A compatible gravity-driven organoid perfusion (GDOP) platform for drug screening with sensitivity and toxicity process evaluation

Abstract

High-throughput experiments, unidirectional fluid replacement, real-time process monitoring, and simultaneous drug sensitivity and toxicity tests are hard to achieve on most existing tumor organoid chips. Here, we developed a gravity-driven organoid perfusion (GDOP) platform facilitating scalable throughput and supporting drug sensitivity and toxicity assessment on organoids. The unidirectional perfusion capability and optimized operational parameters of the GDOP chip were validated through fluid dynamics simulations. Using this platform, we successfully established uniform on-chip triple-negative breast cancer (TNBC) organoids, with endpoint detection results aligning closely with clinical diagnosis. Throughout the drug treatment process, we monitored and then analyzed the morphological and grayscale changes of the organoids. The sensitivity and toxicity tests revealed the optimal concentration range for the 3 chemotherapeutic drugs. In addition, on-chip brain organoids were established, which lays a feasible foundation for future drug toxicity tests of complex organoids. The GDOP platform, combined with its integrated evaluation method, provides a powerful and reliable approach for advancing organoid-based researches.

A gravity-driven organoid perfusion (GDOP) chip is developed to establish uniform TNBC organoids for three-drug sensitivity testing via morphological and grayscale changes, and to generate brain organoids supporting further drug toxicity research.

Introduction

According to the data from the World Health Organization, cancer is the second largest cause of death in the world, wherein lung cancer, female breast cancer, and colorectal cancer have the top three incidences¹. Due to the tumor heterogeneity, drug resistance, impact of tumor microenvironment, difficulty in early-stage diagnosis, metastasis, side effects, and medical limitations, it is incredibly challenging to cure cancer with fixed therapies. Specific cancer models for therapy optimization need to be developed according to cancer types^2–4. Organoid models have been proven to be capable of performing chemotherapeutic drug screening and replacing animal models to a large extent⁵. Drug screening experiments performed on organoids established from clinical tumor samples are relatively more accurate since those organoids have great parental resemblance⁶. Moreover, Operation convenience, high throughput, and precise quantification are required in most drug screening experiments, and organoid drug screening with microfluidic chips as the carrier is the current mainstream⁷. By mimicking the tumor microenvironment and realizing high-throughput screening, microfluidic chips can be used to efficiently and accurately evaluate the drug treatment’s efficacy on tumor cells. Its low sample consumption and real-time monitoring compatibility further improve the efficiency of drug screening and the feasibility of personalized therapy⁸.

For example, Nuttakrit Limjanthong et al. designed a gravity-driven microfluidic perfusion system based on a slow-tilting table with no pump or connecting tube and successfully cultured cells for 3 days⁹. Wang et al. developed a delicate pump-less platform that combines gravity-driven fluid flow and a rock motion to achieve reciprocating flow between two reservoirs for culture media recirculating, while the flow in the perfusion channel was continuous and unidirectional. Vascular endothelial cells cultured with the platform for 5 days showed changes matching cell responses to unidirectional laminar flows¹⁰. Lee et al. designed a gravity-driven flow system without tubing connections. Bidirectional flow of culture media was achieved, and human umbilical vein endothelial cells were cultured in the chip using circulating culture media¹¹. However, the above microfluidic chip, based on the gravity-driven media replacement method, can only realize 2D cell culture and cannot be applied to organoid establishment. Moreover, the shaker-based system makes it hard to achieve process monitoring.

