Research ResourceCancer

Patient-derived tumor-like cell clusters for drug testing in cancer therapy

See allHide authors and affiliations

Science Translational Medicine  24 Jun 2020:
Vol. 12, Issue 549, eaaz1723
DOI: 10.1126/scitranslmed.aaz1723

Little clusters with big promise

Despite recent advances in medicine, identifying the optimal treatment regimen for each patient with cancer remains difficult and often imprecise. There are now multiple methods for analyzing a tumor’s drug sensitivity, including tumor organoids and patient-derived xenografts, but each has its own drawbacks such as a lack of tumor stroma or the time required to obtain results. The approach designed by Yin et al., patient-derived tumor-like cell clusters, aims to overcome some of these shortcomings by using ex vivo culture of tumor cells together with their stroma. Initial testing of this method has shown promising results when applied to patients with several different tumor types.


Several patient-derived tumor models emerged recently as robust preclinical drug-testing platforms. However, their potential to guide clinical therapy remained unclear. Here, we report a model called patient-derived tumor-like cell clusters (PTCs). PTCs result from the self-assembly and proliferation of primary epithelial, fibroblast, and immune cells, which structurally and functionally recapitulate original tumors. PTCs enabled us to accomplish personalized drug testing within 2 weeks after obtaining the tumor samples. The defined culture conditions and drug concentrations in the PTC model facilitate its clinical application in precision oncology. PTC tests of 59 patients with gastric, colorectal, or breast cancers revealed an overall accuracy of 93% in predicting their clinical outcomes. We implemented PTC to guide chemotherapy selection for a patient with mucinous rectal adenocarcinoma who experienced recurrence with metastases after conventional therapy. After three cycles of a nonconventional therapy identified by the PTC, the patient showed a positive response. These findings need to be validated in larger clinical trials, but they suggest that the PTC model could be prospectively implemented in clinical decision-making for therapy selection.


Precision oncology seeks to identify effective therapeutic strategies for cancer patients on an individual basis. Omics technologies are widely used in precision medicine: They have identified key modulators in signaling pathways and furthered understanding of cancers from statistical and macroscopic perspectives (1, 2). However, the clinical application of omics technologies has been limited: A large proportion of cancer patients do not have any actionable mutations (3); even for the patients with targetable genomic alterations, treatment design has been difficult because of the lack of effective practical models that optimize clinical outcomes.

Increasing evidence has suggested that patient-derived tumor models faithfully recapitulate human tumor biology and predict potential clinical responses. Patient-derived tumor xenografts (PDXs) can provide a more accurate measurement of the clinical potential of drugs than traditional methods because PDXs retain the idiosyncratic characteristics of different tumors from individual patients (46). In addition, three-dimensional (3D) cultures and organoids cultured from healthy and tumor tissues from patients with cancer have made great progress using Matrigel as an extracellular matrix (ECM) substitute (7, 8). Tumor spheres or spheroids are spherical aggregates of tumor cells, which have important advantages over cells in 2D monolayer culture and reflect many of the properties of solid tumors (9, 10). Patient-derived organoids (PDOs) have been used to perform high-throughput drug screening and to model the treatment responses of metastatic gastrointestinal cancers (11, 12). Furthermore, organoids generated by human embryonic stem cells or induced pluripotent stem cells may circumvent the limited availability of high-quality human primary materials (1316).

Although these models are promising, their potential clinical application is subject to three major limitations. First, the model is required to inform patient-specific treatment decision-making within a clinically meaningful time window; such a time window is generally less than 2 to 3 weeks for neoadjuvant and postoperative chemotherapy. Matrigel or ECM substitutes are added to mimic the in vivo tissue microenvironment in 3D cultures and organoids. Epithelial cells, once encapsulated within Matrigel, would be enriched mainly through proliferation and differentiation. Thus, it would be difficult to generate enough organoids for drug testing within 2 to 3 weeks from small tissue samples. Second, the standardization of cell culture conditions and drug concentrations is needed to facilitate the application of the model in clinical settings. PDOs demonstrate tissue-specific features, and huge efforts have to be made to develop, validate, and standardize culture protocols for different cancer subtypes (1719). In addition, undefined extrinsic factors encompassed in Matrigel can influence the experimental outcomes and reproducibility (20). Last, drug responses of the model need to accurately reflect those of the patient. One of the intrinsic limitations of PDOs is the lack of stromal cells, whereas PDXs may undergo mouse-specific tumor evolution.

Addressing the limitations of previous technologies, we emphasized a strategy of long-term maintenance and expansion of primary tumor cells in Matrigel-free conditions. We optimized the culture media and conditions to maintain the integrity of the dissociated primary tumor cells (including tumor epithelial cells, macrophages, and fibroblasts) while generating enough materials for drug testing. Here, we established an in vitro tumor model, named patient-derived tumor-like cell clusters (PTCs), which serves as a structural and functional unit recapitulating the original tumors in genotype, phenotype, and drug response. PTCs derived from gastric cancer (GC), colorectal cancer (CRC), or breast cancer samples could simultaneously test 100 to 2000 drugs within 2 weeks for more than 80% of patients who provided surgically resected samples (Fig. 1A). The PTC model has shown more than 93% accuracy in recapitulating patients’ clinical outcomes. Its prospective application to one 56-year-old patient with mucinous rectal adenocarcinoma (MRC) also showed evidence of clinical application potential.

Fig. 1 Long-term maintenance and expansion of PTCs.

(A) An overview of the timeline of PTC culture generation and personalized drug testing. (B) GO overrepresentation analysis in biological pathway was performed on the 963 significant (false discovery rate of <0.1) differentially expressed genes in colorectal tumor tissues and spheres. A P value cutoff of 0.05 and a false discovery rate cutoff of 0.1 were applied to get the enriched GO terms. The length of the bar represented the number of differentially expressed genes in that GO term. (C) Gene expression fold changes of the 963 significant (false discovery rate of <0.1) differentially expressed genes in colorectal tumor tissues and spheres were used as input data for KEGG overrepresentation analysis. Wnt signaling, apoptosis signaling, TGFβ signaling, and cell-cell and cell-matrix interaction signaling are presented as representative pathways. CAM, cell adhesion molecule; ECM, extracellular matrix. (D) The effect of culture substrates’ hydrophobicity on PTC growth. The cell viability was examined by ATP activity (CellTiter-Glo), whereas the hydrophobic property of each substrate was measured as a contact angle. PHB, polyhydroxybutyrate; PDMS, polydimethylsiloxane; COC, cyclic olefin copolymer; PS, polystyrene; COP, cyclo olefin polymer; PMMA, poly (methyl methacrylate); SF-PLA, silk fibroin-poly (l-lactic acid); SF, silk fibroin; GX-01, GeneX, Chip-D-01; SF-PCL, silk fibroin polycaprolactone. (E) Time course of primary cell clustering in the first 120 min and at 48 hours. Scale bar, 50 μm. (F) Bright-field images depicting two CRC patients’ representative PTC phenotypes at the primary and metastasis sites on day 14. Solid clusters with lumens (triangle), solid and compact clusters (arrowheads), and grape-like clusters (arrows) are shown. Scale bar, 50 μm. (G) The numbers of total CRC samples collected and the numbers of PTCs successfully generated. The samples included surgically resected tissue, endoscopically obtained biopsies, and ascites. (H) The numbers of total GC samples collected and the numbers of PTCs successfully generated. The samples included surgically resected tissue, endoscopically obtained biopsies, and ascites. (I) The numbers and proportions of surgically resected samples generating different PTC clusters. (J) Bar diagram showing the numbers of CRC or GC-derived PTCs successfully cultured from poorly, moderately, well-differentiated, or unspecified adenocarcinomas. (K) Distribution of gastric or CRC-derived PTCs grouped by cancer subtypes, including neuroendocrine carcinoma, mucinous adenocarcinoma, or metastases.


