Pharmaco-proteogenomic characterization of liver cancer organoids for precision oncology

Abstract

Organoid models have the potential to recapitulate the biological and pharmacotypic features of parental tumors. Nevertheless, integrative pharmaco-proteogenomics analysis for drug response features and biomarker investigation for precision therapy of liver cancer patients are still lacking. We established a patient-derived LIver Cancer Organoid Biobank (LICOB) comprehensively representing the histological and molecular characteristics of various liver cancer types determined by multi-omics profiling including genomic, epigenomic, transcriptomic, and proteomic analysis. Proteogenomic profiling of LICOB identified proliferation and metabolism subtypes that linked to patient prognosis. High-throughput drug screening revealed distinct response patterns of each subtype that are precisely associated with specific multi-omics signatures. Through integrative analyses of LICOB pharmaco-proteogenomics data, we identified molecular features contributing to drug responses and potential drug combinations for personalized patient treatment. In lenvatinib-resistant organoids, the complementary effect of mTOR inhibitor temsirolimus was proved by treatment modeling and phosphoproteomic analysis. We also provided a user-friendly web portal to help serve the biomedical research community. Our study generated a rich resource to investigate liver cancer biology and pharmacological dependencies, thereby enabling a broad range of biomedical applications including functional precision medicine.

One Sentence Summary:

Integrative pharmaco-proteogenomic profiling of liver cancer organoids identified therapeutic response signatures and predicted drug combinations.

INTRODUCTION

Primary liver cancer is the third leading cause of cancer death worldwide, with hepatocellular carcinoma (HCC) accounting for ~90% of them, followed by intrahepatic cholangiocarcinoma (ICC) and combined hepatocellular-cholangiocarcinoma (CHC) (1). Most liver cancer patients are diagnosed at advanced stages when effective anti-cancer drugs are needed. Due to the high heterogeneity and complexity, most of the drugs targeting liver cancer failed in clinical trials (2, 3). Currently, the majority of approved drugs for HCC are multi-kinase inhibitors, including sorafenib and lenvatinib as the first-line with regorafenib and cabozantinib as the second-line (3). However, the low clinical response rates to these targeted drugs and the inability of genomics to predict sensitive patients make optimal patient selection a major challenge. Systematic characterization of drug response by multi-level genotype-to-phenotype data is supposed to help implement precision therapies for liver cancer.

Inter- and intra-tumor heterogeneity represent a key obstacle in cancer precision therapy. The high heterogeneity between and within hepatic tumors largely determines the sensitivity or resistance of patients, and also exacerbates the failure rate of clinical trials (4). It is thus crucial to develop pre-clinical models which can maintain tumor heterogeneity and in vivo characteristics for efficient drug evaluation and biomarker development. The emergence of organoid models has brought breakthroughs to dissecting tumor heterogeneity and precision oncology. Compared with cancer cell lines of monoclonal origin, organoids consist of different cancer cell clones, and can stably retain the (epi)genetic, transcriptomic, morphological, and pharmacotypic features of the parent cancer (5, 6). Yet, due to the difficulty in establishing liver cancer organoids, especially HCC organoids, which shows a much lower success rate (~ 30%) than other cancer types (7–11), it still remains a major hurdle to establish a large-scale patient-derived liver cancer organoid biobank. Despite the limited number of hepatobiliary cancer organoids in published studies, their efficacy for drug testing and their similarity with original cancer tissues were well-documented (7–11). Nevertheless, there are currently no comprehensive in vitro models that can fully recapitulate the proteogenomic features of liver cancer tissues. Furthermore, systemic multi-omic characteristics and artificial intelligence-facilitated drug screening in liver cancer models are still lacking.

Here, we established a biobank of 65 human liver cancer organoids, which was demonstrated to well preserve the molecular and phenotypical characteristics of parental cancer tissues. We performed comprehensive pharmaco-proteogenomic profiling of the liver cancer organoids by screening a series of drugs used in the clinic or under development, and proteogenomics-based feature modeling achieved high prediction accuracy for drug responses. Particularly, we also provided potential drug combination therapies by utilizing the proteogenomic data and drug response datasets, providing guidance for clinical patient selection and drug combination therapy.

RESULTS

Organoids recapitulate histological and molecular characteristics of primary cancer tissues

We collected surgically resected primary liver cancer tissues to generate a LIver Cancer Organoid Biobank (LICOB) (Fig. 1A). The 6–12 passages of the successfully generated organoids were used for further analysis. For each organoid, the culture interval used in multi-omics analyses and drug screening experiments did not exceed 1–2 passages. Totally, 65 organoids were established from 57 patients, including HCCO (n = 44), ICCO (n = 12), CHCO (n = 4) and hepatoblastoma organoids (HBO) (n = 5). Among them, 40 (62%) were HBV positive, and 13 (20%) were derived from 2 to 4 tumor subregions of the same patients (Fig. 1B and table S1). These organoids showed cancer type-specific morphologies (Fig. 1C), and recapitulated the histological features of the original tissues and cancer types (Fig. 1, D and F), as previously reported (7–11). Likewise, HCCOs, ICCOs and CHCO/HBO expressed HCC-specific, ICC-specific and both markers respectively (Fig. 1, E and F). The average doubling time of LICOB was significantly longer than that of liver cancer cell lines in the “LIMORE” study (12) (fig. S1A). In immune-deficient mice, about 66% of LICOB gave rise to xenograft tumors within 10 weeks with similar histopathology to the corresponding cancer tissues (fig. S1, B and C). We performed proteogenomic profiling of these organoids and parental tissues tissues using reduced-representation bisulfite sequencing (RRBS), whole-exome sequencing (WES), RNA-seq, and label-free proteomics, and then conducted high-throughput drug screen (Fig. 1A).

