Facts and figures of XENTURIONPDXs provide a nearly unlimited source of high-quality, propagatable material for generating ex vivo preclinical models. Drawing on our experience with PDX establishment and characterization14,15,16,17,18, we have created a biobank of matched PDXs and PDXTs that reflects the biological and clinical diversity of metastatic CRC. The primary purpose of this biobank is to streamline the identification, screening, and prioritization of anticancer agents in the preclinical development pipeline to expedite the selection process before embarking on more resource-intensive in vivo validations. While the characterization of CRC tumoroid collections derived directly from patient tumors is extensive1,4,9,10,11,12, limited information exists regarding the molecular and biological fidelity of PDXTs. To address this knowledge gap, we performed a systematic comparison between paired PDXs and PDXTs in terms of mutational profiles, gene copy number architecture, transcriptomic features, and responsiveness to standard-of-care therapy. This comparative effort was also leveraged to extract genes that exhibited concordant modulation under drug pressure in both PDXs and PDXTs, with the aim to pinpoint hits potentially involved in tumor adaptation to therapeutic stress. Following hit nomination, a stepwise drug screen for actionable targets was conducted in PDXTs, and surviving candidate compounds were finally tested in vivo. The workflow of XENTURION characterization is illustrated in Fig. 1A.Fig. 1: Facts and figures of XENTURION.A Schematic overview of XENTURION experimental design. Matched PDXTs and PDXs were subjected to comparative mutational, gene copy number and transcriptomic analyses. Molecular annotation was paralleled by systematic assessment of ex vivo and in vivo response to cetuximab. Post-cetuximab transcriptomic profiles were leveraged to extract upregulated genes potentially involved in adaptive resistance to EGFR blockade. Compounds against candidate targets were tested in a stepwise drug screen, and those that proved effective in PDXT assays underwent final validation in PDXs. B, C Success rate in the early derivation of tumoroid lines according to the nature of the sample of origin (B) or the number of derivation attempts (C). When early-stage tumoroids were derived from different originating samples (e.g., fresh and frozen PDX explants), success rates were computed for models derived from freshly explanted tumors. SR success rate. D Number of validated tumoroids according to the number of freeze-thaw cycles. E Main clinical and molecular features of the starting population from which tumoroid derivation was attempted. The circus plot includes all quality-checked cases with successful validation (n = 133), those that failed validation (n = 24), early-stage cases for which validation was not performed (Not perf, n = 29), and those that failed early derivation (n = 57). F female, M male, MSI-H microsatellite instability high, MSS microsatellite stability, NA not available, WT wild-type. F Odds ratios of a multivariate logistic regression with success status of PDXT early derivation and validation (1, successful, n = 129; 0, failed, n = 73) as dependent variable and several clinical and molecular annotations as independent variables. Red color indicates that the independent variable has a negative effect on the validation rate; blue color indicates the opposite. The only continuous variables are stage and age at collection; all other variables are binary. Confidence interval of odds ratios, 95%. Panel A was partly generated using adaptations of open-access pictures released under Creative Commons Attribution Licenses; see Figure preparation in the Methods for credits and details. Source data are provided as a Source Data file.Between March 2015 and September 2021, a total of 267 CRC liver metastases from 260 patients were processed for tumoroid derivation (for seven cases, two liver metastases from the same patient were available). Information on treatment history was accessible for 255 donor patients (Supplementary Data 1). Of these patients, 174 (68.2%) had received prior chemotherapy. Treatments included an oxaliplatin-based regimen (n = 117, 45.9%), an irinotecan-based regimen (n = 34, 13.3%), a sequential or combination therapy with both oxaliplatin and irinotecan (n = 20, 7.8%), or capecitabine alone (n = 3, 1.18%). Chemotherapy was administered together with targeted agents (the anti-EGFR antibodies cetuximab or panitumumab and/or the anti-VEGFA antibody bevacizumab) in 85 patients (33.3%).To minimize alterations in the biology of tumors and prevent biased selection of specific growth dependencies, we standardized culture conditions that sustained long-term growth of tumoroids in a minimal medium containing EGF as the sole exogenous growth factor. The chosen EGF concentration (20 ng/ml) was adjusted to secure tumoroid proliferation on a population scale. This approach draws inspiration from previous work aimed at defining optimal culture conditions for human and mouse CRC tumoroids19,20 and aligns with the notion that CRC tumoroids gradually become independent from niche signals during cancer progression21. For preliminary inclusion into the biobank, each PDX-tumoroid pair had to show matching identity with the original material by genetic fingerprinting, negativity for human and mouse pathogens, and a histology congruent with CRC phenotypes. This approach resulted in the inclusion of 243 models; 19 cases were excluded due to discordant DNA fingerprinting between the PDX-tumoroid pairs and the original patient sample (see Methods); four were diagnosed as lymphomas by histopathological evaluation; and one was excluded for technical reasons (deterioration of archived material) (Supplementary Data 2).We categorized tumoroids as ‘early-stage’ when their initial propagation cultures could be expanded to a minimum of 200,000 viable cells for cryopreservation, typically achieved after three rounds of cell splitting. The vast majority of samples (211/243, 87%) were processed using freshly explanted PDX tumors as the only source, achieving an 80% success rate (169/211) (Fig. 1B and Supplementary Data 2). In the few instances where PDXT derivation was attempted from frozen PDX tumors, the success rate was lower (5/10, 50%) (Fig. 1B and Supplementary Data 2). Differences in the generation of early-stage PDXTs were also evident when tumoroids were derived in parallel from fresh and frozen material from the same PDX; in particular, among eight PDXs with both fresh and frozen tumor fragments available, early-stage PDXTs were successfully established solely from fresh tissues in four cases and from both fresh and frozen tissues in two cases, resulting in an overall success rate of 75% (6/8) for fresh explants and 25% (2/8) for frozen material (Fig. 1B and Supplementary Data 2). Although the number of early-stage PDXTs from frozen PDX tumors is limited, these findings suggest that freshly explanted tumors may be more conducive to PDXT initial propagation than frozen material. XENTURION also includes 13 tumoroids directly derived from fresh human specimens after surgery; in this subgroup, the success rate in the production of early-stage tumoroids was markedly lower (5/13, 