Establishment of PTC in vitro tumor models for GC and CRC

3D cultures without Matrigel or spheres have low culture rate, 37.5% (6 of 16) (fig. S1A). To achieve the long-term maintenance and expansion of primary tumor cells in Matrigel-free conditions, we had to optimize both culture media and microenvironments. We first performed mRNA sequencing on four spheres and their original tumor tissues derived from one patient with CRC. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis of the 963 significantly differentially expressed genes (false discovery rate of <0.1) found that spheres showed dysregulation of major modulators in Wnt/β-catenin, transforming growth factor–β (TGFβ), apoptosis, cell-cell interaction, and cell-matrix interaction signaling pathways (Fig. 1, B and C, and data file S1). For example, DKK4 was an antagonist of the Wnt/β-catenin signaling pathway (21). DCN blocked the activity of TGFβ (22), FST was a negative regulator of TGFβ-induced epithelial to mesenchymal transition, and Id3 was the downstream modulator of TGFβ (23, 24).

We then tested a panel of growth factors and inhibitors involved in those signaling pathways (table S1). For example, the essential factors of the R-spondin medium, such as epidermal growth factor (EGF), fibrolast growth factor, R-spondin1, TGFβ inhibitor (A83-01), and p38 inhibitor (SB202190), supported the formation of both GC and CRC PTCs. However, CHIR99021 and recombinant Wnt3a promoted the formation of GC PTCs but had no effect on CRC PTCs; hepatocyte growth factor was critically important to culture both GC and CRC PTCs with a high success rate. Last, the substrate hydrophobicity had a substantial effect on the growth of PTCs as well (Fig. 1D). A Teflon-coated substrate demonstrated the highest success rate. Thus, a Teflon-coated culture plate and 100-well chip were chosen for subsequent culture and high-throughput drug testing in this work, respectively.

We next dynamically studied the generation of PTCs by a high-content imaging system and found that PTCs mainly resulted from the migration and self-assembly of dissociated primary cells into clusters (Fig. 1E and movie S1). The number of clusters generated through PTCs was sevenfold that of conventional spheres, and the success rate of generating PTCs (75%) was higher than that of spheres (37.5%) (fig. S1A). PTC growth could be divided into three stages: exponential, stationary, and decline phases (fig. S1B). It seems that the PTCs accomplished self-assembly within the first 2 days of culture, and the number of self-assembled PTCs remained almost constant afterward (Fig. 1E and fig. S1C). As indicated by EdU (5-ethynyl-2’-deoxyuridine) assays, the decline phase of PTCs was delayed by at least 2 to 3 weeks compared to that of the spheres (fig. S1C). PTC also showed higher proliferating activity than spheres, that is, a fourfold or above increase of the total cluster/sphere area (fig. S1B).

PTCs from different patients or even tumors of the same patients at different sites showed distinct types of clusters, indicating inter- and intratumor heterogeneity (Fig. 1F and fig. S1D). For example, in PTCs generated from primary tumors of CRC patient CS-H-210, we observed at least three types of clusters including solid and compact clusters of different sizes, solid clusters with lumens, and “grape-like” clusters; in peritoneal metastatic nodes, most clusters were solid and compact (Fig. 1F). In addition, PTCs demonstrated distinct phenotypes and culture speed from PDOs (fig. S1, E and F). A group of PTCs and PDOs were generated from the same 10 surgical samples. Although 9 of 10 grew out in both cases, it took different amounts of time to achieve enough clusters or organoids to perform drug testing. In general, PTCs took less than 1 week, and PDOs needed 1 month or so (fig. S1E). PDOs from CS-T-858 showed a number of observable thin-walled cystic structures within 14 days, whereas PTCs from CS-T-858 presented thousands of compact clusters devoid of a lumen within 5 days (fig. S1F). PDOs were difficult to image and divide evenly into different wells for the subsequent drug testing because they were encapsulated within Matrigel.

A total of 258 fresh gastric or colorectal tumor samples (200 surgically resected tumor tissues, 31 endoscopic biopsies, and 27 ascites samples) were collected for PTC establishment with an overall success rate of 72.8% (188 of 258) without any preselection, including eight cases with bacterial contaminations and 41 poor-quality samples such as tissues with massive necrosis (Fig. 1, G and H). Although biopsy samples obtained through endoscopy contained smaller amounts of tissue, the corresponding PTCs had a higher success rate than those generated from surgically resected samples (83.9 and 72.5%, respectively) (Fig. 1, G and H). The number of PTCs ranged from 500 to 100,000 within 1 week, depending on sampling approaches (Fig. 1I and fig. S1G). Furthermore, PTCs could be easily adapted to study cancer subtypes that had been impossible or difficult to establish organoids for (17), such as poorly or moderately differentiated adenocarcinoma, neuroendocrine carcinoma, mucinous adenocarcinoma, and metastatic samples (Fig. 1, J and K).

Essential role of stromal cells in PTC assembly

The PTCs demonstrated high morphological similarities with the GC and CRC samples (Fig. 2A and fig. S2A). For example, GS-L-2281 PTC, derived from diffuse-type gastric carcinoma, reconstituted the histologic appearance of signet ring cells of its parental tumor (fig. S2A). Immunofluorescence data revealed that most of the clustering cells (55 to 90%) were CK8/18+/E-cadherin+ epithelial cells (Fig. 2B). PTCs showed expression of markers of cancer stem cells, such as S100A4 and ALDH1A1 (Fig. 2B). These data were further validated by immunohistochemical staining (fig. S2B). PTCs also stained positive for tissue-specific markers. CRC PTCs showed the expression of MUC2 and CDX2 (12, 17, 25), whereas PTCs derived from GC samples showed the expression of CDX2, PDX1, and SOX2 (13, 25) (Fig. 2C and fig. S2B).

Fig. 2 Cellular components of PTCs and essential roles of stromal cells in PTC assembly.

(A) Hematoxylin and eosin staining comparing the CRC and GC PTCs to their matching patient biopsies. Scale bars, 100 μm. (B) Immunofluorescent staining of PTCs on day 14 for epithelial markers E-cadherin and CK8/18 and stem cell markers ALDH1A1 and S100A4. Nuclei were counterstained (blue, Hoechst). The images shown are from CS-A-9518: moderately differentiated CRC. Scale bar, 100 μm. (C) Immunofluorescent costaining of PTCs for CRC and GC marker CDX2 (red) and epithelial marker CK8/18 (green) on day 14. Nuclei were counterstained (blue, Hoechst). The images shown are from CS-M-9885 and GC-P-6802: moderate CRC and GC adenocarcinoma, respectively. Scale bars, 100 μm. (D) Immunofluorescent staining of two CRC PTCs (CS-G-1010 and CS-C-8218) and two GC PTCs (GS-B-4512 and GS-G-1132) on day 14 for epithelial markers CK8/18 (green), fibroblast cell marker FAP (purple), and immune cell marker CD45 (red). Nuclei were counterstained (blue, Hoechst). CS-G-1010 and CS-C-8218: moderately differentiated CRC; GS-B-4512 and GS-G-1132: poorly and moderately differentiated adenocarcinoma GC, respectively. Scale bars, 100 μm. (E) RNA sequencing of 21 individual PTC clusters derived from three patients with moderately differentiated CRC (CS-H-5811, CS-C-4268, and CS-G-8278). Cell marker genes were selected from CellMarker database for each cell type. Solid clusters with lumens (lumen), solid and compact clusters (solid), and grape-like clusters were shown in Fig. 1F. More detailed information on marker genes is found in fig. S2C. (F) Quantitative analysis of the proportion of epithelial cells (CD326+), fibroblast cells (FAP+), and immune cells (CD45+) in two PTC samples (patients CS-T-0090 and GS-W-9020) on day 14 by flow cytometry. CS-T-0090: poorly and moderately differentiated CRC adenocarcinoma; GS-W-9020: not-specified GC. IgG, immunoglobulin. (G) Role of stromal cells in assembly of PTCs. Left: Schematic depicting the culture of PTCs, CD11b cells, EpCAM+ cells, and reconstitution group. Tissue-specific antibody-coated magnetic beads were used to remove macrophages or enrich EpCAM+ cells after the dissociation of tumor samples into single cells. The total dissociated cells were equally split into four parts such that the initial total number of cells in each condition (PTC, CD11b, EpCAM+, and reconstitution group) was the same. All CD11b, EpCAM+, and reconstitution groups followed the same culture and drug test procedures as the PTC group. Right: Representative images of PTC, CD11b, EpCAM+, and reconstitution cell groups from patient CS-T-089 on day 5. CS-T-089: moderate adenocarcinoma CRC. Scale bars, 100 μm. (H) Drug-response profiles of PTC, CD11b, or EpCAM+ cells from patient CS-W-242 on day 7 (n = 3, means ± STD). CS-W-242: poorly differentiated adenocarcinoma CRC. (I) Dynamic drug responses of PTC, CD11b, or EpCAM+ cells from patient CS-W-242 to fluorouracil from days 0 to 7 (n = 3, means ± STD).