Fig. 1. LICOB Establishment and Comparison with Primary Liver Cancers.

(A) Schematic workflow of pharmaco-proteogenomic analysis of liver cancer organoids. (B) Numbers of organoid models in LICOB. (C) Representative brightfield images of LICOB (Scale bar: 100 μm). (D) Representative H&E staining of organoids, xenografts and their original tumors (Scale bar: 50 μm). (E and F) Immunofluorescence (E) and immunohistochemistry (F) analyses for the indicated markers in LICOB (Scale bar: 50 μm). (G) Mutational landscape of LICOB and paired cancer tissues (n = 48 pairs). (H) Spearman correlation coefficients of organoids with paired or unpaired tissues across different omics. (CNV: 48 pairs, RNA-seq: 43 pairs, RRBS: 30 pairs, Proteomics: 22 pairs) (Wilcoxon rank-sum test).

First, we compared the somatic mutations, copy number variations (CNVs), DNA methylation, transcriptomic and proteomic profiles between organoids and paired cancer tissues (table S1). The tumor purity estimated from the WES data for the organoids approached 100%, which was significantly higher than that for the corresponding cancer tissues (fig. S1D). The organoids and their parental tissues showed similar mutation frequencies of genes, including recurrently mutated genes in both HCC and ICC including TP53, KMT2C, RB1, and PBRM1, HCC-specific mutated gene CTNNB1, and ICC-specific mutated genes KRAS and BAP1 (Fig. 1G), consistent mutation clusters by clonal analysis, and equal allele frequencies for most genes (fig. S1E and table S2). Importantly, organoids derived from the same patient displayed homologous clonal mutations and subclonal architecture (fig. S1F), indicating that the organoids could capture the intratumor multi-clonal diversity of liver cancer. Unsurprisingly, the organoids accumulated a few mutations that were underrepresented in cancer tissues (fig. S1G), including SCAF11, TUBGCP6, RBM33, PLEC (fig. S1H), and genes involved in AMPK and mTOR pathways (fig. S1E), suggesting the in vitro selection for tumor subclones during organoid culture. For CNV, DNA methylation, mRNA, and protein profiles, the correlation coefficients for the paired organoid-tissue samples were significantly higher than those of unpaired samples (Fig. 1H and fig. S1I). The organoids clustered with liver cancer samples from TCGA, CPTAC and paired tissues in unsupervised analysis of transcriptomic profiles, whereas the cancer cell lines from LIMORE formed a different cluster (fig. S1, J and K), indicating LICOB organoids better represent liver cancers than cell lines. Functional enrichment analysis demonstrated that genes associated with metabolism pathways in liver cancers were well persevered in LICOB, while immune and inflammatory response pathways were down-regulated due to the lack of microenvironmental cells in the organoids (fig. S1L). Taken together, these results indicated that LICOB recapitulated the histological and molecular features of the original cancer tissues and may serve as reliable models.

We next explored whether LICOB could represent the inter-tumor heterogeneity of liver cancer. Likewise, HCCOs and ICCOs showed cancer type-specific mutations and copy number alterations (fig. S2, A and B). It is worth noting that the mutation frequency of TP53 was higher and CTNNB1 was lower in LICOB models compared with TCGA primary HCC tissues, possibly due to the in vitro culture selection and the fact that most LICOB models were derived from HBV-positive liver cancers which showed higher TP53 and lower CTNNB1 mutation rate. Moreover, the driver gene mutations, CNV, and mRNA profiles were highly consistent in organoids at early and late passages, indicating the genetic stability during long-term in vitro culture (fig. S2, C and D). The multi-omics profiles showed higher similarity in the organoids from the same individuals (fig. S2E). ICCOs showed distinguished multi-omic features from other cancer types, with preferential expression of different liver cancer signatures among them (fig. S2, F and G). These results demonstrated that LICOB models preserved intrinsic molecular traits and diversity of different liver cancer types.

Multi-omic classification of LICOB identifies four subtypes

We performed consensus clustering for the LICOB based on the five omics data, and identified 4 subtypes with distinct molecular characteristics (Fig. 2A). Notably, all the ICCOs were classified into Cluster 3 with elevated RAS signaling and cell junction (Fig. 2, A and B), implying an ICC-dominant feature of Cluster 3, which was thus denoted as L-ICC. Cluster 2 were predominant in cell cycle and MAPK pathways, and thus given the name of L-PL to denote proliferation subtype (Fig. 2, A and B). DNA methylation levels of pluripotent-related genes were significantly lower, and cancer stem cell marker CD44 was upregulated in L-PL (Fig. 2, A and B), with significantly shorter doubling time of L-PL than other subtypes (fig. S2H). Cluster 1 and Cluster 4 enriched lipid-metabolism and drug-metabolism pathways respectively, and hence were labeled as L-LM and L-DM respectively (Fig. 2, A and B). Consistently, L-LM showed accumulated mutations in APOB and increased mRNA expression of CPS1 and PLA2G2A, while L-DM showed higher mutations in ALB, PLEC and elevated protein expression of UGT1A1 and ALDH1A1 (Fig. 2A). Apart from ICCOs, the subtypes identified in LICOB dataset were generally consistent with previous HCC tissue-based clustering results (Fig. 2C). For example, L-PL matched well with the proliferation subtype from Gao’s and Jiang’s studies and Hoshida S1 cluster, L-LM mostly overlapped with the metabolism subtype, whereas L-DM was closer to the intermediate subtype (13–17) (Fig. 2C). The molecular features were further confirmed by the principal component analysis (PCA) and pathway enrichment analysis of the three subtypes, among which L-LM and L-PL were well defined by PC1, while L-DM were distinguished by PC2 plus PC3 (fig. S2I). PC1-negatively and -positively correlated proteins were enriched in lipid homeostasis and amino acid catabolic process versus cell adhesion regulation and interferon signaling respectively, in concordance with the multi-omic features of L-LM and L-PL (fig. S2, J and K). Proteins negatively correlated to PC2 plus PC3 were enriched in xenobiotics metabolism by cytochrome P450 and glutathione metabolism, corresponding to L-DM specific features (fig. S2, J and K).

