000 most variable genes (91). These correlations were compared to a background distribution generated by randomly pairing PDOs and tumors from the full cohort, ensuring that the true pairs were excluded. Welch two-sample t test was used to statistically test whether the true and randomly paired distributions were significantly different.

Clonal evolution analysis

All nonsilent SNV and truncating mutations, copy number values, and tumor purities (inferred from DNA) were used for the clonality analysis. PyClone-VI (25) was applied to these datasets to infer the clonal clustering of mutations in a paired analysis of parent tumors and PDOs. For genomic locations where mutations were not detected in a sample within the pair, read depths were extracted using bcftools (v1.9). Paired clonal clusters and their corresponding cancer cell fractions, along with the VAFs of these mutations in tumor and PDO pairs, were analyzed to understand clonal evolution through clonal ordering and tree visualizations using ClonEvol (92).

For cohort level analysis, PyClone-VI was run cohort-wide (n = 135 pairs) using somatic SNVs, with fixed copy number settings (major allele = 2 and minor allele = 0) due to incomplete CNA data (25). Tumor purity estimates were incorporated where available; for a total of nine tumor-PDO pairs (five PDOs and seven tumors), purity data were not available; for these, we assumed tumor purity to 100% for analysis. PyClone ran successfully on 132 tumor-PDO pairs. We treated each mutation cluster identified by PyClone-VI as a putative clone based on the assumption that variants grouped by similar cellular prevalence represent a shared clonal lineage. To evaluate clonal divergence, we calculated the absolute difference in cellular prevalence between tumor and PDO for each cluster and categorized clusters as highly concordant (<10% difference), moderately concordant (10 to 29%), or divergent (≥30%). For each tumor-PDO pair, we computed the proportion of clusters in each category and stratified results by SNV concordance status. Clusters were stratified by clone size (based on number of associated mutations): small (<5 mutations), medium (5 to 20 mutations), and large (>20 mutations). Association between clone size and concordance category was tested using chi-square test. Logistic regression was applied to compute ORs for the relationship between clone size and divergence [divergent (1) versus not divergent (0)], with large clones as the reference group. The analysis was done in R version 4.4.2. To assess the fate of dominant clones, we identified clusters with >10% cellular prevalence in the tumor and evaluated their representation in the matched PDO. We calculated fold change between PDO and tumor cellular prevalence to quantify expansion, contraction, or loss of dominant tumor clones in PDOs. Clones were labeled as expanded (>1.25× increase in PDO), contracted (<0.75×), unchanged (0.75 to 1.25×), or lost (undetected in PDO).

PDO area tracking with Incucyte

A total of 1000 cells were plated in a 10-μl cell culture medium/Matrigel mix (1:2, v/v) in a 96-well plate. Material catalog numbers are in table S15. After polymerization at 37°C for 30 min, 15 μl of tissue-specific medium was added per well. Drugs were introduced 72 hours later in 15 μl of fresh medium (total 30 μl) using a nine-point dilution series. PDO growth was monitored via daily bright-field imaging with the Sartorius Incucyte S3, and cell area was analyzed using the affiliated software. All wells were normalized to PDO size to assess drug response.

PDO staining and microscopy

A total of 5000 cells were plated in 1% collagen I–coated, black 96-well plates and incubated for 48 hours. Material catalog numbers are in table S15. Cells were then irradiated at 10 Gy using the RS 2000 x-ray irradiator (Rad Source Technologies) and fixed with 4% paraformaldehyde in PBS. Cells were permeabilized and blocked in 0.1% Triton X-100 and 2% bovine serum albumin in PBS for 30 min, followed by a 2-hour incubation at 1:400 dilution with antibodies against RAD51, phospho-histone H2A.X, or geminin. After washing, cells were incubated at 1:500 dilution with Alexa Fluor–conjugated secondary antibodies for 1 hour and then mounted with ProLong Diamond Antifade with 4′,6-diamidino-2-phenylindole. Images were acquired on a Zeiss LSM 880 confocal microscope (63× objective), with exposure settings adjusted to prevent oversaturation. At least 100 nuclei per condition were quantified, and foci were counted using Fiji ImageJ (v2.14.0/1.54f).