PTCs contained FAP+ (fibroblast activation protein–positive) fibroblast cells and/or CD45+ immune cells (Fig. 2D). The cell components in PTCs also showed high inter- or intrapatient heterogeneity. For example, the CS-C-8218 PTCs had no detectable CD45+ cells, and the GS-G-1132 PTCs had fewer fibroblast cells and immune cells than the GS-B-4512 and CS-G-1010 PTCs (Fig. 2D). Furthermore, 21 individual PTC clusters derived from three patients with CRC were selected for whole transcriptome sequencing. Marker expression analysis indicated the presence of VIM+/FAP+ fibroblasts and CD68+/CD45+ immune cells in most PTCs (Fig. 2E and fig. S2C). CD44+/S100A4+ stem cells existed in all 21 PTC clusters (Fig. 2E). EpCAM (epithelial cell adhesion molecule)–dominant PTCs had relatively lower FAP expression, whereas FAP-dominant PTCs, mostly grape-like clusters, had relatively lower EpCAM expression (Fig. 2E). Furthermore, we used flow cytometry assays to quantitatively analyze the proportions of epithelial, immune, and fibroblast cells in six PTCs (Fig. 2F and fig. S2D). We found that the proportions of nonepithelial cells varied among patients, with fibroblast cells ranging from 1.6 to 36.4% and immune cells from 4.1 to 17.5% (Fig. 2F and fig. S2D). PTCs’ drug response seemed to be unchanged in the first 3 weeks (fig. S2E).

We next investigated the essentiality of stromal cells in PTC assembly. Using CD326 (EpCAM) MicroBeads, we confirmed that no assembly of PTCs was observed after the removal of tumor epithelial cells from dissociated primary tumor cells (fig. S2F). Then, we generated two more groups of cells, CD11b and EpCAM+: the former represented cultures assembled from dissociated primary cells after the removal of CD11b+ macrophages, whereas the latter included only tumor epithelial cells (Fig. 2G). We compared these two groups with the PTCs generated from the same five patients (CS-T-089, CS-W-242, CS-T-423, GS-T-688, and GS-T-925) (Fig. 2G and fig. S2G). Both CD11b and EpCAM+ groups produced fewer, smaller clusters, and almost no clusters were generated from patient GS-T-925 samples without macrophage or stromal cells (fig. S2, F and G). To exclude the possibility of interference from cell isolation, we performed reconstitution experiments by mixing EpCAM+ cells and the remaining cells and found no difference in cluster growth between the reconstituted groups and PTC groups (fig. S2F). In addition, the drug response showed distinct patterns between EpCAM+ group and PTCs, suggesting that stroma cells contributed to the drug resistance (Fig. 2, H and I, and fig. S2, H and I). Together, these results indicated that macrophages and other stroma cells played important roles in PTC assembly.

Genomic consistency of PTC with the original tumors

Genomic DNA was isolated from the PTC cells, tumor samples, and matched normal samples for target sequencing and low-pass whole-genome sequencing (WGS). Somatic mutations and copy number variations (CNVs) were called on the basis of target sequencing and the low-pass WGS, respectively. Both were concordant between the PTC cells and the tumor samples. The proportions of common somatic mutations ranged from 0.57 to 1, with a mean of 0.86 and a standard deviation (STD) of 0.13 (Fig. 3A and fig. S3, A and B). Correlations of the CNVs between the PTC cells and tumor samples ranged from 0.52 to 0.96, with a mean of 0.78 and an STD of 0.13 (fig. S3C). Many samples of the PTC cells and tumor samples had nearly identical CNV patterns (Fig. 3B). The observed minor genomic alterations between the PTC cells and tumor samples could be bona fide genomic differences or artifacts caused by insufficient detection power. Common low-frequency subclonal mutations usually have a lower detection power and are more likely to be misclassified as private mutations in PTC cells or tumor samples; in fact, when we increased the variant allele frequency to at least 20 from 10%, the mean proportion of common somatic mutations increased to 0.93 with an STD of 0.17 (data files S2 and S3).

Fig. 3 Genomic comparison between tumor biopsies and corresponding PTCs.

(A) Overview of the identified somatic mutations. Nonsense mutations included all nonsense single-nucleotide variants (SNVs), SNVs at splicing-sites, and frameshift indels. Nonsynonymous mutations include nonsynonymous SNVs and in-frame indels. Nonsynonymous and nonsense means the gene had both nonsense and nonsynonymous mutations. T stood for tumor. Culture time of different PTCs: 4 days (568, 378, 658, 443, 869); 7 days (222, 276, 661, 450); 14 days (195, 573, 216, 978). (B) Overview of CNVs. The rows represent different samples, and the columns represent genomic positions from chromosomes 1 to 22. Colors in the plot correspond to the estimated log2 copy ratio of the genomic regions. The four bottom panels show local views of four cancer genes frequently altered in CRC and GC. MB, Megabase pair. (C) The somatic mutations in tumors and in replicated PTCs. T, tumor. PTC0 to PTC5: PTCs in different wells. (D) The CNVs in tumors and in replicated PTCs.

The discovered genomic alternations largely agreed with previously reported mutations typical for CRC and GC. From the genes covered in our gene panel, TP53, KRAS, PIK3CA, SMAD4, and NRAS were the top five most frequently mutated driver genes in CRC (26), whereas TP53, SMAD4, and PIK3CA were the top three most frequently mutated genes in GC (27). Of the eight patients with CRC in this study, seven had nonsynonymous TP53 mutations, three had KRAS mutations, three had SMAD4 mutations, one had a NRAS mutation, and none had PIK3CA mutations; PTC cells and the corresponding tumor samples had the same mutations in these four driver genes (fig. S3A). Among the five patients with GC, we observed two patients with mutations in SMAD4, one patient with mutations in TP53, and one patient with mutations in PIK3CA. PTC cells and tumor samples shared the same mutations in the three genes (fig. S3A). The typical chromosomal arm CNVs (>50% chromosomal arm amplified or deleted) in CRC and GC were also observed in PTC cells (26, 27). For example, many PTC samples had copy number gains of chromosomes 1q, 7, 8q, 13q, and 20, as well as copy number losses of 1p, 5q, 8p, 14q, 15q, 17p, and 18. Recurrent focal amplifications and deletions were also observed in many patients, including amplifications of ERBB2 and EGFR and deletions of PTEN and SMAD4 (Fig. 3B). In summary, genomic analyses revealed that PTCs maintained the key genomic features of the primary tumors.

