Introduction

Epithelial ovarian cancer (EOC) is the most lethal gynaecological cancer in the Western world, with 314,000 new cases and 207,000 deaths reported in 20201 and incidences are expected to rise2. High-grade serous carcinoma (HGSOC) is the most common and aggressive subtype, accounting for approximately 70% of EOC cases. HGSOCs are characterized by TP53 mutations and genome instability and are often diagnosed at a late stage, resulting in low survival- and high recurrence rates3.

Standard treatment for HGSOC consists of either primary surgery followed by six cycles of platinum-based chemotherapy (carboplatin/paclitaxel), or interval debulking preceded by three cycles of neoadjuvant carboplatin/paclitaxel chemotherapy and followed by another three cycles of adjuvant chemotherapy with the same drugs4. Subsequent maintenance therapy with either bevacizumab and/or a PARP inhibitor has further improved the clinical outcome5,6,7,8. Although approximately 75% of the patients demonstrate a complete clinical response after completion of primary treatment, including cytoreductive surgery and platinum-based chemotherapy, 70–80% of these patients will relapse and the EOC 5-year survival rate is only 35–57%9. Moreover, around 20% of patients with primary disease do not respond to this treatment at all10. The treatment outcome and quality of life for EOC patients could be improved by identifying (platinum) therapy resistant patients before treatment initiation and evaluating the patient’s sensitivity to alternative treatments, while for refractory patients, the systemic treatment and unnecessary toxicity could be spared. In addition, personalized decision-making on primary versus interval debulking could be applied when therapy responses are known upfront.

Although predictive biomarkers have been developed, there still is an unmet need to better predict responses to specific therapies11,12,13. Patient-derived organoids (PDOs) have been shown to represent patient tumour biology and mimic patient responses to drug treatments in various cancer types14,15,16,17. PDOs have been developed for ovarian cancer18,19,20,21,22. Although these models have proven to recapitulate clinical responses to treatment, they lack the original tumour microenvironment (TME), tissue heterogeneity is lost during expansion and model generation often takes too long to be implemented in clinical practice23. Direct ex vivo testing of tumour tissue obtained from patients offers a representative TME, minimal loss of heterogeneity24 and, most importantly, allows short turnaround times to support clinicians in treatment decision-making.

The current study investigated the predictive value of an ex vivo 3D micro-tumour testing platform for treatment-naive EOC patients that subsequently received platinum-based chemotherapy (Fig. 1A). The study aims to examine the opportunity for ex vivo drug sensitivity testing to support decision-making to improve treatment outcomes and reduce overtreatment and unnecessary side effects.

Fig. 1: Graphical abstract and sample inclusion.
figure 1

A Graphical overview of the study. Primary ovarian cancer patients with malignant ascites drainage were included in the study. Patients received neo-adjuvant drug treatment (carboplatin/paclitaxel) and response monitoring was performed by CA125 measurements and CT-scans to measure target lesion size. In the lab, micro-tumours were isolated for drug sensitivity testing. Results from ex vivo drug sensitivity testing in the lab were predictive for patient treatment outcome in the clinic. B Overview of sample inclusion criteria for study dataset. CV = coefficient of variation (stdevneg con/meanneg con). C Correlation of technical replicates (different plates from the same sample) for carboplatin (top, n = 18) and paclitaxel (bottom, n = 16) treatments.

Full size image

Results

Establishment of an ex vivo 3D micro-tumour testing platform

EOC patients with malignant ascites samples were included in the study (117 samples from 104 different patients, Fig. 1B). 24 samples did not pass quality control (QC) criteria due to lack of tumour material and were excluded. In the remaining samples an 80% technical success rate was obtained (%CV < 25%, high 3D gel quality, effect of positive control treatment observed), resulting in a final study dataset of 74 samples (from 65 patients) from which micro-tumours were isolated. Within this study dataset, there was no drift in drug sensitivity over time (Supplementary Fig. 2). Furthermore, high reproducibility was obtained between technical replicates for both carboplatin and paclitaxel treatments (identical treatments on separate plates, Fig. 1C). Taken together, these results validated the establishment of a robust and reliable platform for testing heterogeneous patient samples.