Fig. 2. Molecular Subtypes of LICOB.

(A) Consensus clustering based on multi-omics data revealed four subtypes. Each column represents an organoid sample and rows indicate molecular features. (B) ssGSEA scores of selected gene sets in each LICOB multi-omics cluster and each type of omics data (*FDR < 0.05, Wilcoxon rank-sum test). (C) Comparisons of the HCCO subtypes in LICOB with previous subtyping results from HCC patient tissues (Fisher exact test). (D) Kaplan-Meier curves for overall survival in CPTAC HCC cohort clustered according to the multi-omics signatures of organoid subtypes (Log-rank test). Pathways were enriched by GSEA for cancer hallmarks. (E) Quadrant plot depicting the alteration of 6,729 genes simultaneously detected by transcriptome and proteome in L-LM compared with L-DM. (F) The GO enrichment analysis of differentially expressed genes between L-LM and L-DM.

We conducted HCC tissue-based feature scoring between organoid models (LICOB) and cell line models (LIMORE) (12), and found that compared to the evenly distributed metabolism (S-Mb) and proliferative (S-Pf) features in LICOB, LIMORE mainly manifested a proliferation (S-Pf)-dominated transcriptomic feature (fig. S2L). Evaluation of LIMORE cohort using multi-omic features derived from LICOB cluster revealed that the proliferative feature (L-PL) score was significantly higher than metabolic (L-LM and L-DM) scores in LIMORE cohort (fig. S2M), demonstrating a more pronounced proliferative feature in LIMORE cohort rather than metabolic characteristics in LICOB cohort. These results suggest that organoids can better simulate the molecular subtyping of liver cancer tissues.

To further explore the prognostic value of the LICOB subtypes, we divided 159 HCC patients from CPTAC cohort (16) into three subgroups using the proteogenomic features of LICOB clustering, and found L-PL had the worst survival, while L-LM showed the best prognosis, sticking to the original study that proliferation and metabolism subtypes harbored the worst and best survival respectively (Fig. 2D). L-DM with upregulated drug metabolism showed an intermediate survival inferior to L-LM (Fig. 2D), possibly owing to the different metabolism flux between the two subtypes. Enrichment of differentially expressed genes demonstrated up-regulation of glycolysis and lipid metabolic pathways in L-LM, as compared with pentose phosphate metabolism and glutathione metabolism in L-DM (Fig. 2, E and F, and table S3 and S4). The unique metabolic program may contribute to the worse prognosis of patients with L-DM features, and the discovery of specific drug targets will benefit such patients.

G6PD mediates metabolic reprogramming in L-DM subtype

To explore the potential therapeutic targets, we performed differential protein expression analysis of previously reported druggable targets in HCC (16) and found that glucose-6-phosphate dehydrogenase (G6PD) was significantly upregulated in L-DM compared with the other subtypes (Fig. 3, A and B). The mRNA and protein expressions of G6PD were highly correlated in both LICOB and CPTAC HCC datasets (fig. S3A). Patients with higher G6PD expression levels in the CPTAC and TCGA HCC datasets showed significantly worse survival than those with lower levels (Fig. 3C and fig. S3B), ascertaining G6PD as a tumor-promoting factor (18).

Fig. 3. G6PD as a potential drug target in L-DM.

(A) Differential protein expression of the potential drug targets in LICOB subtypes. (B) The protein and mRNA expression of G6PD between L-LM and L-DM (Wilcoxon rank-sum test). (C) Kaplan-Meier curves for overall survival based on the lower and upper quantile expression of G6PD protein in the CPTAC HCC dataset (Log-rank test). (D) ssGSEA scores of indicated pathways among LICOB clusters. (E) ssGSEA scores of glycolysis pathways. (F) mRNA expression of PPP and glycolysis related genes in each LICOB sample. (G) Pearson correlations between the expression of G6PD with indicated protein. (H) Western blot analysis of G6PD in HCCO8 and HCCO12 from L-DM with G6PD silencing (sh-G6PD) or control (sh-NT). (I) Proliferation of indicated organoids (Two-way ANOVA). (J) Areas of indicated organoids (Student’s t-test). Organoid numbers (shNT vs sh-G6PD): HCCO8 (n = 1049 vs n = 3496); HCCO12 (n = 476 vs n = 897). (K) Boxplots demonstrating indicated metabolites in indicated organoids (Student’s t-test). (L) Dose-response curves of G6PD inhibition for organoids from L-LM and L-DM (Wilcoxon rank-sum test). (M) Selected multi-omic features among LICOB subtypes. (N) Spearman correlations between MYC CNV and expression of proteins from glycolysis (blue dots) and PPP pathways (red dots) with significantly correlated genes labeled. (O) Schematic diagram depicting the function of G6PD in metabolic flux shift in L-DM.