We also divided PTCs into a multi-well chip and investigated the genomic consistency by comparing the genomic and transcriptomic profiles between PTCs in different wells and their original tumors. The somatic mutations, CNVs, and transcriptome profiles showed high consistency, indicating the similarity of the proportions and compositions of the PTCs among wells (Fig. 3, C and D, and fig. S3D). With the exception of sample 222, pairwise overlap analysis showed that the proportions of common somatic mutations between different wells were always greater than 0.8 (fig. S3E). The CNV correlations between the cells had a mean of 0.91 and an STD of 0.09 (fig. S3F). The gene expressions between different wells were highly correlated, with many pairs of wells having correlations close to 1 (fig. S3G).

Developing PTC as a tool for personalized drug testing

To identify the optimum therapeutic option for individual patients, we designed a personalized drug testing system based on PTC assay: We separated thousands of PTCs into a multi-well chip and evaluated drugs in different wells. In clinical practice, the Response Evaluation Criteria in Solid Tumors (RECIST) criteria are widely used to judge the effectiveness of individual treatment, which is highly related to image-based assessment (28). To inform treatment design, the PTC assay accordingly defined an image-based approach to evaluate drug effect, fixed a cell viability cutoff, and then determined drug efficacy concentrations (Ecs) (Fig. 4A).

Fig. 4 The establishment of the PTC system for personalized drug testing.

(A) Three key issues involved in the development of the PTC system: The evaluation of drug effect, the definition of the cutoff, and the determination of Ec. (B) Effect of the cluster numbers on the drug-response consistency. Y axis: coefficient of variation (CV). X axis: number of clusters used. (C) The distribution of CVs in the drug-sample pairs in the gastrointestinal group. n = 339 drug-sample pairs. (D) The distribution of the Pearson correlation coefficients in a 1000-fold random sampling strategy for two results in three parallel drug sensitivity tests. n = 339. (E) PTCs derived from patient GS-X-0805 were split into 461 wells for drug testing. Left: Heat map summarizing the results of the PTC-based drug testing assays using a custom library of 49 drugs tested in clinical trials. Three concentrations of each drug were used in triplicate: 0.1, 1, and 10 μM. Right: Concordance between viability readings obtained from testing assays (49 compounds tested, 3 replications).

First, we evaluated drug effect by measuring the area of all PTC clusters in a well. The cell clusters were photographed at days 0 and 7 in a well. Only clusters with diameters of more than 40 μm at both time points were selected to calculate the cluster areas. The cell viability after applying a drug A was calculated by the following formulapAi=SAi,t1/SAi,t0,pA=1ni=1npAiwhere S is the sum of cluster areas in a well, n is the number of replicates, and t0 and t1 are the time points when the areas are measured. We used the cell viability value of the negative control (NC) with vehicle treatment, calculated the same way as a quality control. If the cell viability score pNC was less than 0.9, indicating that the PTC was possibly in the decline phase, then the PTC test was discarded.

Second, we determined the PTC cutoff according to the RECIST criteria (28). For the purpose of analysis, we collapsed the RECIST criteria into two groups using 0.7 as a cut-off value and defined complete response (CR) or partial response (PR) as effective and stable disease (SD) or progressive disease (PD) as otherwise. Accordingly, the cell viability was divided into the following two categories: The drug was not effective if pA ≥ 0.7, and the drug was effective if pA < 0.7 (Fig. 4A and fig. S4A).

Last, we determined the PTC Ec value of the drug A according to its clinical efficacy. For any precision oncology method that could accurately predict a patient’s clinical response, including the PTC model, its predicted efficacy rate of a drug should be consistent with the overall response rate of this drug among patients. In our PTC model, the predicted efficacy rate was determined by the cell viability cutoff and the drug’s Ec. Thus, after fixing the 0.7 cut-off value, we determined the Ec of a drug as the concentration such that the effective rate in PTC assay was closest to the overall response rate of this drug previously reported in relevant clinical trials (Fig. 4A). We used 20 to 30 patients’ PTC samples as the training cohorts to determine the Ec values of 13 drugs. For example, a group of clinical trials showed that overall response rate was about 11 to 20% for 5-fluorouracil in advanced GC or CRC (2931). Accordingly, 5-fluorouracil Ec was set to 2 μmol/liter, at which the efficacy rate was 18.1% (6 of 33). The Ec of all drugs used in this study was determined (fig. S4B) and summarized in table S2.

Preclinical evaluation of PTC as a tool for personalized drug testing

Whether PTCs can be used as an informative tool for precision oncology depends on two factors: PTC cell viability consistency across wells and the consistency between PTCs’ prediction results and the patients’ responses in the clinic. The latter would be investigated by using two validation cohorts in the following sections.

We first assessed the cell viability consistency across the wells. We investigated the effects of the cluster numbers in a well on the drug-response inconsistency because of tumor heterogeneity. We found that when a well contains 30 to 50 clusters, the PTC drug test results have coefficients of variation (CVs) less than 0.2 in most cases (Fig. 4B). To validate this finding, we performed large-scale cell viability tests. PTC samples from 28 patients were used for drug testing (15 to 22 drugs per patient). Three replications were performed for each drug-sample pair. As a result, we obtained 339 drug profiles in total. Cell viability consistency requires that replicates of the same drug-sample pair render close and highly correlated results (Fig. 4, C and D). We calculated the CVs of the three experiments for each drug-sample pair and found that 91.74% CVs were less than 0.5 (Fig. 4C), indicating that the results of three experiments for most drug-sample pairs were close to each other. In addition, we randomly selected two experiments from each drug-sample pair and calculated the Pearson’s correlation coefficient between the selected experiments. This process was repeated 1000 times. We found that the correlations were consistently larger than 0.8, with a mean of 0.84 and an STD of 0.036 (Fig. 4D and fig. S4, C and D), which indicated that results from different experimental replications for the same drug-sample pair were highly consistent.

It should be pointed out that only a small proportion of patient-derived tumor sample is needed for PTC drug testing, and the majority would be used for pathological examination and other analysis. More than 84% (122 of 145) patients providing surgically resected samples had enough material for testing at least 100 drugs; for example, more than 460 PTC assays were performed with high reproducibility starting with the 0.25 g of tumor sample patient GS-X-0805 provided in the concentrations of 0.1, 1, and 10 μM according to the results of Fig. 4 (A and E) and fig. S4B. Even in cases where only endoscopic biopsies were available, 10 to 100 drugs could be tested in 94% of these cases.

PTC prospectively and accurately recapitulating the clinical outcomes of patients with gastrointestinal cancers

We next investigated the consistency between the cell viability predicted by PTC and patients’ real clinical responses in a validation cohort of gastrointestinal cancer. After obtaining the ethical approval from Peking University Cancer Hospital Ethics Committee, 39 eligible patients with clinical stage III/IV GC or CRC were prospectively enrolled for the preclinical evaluation (Fig. 5A). PTCs were successfully established from 30 patients, and 6 patients were subsequently excluded from the analysis because of patient death, loss to follow-up, heterogeneity of drug responses of PTCs from different sites, or regimen adjustment during the treatment. From the 24 patients included in the analysis, 1 had gastroesophageal junction cancer, 11 had GC, and 12 had CRC. The PTC tests and the treatments of patients were independently performed (see Materials and Methods for more details). Detailed patient clinicopathological characteristics are outlined in data file S4.

Fig. 5 PTCs reflect the clinical outcomes of representative patients with gastric and CRCs.