Isolated micro-tumours recapitulate patient tumour material

In line with expected tumour heterogeneity, H&E staining on the micro-tumours demonstrated patient-specific morphologies including differences in size, compactness and cystic morphology (Supplementary Fig. 3). High expression levels of common HGSOC markers (PAX8 and WT1) were observed in both the tumour samples evaluated in the participating hospitals as in the isolated micro-tumours, confirming the presence of tumour cells in the 3D ex vivo micro-tumour sensitivity assay (Fig. 2A, B, Supplementary Fig. 4A). Importantly, positivity scores did not correlate with drug sensitivity, showing drug sensitivity profiles were not directed by differences in tumour purity (Supplementary Fig. 4B). Comparable EpCam and Ki67 positivity were observed at assay start and end point confirming micro-tumour survival throughout the assay (Supplementary Fig. 5). Strong correlation was obtained for WT1 protein expression comparing original tumour tissue with matched isolated micro-tumours (Fig. 2C). PAX8, a transcription factor essential in embryogenesis and known to be highly expressed in EOC25, demonstrated elevated expression levels in all original tumour samples but only in a subset of freshly isolated micro-tumours (Fig. 2C). For most samples PAX8 expression levels were restored after culturing (Fig. 2D), suggesting that PAX8 protein levels were only temporarily decreased during transportation and/or processing and re-expressed during ex vivo testing.

Fig. 2: Isolated micro-tumours recapitulate patient tumour material.
figure 2

A Ovarian cancer marker expression in original tumour material determined in the clinic (left, n = 34) and isolated micro-tumours (right, n = 43). Percentage of samples positive for each marker is indicated on the top. B Representative images of patient and micro-tumour pathological stainings. C Tumour marker positivity comparing original tumour material to isolated micro-tumours. Numbers indicate the number of samples per combination of original/micro-tumour score. D PAX8 positivity score at the start and end point of the ex vivo drug sensitivity assay. Percentage of samples positive is indicated on the top.

Full size image

Since the BRCA mutation status is routinely measured in the hospital, we compared the BRCA mutation profile of the original tumour with the isolated micro-tumours for a subset of patients (Supplementary Fig. 6A). Importantly a high correlation was observed for these measurements, with 1 patient being classified as mutant in the original tumour and normal in the isolated micro-tumours. In addition, we correlated the BRCA status obtained in the hospital to functional ex vivo responses to olaparib (Supplementary Fig. 6B). As expected, most patients with a BRCA mutation demonstrated ex vivo responses to olaparib. Of note, the patient classified as BRCA mutant without response to olaparib was classified as normal in the micro-tumours (left top sample, Supplementary Fig. 6A). A small subset of the patients without BRCA mutation was responding to olaparib, which could be caused by alternative aberrations in the DNA repair pathways.

An advantage for ex vivo testing compared to PDOs is the presence of the TME. Indeed, different subtypes of T cells and macrophages were identified after isolation (Supplementary Fig. 7). Although most immune cells were located in the stroma, immune infiltration in the micro-tumours was also observed (Supplementary Fig. 7B) Similar to micro-tumour morphology, patient heterogeneity was observed for both immune cell presence and tumour infiltration (Supplementary Fig. 7).

Establishment of a clinical correlation cohort

To demonstrate the predictive value of the platform, patient-specific drug sensitivity profiles obtained from ex vivo 3D micro-tumour testing were correlated to clinical responses in HGSOC patients eligible for first-line treatment with carboplatin and paclitaxel (Fig. 3A).