G6PD is a rate-limiting enzyme in the pentose phosphate pathway (PPP) parallel to glycolysis, and is essential to generate NADPH for redox regulation and providing substrates for nucleotide synthesis (19). The ssGSEA analysis showed drug/xenobiotic metabolism, nucleotide metabolism pathway, and biological oxidations were enhanced in L-DM, while glycolysis was downregulated and hypermethylated (Fig. 3, D and E, and fig. S3C). Thus, the abnormally high expression of G6PD might rewire metabolism from glycolysis to the PPP in L-DM. The PPP branches from glucose 6-phosphate (G6P), which is irreversibly catalyzed by G6PD, and produces NADPH to generate reduced forms of antioxidant molecules like reduced glutathione (GSH) (20). Targeted metabolomics analysis showed that NADPH and GSH were higher in L-DM than in L-LM (fig. S3D), suggesting a strong anti-oxidative stress response in L-DM. While, the precursor G6P and the glycolytic metabolite fructose 6-phosphate were higher in L-LM (fig. S3D). Similar to the transcriptomic and proteomic profiling (Fig. 2F), the metabolite enrichment analysis revealed enhanced glycolysis and gluconeogenesis in L-LM compared to glutathione and nucleotide sugars metabolism in L-DM (fig. S3E). Indeed, key enzymes in the PPP pathway like PGD, TKT and TALDO1 were upregulated in L-DM and showed significantly positive correlations with G6PD protein expression, while the glycolic enzymes were downregulated in L-DM (Fig. 3, F and G, and fig. S3F). These results hinted that the enhanced PPP flux could provide reducing equivalent and backbones for nucleotide synthesis supporting redox homeostasis and cell growth.

Therefore, we asked whether inhibiting G6PD could exert a tumor-suppressing role in L-DM. The proliferation of HCCO8 and HCCO12 within L-DM was strongly suppressed upon G6PD knockdown (Fig. 3, H and I). During long term culture, their 3D growth was also blocked by G6PD knockdown, as shown by the decreased average size of organoids and significantly increased ratios of organoids with small area (Fig. 3J, fig. S3, G and H). Unsurprisingly, G6PD knockdown significantly reduced NADPH and GSH production (Fig. 3K), suggesting that proliferation inhibition may result from redox homeostasis dysregulation. Compared to organoids from L-LM, the organoids in L-DM were more sensitive to the selective G6PD inhibitor G6PDi-1 (21) (Fig. 3L), confirming the therapeutic potential of targeting G6PD.

Meanwhile, MYC amplification was also a key feature in L-DM (Fig. 2A and 3M), with simultaneously enhanced MYC and G6PD expression and positive correlations between them (fig. S3, I and J). MYC amplification positively and negatively correlated with expression of proteins in the PPP pathway and glycolysis pathway respectively (Fig. 3N). In concordance with the elevated NRF2 pathway in L-DM (Fig. 2F), the expression of NRF2, a G6PD transcription activator (22, 23), was the highest in L-DM among the LICOB dataset (fig. S3K). NRF2, an essential regulator of antioxidant response that drives cancer progression, can be upregulated by MYC (24). Therefore, MYC amplification might activate the transcription of G6PD by NRF2 and then support cell growth and antioxidant defense through enhancement of PPP flux to potentiate survival and progression under oxidative stress (Fig. 3O). These results affirmed the therapeutic value of G6PD and highlighted the feasibility of multi-omics data-driven therapeutic discovery.

Heterogeneity of drug response in LICOB correlates with molecular subtypes

We next performed high-throughput drug screening of 76 drugs using the LICOB cohort (Fig. 4A, fig. S4A, and table S5). Drug responses of organoids were determined by calculating area under the curve (AUC), half-maximal inhibitory concentration (IC50), and maximal effect level (Emax) (25) (tables S6–S8), which showed good concordance with each other (fig. S4B). High reproducibility between experiments was also observed from the biological replicates of response to sorafenib in randomly selected organoids (fig. S4C).

Fig. 4. Heterogeneous Drug Response in LICOB.

(A) Mechanisms of action for the 76 drugs. (B) Average AUC values between organoids from HBV positive and HBV negative patients of each drug category (Wilcoxon rank-sum test). (C) Representative scatterplot of AUC values of drugs with shared molecular targets (Pearson correlation test). (D) Average AUC values among the four subtypes for the 76 drugs. The AUC values were Z-score normalized by row. (E) Comparisons of AUC values for the chemotherapy drugs among the four subtypes (Student’s t-test). (F) ssGSEA scores for Reactome Cell Cycle Mitotic pathway among the four subtypes (Student’s t-test). (G) Multi-omic features associated with the responses of FGFR inhibitor (BGJ398 and PD173074) or MET inhibitor (tivantinib). The bar plot showing the Spearman correlation between the AUC of the drug and the feature values in each cluster. For BGJ398 and PD173074, the correlation coefficients were averaged. (H) The distribution of mean AUC values between S-ICCOs and R-ICCOs (Wilcoxon rank-sum test). (I) Kaplan-Meier curves for overall survival in CPTAC cohort based on the signatures of L-ICC clusters (Log-rank test). (J) Comparisons of R-ICCOs and S-ICCOs clusters in L-ICC using representative pathways reported in CPTAC cohort. (K) Multi-omic features associated with the responses to afatinib. (L) ssGSEA scores of indicated gene sets in S-ICCOs or R-ICCOs based on the proteomics data.