38%). Finally, for one case, early-stage tumoroids were successfully obtained from both fresh PDX explants and the original patient sample (Fig. 1B and Supplementary Data 2).Overall, XENTURION comprises 186 early-stage tumoroids of metastatic CRC, with a success rate of 77% (186/243); the collection is predominantly represented by PDXT lines (181/186, 97%), each with paired PDXs available (Fig. 1B and Supplementary Data 2). In most models, a single derivation procedure sufficed to yield early-stage tumoroids, with a success rate of 60% (147/243) (Fig. 1C). In cases where early-stage tumoroid establishment failed after the first attempt, two or more additional rounds were performed if PDXs were available. The success rate of early-stage tumoroid establishment showed a proportional decrease with attempt repetition: 44% after the second attempt, 36% after the third attempt, and 29% after the fourth or subsequent attempts (Fig. 1C). Hence, we can reasonably conclude that tumoroids of metastatic CRC that do not grow in culture after the first derivation round are less likely to give rise to early-stage models.To qualify as validated models capable of long-term recovery and expansion, early-stage tumoroids underwent a minimum of three freeze-thaw cycles. DNA fingerprinting-based identity checks and microbiologic tests for Mycoplasma detection were conducted after each cycle. Of the 145 early-stage tumoroids subjected to at least three freeze-thaw cycles, 121 (83%) passed validation (Fig. 1D and Supplementary Data 2). Notably, the lack of recovery after the first freeze-thaw cycle proved to be a reliable indicator, identifying 92% (22/24) of cases that would not withstand additional ‘rescue’ cycles and, therefore, would fail validation. Conversely, almost all early-stage tumoroids that successfully recovered after the first freeze-thaw cycle proceeded to complete validation in subsequent cycles (121/123, 98.37%) (Fig. 1D and Supplementary Data 2). For this reason, we admitted in the final collection of validated cases additional models that had survived two freeze-thaw cycles (seven cases) or one cycle (five cases).Based on these selection criteria, an initial version of the collection encompassed a total of 133 validated tumoroids: 129 PDXTs (with paired PDXs) and four tumoroids directly derived from donor patients (Fig. 1D and Supplementary Data 2). PDXTs and their matched PDXs with sufficient and good-quality nucleic acid material were processed to obtain mutational, gene copy number and transcriptomic profiles. Of note, one validated PDXT model was identified as an anal squamous cell carcinoma after subsequent transcriptomic analysis and post hoc pathological examination. Therefore, the final count of the definitive XENTURION collection comprises a total of 128 fully validated PDX/PDXT pairs, with a complete dataset for all molecular dimensions available for 114 paired siblings.The key clinical and molecular attributes of the samples that fed into XENTURION, including primary tumor sidedness and stage, patients’ sex, age and exposure to therapy before sample donation, DNA microsatellite status, and the presence of clinically relevant driver mutations, are summarized in Fig. 1E. To explore whether the process of tumoroid derivation led to over- or under-representation of these features in XENTURION compared to the starting population, we assessed their relative distribution in validated models versus those failing early-stage derivation or validation. Enrichment analysis revealed that early-stage derivation or validation of metastatic samples with primary tumor location in the right colon was less successful than expected by chance (P = 0.005, odds ratio [OR] = 0.38, 95% confidence interval [CI] 0.19–0.74]) (Fig. 1F). This observation may be attributed to the fact that left-sided CRC tumors, which are usually more dependent on EGFR signaling22, were more stimulated to grow in the presence of the EGF ligand present in the culture medium compared to right-sided tumors. A similar enrichment among samples that failed to be established was observed for tumors harboring KRAS mutations (P = 0.019, OR = 0.46, 95% CI 0.23–0.88) (Fig. 1F). This result was unexpected, given that KRAS mutant CRC tumors are generally more aggressive than KRAS wild-type tumors23,24, and ectopic introduction of mutant KRAS promotes – rather than contrasts – the expansion of CRC tumoroids25. We suspected that the higher representation of KRAS mutant cases among tumoroids that did not survive initial derivation or validation might be related to a procedural bias linked to the timing of tumoroid generation. Indeed, PDXTs from KRAS wild-type tumors were more often derived from late-passage (more than three) PDXs, typically from large cohorts propagated in vivo multiple times to obtain sufficient replicas for testing with the anti-EGFR antibody cetuximab. Since mutant KRAS confers resistance to cetuximab26, PDXs with KRAS mutations were not repeatedly expanded for cetuximab treatment, and tumoroids were more frequently generated from smaller cohorts at earlier passages. To test this hypothesis, we computed the PDX passage at which tumoroids were derived along with the two significant enrichments shown in Fig. 1F (KRAS mutations and tumor right-sidedness) using logistic regression. In multivariate analysis, the odds ratio of tumor sidedness maintained statistical significance (P = 0.025, OR = 0.32, 95% CI 0.12–0.86), whereas the odds ratio of KRAS mutations became not significant (P = 0.313, OR = 0.64, 95% CI 0.26–1.53) (Supplementary Fig. 1). Accordingly, there was a trend for late-passage PDXs to be more likely to yield validated tumoroids than early-passage PDXs (P = 0.097, OR = 1.29, 95% CI 0.97–1.76) (Supplementary Fig. 1). This supports the assumption that PDX passaging rather than KRAS mutations impacted PDXT stability. This observation is in line with our finding that fresh samples from patients (never passaged in mice) were less susceptible to grow in culture (Fig. 1B), suggesting that serial mouse engraftment eases the adaptation of cancer cells to long-term propagation ex vivo.Mutational and gene copy number analysis of paired PDXTs and PDXs reveals substantial model concordanceWe performed targeted next-generation sequencing of 116 relevant CRC genes27 to detect small somatic alterations (single nucleotide variants [SNVs] and indels) in a set of 125 sibling pairs comprising validated PDXTs and matched PDXs with sufficient and good-quality DNA material. The overall distribution of variant allele frequencies (VAFs) and the number of identified variants showed consistency between PDXTs and PDXs (P = 0.09 and P = 0.72, respectively, by Kolmogorov-Smirnov test) (Supplementary Fig. 2). At the level of individual genes, the vast majority of mutations were conserved (Fig. 2A). A statistically significant imbalance, with a higher representation of gene mutations in PDXTs, was observed only for APC (P = 0.021 by χ2 test) and KAT6A (P = 0.046 by χ2 test). However, manual curation of sequencing reads confirmed the presence of APC and KAT6A mutations in 21/24 (87.5%) and 4/6 (66.7%) of the corresponding PDXs, respectively (Supplementary Fig. 3). The overall consistency of the mutational landscape between the PDX-PDXT pairs extended to the level