Park et al. described a 3D plastic cytotoxicity assay device to assess the killing ability of lymphocytes in a 3D microenvironment by spatiotemporal analyzing the lymphocytes and cancer cells embedded in 3D extracellular matrix (ECM)¹². Yoshimitsu et al. developed a microfluidic perfusion culture system with a microchamber array that can be applied to self-renewal and differentiation studies of hiPSCs (human induced pluripotent stem cells). Immunohistochemical (IHC) analysis showed that the status of hiPSCs could be successfully controlled, and the effects of three antitumor drugs on hiPSCs were comparable between microchamber array and 96-well plates¹³. Lim and Park developed a microfluidic culture device with a concentration gradient generator that enabled cells to form spheroids in microwells and perfused in a cancer drug gradient for 3 days. The results show that the spheroid’s number, roundness, and cell viability were inversely proportional to the drug concentration¹⁴. Choi et al. designed a microfluidic device and established pancreatic ductal adenocarcinoma organoids in a microwell array. The responses to several chemotherapeutic drugs and the effectiveness of immunotherapy comprised of NK cells in combination with a novel immunomodulator were tested using the microfluidic cultured organoids¹⁵. The feasibility of microfluidic approaches in difficult-to-culture cancer spheroids cultivation was explored by Dadgar et al. Multichamber microfluidic devices with integrated microvalves were fabricated to enable serial seeding of chambers. Parallel testing of several drug concentrations was performed on ovarian cancer spheroids by passive-driven perfusion¹⁶. Sugiura et al. developed an external pressure-driven perfusion culture chip with an 8 × 5 microchamber array for parallel drug cytotoxicity assays. The cytotoxic effect assessments of seven anticancer drugs were successfully carried out¹⁷. Prince et al. developed an external pressure-driven microfluidic platform with an integrated gradient generator and fluid dynamic micropores. This device was used to culture cancer cells and test their response to drugs under different concentrations¹⁸. Although the above microfluidic systems can realize cultivation and drug testing of cells and/or organoids, most of them need complex tubing settings and/or external passive driving, which impedes their further applications in high-throughput drug screening (Supplementary Table S1). It is also inconvenient to obtain the endpoint viability of organoids directly. Among the current microfluidic systems, finding an accurately designed chip that can easily realize on-chip culture and monitor of tumor and regular organoids, drug sensitivity tests, and toxicity assessments is still demanding and challenging.

Triple-negative breast cancer (TNBC) is a breast cancer subtype with strong drug resistance, lack of effective clinical drugs, and more likely to recur, leading to a low five-year survival rate (77%) compared to breast cancer in general (91%)^19–21. This study established a tumor organoid model from the clinical TNBC sample on GDOP platform. The GDOP microfluidic chip that can be used for high-throughput breast cancer drug screening, automatic media replacement, drug sensitivity (targeting tumor organoids), and drug toxicity (targeting conventional organoids) evaluation, compatible with both process monitoring and endpoint detection, was thoroughly tested and discussed. The chip’s structure and treatment procedures were well designed to depress air bubbles and advance 3D organoid formation. The flow field coupled with particle distribution inside the chip was simulated and analyzed to optimize the uniformity of organoid dimensions and verify the sufficiency of nutrient delivery. Breast cancer cell line T47D was first used to validate chip parameters and functions. Then, TNBC organoids and 3D spheroids of regular breast cell line MCF10A were successfully established, monitored in-situ, and systematically tested with 3 chemotherapeutic drugs in the chip. Studying the neurotoxicity of drugs is also highly important²², and hiPSCs derived brain organoid model was established with the same chip, which further demonstrates the GDOP platform’s adaptability, stability, and robustness.

Results

GDOP microfluidic chip

Evaluating the culture and drug screening applicability of the microfluidic system for breast cancer organoids and brain organoids is one of the main aims of this study. Figure 1A is a physical photograph of the GDOP chip. As shown in Fig. 1B, the GDOP chip has an overall size of 37 × 44 × 13 mm. The chip is equipped with six independent perfusion channels. Each channel has 8 culture chambers with a U-shape bottom.

Fig. 1. Schematics of the GDOP chip.

A Physical photograph of the GDOP chip. B Structures of the GDOP chip; C Illustration of one channel in the chip and the details of the air trapping structure design; D Manufacturing and surface treatment procedures; E On-chip drug screening process (detailed description of the live-cell imaging system see Supplementary Fig. S2); F Establishment of hiPSCs induced brain organoids.

Culture media reservoirs are located at the left end of the chip, and waste containers are located at the right end. A serpentine flow resistance channel connects the culture media reservoir with the culture chamber channel (Fig. 1C). An air bubble trapping structure is provided at the junction between the flow resistance channel and the culture channel. If there are air bubbles that appear in the flow channel during cell seeding and/or fluid replacement processes, the bubbles will stick in the trapping structure because of buoyancy. The dead volume chamber is set to collect the cells left at the flow resistance channel and the lower end of the media reservoir during cell seeding processes. Therefore, the size uniformity of organoids in the culture chambers will not be affected by the cell residues flowing into the first chamber during fluid replacement processes. The reservoir can contain 100 μL culture media, and the head height of the reservoir is higher than the head height of the waste container. The chip can realize unidirectional automatic perfusion culture of organoids only by gravity without an external driving force. The assembly process of the chip is shown in Fig. 1D with details described in the experimental method section. Experiments and influence analysis of Pluronic F-127 solution concentration differences on hiPSCs growth were performed, and the results are presented in Supplementary Fig. S1.