The drug response patterns were diverse among LICOB models (fig. S4, D and E), in which HBV-positive organoids were generally more resistant to chemotherapy than HBV-negative ones (Fig. 4B). Most of the organoids showed similar responses to drugs against the same targets, such as TOP2 inhibitors doxorubicin and epirubicin (R = 0.91, P < 2.2e-16), HDAC inhibitors belinostat and vorinostat (R = 0.8, P = 5.5e-15), and BET PROTAC inhibitors dBET6 and ARV-771 (R = 0.77, P = 1.5e-13) (Fig. 4C). However, several agents targeting the same pathway displayed apparent inconsistency, such as PARP inhibitors olaparib and talazoparib (R = 0.38, P = 0.019) (Fig. 4C). Almost all organoids showed resistance to olaparib, whereas several HBOs and ICCOs were sensitive to talazoparib (Fig. 4C and fig. S4, D and E), suggesting the necessity to compare the efficacy and underlying mechanisms of different drugs against the same targets. To explore whether the LICOB pharmaco-proteogenomic data is consistent with the well-studied drug resistance mechanism, we investigated lenvatinib response and EGFR-related pathways which conferred resistance to lenvatinib treatment in HCC (26, 27). We found lenvatinib resistance (high AUC) was positively correlated with high expression levels of proteins involved in EGFR tyrosine kinase inhibitor resistance pathway and negatively correlated with DNA methylation of associated genes (fig. S4F). More interestingly, organoids with higher protein expression or lower DNA methyaltion of EGFR tyrosine kinase inhibitor resistance pathway were more resistance to lenvatinib as evidenced by higher AUC (fig. S4G). These data collectively suggested LICOB as a powerful model to study the drug resistance mechanism.

Distinct drug response patterns were observed among the 4 LICOB subtypes (Fig. 4D). BEZ235 and Temsirolimus, inhibitors targeting PI3K-AKT-mTOR pathway, were sensitive in proliferation subtype L-PL and resistant in the metabolism subtypes L-LM and L-DM (fig. S4H). Specifically, L-PL showed the highest sensitivity to chemotherapy drugs (Fig. 4, D and E), with the uppermost activity of the cell cycle pathway than the other subtypes (Fig. 4F). L-LM and L-DM organoids showed more similar responses to the tested compounds. For example, they were generally more sensitive to the multi-kinase inhibitors than the other subtypes, where L-LM and L-DM organoids showed preference to regorafenib and lenvatinib respectively (fig. S4I). Likewise, both L-LM and L-DM showed higher sensitivity to tivantinib (Fig. 4D). Tivantinib is a selective c-MET inhibitor showing promising effects on HCC in a phase 2 trial (28). We did not observe mutations, amplifications, or elevated mRNA and protein expressions of MET in L-LM and L-DM (fig. S4, J and K). However, genes required for MET activation showed higher CNV, elevated mRNA and protein expressions in both subtypes (Fig. 4G), including MET ligand HGF and its activator HPN (29), USP8 that stabilizes MET by deubiquitination (30), DOCK7 that are required for MET activation (31), and downstream effector CRKL that bridges signal transduction to intracellular pathways (32). Thus, tivantinib efficacy was possibly associated with the upregulation of genes required for MET activation and functioning rather than MET expression. Indeed, two phase 3 studies have reported no significant efficacy of tivantinib in HCC patients with highly expressed c-MET (33, 34), indicating a multi-omics biomarker-driven patient selection may help identify tivantinib-sensitive subpopulations. Not surprisingly, several inhibitors showed different efficacies between L-LM and L-DM (Fig. 4D). L-DM showed the highest sensitivity to FGFR inhibitors BGJ398 and PD173074, and proteogenomic analysis revealed the activation of FGFR pathway by the corresponding gene hypomethylation, amplification, and elevated mRNA and protein expression in this subtype (Fig. 4D and 4G).

Notably, L-ICC organoids were resistant to tyrosine kinase inhibitors compared to the HCC subtypes (fig. S4I), highlighting distinct drug response patterns between HCC and ICC. The drug responses within L-ICC were also heterogeneous, where ICCO2, ICCO5, ICCO10, ICCO11, and ICCO12 were generally more resistant than others based on the sensitivity score and were thus named R-ICCOs, with the others denoted as S-ICCOs (Fig. 4H and fig. S4D). We clustered the CPTAC ICC patients (35) according to the proteogenomic features of R-ICCOs and S-ICCOs and identified two subgroups with significantly different survival (Fig. 4I). Likewise, by comparing the proteomic features of L-ICC organoids to ICC tissues from the CPTAC cohort at the pathway level, we found that R-ICCOs were close to the original poorest survival subtype (S1) (Fig. 4J). Proteogenomic analysis indicated that the EGFR-TKI-resistant features including TP53 and KRAS mutations, ERBB2 and MAP3K3 amplifications, enhanced mRNA expression of PTPN11 and MAP3K2, and interferon signaling pathways were enriched in R-ICCOs (Fig. 4, K and L). Together, these results demonstrated the heterogeneous drug responses in LICOB were associated with specific multi-omics signatures.

Predictive modeling of drug responses with proteogenomic data

To identify robust molecular features contributing to drug responses, we constructed elastic net regression models using multi-omics data (fig. S5A). The performance of our prediction model was evaluated by Pearson’s correlation coefficient, cosine similarity, and mean square error (MSE) between predicted AUCs and measured AUCs in the testing dataset (Fig. 5, A and B, fig. S5B, and table S9).

Fig. 5. Pharmaco-proteogenomic Analysis in LICOB.

(A) Circos plot showing the Elastic net prediction model for each of 76 drugs. The first inner circle represents the sensitivity of each LICOB sample, blue and pink colors represent the top sensitive and resistant samples. The second inner circle indicates AUC values for each sample. The third inner circle represent the multi-omics or single-omic data types showing the best prediction accuracy, and the outermost circle indicate the number of features in each prediction model. (B) Comparisons of drug prediction accuracies evaluated by cosine similarity and MSE. (C) Pearson correlations between predicted AUC values and measured AUC in the testing set of LICOB for indicated drugs. (D) Multi-omic features with the best prediction accuracy for sorafenib. Barplot on the left shows the coefficients of features in the model. (E) Pathways enriched by multi-omic features representing sorafenib resistance (Hypergeometric test, Benjamini-Hochberg adjusted). (F) The spearman correlation between predicted AUC values and measured AUC for lenvatinib in GDSC. (G) Multi-omic features with the best prediction accuracy for lenvatinib. Barplot on the left shows the coefficients of features in the model. (H and I) Pearson correlations between the activities of indicated pathways and AUC values of LICOB models to lenvatinib.