of specific mutations. The Jaccard similarity coefficient, used to quantify this concordance, was markedly higher for matched models than unmatched models (median matched, 1.00, interquartile range [IQR] 0.66–1.00; median unmatched, 0.00, IQR 0.00–0.00; P < 2.2e-308 by two-tailed Mann–Whitney test) (Fig. 2B). Importantly, the extent of mutational concordance between PDXTs and PDXs mirrored that of a recent comparison involving 536 original patient tumors and matched PDXs across 25 cancer types28. This indicates negligible divergence between pre-derivation samples, PDXs and PDXTs when considering the general mutational repertoire.Fig. 2: Comparative landscape of somatic single nucleotide variations and indels in paired PDXTs and PDXs.A Common and private alterations in 124 pairs of matched PDXTs and PDXs. One pair for which mutational data were available was excluded because no alterations with VAFs > 0.05 were detected. Genes without any alteration in the whole cohort were removed. The top barchart shows the total number of mutations for each sample. The barchart on the right shows the percentage of mutations for each gene in the cohort. B Jaccard similarity indexes of somatic alterations between 124 matched PDXs and PDXTs. C Gene-level population frequencies of mutational alterations in PDXTs versus those detected in the TCGA dataset or the MSK-IMPACT dataset; the inset shows that the correlation is not driven solely by genes with high mutational frequencies. Source data are provided as a Source Data file.Subclonal variants that are poorly represented in PDXs may become dominant in PDXTs if they confer a growth advantage in culture. To study the clonal composition of PDX-PDXT pairs, we examined the prevalence of shared mutations with VAFs higher than 0.05 for genes with at least five alterations in the collection. No significant differences in allele frequencies emerged from this analysis (P = 0.66 by paired t-test) (Supplementary Figs. 4 and 5), indicating that intratumor clonal heterogeneity was substantially preserved in PDXTs with respect to originating PDXs. In some cases, the allele frequencies of alterations in frequently mutated tumor suppressor genes (for example, APC and TP53) were 1 in PDXTs and slightly lower in the paired xenografts (Supplementary Fig. 4), suggesting subtle defects in the filtering procedure of mouse reads deriving from host stromal contamination.Next, we compared the frequency of gene alterations in our collection with two large datasets of human samples from CRC patients: TCGA, mainly consisting of primary tumors27, and MSK-IMPACT, predominantly composed of metastatic samples29. We found significant correlations between PDXTs and both clinical datasets (Pearson coefficient, 0.93, P = 1.46e-51 for TCGA; Pearson coefficient, 0.96, P = 2.37e-64 for MSK-IMPACT) (Fig. 2C) as well as between PDXs and these datasets (Pearson coefficient, 0.92, P = 4.24e-49 for TCGA and 0.95, P = 4.57e-62 for MSK-IMPACT) (Supplementary Fig. 6). It is worth noting that colorectal tumors with microsatellite instability (MSI) are typically less prevalent in metastases compared to primary tumors, consistent with their generally better prognosis30. For example, MSI tumors account for approximately 12% of cases in the TCGA dataset of primary tumors27, 4% in the MSK-IMPACT collection of metastatic lesions29, and 2.46% (6/243) in XENTURION metastatic PDXs (Fig. 1E). Despite this difference, the above comparisons underscore that XENTURION reflects the overall mutational landscape of patient cohorts and point to substantial similarity in mutational frequencies between primary and metastatic CRC tumors. A high level of genomic concordance between primary and metastatic colorectal tumors has already been documented in an MSK-IMPACT comparative analysis of recurrently mutated genes29.We then investigated whether the PDXT validation protocol could lead to the enrichment or depletion of defined variants. By applying univariate logistic regression models to predict validation, we considered the presence or absence of any SNVs or indels in PDXTs as independent variables and the validation status as the dependent variable. Among genes mutated in at least five tumoroids, only mutations in the CTNNB1 gene (encoding β-catenin) were significantly over-represented in PDXTs that failed validation (P = 0.002, OR = 0.067, 95% CI 0.01–0.40) (Supplementary Fig. 7 and Supplementary Data 3). Both CTNNB1 and APC mutations result in constitutive activation of the Wnt pathway, which sustains CRC proliferation. However, mutant β-catenin is known to be more modulatable by exogenous Wnt stimulation than mutant APC31. Since PDXTs were cultured in the absence of Wnt agonists, it is conceivable that CTNNB1 mutant samples are less fit to grow in a nutrient-poor medium compared to APC mutant samples.Most colorectal tumors display chromosomal instability, a condition that may be intensified by evolutionary bottlenecks such as those introduced during tissue culture propagation. To examine whether copy number changes materialized in our models following ex vivo culturing, we conducted low-pass whole genome sequencing in the same 125 PDX-PDXT pairs used for mutational profiling. This analysis revealed a high consistency in copy number variations between PDXTs and the corresponding PDXs compared with unmatched samples (Pearson coefficient between segmented log ratios: median matched, 0.89, IQR 0.83–0.94; median unmatched, 0.40, IQR 0.30–0.50; P = 1.92e-81 by two-tailed Mann–Whitney test) (Fig. 3A). In PDXTs, the overall landscape of chromosomal alterations paralleled observations in patients27. In particular, whole-arm copy number gains were detected in chromosomes 7, 13 and 20, and long-arm specific gains were detected in chromosome 1 and 8; whole-arm losses occurred in chromosome 18 (where the SMAD4 gene lies) and in the short arms of chromosomes 1 and 8 (Fig. 3B). Accordingly, the population frequencies of copy number alterations (CNAs) at the gene level, obtained with GISTIC, showed positive correlations between PDXTs and the TCGA or MSK-IMPACT patient cohorts (Pearson coefficient, 0.92 [gains] and 0.87 [losses] for TCGA; 0.80 [gains] and 0.88 [losses] for MSK-IMPACT; P < 1e-230 by two-tailed Mann-Whitney test for both comparisons) (Fig. 3C). Similar correlations were observed between PDXs and the patient cohorts (Pearson coefficient, 0.92 [gains] and 0.90 [losses] for TCGA; 0.84 [gains] and 0.89 [losses] for MSK-IMPACT; P < 1e-230 by two-tailed Mann-Whitney test for both comparisons) (Supplementary Fig. 8). In summary, our data suggest that PDXTs generally retain the mutational and genomic structure of parental PDXs. Moreover, the distribution of major mutational drivers and CNAs observed in XENTURION PDXTs and PDXs is largely superimposable to that of human CRC samples.Fig. 3: Comparative copy number architecture in paired PDXTs and PDXs.A Distribution of Pearson correlation coefficients between copy number profiles of matched (n = 125) and unmatched (n = 15,500) pairs of PDXTs and PDXs. B Autosomal copy number profiles of PDXTs (n = 125), expressed as segmented log2 ratio of the normalized read depth. Red and blue colors indicate gain and loss events, respectively. C Gene-level population frequencies of gain or loss events, as