Clinical-origin TNBC organoids drug sensitivity experiments and commercial MCF10A 3D cell spheroid toxicity tests have been achieved by seeding the two kinds of cell suspensions inside the chip after subculture, monitoring the spheroid size (equivalent diameter from projected area) and grayscale changes due to the process of sample growth and chemotherapy drug treatment, and testing the viability of the organoids by luminescence test at the end. In addition, the establishment and long-term culture from hiPSCs to brain organoids have been realized on the chip. The process is monitored by bright-field microscopy every day. The neural development and brain development are identified by qPCR and IF. The process is shown in Fig. 1E, F.

Gravity-driven fluid replacement experiments and hydrodynamic analysis

The size uniformity of organoids, the elimination of air bubbles, and the delivery method of nutrients can directly affect the consistency and effectiveness of drug screening²³. Through fluid dynamic simulation, the experimental parameters can be optimized, experiment efficiency can be improved, cell experiments can be reduced, and therefore, the entire cost can be saved. First, a simulation of the seeding process is executed. Due to the action of gravity and drag force, the uniformity of seeding cell distribution is mainly affected by the seeding speed. The seeding process simulation is conducted in two steps. One is to find the speed range that air cavity will not appear in the corners of the chip, as shown in Fig. 2A. When the seeding speed is less than 50 μL/s, the air cavity will not appear. The other is the particle tracking simulation of cell movement and distribution in each chamber (Fig. 2B). The specific cell distribution data are obtained and analyzed (Fig. 2C). The coefficients of variation (CV) of the cell numbers in each chamber at different seeding speeds are calculated. It can be obtained that when the cell seeding speed is 30 μL/s, the cells can be relatively evenly distributed in each chamber with a CV of 4.3%. In subsequent experiments, a seeding speed of 30 μL/s was adopted, resulting in more uniform organoids.

Fig. 2. Gravity-driven fluid replacement experiment and hydrodynamic analysis.

A Seeding flow rate-dependent air cavity phenomenon simulation; B Particles (cells) tracking simulation of the chip while cells seeding; C Numbers of particles (cells) in each chamber with different seeding flow rates; D Remaining culture media volumes in the reservoirs over time by experiments while gravity-driven fluid replacing(The experimental process is recorded in Supplementary Fig. S3); E Average inlet flow rate over time by experiments while gravity-driven fluid replacing; F Simulation of fluid flow field in the chip while gravity-driven fluid replacing; G Culture media flow speeds in each chamber while inlet flow rate is 0.08 μL/s.

Next, regular culture media replacement experiments are performed with the microfluidic chip. The amount of the culture media left in the reservoir is monitored over time. The initial culture media volume is 100 μL, and the remaining media in the last is less than 20 μL (a few media remains to ensure that no air will enter the inlet channel after evaporation). The entire fluid replacement process lasts for 900 s (Fig. 2D). By integrating the flow velocity (Fig. 2E), it can be found that the flow rate of the entire fluid replacement process gradually decreases and there is a very minimum drop within concerned time range. Therefore, the flow rate can be assumed as “relatively stable” over time in this study (Supplementary Note S1). The 4 mm height difference between the bottom of the fluid reservoir and that of the waste container ensures that the fluid replacing flow rate is always maintained at a high speed. Unidirectional perfusion can mimic in vivo blood flow, which provides a continuous supply of fresh media with necessary nutrients and oxygen provision, as well as effective metabolic waste elimination. Stable micro-environmental conditions (pH, osmotic pressure, etc.) can also be maintained. In addition, unidirectional perfusion supports drug delivery tests, reduces the risk of contamination, and contributes to cell polarization and structure formation. It plays an indispensable role in maintaining healthy organoid growth, accelerating functional expression, and widening the applications (e.g., drug screening and toxicological testing)²⁴. To achieve the unidirectional perfusion culture, flow in the entire culture region should be laminar, eddy-free, and turbulent-free. Take the steady flow rate (0.08 μL/s) from the experimental results (Fig. 2E) as the inlet flow rate to simulate the velocity distribution in the chamber. A probe is set at 300 μm above the chamber bottom (organoid height), then the flow line (Fig. 2F) and fluid velocity (Fig. 2G) in the chambers are obtained. It can be proved that the flow in the culture chamber is laminar without vortex, and the flow rates in each chamber are about the same.