The predicted features were closely tied to drug action mechanism and would provide new insight to guide clinical usage (table S10). The sensitive features of IGF1R inhibitor OSI-906 were highly enriched in the IGF signaling pathway and p53 pathway (fig. S5, C and D). Drug metabolism CYP450, biological oxidations, and TNF signaling pathway enriched genes were predicted to contribute to OSI-906 resistance (fig. S5, C and D). The sensitivity to PI3K inhibitor BEZ235 was predicted to be correlated with higher insulin receptor substrate 2 (IRS2) and pyruvate dehydrogenase phosphatase 1 (PDP1) protein expression (fig. S5E). IRS2 was reported to activate PI3K and regulate aerobic glycolysis in tumor cells (36), while PDP1 plays a critical role in catalyzing pyruvate metabolism (37). The predicted sensitivity features to BEZ235 were enriched in PI3K-AKT activation pathway and pyruvate dehydrogenase PDH complex, concordant with the action mechanism of BEZ235 and indicating a potential role of glucose homeostasis on PI3K inhibitor response (fig. S5F). We also found that BEZ235 resistance was associated with hypomethylation of ABCA5, a member of the ATP-binding cassette (ABC) transporter responsible for drug efflux, indicating that ABC transporter contributed multidrug resistance (MDR) should be a concern in clinical applications (fig. S5, E and F).

Among the approved liver cancer therapeutics, the second-line drug regorafenib achieved the best prediction results (R = 0.78), followed by the first-line drug sorafenib (R = 0.69) and lenvatinib (R = 0.68) (Fig. 5C). For sorafenib, the features associated with higher sensitivity included hypermethylation of MYEOV, higher mRNA expressions of EFNA2, TINAG, and UPK3A, and decreased IGF1R protein level (Fig. 5D). MYEOV has been proved to promote cancer progression and metastasis (38), while its function in HCC remains unexplored. The resistance features were enriched in tyrosine and intracellular kinase pathways including IGF1/IGF1R, EGFR, FGF11, RAS, and ERK signaling pathways (Fig. 5E). In addition, HDAC pathway and endocytosis were also correlated to sorafenib resistance. We also applied our elastic net model to predict sorafenib responses in GDSC database (39, 40) and found the predicted AUC correlated well with observed ones in HCC cell lines (Fig. 5F). For regorafenib, the sensitive response features were enriched in autophagy which has been proved to be induced by regorafenib through AMPK activation (fig. S5, G and H) (41). The resistant features for regorafenib were enriched in PI3K-AKT pathway which was reported to restrict the antitumor effect of regorafenib (42) (fig. S5, G and H).

Exploration of the predicted features found that lenvatinib resistance positively correlated with higher mRNA expression of KRT19 and NDN and thromboxane pathway, while negatively correlated with higher protein levels of ADH1C and CES1 and VEGF pathway (Fig. 5, G–I). Although all three drugs are multi-targeted tyrosine kinase inhibitors, unlike sorafenib and regorafenib, lenvatinib mainly inhibits angiogenesis by targeting pro-angiogenic molecule VEGFR1–3 and FGFR (3). The lack of endothelial cells may render LICOB less sensitive and harder to predict response to lenvatinib (fig. S4, D and E).

Identification of combination regimens to promote lenvatinib response

Based on somatic mutations or proteomic data, multiple synergistic drug combination opportunities have been calculated and validated in preclinical and clinical studies (43, 44). We here developed a network-based method to predict candidate drug combinations in LICOB cohort using the proteogenomic data (Fig. 6A and fig. S6A). We identified all possible drug combinations for each of the 76 tested drugs (fig. S6B and tables S11–S12), and found that lenvatinib may have better combination effects with YM155 or temsirolimus (Fig. 6B and fig. S6, B and C). Since temsirolimus targeting mTOR has been approved for clinical use in cancer treatment (45), we next focused on its combination with lenvatinib. Temsirolimus and lenvatinib interact through EGFR inhibitor resistance and TGF-beta signaling pathways, where the correlations of these pathway scores with lenvatinib and temsirolimus sensitivity were negative and positive respectively (Fig. 6B and fig. S6C). Likewise, temsirolimus also ranked as the top complementary drug for tivantinib (fig. S6D). These complementary inhibitions can be observed in (epi)genome, CNV, and transcriptome (Fig. 6C and fig. S6C). Consistently, a recent high-throughput drug combination test also proved the high synergy in targeting receptor tyrosine kinase and PI3K-mTOR pathway across breast, colon and pancreatic cancer cell lines (46). We also found that HCC organoids resistant to either lenvatinib or temsirolimus were predicted to be sensitive to their combination (Fig. 6D). Even ICCOs that showed nearly no response to lenvatinib tended to have higher scores for such combination (Fig. 6D).

Fig. 6. Drug Combination Prediction and Validation in LICOB.

(A) The workflow of drug combination prediction. (B) The drug-pathway network map for lenvatinib. Orange nodes represent predicted drugs with combination efficiency with lenvatinib, and node sizes indicate combination scores with lenvatinib. Green nodes indicate pathways connecting each pair of drug combination. (C) The predicted scores for the combination of lenvatinib and other drugs. The dot size represents the score contributed by different omics data. (D) The predicted scores of lenvatinib and temsirolimus combination for all the LICOB samples. (E) Heatmap showing the normalized abundance of phosphorylation sites, referred as phosphorylation level, in organoids treated with indicated drugs. (F) Heatmap showing lenvatinib-temsirolimus combination dose-response matrix in HCCO31. (G) Western blot analysis of the MEK-ERK cascade and AKT activation. (H and I) Drug efficacy on tumor growth in xenograft model of HCCO31 (H) and lenvatinib-resistant PDX models (I) Representative tumor xenograft images at the end of treatment (Scale bars: 1 cm) and tumor growth curves of each group are shown (Data are mean ± s.e.m., Two-way ANOVA).