identified by GISTIC, in PDXTs versus those detected in the TCGA or the MSK-IMPACT datasets. Amp amplification, Del deletion. Source data are provided as a Source Data file.PDX genomic patterns are largely preserved during PDXT serial passagingTo conduct a more precise investigation of potential genomic pattern alterations that may mark the transition from PDXs to PDXTs and to explore whether these patterns change with PDXT serial passaging, we generated whole exome sequencing data for a subset of 23 trios. These trios comprised donor PDXs, early-passage (third passage) PDXTs, and late-passage (passages from eight to 12) PDXTs from the same patient. Models were selected based on the representativeness in the distribution of high-frequency mutations (APC 14/23, 60.8%; TP53 16/23, 69.6%; KRAS 9/23, 39%) and the availability of quality-checked DNA in sufficient quantities for library preparation, including matched normal DNA as a reference for high-confidence annotation of somatic mutations and copy number variations. This deeper analysis in a more restricted subset of models confirmed the overall mutational and gene copy number concordance between paired PDXs and PDXTs observed in the larger cohort of 125 models, as previously evidenced through targeted sequencing and low-pass whole genome sequencing. The median Jaccard similarity coefficient for mutations with VAF > 0.05 was markedly higher for matched PDXs and early-passage PDXTs than for unmatched models (matched, 0.77, IQR 0.64–0.86; unmatched, 0.00, IQR 0.00–0.00; P = 7.1e-39 by two-tailed Mann–Whitney test) (Supplementary Fig. 9A). Likewise, the median Pearson correlation of copy number profiles was higher for matched models compared with unmatched samples (normalized depth ratio for matched samples, 0.94, IQR 0.91–0.95; unmatched, 0.55, IQR 0.46–0.62; P = 4.831e-16 by two-tailed Mann–Whitney test). We did not observe recurrent copy number gains or losses when comparing early PDXTs to matched donor PDXs (Supplementary Figs. 9B and 10A).The substantial similarity in the mutational and copy number landscape exhibited by matched PDXs and early-passage PDXTs was maintained in late-passage models. High concordance was observed in early- and late-passage PDXT pairs when considering single-nucleotide alterations (median Jaccard index for matched pairs, 0.75, IQR 0.59–0.85; unmatched pairs, 0.00, IQR 0.00–0.00; P = 7.6e-45 by two-tailed Mann–Whitney test) (Fig. 4A) as well as gene copy number (median Pearson correlation between normalized depth ratios for matched samples, 0.93, IQR 0.88–0.95; unmatched, 0.56, IQR 0.50–0.62; P = 4.887e-16 by two-tailed Mann–Whitney test). No specific events of copy number gain or loss were detected in late- versus early-passage PDXTs (Fig. 4B and Supplementary Fig. 10B). These results are consistent with prior evidence showing that copy number profiles remain largely stable during PDX engraftment and serial passaging18.Fig. 4: Comparative genomic landscape in matched early- and late-passage PDXTs.A Jaccard similarity indexes of somatic alterations (VAFs > 0.05) between 23 matched early- and late-passage PDXTs. The Jaccard index for model CRC1460 is 0, likely due to low tumor mutational burden (TMB) (1.96 mutations [muts] per mega base pairs [Mbps] in early-passage PDXTs and 2.13 muts/Mbps in late-passage counterparts; median TMB for all-early passage PDXTs, 5.4 muts/Mbps, IQR 4.5–6.1; median TMB for all late-passage PDXTs, 5.7 muts/Mbps, IQR 4.8–6.7). CRC1460 low TMB, coupled with variant annotation limited to PCGR tiers ≤ 3 (enriched for mutations with stronger potential relevance for cancer, see “Methods” section), resulted in detecting only a single alteration exclusively in the late-passage PDXT. B Comparison of autosomal copy number profiles between 23 matched early- and late-passage PDXTs. ‘No changes’ refers to stable or quasi-stable regions. ‘Losses in late-passage PDXTs’ are defined as loci with copy number ≤1 in the late-passage PDXTs and ≥2 in the early-passage counterparts. ‘Gains in late-passage PDXTs’ are defined as loci with copy number ≥5 in the late-passage PDXTs but not in the early-passage counterparts. The genome is represented by 100k base pair long bins; each row in the heatmap represents a pair of matched early- and late-passage PDXTs. C Comparison of LOH events between 23 matched early- and late-passage PDXTs. ‘No LOH’ indicates regions without LOH in both early- and late-passage pairs. ‘LOH newly detected in late-passage PDXTs’ indicates regions with newly acquired LOH events in the late-passage PDXTs that are not present in the early-passage counterparts. ‘LOH no longer detected in late-passage PDXTs’ indicates regions with LOH events detected in the early-passage PDXTs that are no longer detected in the late-passage counterparts. ‘Common LOH in both early- and late-passage PDXTs’ indicates regions with LOH events shared between sibling pairs. The genome is represented by 100k base pair long bins; each row in the heatmap represents a pair of matched early- and late-passage PDXTs. D Number of clusters (clones) inferred by PyClone-vi in 23 matched early- and late-passage PDXTs. Source data are provided as a Source Data file.We also examined potential loss of heterozygosity (LOH) events that may have occurred during PDXT derivation from PDXs or during tumoroid passaging by analyzing the frequency of minor alleles for heterozygous germline single-nucleotide polymorphisms (SNPs). When comparing parental PDXs to early-passage PDXTs, we did not observe any recurrent LOH gained during PDXT generation, only minor sporadic events (0.48% of the studied genomic regions). Conversely, there was a relatively high frequency of LOH events shared between donor PDXs and PDXTs (21.04%) (Supplementary Fig. 9C). Similarly, comparisons between early- and late-passage PDXTs revealed a low incidence of LOH events gained in late-passage models (0.19%) and a higher occurrence of shared LOH events (21.46%) (Fig. 4C). In both comparisons, we noted a small percentage of errors in LOH calls due to regions with a lost allele in PDXs not detected in early-passage PDXTs (0.16%), or regions with a lost allele in early-passage PDXTs not detected in the late-passage counterparts (0.07%). These infrequent errors are likely attributable to intrinsic noise in the segmentation and calling procedure.Finally, we explored tumor heterogeneity across matched models from the same patient by estimating variations in the number of subclones using PyClone-vi. This tool infers the number of clones from single-nucleotide variants, small indels and CNAs, and produces an output that includes the probability of assignment of a mutation to a specific subclone (thereby identifying a subclone as a cluster of mutations) and the number of mutations belonging to a cluster32. By applying filters with a 50% probability threshold and a minimum of 10 mutations for cluster definition, we found that all models were oligoclonal, with most consisting of 2–3 clusters. Discernible yet modest variations in clonal architecture were observed in approximately 50% of the analyzed trios. Among PDXs and early-passage tumoroids, 13 models maintained the same number of clusters, while five early-passage tumoroids gained one cluster and five lost one cluster compared with