Fig. 3: Clinical correlation study overview.
figure 3

A Correlation outline. The ex vivo drug sensitivity test was finalized within 2 weeks after sample collection (before patient treatment initiation) and drug responses were quantified by calculating AUCs from dose-response curves. Patient responses were monitored by CA125 measurements before treatment and after 2–4 cycles of chemotherapy (before interval debulking). Taking these measurements, patient CA125 decay rate was calculated and correlated to ex vivo drug responses. B Overview of patient inclusion criteria for clinical correlation cohort. C Ex vivo dose response curves of micro-tumours treated with carboplatin. Grey lines represent all patients included in the clinical correlation cohort. Example sensitive (green) and resistant (blue) patients are highlighted and representative images of 3D micro-tumours after treatment are shown. Points represent mean and standard deviation of technical replicates. D Same as C, but then for paclitaxel.

Full size image

Patients were included in the clinical correlation cohort based on subtyping, NACT treatment and CA125 measurements (Fig. 3B, Table 1). The obtained dose-response curves demonstrated a high diversity in carboplatin and paclitaxel sensitivity for the different patients included in the study (Fig. 3C, D). For example, VOA_063 and VOA_027 were resistant and sensitive to carboplatin and paclitaxel treatment, respectively (Fig. 3C, D).

Table 1 Patient characteristics of the clinical correlation cohort
Full size table

As expected, patients bearing BRCA mutations showed good clinical responses (high CA125 decay rate) (Supplementary Fig. 8A). However, BRCA mutations are only prevalent in a small subset of HGSOC patients and chemotherapy responses are dispersed in BRCA wildtype patients. CA125 baseline levels and Ki67 expression were a poor predictor for patient clinical responses (Supplementary Fig. 8B, C), confirming the need for alternative models and/or biomarkers.

Ex vivo micro-tumour responses correlate to clinical patient responses

We identified a strong correlation between the clinical and predicted (ex vivo) CA125 decay rates (Fig. 4A, r = 0.77). Similar to observed CA125 decay rates, predicted CA125 decay rates were significantly related to CA125 decay categorization according to gynaecological cancer intergroup (GCIG) guidelines and change in target lesion size (Fig. 4B, C for predicted decay rates, Supplementary Fig. 9A, B for observed decay rates).

Fig. 4: Ex vivo micro-tumour responses correlate to clinical patient responses.
figure 4

A Correlation between actual and predicted CA125 decay rates. Correlation: 0.77 (CI95: 0.63 - 0.87), r-squared:0.59, LOOCV correlation:0.71, LOOCV r-squared:0.51). B Predicted CA125 decay rates for GCIG response categories (CR complete response, PR partial response, SD stable disease). Statistical significance was determined using a Mann-Whitney U test (** p < 0.01). C Correlation of change target lesion size (%) between diagnosis and interval (before debulking) to predicted CA125 decay by ex vivo testing. Pearson correlation coefficient between predictedCA125 decay and the change in target lesion size: -0.49 (CI95: 0.83 - -0.07). D PFS rates for patients having high and low predicted CA125 decay rates (median split). Left graph: all patient included in the clinical correlation cohort (n = 37), middle: patients that did not receive interval debulking (n = 11), right: patients that did receive interval debulking (n = 26).

Full size image

Finally, predicted ex vivo CA125 decay rate was correlated to patient progression-free survival (PFS). When grouping patients on ex vivo predicted CA125 decay rate (median split) significantly prolonged PFS was observed in the patients with high predicted decay rates, compared to low predicted decay rates (Fig. 4D, left panel). This difference was only identified in patients with interval debulking (Fig. 4D, right panel). Patients without interval debulking have very short survival independent of chemosensitivity (Fig. 4D, middle panel), showing that both debulking and sensitivity to chemotherapy are critical for long term survival of ovarian cancer patients.

Overall, these results demonstrate a high predictive value of the ex vivo 3D micro-tumour testing platform for first-line platinum-based therapy response in ovarian cancer patients.