To validate the prediction, we performed a comparative study by treating four selected organoids showing high predictive scores. Six hours post-treatment, we collected the organoids for global phosphoproteomics profiling including 23,754 quantified phosphosites (table S13). Notably, organoids treated with either lenvatinib or temsirolimus showed independent down-regulation of phosphosites involved in the pathways targeted by either drug (including mTOR, VEGF and PDGF pathways), while their combination led to universal down-regulation of almost all the above phosphosites compared with DMSO (Fig. 6E), indicating the effectiveness of drug combination. Then, in vitro experiments on ICCO10, ICCO2, HCCO3 and HCCO31 validated that lenvatinib alone showed limited effects on cell proliferation inhibition, while the addition of temsirolimus exhibited a strong synergistic inhibition effect (Fig. 6F, fig. S6, E and F). Lenvatinib alone slightly reduced MEK/ERK activation while enhancing the phosphorylation of AKT (Fig. 6G and fig. S6G). The addition of temsirolimus enhanced the inhibition of MEK/ERK and significantly reversed the activation of AKT (Fig. 6G and fig. S6G), further demonstrating a potential synergistic mechanism for their combination. Mice bearing HCCO31 xenograft showed significant growth inhibition when treated with lenvatinib plus temsirolimus than either alone (Fig. 6H). Of note, lenvatinib plus temsirolimus showed significant therapeutic effects in PDX models derived from a lenvatinib-resistant HCC patient (Fig. 6I), indicating the promising synergy of two drugs in vivo. Similarly, we also validated the predicted synergistic effect of lenvatinib and ganetespib in vitro (fig. S6H). Altogether, the predicted drug combinations using pharmaco-proteogenomic data may provide informative evidence for further clinical application.

Finally, to facilitate the utilization of the proteogenomics and pharmacological data in our LICOB cohort to a broad biomedical community, we developed an interactive website (http://cancerdiversity.asia/LICOB/) for rapid and comprehensive data exploration (fig. S7).

DISCUSSION

Proteogenomic characterization of patient tissues has deepened our knowledge of genotype-phenotype correlations in liver cancer, identifying potential therapeutic targets and cancer subtypes (14, 16, 35, 47). Further pharmaco-proteogenomic analysis of optimal pre-clinical models will boost the therapeutic discovery and biomarker identification to fill the gap in clinical translation. Herein, we pioneeringly present the integrative pharmaco-proteogenomic analyses with our newly established liver cancer organoids, which advances predictive modeling toward a systematic representation of the biological mechanisms underlying drug responses and potential drug combinations, a critical direction for functional precision medicine.

To validate our LICOB as a faithful model of liver cancer biology, multi-platform comparisons with the original tissues and public datasets were conducted. In addition to the retention of morphology, multi-omics characteristics, and tumor heterogeneity, the LICOB cohort recapitulated previous cancer tissue-based molecular subtypes, especially the proliferation and metabolism subtypes of HCC as well as the poorest/best prognostic subtypes of ICC (16, 35). In contrast, patient-derived HCC cell lines from LIMORE cohort and a panel of 34 commercial liver cancer cell lines was only similar to the proliferation subtype of HCC (48). Of note, within the LICOB cohort, the previous metabolism subtype of HCC was further sub-classified into L-LM and L-DM with distinct signaling and metabolic traits. L-DM was featured by the metabolic reprogramming from glycolysis to PPP for redox homeostasis maintenance, possibly through the MYC-NRF2-G6PD axis, where G6PD was proposed as an effective therapeutic target. Thus, our LICOB models could recapitulate the intrinsic characteristics of derived cancer tissues, provide new treatment opportunities, and are ideal candidates for in vitro drug screening.

The limits of therapeutic decision making based on mutational profiling alone are well recognized (49), and transcriptomic profiling neither reliably predicts the abundance or functional status of their corresponding proteins. Indeed, we showed that drug responses were determined by integrative proteogenomic features of the organoids, while cell line data in GDSC favored genomic features than gene expression profiles (39), and CPPA claimed an increased predictive power by protein levels (44). This was possibly caused by limited omics data types generated in the previous studies. As for liver cancer, druggable somatic mutations were even rare, especially in HCC, and we also showed that tumors with certain mutations may not respond well to the corresponding drugs. It should be noted that proteomics was limited by the detection rate of current strategies compared with transcriptomic and genome-wide DNA methylation profiling by next generation sequencing (50). Thus, the heterogeneity and complexity of liver cancer require combined multi-omics data for drug selection and biomarker discovery using a large collection of models that optimally mimic human cancer.

We showed that reasonable drug combinations based on baseline pharmaco-proteogenomic data could achieve better efficacy than the single treatment of individual drugs. This combinatorial drug treatment can be especially effective in targeting heterogeneous tumors that harbor subclones bearing different kinds of somatic mutations or aberrant gene expressions. In lenvatinib-resistant organoids, the complementary anti-tumor effect of temsirolimus was proved by calculating combination score, phosphoproteomic analysis, and treatment modeling. Apart from the flexibility in designing drug combinations, our data-driven LICOB models have the potential to incorporate additional levels of molecular information like global phosphoproteomics and longitudinal response signals, as exemplified by the pre- and post-treatment kinase phosphorylation changes on lenvatinib plus temsirolimus. Optimistically, if the model is clinically applied, the patient treatment decision may be refined by fine-tuning the model and taking all relevant information into consideration (51).