originating PDXs (Supplementary Fig. 9D). In comparisons of late- versus early-passage PDXTs, 11 models exhibited the same clonal organization, seven late-passage PDXTs gained one or two clusters, and five lost one or two clusters (Fig. 4D). These findings suggest no overt differences in subclonal heterogeneity between PDXs and PDXTs, as well as between early- and late-passage PDXTs. Overall, the various levels of genomic analysis in parental PDXs, early-passage PDXTs, and late-passage PDXTs did not reveal substantial drifts caused by the experimental workflow, while providing a more detailed landscape of the molecular features characterizing XENTURION models throughout their lifespan.PDXT transcriptional identity retains fidelity to corresponding PDXs and is stable over timeTranscriptomic data were used to compare the gene expression profiles of 21 surgical specimens from donor patients (human liver metastases, HLMs), 119 PDXs from which PDXTs were successfully derived and validated, and 124 validated PDXTs, all chosen for meeting standard RNA quality criteria for sequencing. Based on gene ontology (GO) enrichment analysis of differentially expressed genes, gene signatures related to cellular division and DNA replication were more abundant in PDXs than HLMs, consistent with the faster growth rates of xenografts compared with those of tumors in patients33 (Supplementary Data 4). Pathways downregulated in PDXs versus HLMs were associated with innate and adaptive immunity and stromal remodeling (Supplementary Data 4), as expected for models grown in immunocompromised animals and in agreement with the observation that human stromal cells are replaced by mouse components soon after tumor implantation34,35. Being derived from PDXs, PDXTs predictably showed similar upregulated and downregulated pathways with respect to HLMs (Supplementary Data 4).Comparative analysis of xenografts and validated tumoroids revealed that gene signatures of steroid, retinoid and fatty acid metabolism were more expressed in PDXTs than in PDXs (Supplementary Fig. 11 and Supplementary Data 4). This may be attributed to metabolic adaptations to the culture conditions and is in line with previous results obtained in a smaller set of 19 CRC PDX/tumoroid sibling pairs36. A cluster of gene sets functionally related to cellular response to innate immunity pathways stood out as significantly downregulated in PDXTs compared to PDXs (Supplementary Fig. 11 and Supplementary Data 4), likely as a result of the depletion of host innate immune cells during the transition from in vivo tumors to ex vivo cultures. Additionally, to gain deeper insights into the molecular characteristics that may impact tumoroid establishment, we analyzed differentially expressed genes between the 119 PDX samples with accompanying RNAseq data that had successfully produced validated PDXTs and 49 PDX samples from which tumoroid early derivation or validation had failed. This analysis identified 328 downregulated genes and 113 upregulated genes in PDXs that successfully originated tumoroids compared to unproductive PDXs. GO annotation of the downregulated genes indicated an enrichment for signatures associated with epithelial squamous differentiation (‘keratinocyte differentiation’ GO term, adjusted P = 1.17e-6 by one-tailed Fisher’s exact test), whereas upregulated genes exhibited features related to extracellular matrix components (‘collagen-containing extracellular matrix’ GO term, adjusted P = 3.76e-2 by one-tailed Fisher’s exact test) (Supplementary Fig. 12A and Supplementary Data 4). This suggests that PDX models with low expression of epithelial differentiation markers and high expression of extracellular matrix molecules are more likely to establish tumoroids.To investigate the transcriptional fidelity of PDXT-PDX pairs, we first considered a subset of 79 ‘super-matched’ samples selected based on genealogical proximity (i.e., pairs made of an early-passage PDXT with its nearest ancestor PDX). This analysis showed high consistency in transcript abundance between samples, with an intramodel Pearson correlation coefficient significantly greater than intermodel correlations (median matched, 0.83, IQR 0.80–0.85; median unmatched, 0.62, IQR 0.57–0.66; P = 3.477e-52 by two-tailed Mann–Whitney test) (Fig. 5A). We then extended the survey to a more variegated set of sample families, which included early and late propagations from the same PDXT and one or more matched PDXs grown in more distant generations of mice. In this larger set, consisting of 116 PDX/PDXT sibling models for a total of 308 data points (including replicates), the similarity between matched PDXTs and PDXs was confirmed; in particular, the intramodel Pearson correlation coefficient was significantly higher than intermodel correlations (median matched, 0.78, IQR 0.75–0.81; median unmatched, 0.52, IQR 0.50–0.54; P = 1.50e-39 by two-tailed Mann–Whitney test), and 80/116 (69%) models derived from the same originating tumor proved to belong to the same cluster by unsupervised hierarchical clustering (Supplementary Fig. 12B). We noted that the transcriptional profile of tumor CRC1241, which had a very high PDXT-PDX correlation (Pearson coefficient, 0.90), was different from the rest of our cohort (average Pearson coefficient, 0.31) (Fig. 5A). Accordingly, a deep-learning tool that uses RNA gene expression data to infer a tumor’s primary tissue of origin37 predicted the tumor as a cervical squamous cell carcinoma (Supplementary Data 5), and post hoc pathological revision cataloged it as an anal squamous cell carcinoma. For this reason, we excluded CRC1241 from further analyses.Fig. 5: Comparative gene expression profiles and transcriptional subtype assignment in paired PDXTs and PDXs.A Pearson correlations of gene expression profiles in matched PDXs and PDXTs. Pearson correlation coefficients were calculated for matched (n = 79) and unmatched (n = 6.162) pairs. B Pearson correlations of gene expression profiles in matched early- and late-passage PDXTs. Pearson correlation coefficients were calculated for matched (n = 23) and unmatched (n = 506) pairs. C CMS and CRIS subtype assignment in 79 pairs of matched PDXTs and PDXs. NC non-classified. Source data are provided as a Source Data file.To explore whether tumoroid propagation leads to any transcriptional drift, we examined differentially expressed genes between the early- and late-passage PDXTs that were also utilized for the genomic comparisons. Only 30 genes were significantly downregulated in late-passage tumoroids compared to their early-passage counterparts, with no genes found to be upregulated. This finding indicates a substantial conservation of transcriptomic profiles across serial passaging. This consistency was further corroborated by correlation analyses (median Pearson coefficient, 0.920; IQR for matched models, 0.86–0.94; IQR for unmatched models, 0.39–0.52) (Fig. 5B).Gene expression profiling has been recently deployed to develop CRC classifiers with prognostic and predictive significance. The Consensus Molecular Subtypes (CMS) classifier was built on whole-tumor transcriptomes (including cancer cell and stromal/immune transcripts)38, whereas the CRC Intrinsic Signature (CRIS) classifier utilized PDX gene expression datasets to derive