Ex vivo micro-tumour testing platform provides patient-centred decision support

In addition to the standard of care (SoC) first-line treatment (carboplatin and paclitaxel), the ex vivo drug sensitivity to second-line treatments (gemcitabine, doxorubicin, topotecan, olaparib and niraparib) was evaluated (Fig. 5A). Similar to the responses to carboplatin and paclitaxel, high variation in patient responses was observed for each of the treatments and patient-specific drug sensitivity profiles were identified (Fig. 5A, B). Some patients demonstrated high sensitivity (VOA_027_A_I) or resistance (VOA_038_A_I) to all tested drugs, while others showed a more diverse sensitivity profile. For example, VO_167_A_I was resistant to all treatments except for doxorubicin, while VO_091_A_II demonstrated intermediate sensitivity to all chemotherapeutic drugs but strong resistance to carboplatin (Fig. 5B). The results are listed as a relative score in an integrated patient report (Fig. 5C).

Fig. 5: Ex vivo micro-tumour responses to support patient treatment decision making.
figure 5

A Heatmap showing ex vivo patient responses to first and second-line treatments. Colour represents AUC normalized per treatment. B Ex vivo dose response curves of micro-tumours treated with carboplatin, paclitaxel, gemcitabine, topotecan and doxorubicin, respectively. Example patients for patient report (see Fig. 5C) are highlighted. Points represent mean and standard deviation of technical replicates. C Responses to measured treatment options for a specific patient relative to the full cohort of patients. Each horizontal green-to-blue bar indicates the range of AUCs measured for a particular drug in the entire cohort of patients. The vertical line indicates the patient’s percentile within this range.

Full size image

Discussion

Here, we present an ex vivo 3D micro-tumour testing platform that demonstrates precise treatment effect prediction of first-line SoC chemotherapy before treatment initiation. Ex vivo predictions were validated by multiple clinical parameters (CA125 decay, target lesion size and PFS). The platform identifies patient-specific sensitivity profiles by integrating and modelling carboplatin and paclitaxel ex vivo sensitivity and upon validation could provide clinical decision support prior to start of chemotherapy. Moreover, additional drugs can be tested in parallel, allowing clinicians to gain insight into alternative treatments next to the SoC. Results are generated within 2 weeks after sample collection, which aligns with the available time frame for treatment decision-making.

In the current study, a strong correlation was observed between predicted and clinical responses indicating that ascites-derived micro-tumours are representative for the patient’s cancer and its sensitivity to treatment. Reduced predictive accuracy for some patients likely reflects biological variation expected from fresh tissue tumour samples and clinical measurements. No correlation between absolute predictive error and potential confounding factors (e.g., age, number of treatment cycles, %CV, CA125 baseline and interval levels, CA125 interval, ascites volume, BRCA status, FIGO stage, WHO score, history of other malignancies, measurements on input material including size and number of micro-tumours) was found (analysis and data provided in Supplementary Data 4 and the Supplemental Code). Correlation coefficients could potentially be further improved when increasing standardization of clinical measurements such as fixed timing of repeated CA125 evaluations before each treatment, longer follow up after treatment completion and thorough evaluation on the effect of the treatment regimen (doses and intervals) on CA125 decay rates. In addition, correlation coefficients could be improved by further optimization and automation of the assay pipeline including the imaging-based analysis, increasing tumour purity after isolation and adding tumour representation markers to the predictions. Importantly, future studies would need to evaluate the correlation coefficient when including patients with a wider range of treatment intervals and treatment cycles.

A limitation of the current study is that only ovarian cancer patients with ascites containing sufficient tumour material could be included. However, in 92% of all HGSOC ovarian cancer patients ascites is present26. Additional improvement of the technical aspects of the assay will need to increase the technical success rate (currently 80%) and thereby the number of patients with successful drug response predictions. Example improvements are reducing the amount of tumour material needed per well, preselecting tumour material to prevent mesothelial interference and combining carboplatin and paclitaxel in single wells. Implementation of the use of solid tissue samples derived from biopsies or primary debulking procedures would further increase the number of eligible patients for predictive ex vivo tumour testing and would allow extension of the platform to for example low-grade ovarian cancers that are less prone to accumulate malignant ascites26,27. In addition, malignant pleural fluid samples could be used to further increase the number of patients included.