(A) The patient selection flowchart. (B) Box-plot of the drug test results for 24 patients in the CR/PR group (n = 14) and the SD/PD group (n = 15) on the basis of PTCs. The patient GS-J-251 received two treatments. The patients CS-G-907 and CS-W-570 each received three treatments. Middle bar, median; dotted line, cell viability cutoff of 0.7. P = 2.0 × 10−7. (C) A 2 × 2 contingency table and a heat map summarizing in vitro drug responses based on PTC testing and clinical outcome based on RECIST. n = 29. (D) Treatment and procedure timeline for patient GS-J-251. CEA and CA72-4 values are shown at different time points. The values highlighted in red indicated the efficacy of the paclitaxel. (E) The drug-response profile of PTCs from patient GS-J-251. Blue arrow indicates SOX therapy, and red arrowhead indicates paclitaxel therapy. Fluorouracil (equivalent to tegafur, the effective ingredient in S-1, with regard to cell viability) was used in the test. Cis, cisplatin. Doc, docetaxel. Epi, epirubicin; Eto, etoposide; Flu, fluorouracil; Iri, irinotecan; Pac, paclitaxel; Oxa, oxaliplatin. (F) Left: Fused PET-CT image and MIP (maximum intensity projection) image of patient GS-J-251 after postoperative chemotherapy. Both images showed increased uptake of 18F-FDG with a SUVmax of 3.5 in the mildly thickened peritoneum on 28 February 2018. Right: After two perfusions of paclitaxel, the metastatic node (indicated by a red arrow and blue circle) disappeared on 02 August 2018. (G) CT images of patient GG-S-331 before and after chemotherapy. The size of the metastatic lymph node, indicated by red arrow, decreased from 1.9 × 1.5 to 1.0 × 0.6 cm. Red box represented the enlarged view of the tumor. (H) Microscopic image of gastric tissues of patient GG-S-331. Morphology indicated grade 2 tumor regression. Scale bar, 300 μm. (I) The drug-response profile of PTCs from a biopsy sample from patient GG-S-331 (n = 3, means ± STD). The blue arrow indicates a significant effect of the combination of oxaliplatin (Oxa) and fluorouracil (Flu) (P = 0.01).

We found that the CR/PR group had a significantly lower cell viability score than the SD/PD group (Wilcoxon’s test, P = 2.0 × 10−7; Fig. 5B and data file S4). Using 0.70 as the cut-off value, the prediction accuracy by PTC for the drug effectiveness was 100% (14 of 14) for the CR/PR group and 93.3% (14 of 15) for the SD/PD group, with an overall accuracy of 96.6% (28 of 29) (Fig. 5C).

We present the detailed treatment process of two of the 30 patients. PTC predictions and clinical outcomes of other patients can be found in fig. S5 and data file S4. Patient GS-J-251 was diagnosed with signet ring cell gastric carcinoma and further identified to have peritoneal metastasis (pathological stage: pT4aN3aM1). The patient received four cycles of S-1 (the effective ingredient tegafur is equivalent to fluorouracil) and oxaliplatin (SOX) combination postoperatively and two rounds of intraperitoneal hyperthermic perfusion with paclitaxel alone (Fig. 5D). The corresponding PTC showed no response to the combination of fluorouracil and oxaliplatin (Fig. 5E). Positron emission tomography–computed tomography (PET-CT) showed a large number of peritoneal and retroperitoneal lymph nodes with high metabolism and pelvic effusion, indicating peritoneal metastasis (Fig. 5F). The patient showed a modest response to paclitaxel, as evidenced by the decrease in two tumor markers (CEA and CA72-4) after two intraperitoneal hyperthermic perfusions (Fig. 5D) (32). This result was consistent with the PTC prediction. Because of the positive response from the intraperitoneal perfusion and the PTC’s prediction, the patient switched to paclitaxel regimen in February 2018, and peritoneal nodes (SUVmax 3.5) disappeared after two perfusions of paclitaxel (Fig. 5F). Until 19 February 2020, the patient was alive.

Patient GG-S-331 was diagnosed with moderately differentiated stage III gastric adenocarcinoma based on an abdominal CT examination. The patient provided endoscopic biopsies for the generation of PTCs before neoadjuvant chemotherapy. Given the small biopsy, it was challenging to generate a sufficient amount of cells to perform personalized drug tests within 2 to 3 weeks. This patient received neoadjuvant chemotherapy with the SOX regimen. After three cycles of chemotherapy treatments, a metastatic lymph node markedly shrank (Fig. 5G), which was further confirmed by pathologic assessment (grade 2 tumor regression) (Fig. 5H). These clinical outcomes were in line with the PTC prediction (Fig. 5I).

PTCs accurately predicting pathological CR or nonresponsive patients with breast cancer

We next examined whether PTCs could be adapted to different cancer types besides CRC or GC and performed validation on an independent cohort of patients with breast cancer. To this end, we successfully generated 140 breast cancer PTCs from 153 patients who provided 177 fresh tumor tissues through lumpectomy or puncture biopsy, with different success rates for luminal A (94%), triple-negative (83%), luminal B (70%), and HER-2–positive (70%) (Fig. 6, A and B). Among them, six patients with each tumor type provided two to four surgical samples from primary tumors, and all samples, except one, successfully generated PTCs. Five patients provided two surgical samples each, one from primary tumors and another from lymph nodes. All primary samples successfully generated PTCs, and three lymph node samples generated PTCs. Four patients provided two endoscopic samples each, one from the primary tumor and another from lymph nodes, and two pairs of endoscopic samples successfully generated PTCs. The breast cancer PTCs recapitulated the receptor status and the conserved multicellular organization of the original breast cancer samples (fig. S6, A and B).

Fig. 6 PTCs reflect the clinical outcomes of representative patients with breast cancers.

(A) 140 PTCs were generated from 177 samples. The samples included surgically resected tissue, needle biopsies, and ascites. The overall success rate was 79%, and the success rates for the three different types of samples were 79, 78, and 100%, respectively. (B) The numbers and success rates for PTCs derived from different breast cancer subtypes. (C) The patient selection flowchart. (D) Boxplot of the drug test results of 35 patients for the CR/PR group (n = 25) and the SD/PD group (n = 10) based on PTCs. Middle line, median; dotted line, cell viability cutoff of 0.7. P = 0.01. Patients BH-P-7273 (7273), BL-P-2778 (2778), and BL-P-9204 (9204) showed consistence between PTC prediction and M&P classification. The drug-response prediction for BL-P-1253 (1253) was incorrect, which might be because of high heterogeneity in PTCs derived from the limited puncture sample. (E) A 2 × 2 contingency table and a heat map of in vitro drug responses based on PTC test and clinical outcome based on RECIST. n = 35. (F) PTC drug response and clinical imaging data for patient BH-P-7273. Left: The drug-response profile of PTCs. The blue arrow indicates the combined therapy of docetaxel, carboplatin (Car), and trastuzumab (Tra) used for the treatment of the patient. Right: MRI and histopathological examination. Scale bar, 200 μm. (G) PTC drug response and clinical imaging data for patient BL-P-2778. Left: The drug-response profile of PTCs. The blue arrow indicates the combined therapy of docetaxel and fluorouracil used for the treatment of the patient. Right: MRI and histopathological examination. Scale bar, 200 μm. (H) PTC drug response and clinical imaging data for patient BL-P-9204. Left: the drug-response profile of PTCs. The blue arrow indicates the combined therapy of docetaxel and epirubicin used for the treatment of the patient. Right: MRI and histopathological examination. Pir, Pirarubicin. Scale bar, 200 μm. (I) The distribution of PTC drug efficacies along M&P classification (Spearman correlation = −0.60, P = 0.00030). The CR/PR group (red) and the SD/PD group (blue) are highlighted. The M&P = 1 and M&P = 5 cut-off lines were set according to receiver operating characteristic curves in fig. S9 (H and I), respectively. The cutoffs were determined by Youden’s index.

After obtaining the ethical approval from Peking University People’s Hospital Ethics Committee, 50 eligible patients with breast cancer naïve to treatment were prospectively enrolled for the consistency assessment (Fig. 6C). The inclusion criteria are listed in Materials and Methods. PTCs were successfully established from 42 patients, and 7 patients were subsequently excluded from the analysis because of patient death, loss to follow-up, heterogeneity of drug responses of PTCs from different sites, or regimen adjustment during the treatment. Among the 35 patients included in the analysis, 10 had triple-negative breast cancer, 8 HER-2+, 7 luminal A, and 10 luminal B. Detailed patient clinicopathological characteristics are outlined in data file S5.