cancer cell-specific subtypes35. We first assigned each PDXT and PDX of the ‘super-matched’ 79 pairs to a CMS or CRIS subtype. Plausibly, many models failed CMS categorization due to lack of human stroma, which greatly contributes to CMS subtype assignment34,35,39. Conversely, all models received a CRIS designation (Fig. 5C). We then evaluated the consistency in subtype assignment using a tailored version of the Jaccard index, whereby the number of models with the same subtype in matched PDXs and PDXTs was divided by the total number of models assigned to that subtype. This index revealed good overall correspondence, with average values across subtypes of 0.44 for both CMS and CRIS. At the level of individual subtypes, a general stability in class assignment was observed with the exception of CMS4 (Fig. 5C). Specifically, the consistency index between PDXs and PDXTs was 0.69 for CMS1; 0.48 for CMS2; 0.40 for CMS3; 0.18 for CMS4; 0.45 for samples that failed CMS classification; 0.43 for CRIS-A; 0.33 for CRIS-B; 0.55 for CRIS-C; 0.41 for CRIS-D; and 0.45 for CRIS-E. The poor performance of PDX and PDXT class assignment to CMS4 is expected, as CMS4 characteristics are dominantly driven by human stromal transcripts that are absent in XENTURION models. PDXTs therefore display representative gene expression profiles that define their identity with parental PDXs and allow their classification into RNA expression-based CRC subtypes without a substantial culture bias.PDXT sensitivity to cetuximab is concordant with PDX response in a large-scale population trialThe anti-EGFR antibody cetuximab is a standard-of-care treatment with demonstrated clinical benefit in patients with inoperable RAS/RAF wild-type metastatic CRC40. We and others have used PDX-based resources to identify determinants of responsiveness and resistance to cetuximab and to nominate novel druggable targets for cetuximab-resistant tumors14,15,16,17,36,41. As a consequence, a large part of XENTURION’s PDXTs were derived from PDX models for which annotation of sensitivity to cetuximab was available. We leveraged this information to investigate how and to what extent PDXTs may act as functional ex vivo surrogates of therapeutic profiles in xenografts.Sensitivity to cetuximab was assessed in 119 validated PDXTs with growth characteristics and manipulability suitable for pharmacologic experiments. Each model was plated at three different cell densities (1250, 5000 and 20,000 cells/well) in a 96-well format and cultured for one week in the presence or absence of 20 μg/ml cetuximab without EGF (which competes with the antibody for receptor binding). The selected antibody concentration was based on existing literature data and represented an intermediate dose reported to achieve an inhibitory plateau across CRC cell lines and tumoroids9,17,36,42,43. The response was determined by measuring the ratio between treated and untreated cells and using as readouts endpoint luminescent ATP content and longitudinal cell imaging, for a total of 6174 measurements. A coefficient of variation (CV) was calculated to evaluate consistency between biological triplicates. Based on these CV measurements, 116 PDXT models with above-threshold inter-experiment consistency (see Methods) were selected for further analyses.The experimental setting was overall robust and reproducible, as documented by the significant correlations between luminescence-based and imaging-based detection for all cell plating densities (Pearson coefficient, 0.52; P = 1.7e-8 [1250 cells]; 0.65; P = 1.4e-15 [5000 cells]; 0.77; P = 5.5e-24 [20,000 cells]), and resulted in a graded distribution of responsiveness to cetuximab treatment (Fig. 6A). We then used linear regression models to compare all these PDXT measurements with the in vivo tumor response (defined as the relative volume change after three weeks of treatment) in 79 matched PDXs for which cetuximab therapeutic annotation was available (Supplementary Data 6). Also in this case, correlations were all positive and significant, with ATP values for the 5000 cell-plating density showing the best performance (Pearson coefficient, 0.56, P = 9.9e-8) (Fig. 6B and Supplementary Fig. 13A). Additionally, a significant, albeit slightly lower, concordance in cetuximab response between PDXs and PDXTs was observed when in vivo therapeutic sensitivity was measured by calculating tumor growth inhibition (TGI) scores, which compare tumor growth in the cetuximab-treated group to tumor growth in mice exposed to placebo (Pearson coefficient, 0.47; P = 5.5e-4) (Supplementary Fig. 13B and Supplementary Data 6). For this analysis, data were available for only 51 PDXs because some mice in the placebo arms had to be euthanized before the end of the three-week monitoring period due to reaching the humane endpoint. We attribute the reduced statistical significance observed with this alternative metric to the smaller sample size of the PDX-PDXT pairs included in the analysis.Fig. 6: Comparative annotation of cetuximab response profiles in paired PDXTs and PDXs.A Correlation of cetuximab response between values of endpoint ATP content (relative cell number) and values obtained by longitudinal cell imaging (relative tumoroid total area) in 116 PDXTs plated at different cell densities and treated with cetuximab (20 μg/ml) for one week. Each dot represents one single experiment performed in biological triplicate. Responses were assessed in 116 models for 5000 and 20,000 cells/well, and 102 models for 1250 cells/well. The shaded area represents the confidence interval of linear model prediction, 95%. B Correlation of cetuximab response in 79 pairs of matched PDXTs and PDXs. Response in PDXTs was evaluated as the ratio of viable cells after one week of treatment (20 μg/ml cetuximab, 5000 cells/well in a 96-well format) to untreated controls; response in matched PDXs implanted in both male and female NOD-SCID mice was evaluated as the percentage of tumor volume variation after three weeks of treatment (20 mg/kg, intraperitoneal injection twice a week) compared with tumor volume the day before treatment initiation. The shaded area represents the confidence interval of linear model prediction, 95%. C ROC curve showing the performance of PDXT-based results in predicting cetuximab response in vivo in 79 pairs. AUC, 0.81; responders (target prediction), 17; non-responders, 62. D EGFR KO scores for 13 PDXTs, distributed according to cetuximab sensitivity (black dots). Results are the average of the mean effect size of two sgRNAs against EGFR in two independent experiments, each performed in biological triplicates (with the exception of CRC0148, which was tested in three independent experiments). KO knockout, WT wild-type, amp amplification, mut mutation. Source data are provided as a Source Data file.Next, we analyzed cetuximab response in relation to genetic biomarkers known to confer resistance to cetuximab in patients. In line with clinical observations, tumors harboring KRAS, NRAS or BRAF mutations were generally refractory to EGFR blockade in both platforms (15 models out of 18 had a luminescence ratio > 0.7 when tested as PDXTs and a growth increase of more than 70% when tested as PDXs) (Fig. 6B). However, we also found some discrepant examples. First, one KRAS Q61H mutant