Shuford et al. previously reported predictive testing using resected tumour tissue from treatment-naive ovarian cancer patients28. Here, we establish reproducible predictive testing based on ascites derived tumour tissue. The disadvantage of using solid tissue is the low inclusion rate. About half of the HGSOC patients receive neo-adjuvant treatment and will therefore often lack tumour material before treatment initiation unless undergoing additional diagnostics like a laparoscopy. In contrast to the study of Shuford et al., which categorized tumour responses, our platform predicts chemotherapy sensitivity as a continuous variable. This is more robust to sampling biases and covers the full range of sensitivity profiles. Next, Senkowski et al. demonstrated a correlation of HGSOC PDO chemotherapy response to CA125 decay rates using a limited number of samples (n = 10)22. Although the correlation looks promising, PDO establishment took 89 days on average, which is not in line with clinical decision making. Ding et al. reported development of a micro-organosphere assay to measure patient responses29. Although feasible within clinical timelines, the platform lacks image-based analysis limiting interpretation of drug responses and evidence for correlation to clinical results is limited (8 patients) and only reported for colorectal cancer patients.

Although BRCA status is known to correlate to chemosensitivity30, we are not aware of an available clinical test to predict systemic treatment outcome in BRCA wildtype HGSOC patients before start of treatment. Over the last few years multiple efforts have been made to develop a predictive platform for (ovarian cancer) chemosensitivity in different domains and especially a rise in the development of multi-omics platforms has been observed28,31,32,33,34. Although very helpful, clinical implementation of these platforms might be difficult due to the density of tissue, data and complex analyses needed (RNA sequencing, DNA sequencing, radiomics, collection of tumour tissue, blood sampling, etc). In contrast, the platform presented represents a single functional assay simplifying integration into the healthcare system.

The current study explored the predictive value of ex vivo 3D micro-tumour testing for first line chemotherapy responses for primary ovarian cancer patients. In this treatment line, platinum-based therapy (carboplatin) is considered the backbone treatment and the response rate is rather high (~75%)9. Initial application of the platform at this disease stage would therefore likely be to select the best the combination therapy (e.g., gemcitabine or doxorubicin instead of paclitaxel) to further optimize the response rate.

In addition, it would be of high value to evaluate whether 3D ex vivo micro-tumour testing can produce predictive information that could support treatment decisions in later treatment lines (e.g., predict response for treatment refractory patients, or platinum insensitive patients that progress within 6 months). For application in recurrent disease, additional studies are currently being performed to identify if results obtained in primary disease are applicable and to validate clinical correlations using ascites and/or biopsies from recurrent patients. To prevent false negatives resulting in treatment withdrawal from patients that would have benefitted, strict cut-offs could be applied for selection of resistant patients.

Next to chemotherapies, good predictive biomarkers are also lacking for immunotherapies. This limits the introduction of these therapies for indications that typically show low response rates, like ovarian cancer. Ex vivo micro-tumour testing keeps the tumour cells in their native TME and TME components present in ascites have shown to be important for ovarian cancer progression35,36. Ongoing studies correlating ex vivo immune responses to clinical patient responses will reveal the potential of the platform to improve stratification of patients for immunotherapies.

Furthermore, the platform could be applied in a broader disease setting. For example, ascites accumulation is observed in many cancer types including breast, colon, gastrointestinal, pancreas and uterine cancer, highlighting the potential application of the platform in other cancer types. A comparable assay setup starting from pleuritis carcinomatosa samples and transurethral resections is currently evaluated. This could provide opportunities for additional patients and indications.

Altogether, our study demonstrated the significant value of ex vivo functional tumour testing to enable better stratification of ovarian cancer patients for effective therapy options. Application of the platform in the clinic could result in personalised and more effective treatment selection as well as prevention of overtreatment to reduce side effects and increase quality of life. Moreover, improved therapy response prediction could potentially support patient stratification for clinical trials and thus speed up the development of novel drugs for EOC patients. Prospective validation in larger patient series will further establish the value of implementation of ex vivo functional testing for ovarian cancer patients in the clinical routine, upon diagnosis as well as for platinum-resistant disease.