PDO drug screening

A total of 1200 cells were plated in a 8-μl cell culture medium/Matrigel mix (1:2, v/v) in a 384-well plate. Material catalog numbers are in table S15. After a 30-min incubation at 37°C for polymerization, 15 μl of tissue-specific medium was added per well. Drugs were added 72 hours later in 15 μl of fresh medium (total volume, 30 μl) and incubated for 168 hours. Drugs from MedChemExpress or SelleckChem and were diluted per the manufacturers’ instructions. For single-agent testing of talazoparib, we did a 10-point, threefold serial dilution ranging from 0.2 nM to 10 μM. For the combination screening, we did a six-point, threefold serial dilution, with the highest concentration of talazoparib remaining at 10 μM. The concentration ranges of the combination drugs were adjusted to ensure that their respective median inhibitory concentration values were included within the screening range. For most PDO lines, the highest concentrations were as follows: 500 μM temozolomide, 10 μM ceralasertib; 33.3 nM prexasertib, 33.3 nM adavosertib, 10 μM AZD7648, and 33.3 nM gemcitabine. One PDO line was tested with a maximum concentration of 33.3 nM ceralasertib due to its high sensitivity to the drug in single-agent testing. All experiments were done at least three times with technical and biological replicates. Luminescent viability was assessed using CellTiter-Glo 3D, and luminescence was read using the Biotek Synergy Neo2 Plate reader. Data analysis was performed running nonlinear regression in GraphPad Prism 9. As the reported C_max for talazoparib was too low to induce sensitivity (14), viability at the maximum dose (10 μM) was used to classify sensitivity: viability of >25% as resistant and ≤25% as sensitive. Synergy scores were calculated using SynergyFinder+ based on the Zero Interaction Potency model, with scores of >10 indicating synergy, <−10 indicating antagonism, and between −10 and 10 suggesting an additive effect.

Germline mutation identification for PARPi-treated PDOs

Germline mutation in the PDOs used in the PARPi screening were identified through various methods, including clinical sequencing reports, Exact-1 pipeline, and WGS. Clinical sequencing reports were generated by platforms such as Ambry, MSK-IMPACT, GUARDIAN360, 50 Gene, and Foundation One. Germline somatic variants identified through WES were derived from blood or saliva samples, following the pipeline as previously described (93). Germline data from selected WGS analysis were obtained from samples included in a separate study (94).

DepMap analysis

Drug sensitivity data [log₂ fold change (log2FC), release 24Q2] for talazoparib and associated mutation and expression data (release 25Q3) were retrieved from the DepMap project. Talazoparib log2FC data were available for a total of 562 cell lines, of which 553 had expression data and 557 had mutations data available. For categorical comparisons, cell lines were classified as sensitive (log2FC in the bottom 25th percentile; n = 136) or resistant (log2FC >0; n = 118). DGE analysis was performed between sensitive and resistant groups using DESeq2, controlling for tumor lineage as a confounder (design: ∼OncotreeLineage + Talazoparib_Response). GSEA was subsequently performed using the fgsea package in R, querying hallmark (all pathways) and KEGG, Reactome, and GO Biological Process gene sets subset to pathways matching the terms DNA_REPAIR, HOMOLOG, MISMATCH, MMR, or BER. Correlation between gene expression and talazoparib sensitivity (log2FC) was assessed for a curated panel of HRD/DDR genes using Pearson correlation across all available cell lines (n = 553), with statistical significance defined as FDR < 0.05. To assess associations between specific somatic mutations and drug response, we applied multivariate linear regression analyses associating mutation status (wild-type = 0, mutated = 1) with talazoparib sensitivity (log2FC), correcting for tissue lineage (Drug Response ∼ Mutation + OncotreeLineage). Negative coefficients indicate that the mutation is associated with increased sensitivity (lower cell viability). P values were adjusted for multiple testing using the Benjamini-Hochberg method.

Supplementary Materials

This PDF file includes:

Figs. S1 to S9

Tables S1 to S16