model that had been categorized as a mild non-responder in vivo (relative tumor volume increase after three weeks of treatment, 42.35%) proved to be sensitive in the corresponding PDXTs (treated/untreated luminescence ratio, 0.23) (Fig. 6B). Interestingly, heterogeneous responses of KRAS Q61H mutant metastatic CRC tumors to anti-EGFR antibodies have also been observed in patients44,45. Second, a model with a subclonal cetuximab resistance mutation in the originating tumor (EGFR G465R, VAF 0.195) showed overt resistance in PDXs (relative tumor volume increase after three weeks of treatment, 166.5%) but appreciable sensitivity in the matched PDXT (treated/untreated luminescence ratio <0.3) (Fig. 6B). In this case, the disconnect between PDX and PDXT data is due to sampling bias; the PDXs used for monitoring tumor response to cetuximab in vivo harbored the EGFR G465R mutation, whereas the PDXTs used for the ex vivo drug screen were derived from a sibling xenograft where the alteration was absent. The final element of divergence was found for ERBB2 amplification, which predicts poor response to EGFR inhibition in patients with metastatic CRC46,47. As shown previously14,16, ERBB2-amplified PDXs failed to respond to cetuximab (Fig. 6B); however, this resistant phenotype was only partially recapitulated in PDXTs, with three models out of five displaying a certain degree of sensitivity (treated/untreated luminescence ratio ≤ 0.3) (Fig. 6B). The signaling and transformation potency of HER2 in ERBB2-amplified tumors is tunable by EGF stimulation48. On this ground, we speculate that the HER2 bypass pathway that blunts response to EGFR inhibition was below threshold in some PDXTs due to lack of EGF in the culture medium (thus, tumoroids retained sensitivity to cetuximab); conversely, the widespread availability of murine EGF in PDXs stimulated HER2 signaling to an extent sufficient to impart resistance to EGFR inhibition in vivo.We reasoned that results from this population trial might prove valuable to formalize the predictive accuracy of PDXTs in modeling PDX experiments. The overall area under the curve to distinguish overtly responsive PDXs (relative tumor volume shrinkage after three weeks of treatment > 50%) from those that remained stable or progressed while on treatment was 0.81 (Fig. 6C). Based on this ROC analysis, a luminescence ratio of 0.4 in PDXTs identified 94% of matched PDXs that responded to treatment with tumor regression (Supplementary Data 6). However, the positive predictive value of pharmacologic assays in PDXTs was relatively low (FDR = 0.6), confirming the importance of model-matched in vivo validation during the preclinical phases of drug development. These considerations illustrate the merit of assessing the efficacy of a specific drug in a vast collection of tumoroids to advise the rational selection of models for PDX experiments.Gene editing recapitulates the outcome of pharmacologic target inhibition in PDXTsCancer dependency maps, obtained by perturbing genes with RNA interference or gene editing technologies, have provided a catalog of tumor vulnerabilities with potential clinical actionability49,50. These efforts have been traditionally pursued in immortalized cancer cell lines, but there is now increasing recognition that functional genomics screens in tumoroids would be better representative of cancer biology and diversity51. With this in mind, we sought to explore whether genetic versus pharmacologic inhibition of an index cancer dependency gene results in similar or different effects on PDXT viability. Given the large number of models with known response to cetuximab, EGFR was selected as a target, and CRISPR-Cas9 technology was employed to systematically disrupt the EGFR gene in 13 representative PDXTs with variable sensitivity to cetuximab and proven amenability to lentiviral transduction.Two different sgRNAs targeting EGFR in exon three were independently transduced into Cas9-expressing PDXTs (Supplementary Figs. 14 and 15A, B). Seven days after infection, tumoroids were processed for luminescence-based detection of ATP content. A knockout (KO) score was calculated by intra-model normalization of the viability outputs of EGFR-edited PDXTs to conditions of negligible influence on cell fitness (deletion of a neutral/non-essential gene) or strong influence (deletion of a lethal/essential gene) (see “Methods” section). The consequences of EGFR deletion on PDXT viability were similar for the two sgRNAs (Pearson coefficient, 0.84, P = 3.3e-4) (Supplementary Fig. 15C), supporting robustness and reproducibility of the dataset. Remarkably, the overall outcome of EGFR genetic ablation was significantly correlated to that of cetuximab treatment (Pearson coefficient, 0.78, P = 0.0016), and in some cases a direct quantitative correspondence between the extent of pharmacologic sensitivity and the impact of gene deletion could be observed (Fig. 6D); for example, a RAS wild-type model that proved to be highly refractory to cetuximab treatment was also poorly impacted by EGFR disruption (CRC0151); in a complementary fashion, EGFR KO was severely detrimental in an ERBB2-amplified PDXT that was also particularly sensitive to cetuximab (CRC0080) (Fig. 6D). Collectively, these findings underscore the power and reliability of using genetic approaches in tumoroids for preclinical characterization of drug targets and to interrogate the effects of loss-of-function alterations in cancer-relevant genes.An in silico, ex vivo and in vivo funneling approach identifies actionable co-dependencies that attenuate response to cetuximabThe concordance of molecular profiles and therapeutic annotation in matched PDXTs and PDXs prompted us to embark on a discovery effort aimed to identify adaptive dependencies in models that were sensitive to, but not eradicated by, EGFR inhibition. To do so, we followed a principled approach meant to triage candidate vulnerabilities using sequential selection bottlenecks, with the final aim to nominate only those targets that passed strict validation criteria.We started by analyzing transcriptional responses to drug pressure in cetuximab-sensitive models (33 PDXs and 12 PDXTs) treated with the antibody, with the assumption that some upregulated gene products may adaptively convey compensatory signals to contrast EGFR inhibition. Similar to basal (pre-treatment) profiles, also on-treatment gene expression changes were coherent in PDXTs and PDXs (Pearson coefficient, 0.8; P < 2.2e-308) (Fig. 7A). The list of genes that were upregulated by cetuximab in all PDXs and PDXTs examined was refined by removing low-expressed ones and those that were also modulated by treatment in an independent set of 21 cetuximab-resistant PDXs. The resulting compendium of 1916 genes was restricted to a subset of 119 genes encoding druggable targets, based on a two-tiered selection strategy: (i) a target tractability assessment using published criteria50, such as the availability of compounds in clinical or preclinical development and/or the documentation of an associated response biomarker; (ii) data output from the Drug-Gene Interaction Database (www.dgidb.org), a web resource that provides information on drug-gene interactions and druggable genes. The 119 gene subset was further narrowed down to 