Methods

Patient inclusion

From 2019 to 2023, ovarian cancer patients from the Netherlands were enroled under a Western IRB-approved protocol into the TUMOVCA study (Medical Ethics Committee Leiden The Hague Delft, protocol number IRB P18.032). The IRB reviewed the protocol and confirmed the trial not being subject to the WMO, as patients received standard-of-care treatment and were not going through extra invasive procedures. The study was performed in accordance with the Declaration of Helsinki. The following Dutch hospitals positively assessed local feasibility for execution of the trial according to the approved protocol IRB (P18.032): Alrijne Ziekenhuis (Leiderdorp), Amphia Ziekenhuis (Breda), Catharina Ziekenhuis (Eindhoven), Haaglanden Medisch Centrum (Den Haag), Haga Ziekenhuis (Den Haag), Laurentius Ziekenhuis (Roermond), Leiden University Medical Center (LUMC, Leiden), Maasstad Ziekenhuis (Rotterdam), Nij Smellinghe Ziekenhuis (Drachten), Radboud UMC (Nijmegen), Rijnstate (Arnhem), Reinier de Graaf Ziekenhuis (Delft), Slingeland Ziekenhuis (Doetinchem), Tsjongerschans Ziekenhuis (Heerenveen), VieCuri Medisch Centrum (Venray). All participants were fully informed about the study’s objectives, procedures, and implications, and written informed consent was obtained prior to their inclusion in the study.

Patient samples and clinical data collection

Tumours were characterized by haematoxylin and eosin-stained (H&E) formalin-fixed paraffin-embedded (FFPE) slides and immunohistochemical staining at the local pathology laboratory. HGSOC diagnosis was based on pathology evaluation in the local hospitals. Patients were treated according to the standard of care (SoC) and treatment responses were monitored by CA125 decay, change in target lesions and progression-free survival (PFS). Measurements were performed according to local practice and CT scans were scored according to the RECIST criteria by an independent reviewer. Sample codes are defined by patient code (e.g., VO_xxx) and drainage number (e.g., A_I = first ascites drainage, A_II is second ascites drainage, etc).

Patient inclusion in the clinical and study dataset

Patients and samples were included in the study cohort based on the following selection criteria: 1) availability of malignant ascites, 2) enough tumour material when compared to mesothelial cells and 3) a successful drug sensitivity assay was performed (see Ex vivo drug sensitivity) resulting in a cohort of 74 samples from 65 patients (Fig. 1B).

For the clinical correlation cohort, patients were selected from the study cohort based on the following criteria: 1) primary HGSOC (n = 51), 2) treatment with neo-adjuvant chemotherapy (NACT) (n = 41) and 3) CA125 measurements available (n = 38) (Fig. 3B).

Micro-tumour isolation

Micro-tumours were isolated by centrifugation of ascites and homogenized by filtration (20–200 µm pluriSelect Life Science). To prevent interference of red blood cells, treatment with red blood cell lysis buffer (Invitrogen, Cat# 00-4300-54) was performed. Next, isolated micro-tumours were used for ex vivo drug sensitivity assays and/or immunohistochemistry.