Using 0.70 as the cutoff value, the prediction accuracy of PTC for the drug effectiveness was 92% (23 of 25) for the CR/PR group and 60% (6 of 10) for the SD/PD group, with an overall accuracy of 82.9% (29 of 35) (Fig. 6, D and E). Furthermore, 32 patients received neoadjuvant chemotherapies, and their pathological response was evaluated using the Miller & Payne (M&P) classification (data file S5). Three cases, in which the PTC failed to predict image-based RECIST outcomes, demonstrated the consistence between PTC and M&P classification. Patient BH-P-7273 received two sequential neoadjuvant chemotherapies, mainly including carboplatin for a total of eight cycles. Magnetic resonance imaging (MRI) indicated that the patient had SD, which was inconsistent with the PTC prediction (Fig. 6F). However, histological examination supported that the patient had complete pathological response (M&P = 5). The PTC assay showed that patients BL-P-2778 and BL-P-9204 showed no response to the combinations of docetaxel and epirubicin or of docetaxel and fluorouracil, respectively. MRI indicated that both patients had PR, but histological examination showed that both patients had no pathological response (M&P = 1), consistent with PTC prediction (Fig. 6, G and H). Thus, after the correction by M&P values, the overall prediction accuracy of PTC for the drug effectiveness reached 91.4% (32 of 35) (data file S5 and figs. S7 to S9).

The PTC cell viability score was significantly correlated with the M&P classification (Spearman Correlation = −0.60, P = 0.00030) (Fig. 6I). The PTC prediction of nonresponse patients (M&P = 1) had an 87.5% (seven of eight) accuracy at the cutoff of 0.66 (fig. S9H). The nonresponse rate of the remaining patients (M&P > 1) was 8.7% (2 of 23), much lower than the clinical nonresponse rates, 29.0% (9 of 31) in our data and 15% in the literature (33). The pathological CR (pCR) rate (M&P = 5) could reach 66.7% (8 of 12) accuracy with the cutoff of 0.40 [area under the curve (AUC) = 0.755] (fig. S9I), much higher than the clinical response rates, 35.5% (11 of 31) in our data and 14% in the literature (33). Together, these results indicated that PTCs can not only accurately predict drug responses for patients with treatment-naïve breast cancer but also help clinicians identify potential patients with pCR response or poorest response.

PTC prospectively identifying an effective but nonconventional therapy for a patient with MRC

After PTC-directed therapy showed high prediction accuracy and marked clinical response in patient GS-J-251, we established PTC as a guide in chemotherapy selection for a patient who failed in conventional therapy. Patient CS-G-907 was a 56-year-old man diagnosed with mucinous adenocarcinoma (MRC) with multiple metastases (data file S4 and fig. S10A). MRC, found in 10 to 15% of all primary CRC cases, is generally considered an atypical and unfavorable subtype of CRC. The prognosis of MRC and its clinical responses to treatment remain controversial (3436). The overall diagnosis and treatment procedure of patient CS-G-907 are summarized in Fig. 7A. The 14-day time course for PTC establishment and personalized drug combination testing is shown in fig. S10B. PTC groups 1 and 2 were generated from ascites obtained from the patient before neoadjuvant XELOX chemotherapy and after postoperative XELOX chemotherapy, respectively. Drug tests showed that PTC group 2 was less sensitive than PTC group 1 to the combination of fluorouracil and oxaliplatin (Student’s t test, P = 0.0052) (Fig. 7, B and C). Both PTC groups 1 and 2 were sensitive to the combination of epirubicin, cisplatin, and fluorouracil (ECX), a conventional therapy for patients with breast cancer (Fig. 7, B and C).

Fig. 7 PTC-assisted therapy and clinical outcomes.

(A) The timeline of diagnosis and treatment procedures for patient CS-G-907. Ascites were taken for generation of PTC on 26 July 2017 and 05 May 2018 and were named PTC groups 1 and group 2 chronologically. (B) The drug-response profile of PTC group 1. Arrows indicate the drug responses to the XELOX (blue) and ECX (red) regimens. PTC group 1 was more sensitive to ECX than to XELOX (Student’s t test, P = 0.0029). (C) The drug-response profile of PTC group 2. Arrows indicate the drug responses to the XELOX (blue) and ECX (red) regimens. PTC group 2 was more sensitive to ECX than to XELOX (Student’s t test: P = 0.00046). In addition, PTC Group 2 was significantly more resistant to XELOX than PTC group 1 (Student’s t test, P = 0.0052) and more resistant to ECX than PTC group 1 (Student’s t test, P = 0.028). However, PTC group 2 was still sensitive to ECX. Gem, gemcitabine; Fol, folinate calcium; Pem, pemetrexed; Ral, raltitrexed. (D) MR images of patient CS-G-907 before and after the first three cycles of the XELOX regimen. Red arrowhead indicates the tumor circled by a yellow dashed line. (E) Dynamic changes in tumor biomarkers (CEA, CA125, CA72.4) during XELOX and ECX regimens, as indicated by the line and arrows, respectively. Dotted lines indicate normal values. (F) Selected CT images of omental nodes and ascites before and after treatment with the ECX regimen. Yellow dashed ovals indicate omental nodes (left). Red arrowheads indicate ascites (right).

The patient first underwent neoadjuvant chemotherapy with the XELOX regimen (fluorouracil and oxaliplatin) for three cycles before undergoing laparoscopic low anterior resection. Patient demonstrated a PR after aforementioned neoadjuvant XELOX chemotherapy, as evidenced by MRI and histologic examination of laparoscopic resection samples including primary tumors, regional lymph node metastasis, and peritoneal metastasis at pelvic cavity (Fig. 7D and fig. S10C). This was consistent with results from the PTC test, where PTC group 1 showed a modest response to the XELOX regimen (pA = 0.66).

After surgery, the patient continued to undergo treatment with the XELOX regimen for another five cycles. At the later stage of chemotherapy, however, the patient was diagnosed with multiple peritoneal nodes (SUVmax 4.2) and omental nodes (SUVmax 3.3). The patient also displayed a marked increase in tumor biomarkers CEA, CA72.4, and CA125 and a large amount of abdominal and pelvic ascites (Fig. 7E). Drug test on PTC group 2 confirmed that the patient had developed resistance to the XELOX regimen (pA = 1.10) (Fig. 7C); this drug test also revealed that the patient might be sensitive to the nonconventional therapy ECX (pA= 0.219). After thorough communication with the clinicians, the patient was administered the ECX regimen starting on 15 May 2018. The patient has since completed three cycles of this therapy and has shown positive responses such as peritoneal nodes disappeared, the omental nodes shrank, the volume of ascites decreased, and the tumor markers CEA, CA 125, and CA72.4 were markedly reduced (Fig. 7, E and F).


In the past 50 years, hundreds of human cancer models were developed, including soft agar colony formation assays, “conditional reprogramming” 2D cell culture, patient-derived tumor cells, 3D culture, and PDOs/PDXs. Although promising, those models have many limitations before being implemented in the clinic. To be a valuable reference in clinical decision-making, time, standardization, and accuracy are the three most prominent considerations for any in vitro tumor models. Here, we report a personalized drug testing system called PTC generated from the self-assembly and proliferation of primary tumor cells, which recapitulated the key genomic and cellular features of tumors (also named miniTumors). We optimized both culture medium and microenvironment of those primary cells, including the removal of Matrigel, the defined medium, and the implementation of hydrophobic substrate.