The current study has provided a comprehensive pharmaco-proteogenomic landscape based on a total of 65 liver cancer organoids covering four different liver cancer types, but the sample size was limited by the low success rate in liver cancer organoids establishment. Despite of the high similarity with derived tumor tissues, the organoids in our study still lack the tumor microenvironment-associated immune cells and stroma cells. It is imperative to further improve the tissue dissociation conditions and organoid culture system to reserve the entire tumor microenvironment. In addition, the efficacy of drugs targeting cell cycle may be underestimated in the slow-proliferating organoids due to the insufficient drug treatment time in the current study. Further optimization of the drug screening system is required. Moreover, our findings in liver cancer multi-omics subtyping, drug response features, and drug combination regimen need to be further validated in clinical practice.

In summary, our study provided a framework for annotating patient-derived organoids and drug response, complementing the large-scale clinical liver cancer profiling by TCGA, CPTAC and others. We demonstrated that systematic characterization of molecular features together with biological networks could predict drug response with high accuracy and provide potential drug combinations for personalized treatment. Our interactive and user-friendly web portal offers a rich resource for the oncological community to investigate and thus benefit the clinical utility.

MATERIALS AND METHODS

Study design

The objective of this study was to establish a patient-derived liver cancer organoid biobank and provide pharmaco-proteogenomic information for the development of predictive biomarkers and new therapeutic options for liver cancer. To this end, we have established 65 liver cancer organoids, performed multi-omics analyses including mutation, CNV, DNA methylation, transcriptomics, and proteomics, and conducted high-throughput screening of 76 drugs. Based on the pharmaco-proteogenomic analysis, we constructed high-accuracy drug response prediction models for each tested drug. A network-based method was developed to predict potential drug combinations and identify the combination effect of lenvatinib with temsirolimus in inhibiting tumor growth, which was validated by in vitro and in vivo treatment modeling. Mice were randomly assigned to different treatment groups. The number of replications is indicated in each figure legend.

Clinical specimens

Liver cancer specimens were collected from patients underwent surgical resection at Zhongshan Hospital, Fudan University with signed informed consent forms. The clinical information was collected and summarized in table S1. Each tumor tissue sample was split into several parts and subjected to digestion and subsequent culture for organoid generation, DNA and RNA isolation, or histology analysis. The study was approved by the Research Ethics Committee of Zhongshan Hospital.

Liver cancer organoid generation

Liver cancer tissues were minced and digested at 37 °C in phosphate buffer saline (PBS, Gibco) containing Collagenase Type IV for 30–60 min. The suspension was filtered through a 100-μm cell strainer and centrifuged. The pellet was resuspended in cold organoid culture medium (Supplementary Materials) and then mixed 1:2 with Matrigel (Corning) to reach a density of 4000 cells per 50 μL before seeding into a 24-well culture plate. After solidification, organoid culture medium was added to each well and organoids were cultured in a humidified incubator at 37°C with 5% CO₂.

WES, RRBS, RNA-seq, and proteomics analyses

Genomic DNA and total RNA were extracted and subjected to WES, RRBS, and RNA-seq, and cellular proteins were subjected to DIA proteomic analysis. The detailed processes of sample preparation, sequencing, and MS detection are presented in the Supplementary Materials, and the complete matrix format data used in our analysis were provided in biosino NODE database (OEP003191) through the URL: https://www.biosino.org/node/project/detail/OEP003191. Crucially, we generated four LICOB clusters based on five omics data using concensus hierarchical clustering. Pathways representing each cluster were calculated using ssGSEA then statistically tested.

High-throughput drug screening and analysis

The detailed information of the 76 drugs was provided in table S5. Organoids were recovered from Matrigel and then seeded into 384-well plates. After 72-hour incubation, organoids were treated with drugs of serial dilutions. After another 72-hour drug treatment, CellTiter Glo 3D (Promega) was added to each well followed by measuring the luminescent signals to determine the cell viability. Drug curves were fitted using R package GRmetrics, and AUC, IC50 and Emax values were derived from the fitted curves (see Supplementary Materials).

Pharmaco-proteogenomic analysis by machine learning and drug combination prediction

For drug response prediction, we constructed elastic net regression models using multi-omics data with 1,000 times bootstrapping. The whole dataset was split by 8:2 for training (51 organoids) and testing (13 organoids). We then defined signal entropy to assess the activities of pathways, and bridge two drugs holding positive and negative correlations of their AUC values with pathway activities. Considering the possibility that two drugs may be connected through multiple pathways or data levels, a weighting system was determined to finally identify potential drug combinations. The calculation of pathway activity from signal entropy model and detailed weighting algorithm can be found in Supplementary Materials.

Statistics

Standard statistical tests were used to analyze the clinical data and functional experiment results including but not limited to Student’s t-test, Fisher’s exact test, Kruskal-Wallis test and Log-rank test. All statistical tests were two-sided, and statistical significance was considered when P value or FDR <0.05 unless otherwise indicated. All the analyses of clinical data were performed in R (version 4.1.0).

Data and materials availability:

Liver cancer organoids generated in this study will be available from Liver Cancer Institute, Zhongshan Hospital and can be requested at gaoqiang@fudan.edu.cn. Organoids distribution to third parties requires a Material Transfer Agreement (MTA) and will be authorized by the Research Ethics Committee of Zhongshan Hospital. The original and matrix format data of WES, RNA-seq, RRBS, proteomics, and phosphor-proteomics generated in this study can be viewed in biosino NODE database (OEP003191) through the URL: https://www.biosino.org/node/project/detail/OEP003191. Any additional information required to reanalyze the data reported in this paper is available from the lead contact (gaoqiang@fudan.edu.cn) upon request.