Immunohistochemistry

Freshly isolated micro-tumours were fixed using 4% formaldehyde, washed with phosphate-buffered saline (PBS) and stored at 4 °C until processed for histology. Micro-tumours were paraffin embedded, sectioned, dewaxed (Biosystems, Cat# AR9222) and sections were rinsed with 100% ethanol and wash solution (Biosystems, Cat# AR9590). Antigens were retrieved using Epitope Retrieval Solution 1 (Biosystems, Cat# AR9961). Sections were rinsed with wash solution, blocked (Biosystems, Cat# RE7102), rinsed with wash solution and incubated with primary antibodies. After washing, sections were incubated with Bond Polymer Refine Detection Kit (Biosystems, Cat# DS9800) and counterstained with haematoxylin. Sections were dehydrated and embedded with Pertex before visualising/scanning on a slide scanner VS200 Olympus at x20. The following antibodies were used: anti-EpCam (concentration 1:200, Abcam, ab20160), anti-Ki67 (concentration 1:300, Abcam, ab16667), anti-WT1 (concentration 1.100, Abcam, ab89901) and anti-PAX8 (concentration 1:100, Abcam, ab53490). Positivity was defined by % of positive cells. Positivity score:0 = 0% positive, 1 = 0–5% positive, 2 = 5–25% positive, 3 = 25–50% positive, 4 = 50–100% positive). Binary positivity classification: positive: > 5% of cells expressing marker, negative: < 5% of cells expressing marker). Raw data for immunohistochemical stainings can be found in Supplementary Data 1–3.

Ex vivo drug sensitivity assay

A graphical outline of the complete assay procedure can be found in Supplementary Fig. 1. Assay was marked as successful when there was 1) enough material to perform at least treatment with SoC (carboplatin/paclitaxel) and 2) high assay quality with a coefficient of variation (%CV) < 25 (see all values in Supplementary Data 4) and 3) no other technical failure (high gel quality, treatment effect of positive control)). An overview of sample numbers in different categories can be found in Fig. 1B.

Micro-tumours in a hydrogel (Matrigel, Corning, cat#356231) suspension were seeded in 3D into a 384-well plate (Greiner, cat# 781091) as shown before37,38. Plates were incubated for 30 min at 37 °C, after which tumour culture medium (Advanced DMEM F12 (Gibco, cat# 12634028) with 2 mM Glutamax (Gibco, cat# 35050038), 100 U/mL PenStrep (Gibco, cat# 15140122), 0.01 M HEPES (Gibco, cat# 15630080), 1x antimycotic solution (Sigma-Aldrich, cat# A5955), 1x B27 (Gibco, cat# 17504001), 10 ng/mL Heregulin (ImmunoTools, cat# 11343045), 10 ng/mL hFGF-b (PeproTech, cat# 100-18B), 0.01 mM Y27632 (SelleckChem, cat# S1049) and 10 ng/mL hEGF (PeproTech, cat# AF-100-15) was added.

SoC monotherapies and positive control (Staurosporine, single concentration in technical quadruplicate) drugs were added within 24 h after seeding using the Tecan D300e Digital Dispenser with a randomized layout. Untreated negative control wells (16–20 technical replicates) were used to normalize drug responses and assess technical variation. The following drugs were included: carboplatin (Fresenius Kabi, Cat# 08717371358198, 16 concentrations in technical duplicates), paclitaxel (Selleckchem, Cat# s1150, 16 concentrations in technical duplicates), gemcitabine (Selleckchem, Cat# s1149, 16 concentrations in technical duplicates), doxorubicin (Selleckchem, Cat# s1208), topotecan (Selleckchem, Cat# s1231, 12–16 concentrations in technical duplicates), olaparib (Selleckchem, Cat# s1060, 12–16 concentrations in technical duplicates) and niraparib (MedChemExpress, Cat# HY-10619, 12–16 concentrations in technical duplicates). Following a 6-day exposure, samples were fixed and stained with Hoechst 33258 (Merck Life Science, cat# 94403) and Phalloidin-Atto 565 (Merck Life Science, cat# 94072) overnight at room temperature. Plates were washed 3 times with PBS before imaging.

Imaging was performed by an ImageXpress Micro XLS system (Molecular Devices, Wokingham, UK). Each well was imaged at 4x magnification as a stack of ~34 1024 × 1024 16-bit images with a z-distance of 50μm. Images were segmented and analysed using the Ominer software and the KNIME Analytics Platform (v.4.3.4), (Di et al., 2014; Sandercock et al., 2015). Per well, the number and structure of the objects (micro-tumours, nuclei and cytoskeleton) were quantified and aggregated, resulting in >100 morphological features from alive micro-tumours including number and mean area of micro-tumours and micro-tumour total area. Quality control checks (visual inspection of gel quality, effect of positive control treatment and %CV) were performed.