13 candidates through a shortlisting process that considered target conceptual novelty and translational potential as well as known potency and in vivo bioavailability of the corresponding specific drugs (Supplementary Fig. 16 and Supplementary Data 7).Fig. 7: Drug screen in PDXTs.A Correlation of transcript changes between 33 PDXs exposed to cetuximab for three days (n = 28) or six weeks (n = 6) and 12 PDXTs exposed to cetuximab for three days. The scatterplot shows log-fold changes (LogFC) between treated and untreated samples for 19,716 genes. Color shading reflects differential expression P values, obtained with DESeq2. Statistical analysis by two-sided Wald test followed by Benjamini-Hochberg multiple comparison correction. B Maximum inhibition scores for 13 drugs tested in three PDXTs. PDXTs were treated for 48 h (5000 cells/well) in three independent experiments in biological triplicates. Maximum inhibition score was the difference between the viability of untreated cells and that at maximum drug dosage, normalized against the viability of untreated cells and averaged for results of the three independent experiments. Maximum drug dosage: HTH-02-006, 16 μM; vorinostat, 10 μM; SBI-0206965, 20 μM; vismodegib, 50 μM; A922500, 100 μM; MRX-2483, 1 μM; BP-1-102, 1 μM; OSMI-4, 20 μM; crenigacestat, 5 μM; pemigatinib, 1 μM; tomivosertib, 5 μM; TG003, 50 μM; cilastatin, 0.5 μM. Drug targets are specified. C Cell viability in 12 PDXTs treated for 3 weeks (CRC0059 and CRC0322, 1000 cells/well) or 1 week (all other models) with SBI-0206965 (10 µM), HTH-02-006 (4 µM), and vorinostat (1.25 µM), alone or with cetuximab (20 µg/ml, one-week treatments; 5 µg/ml, three-week treatments). Two independent experiments in biological triplicates were performed. Heatmap signals were normalized to the sum of the values of the corresponding experiments and reported as a fraction of the maximum value reached by single models. Statistical analysis by repeated measures one-way ANOVA followed by Šídák’s multiple comparison test using the aggregated average value of replicates for each PDXT model (n = 12): cetuximab versus cetuximab + SBI-0206965, P < 0.0001; SBI-0206965 versus cetuximab + SBI-0206965, P = 0.0021; cetuximab versus cetuximab + HTH-02-006, P < 0.0001; HTH-02-006 versus cetuximab + HTH-02-006, P = 0.0027; cetuximab versus cetuximab + vorinostat, P < 0.0001; vorinostat versus cetuximab + vorinostat, P = 0.0006. Cet cetuximab, HTH HTH-02-006, SBI SBI-0206965, Vori vorinostat. Source data are provided as a Source Data file.To preliminarily assess whether the prioritized hits were valuable therapeutic targets, we performed short-term (48 h) viability assays using well-characterized chemical inhibitors of the 13 candidates. Each inhibitor was tested in three PDXTs in the absence or presence of cetuximab, with model selection guided by robust cetuximab-induced overexpression of the targeted gene (in most cases, one or more models were used to test more than one drug) (Supplementary Data 8). The compounds were evaluated in a four-point dose-response assay with biological triplicates and in three independent experiments, using the conditions that proved to be the most accurate in recapitulating response in vivo (5000 cells/well in a 96-well format and luminescence-based detection of ATP content as viability readout), for a total of 3510 measurements (Supplementary Fig. 17). Drug doses were calibrated using literature data and publicly available pharmacologic profiles52, with a maximum dose for each inhibitor equal to twice the standard IC50 value. From this first set of assays, three compounds stood out for having the highest maximum inhibition score as single agents (i.e., their maximum dose had the strongest impact on cell viability) (Fig. 7B and Supplementary Data 9): HTH-02-006, targeting NUAK, a member of the AMPK subfamily of serine/threonine protein kinases; vorinostat, targeting histone deacetylases (HDACs); and SBI-0206965, targeting the serine/threonine kinase ULK1. Interestingly, these targets are variably involved in promoting autophagy, a homeostatic mechanism triggered by cellular stress53.Blockade of the 13 candidate targets in combination with cetuximab incrementally reduced cell viability, again with HTH-02-006, vorinostat and SBI-0206965 showing the highest maximum inhibition scores (Supplementary Fig. 17 and Supplementary Data 9). We then explored the interaction between cetuximab and these three top inhibitors under conditions of longer drug exposure to favor the implementation of adaptive reactions over time. This analysis was extended to the full set of 12 cetuximab-sensitive PDXTs with available gene expression data. In all tested models, combination therapy with cetuximab and either HTH-02-006, vorinostat or SBI-0206965 was significantly more effective than single-agent treatments and outperformed cetuximab monotherapy (Fig. 7C).Finally, pharmacologic experiments were translated into the in vivo setting. PDX model CRC0322 was chosen for administration of HTH-02-006 and SBI-0206965 because the corresponding PDXT was particularly sensitive to both drugs; following a similar reasoning, PDX model CRC1331 was chosen because the matched PDXT displayed the highest maximum inhibition score for both vorinostat and HTH-02-006 (Fig. 7B). As expected, both PDX models were strong responders to cetuximab monotherapy (Fig. 8). While single agent-treatment with HTH-02-006, vorinostat or SBI-0206965 had negligible or null effect on tumor growth (Supplementary Fig. 18), the combination of HTH-02-006 or SBI-0206965 with cetuximab proved to be more effective than cetuximab alone in reducing tumor size in CRC0322 PDXs (Fig. 8). In CRC1331, combination therapy of cetuximab with either vorinostat or HTH-02-006 did not significantly outperform cetuximab monotherapy (Fig. 8); however, it is worth noting that response to cetuximab was exceptionally profound in this PDX (mean tumor volume decrease, 93.77%), which may have masked the contribution of the other drug to tumor regression. In essence, these results emphasize the value of integrative molecular and biological data, coupled with vast availability of experimental models, for ‘aggressive’ prioritization of targets with meaningful translational potential.Fig. 8: Drug screen validation in representative PDXs.Tumor volume changes in PDXs implanted in female NOD-SCID mice and exposed to the indicated modalities for 4 weeks. Cetuximab, 20 mg/kg (intraperitoneal injection twice a week); HTH-02-006, 10 mg/kg (intraperitoneal injection twice a day); SBI-0206965, 20 mg/kg (intraperitoneal injection three times a week); vorinostat, 50 mg/kg (intraperitoneal injection three times a week). Dots represent volume changes of PDXs from individual mice, and plots show the means ± SD for each treatment arm. n = 6 for CRC0322 and CRC1331 models exposed to HTH-02-006 + cetuximab; n = 9 for CRC0322 model exposed to cetuximab or SBI-0206965 + cetuximab; n = 10 for CRC1331 model exposed to cetuximab or vorinostat + cetuximab. Tumor volume changes of the placebo arm are shown in Supplementary Fig. 18. Statistical analysis by two-tailed unpaired t test with Welch’s correction. Cet cetuximab, HTH HTH-02-006, SBI SBI-0206965, Vori vorinostat. Source data are provided as a Source Data file.