Matrigel or other ECM substitutes are an essential component for 3D organoids because they mimic the tumor microenvironment. However, these substances contain undefined extrinsic factors that can influence the experimental outcomes and standardization. The protocols of all organoid derivation are highly variable, and organoid-to-organoid reproducibility might be a limitation for clinical applications (20). Counterintuitively, organoids from normal epithelium often grow faster than organoids derived from advanced tumors (37, 38). To ensure outcome validity, various chemicals need to be added or removed to inhibit the growth of healthy cells or to promote culturing of different subtypes of tumor epithelial cells (17, 39). Thus, extra time and efforts are required to standardize the culture of organoids and validate their prediction accuracy before any clinical application (20, 39). Last, organoid cultures generally require a few months before they generate sufficient material for cell viability testing, especially for endoscopic biopsies, and thus can hardly fit into the clinically meaningful time frame.

In this work, we established clear culture conditions for PTCs, drug Ecs, and an efficiency cutoff for accurate clinical prediction. Through tumor tissue’s dissociation and self-assembly, we were able to obtain 500 to 100,000 clusters from each individual endoscopic biopsy or tumor sample within 1 week. We confirmed that, in addition to CK8/18+ epithelial cells, the PTCs contained stromal cells such as macrophages and fibroblast cells; stromal cells played an indispensable role in PTC assembly and drug-response recapitulation. PTCs could predict patients’ clinical responses to targeted agents or chemotherapies with an accuracy of 93.2% within 2 weeks.

We started with the establishment of the PTC technology for gastrointestinal cancers and showed its accuracy in predicting patients’ clinical outcomes. To demonstrate potential applications in other cancers, we extended the PTC to breast cancer. For patients with breast cancer, a complete pathological response after primary chemotherapy is of critical importance (40, 41). The neoadjuvant chemotherapy, mainly based on the patients’ clinicopathological examination, resulted in about 14% pCR rate. The pCR accurate rate based on the PTCs could increase fourfold or so (66.7% versus 14%). Similarly, the PTC prediction of nonresponse in patients achieved 87.5% (seven of eight) accuracy, implying that those patients would have more opportunities to receive effective treatment.

PTCs demonstrated their potential to predict effective therapeutic regimens. For example, in many patients with breast cancer, only one regimen demonstrated effect and the remaining ones showed little or no effect. Epirubicin showed weak effect in patient BH-P-4867, and both docetaxel and trastuzumab showed negligible effects. Similarly, for patients BT-P-8771, BT-P-2890, and BH-P-8752, the combination regimens demonstrated similar drug test results to monotherapy, indicating that the remaining drugs made no or little contribution to the treatment. In contrast, the combination regimens demonstrated better effect than single-drug regimens for patients BL-P-1508, BH-P-6430, BL-P-6784, and BL-P-1608. Especially for BL-P-1508, epirubicin and docetaxel showed no or weak effects, but their combination demonstrated a strong effect.

The role of tumor heterogeneity in drug resistance remains largely unclear (42); PTCs could enable more investigations in this regard because some PTCs consisted of two or more distinct clusters in phenotypes or genotypes. It should be noted that in cases with PTC values less than 0.7 (CR or PR), a large number of drug-resistant clusters existed. Thus, the PTCs could provide an excellent in vitro model for the investigation of tumor heterogeneity and drug resistance. It is likely that once more PTC data associated with clinical therapies become available, mathematical models could be developed to facilitate selection of clinical therapies.

We demonstrated the example of PTC-guided clinical treatment for a patient with tumor evolution. The PTC accurately modeled the dynamic responses of the patient to the XELOX regimen. The patient had a notably positive response to the nonconventional ECX therapy identified by PTC; his metastatic nodes almost disappeared, whereas the volume of his ascites and a number of tumor markers markedly decreased. More than 80% of surgically resected samples and 69% of endoscopic biopsies could generate >5000 PTC clusters, and ~100 drugs could be tested for these samples. Thus, when conventional therapies recommended by National Comprehensive Cancer Network (NCCN) fail for patients, the PTCs could help oncologists directly and accurately identify alternative regimens beyond the NCCN-recommended scope. As long as ascites are available, PTC could dynamically provide therapeutic guidance for the patients, similar to patient CS-G-907 discussed above.

The PTC model has some limitations. Long-time culture and passage of PTCs were difficult. The PTCs’ drug-response patterns could change after 4 weeks of culture, possibly because of the loss of stromal cells. Another technical limitation was that the PTCs cannot test the response of immunotherapy or angiogenesis-based therapies because of the lack of T and endothelial cells. In addition, this work only studied three tumor types and one prospective application. Although these weaknesses will need to be addressed for broader applications, our results suggest that it should be clinically feasible to implement PTCs to deliver a prediction for the outcome of neoadjuvant and conventional chemotherapies for patients with CRC, GC, or breast cancer.


Study design

The study was a multicenter prospective observational study. The objective of the study was to evaluate the feasibility and predictive value of a standardized PTC-based test to distinguish patients with and without response to standard-of-care treatments. The study was carried out in the Peking University Cancer Hospital and Peking University People’s Hospital and was approved by the local ethical review boards. The protocol complies with the Declaration of Helsinki, Chinese Law, and Good Clinical Practice. Clinical responses of the biopsied lesion were evaluated according to the RECIST criteria by radiologists and physicians. Written informed consent was obtained from the patients and/or their authorized representatives.

Two cohorts of patients were recruited in this study. The cohorts of patients with gastrointestinal cancer and breast cancer were accrued at the Peking University Cancer Hospital and the Peking University People’s Hospital, respectively. Patients were not randomized because this was an observational study. Inclusion criteria included adults between 18 and 75 years old; patients with advanced, recurrent, or metastatic gastrointestinal cancer and at least one measurable lesions or malignant ascites; patients with breast cancer naïve to treatment (RECIST 1.1); ECOG 0-2 and eligible for further chemotherapy; expected life span longer than 6 months; and fresh tissue available through either biopsy or surgical resection of the primary or metastatic lesions. Patients with severe organ dysfunction [total bilirubin of >1.5 × upper limits of normal (ULN), alanine aminotransferase and aspartate aminotransferase of >2.5 × ULN, or serum creatinin of >1.5 × ULN and creatinine clearance rate of ≤ 50 ml/min] were excluded. Other exclusion criteria were as follows: pregnancy or lactation; cognitive impairment, mental illness, or poor compliance; and allergy to chemotherapeutic agents. The PTC cell viability was calculated by technicians, and the clinical efficacy was independently evaluated by clinicians. The technicians performing the PTC tests were blinded to the interventions and clinical outcomes of the patients.

Human tumor samples

All human tissue samples were obtained from Beijing Cancer Hospital and Peking University People’s Hospital. Before surgery at the center, all patients provided written informed consent to allow any excess tissue to be used for research studies. The study was approved by the ethical committees. The pathologic status of the specimens was provided by the hospitals.


Treatment plans were discussed at a multidisciplinary team meeting involving surgeons, oncologists, radiologists, pathologists, and other professionals. Patients with gastrointestinal cancers received fluorouracil-based therapy, combined with oxaliplatin or irinotecan. Some patients received oxaliplatin (130 mg/m2) intravenously on day 1 and capecitabine (1000 mg/m2) orally twice daily on days 1 to 14 (XELOX/CAPOX) and repeated every 3 weeks. Some patients received oxaliplatin (85 mg/m2) intravenously on day 1, leucovorin (200 mg/m2) on day 1, and bolus 5-FU (400 mg/m2) on day 1, followed by 5-FU (2400 mg/m2) 46-hour infusion (mFOLFOX6), and repeated every 2 weeks. Some patients received CPT-11 160 to 180 mg/m2 on day 1, leucovorin (200 mg/m2) on day 1, and bolus 5-FU (400 mg/m2) on day 1, followed by 46-hour infusion (2400 mg/m2; mFOLFIRI), and repeated every 2 weeks. Some patients received oxaliplatin (130 mg/m2) intravenously on day 1 and S-1 (80 mg/m2) twice orally daily on days 1 to 14 (SOX) and repeated every 3 weeks. Some patients received oxaliplatin (130 mg/m2) on day 1 and raltitrexed (3 mg/m2) on day 1 (Oxa + Ral) and repeated every 3 weeks. Some patients with gastrointestinal cancers were treated with paclitaxel according to clinically approved regimens. Some patients received paclitaxel (175 mg/m2) intravenously over 3 hours and repeated every 3 weeks. Other patients received paclitaxel (135 mg/m2) intraperitoneal infusion (paclitaxel intraperitoneal chemotherapy). Still, other patients received paclitaxel (85 mg/m2) intravenously day 1 and carboplatin (300 mg/m2) intravenously on day 1 and repeated every 3 weeks. Patient CS-G-907 received epirubicin (50 mg/m2) intravenously on day 1, cisplatin (60 mg/m2) intravenously on day 1, and capecitabine (1000 mg/m2) twice daily on days 1 to 14 (ECX) and repeated every 3 weeks.