Curve fitting was performed for each feature using a 3-parameter hill-curve model between the log10 of the drug concentration and the aggregate response measures normalized to the untreated control. Fitting follows the strategy of Motulsky and Brown39 using the curve_fit method from the SciPy python package (v. 1.15.1). Curves were quality assessed based on a root mean squared error (RMSE) < 0.3 (values per patient, drug and feature can be found in Supplementary Data 4).

Clinical CA125 decay rate

Patient clinical responses were evaluated by change in CA125 blood serum biomarker levels (CA125 decay), which has proven to be a strong predictor for overall and progression free survival in ovarian cancer patients40,41.

CA125 measurements for treatment-naive patients were collected at diagnosis (CA125 baseline). Interval CA125 measurements were taken after 2–4 courses of SoC chemotherapy, prior to any surgery (CA125 interval). The CA125 measurement closest to treatment initiation (baseline measurement) and after platinum therapy but before interval debulking (interval measurement) were used for calculating CA125 decay rates. The date of the first chemotherapy administration was used as the start date.

The CA125 decay rate is the first-order decay between two CA125-measurements and their interval (in days). It is approximately the percentage of decrease per day, so a decay rate of 0.06 means approximately a 6% decrease in CA125 U/ml per day. The decay rate is converted into half-life as follows: T1/2 = ln(2)/decay rate.

Predicted CA125 decay rate

Ex vivo responses to carboplatin and paclitaxel were assessed by calculating the area under the curve (AUC) of fitted dose-response curves using the following features: (carboplatin total cluster area, paclitaxel cluster count and paclitaxel mean cluster area). The CA125 decay rate is predicted using a linear regression model combining cube-root of 1-AUC for each of the features and the CA125 decay rate.

Input data for the model can be found in Supplementary Data 4. A python script to recreate the model, determine the leave-one-out cross-validation performance, determine the 95% confidence interval of the Pearson correlation coefficient, sample discordance analysis and further model performance evaluations can be found in the Supplemental Code.

Target lesion size

Target lesion was assessed by central radiologist review of CT scans from 15 patients. CT scans were taken before starting chemotherapy treatment (baseline) and after 2–3 chemotherapy cycles/before interval debulking. CT scans were assessed according to RECIST 1.1 guidelines. In brief: in the baseline image, up to five target lesions were selected per patient and their longest diameters were summed. This number was compared to the summed diameters of the same lesions in the interval scan. In addition, the interval scan was assessed for unequivocal progression in non-target lesions or the appearance of new lesions.

GCIG classification for CA125 decay

Patients were classified based on the GCIG guidelines as follows: complete response (CR) = CA125 < 35 U/mL at interval measurement, partial response (PR) = CA125 < 50% at interval compared to baseline measurement and higher than 35 U/mL at interval measurement, stable disease (SD) = CA125 >50% and <200% at interval compared to baseline measurement. Patients with progressive disease (PD) = CA125 measurement > 200% at interval compared to baseline measurement were not present in the clinical cohort. Statistical significance for GCIG classification versus predicted CA125 decay was determined using a Mann-Whitney U test.

Progression-free survival (PFS)

PFS was defined as the time between the start of treatment and the occurrence of progression or recurrence. For PFS analysis, all patients from the clinical cohort were included, except for one patient that prematurely deceased due to heart failure (37 patients included). Interval debulking was performed in 26 out of 37 patients. For survival analysis (Kaplan-Meier), the cohort was split at the median predicted CA125 decay rate.

Result reporting

The dose-response curves of measured total area covered by micro-tumours were used to assess the effect of all chemotherapies. Per drug, the AUCs of all patients were ranked and the quantiles were calculated. For an individual patient report, the quantiles of that patient’s tumour ex vivo response were combined into a single overview.