Patients with breast cancer were treated with taxane-based combination regimens in the neoadjuvant setting. Some HER-2–negative patients received docetaxel (75 mg/m2) or paclitaxel (175 mg/m2) and epirubicin (75 mg/m2) intravenously in 3-week cycles. Some HER-2–negative patients received capecitabine (1000 mg/m2) orally twice a day on days 1 to 14 and docetaxel (75 mg/m2) or paclitaxel (175 mg/m2) on day 1. Other HER-2–negative patients received epirubicin (75 mg/m2) and cyclophosphamide (600 mg/m2), followed by docetaxel (75 mg/m2) or paclitaxel (175 mg/m2) in 3-week cycles. Among HER-2–positive patients, some received trastuzumab (8 mg/kg for the first dose and then 6 mg/kg for the second dose) in combination with taxanes in EC-TH regimen. Other patients received TCH and docetaxel (75 mg/m2) and carboplatin (AUC of 6) with trastuzumab (8 mg/kg for first dose and then 6 mg/kg for the second dose) in 3-week cycles. There was also one estrogen receptor–positive patient taking daily toremifene (60 mg/day) orally for neoadjuvant endocrine therapy only.

Validation cohort data collection and analysis

After obtaining the biopsies from patients, we generated PTCs in a week and then performed PTC drug testing for a variety of drugs that clinicians might use (after consulting with clinicians). We got the drug-response data in 2 weeks after obtaining the biopsies. Most patients had not even started the treatments by the time we obtained the PTC test results. Other than the second treatments for patients GS-J-251 and CS-G-907, no information about the PTC test results was provided to patients before and during the treatments. After all the clinical outcomes and PTC drug tests were available, we compared the patients’ clinical outcomes with the PTC cell viability values for the drugs that were used by the patients.

Statistical analysis

For continuous variables, groups were compared using two-sample Wilcoxon’s test (also known as Mann-Whitney test) or two-sample t test. Fisher’s exact test was used to test the independence of two categorical variables. The correlation between two continuous variables was calculated using Pearson’s correlation, and the correlation between a continuous variable and an ordinal categorical variable was calculated using Spearman’s correlation. All statistical analyses were performed in R or Excel. P values were corrected for multiple testing using the Benjamini-Hochberg procedure to calculate the false discovery rate.


Materials and Methods

Fig. S1. Culture and characterization of PTCs.

Fig. S2. Cellular characterization of PTCs.

Fig. S3. Genomic comparison of PTCs with the original tumors.

Fig. S4. Drug and genomic consistency of PTCs in different wells.

Fig. S5. PTCs recapitulating the treatment responses of patients with GC and CRC.

Fig. S6. Characterization of breast cancer PTCs and comparison of PTC cell viability and the treatment responses of patients.

Fig. S7. PTCs recapitulating the treatment responses of patients with triple-negative breast cancer.

Fig. S8. PTCs recapitulating the treatment responses of patients with luminal A or B breast cancers.

Fig. S9. PTCs recapitulating the treatment responses of patients with HER2+ breast cancer.

Fig. S10. Generation of PTCs from the ascites of patient CS-G-907 for drug testing.

Table S1. Summary of compounds tested on CRC or GC PTC media.

Table S2. Efficacy and concentrations of drugs used in this study.

Table S3. Gastrointestinal PTC growth medium.

Table S4. Breast cancer PTC growth medium.

Table S5. The primary antibodies used in immunohistochemistry and immunofluorescence assays.

Table S6. A cancer panel targeting 105 cancer-related genes.

Data file S1. Fold change of 963 significant differentially expressed genes.

Data file S2. All detected mutations between tumors and PTCs.

Data file S3. The overlaps of mutations between tumors and PTCs.

Data file S4. Clinicopathological data for patients with CRC or GC.

Data file S5. Clinicopathological data for patients with breast cancer.

Data file S6. Original data for all the PTC drug tests.

Data file S7. BIC-seq2 segmentations.

Data file S8. CNV correlations.

Movie S1. CRC self-assembly into PTCs.

References (4382)


Acknowledgments: We thank the Institute of Molecular Medicine and National Center for Protein Science at Peking University for confocal microscopy and flow cytometry. We thank W. Guo at Zhejiang University, H. P. Cheng at Peking University, and B. Li at Stanford University for critical comment and discussion. Part of the analysis was performed on the Computing Platform of the Center for Life Science. Funding: This work was supported by the National Key Basic Research Project of China (2018YFA0108101 to J.J.X. and 2016YFC0207705 to R.X.), the National Natural Science Foundation of China (81827809 to J.J.X., 81421004 to J.J.X., 31870805 to J.-F.J., 11971039 to R.X., 71532001 to R.X., and 81773214 to A.W.), the Peking University Clinical Scientist Program (BMU2019LCKXJ011 to J.-F.J.), the Beijing Municipal Administration of Hospitals Clinical Medicine Development of Special Funding Support (ZYLX201701 to J.-F.J.), and “San Ming” Project of Shenzhen City (SZSM201612051 to J.-F.J.). Author contributions: S.Y. and J.J.X. developed and analyzed the PTC cultures. S.Y., J.L., and H.Z. did the high-throughput sequencing, and R.X., Y.X., D. Y., and J.W. analyzed sequencing data. B.Y. and Y.Y. did immunofluorescent staining and flow cytometry. A.W., S.W., Y.L., C.W., L.W., Z.B., X.J., X.G., Y.J., Z.J., Ziyu Li, and J.-F.J. performed surgery, isolated tumor tissues, and organized ethical approval. A.W., S.W., L.T., F.R., Y.Z., D.S., Zhongwu Li, and N.L. did the clinical outcome evaluation. S.Y., A.W., Y.L., S.W., L.W., X.Y., and X.G. helped to collect clinical samples and data. S.Y., R.X., A.W., S.W., J.-F.J., and J.J.X. participated in data analysis and project design. S.Y., R.X., and J.J.X. wrote the manuscript. All the authors reviewed and accepted the contents of the article. S.Y., R.X., A.W., and S.W. equally contributed to this manuscript. J.-F.J. and J.J.X. share senior authorship. Competing interests: J.-F.J. is the medical director of Peking University People’s Hospital. The Peking University has submitted two patents regarding PTC cultures [A method for the culture of primary colorectal cancer cells (201810607216.2); a medium formula for the culture of primary colorectal cancer cells (201810607728.9)]. J.J.X. and H.Z. hold shares of GeneX Health Co. Ltd. All other authors declare that they have no competing interests. Data and materials availability: All data associated with this study are present in the paper or the Supplementary Materials. Distribution of PTCs and deposition of DNA sequencing data in publicly available databases are regulated by the informed consent that participants to this study signed. PTCs, clinical outcome, DNA sequencing data, and safety data on a per-patient level can be obtained through the Institutional Review Board of the Peking University (bme{at} All other codes and materials used in this study are freely or commercially available.

Stay Connected to Science Translational Medicine

Navigate This Article