Peclet number (Pe) is a dimensionless number that describes the relative importance of convection and diffusion effects during mass transfer. When the number is large, it means convection dominates the heat or mass transfer process, whereas when the number is small, it indicates that diffusion is more important²⁵.

$$Pe=\frac{UL}{D}$$ 1

In Eq.1, U represents the flow velocity, L represents the characteristic length, and D represents the diffusion coefficient calculated based on the diffusion constant of the green fluorescent protein (8.7 × 10⁻⁷ cm²/s) and sodium ions (1.33 × 10⁻⁵ cm²/s)²⁵. The Pe values of protein and sodium ions in the culture chamber are 85 and 5.63, respectively, indicating that convection is dominant and nutrients (such as hormones and growth factors) can be transferred passively from upstream to downstream in the microfluidic channel. At the same time, the metabolic waste can be pushed out of the chip. The above results prove that the microfluidic system can achieve unidirectional perfusion culture.

Establishment of T47D cell 3D spheroids in the microfluidic chip

To evaluate on-chip organoids forming performance, we use breast cancer T47D cells for principle verification because they have stable biological characteristics, mature culture methods, and a high correlation with clinical pathology samples. As shown in Fig. 3A–C, the formed 3D spheroids with specific seeding concentration are relatively uniform in size (Cell passage and on-chip spheroid formation see Supplementary Note S2). 3D spheroids in one channel slightly increase in size with the increasing distance between the culture chamber and the fluid inlet. The spheroids in the whole chip (48 chambers) have the smallest CV (6.95%) on their sizes when the seeding concentration is 0.75 × 10⁶/mL (Fig. 3D). Figure 3E is the image of bright-field and live/dead staining results at D4 when the initial seeding concentration is 0.75 × 10⁶/mL. The T47D 3D spheroids in the chip are at healthy status without obvious dead cells. In addition, a long-term on-chip culture is carried out, and the size of the spheroids is monitored. The size of T47D 3D spheroids after forming reduces in the first days and then gradually increases. The culture process lasts for as long as 22 days, which could meet the needs of most tumor organoid research (Fig. 3F).

Fig. 3. Growth status of T47D cells in the chip.

Formed 3D spheroid sizes in each chamber of the chip with 0.5 × 10⁶/mL (A) 0.75 × 10⁶/mL (B) 1 × 10⁶/mL (C) seeding concentrations. D CV of 3D spheroid sizes at different seeding concentrations; E Live/dead staining results and bright-field image of spheroids in one channel of the chip at D4 with initial seeding concentration of 0.75 × 10⁶/mL; F Size changes of a long-term cultured T47D 3D spheroid in the GDOP chip. The camera parameters are the same when taking the bright-field photos in (A, B, C). Scale bars = 500 μm.

Identification of TNBC organoid

It is essential to demonstrate that the patient-derived breast cancer organoids effectively retain the structural characteristics of the original specimens and the subsequent drug screening results are reliable (Tissue dissociation and organoid culture protocols see Supplementary Note S3). The HE and IHC staining results demonstrated that both tissues and tumor organoids exhibited enlarged nuclei with irregular morphology and increased nuclear-to-cytoplasmic ratio. In the organoids, the positive rates of TNBC markers, estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor (HER2), were all less than 1%, indicating negative expressions. The proliferation marker Ki67 expression was approximately 77%, which was closely resembled that in tissue samples (Supplementary Fig. S4). These results indicated that the TNBC organoid model could maintain the expression status of related markers corresponding to the primary tumor sample after subculture.

Drug sensitivity tests on TNBC organoids

Before conducting drug sensitivity experiments of breast cancer organoids, we have verified the feasibility by performing both well-plate based and chip-based tumor drug sensitivity experiments using T47D cell lines (Supplementary Figs. S5 and S6). We treated the organoids with 3 drugs having different working mechanisms. The drugs are (S)-10-hydroxy camptothecin (HY-N0095, MedChemExpress), Toyocamycin (HY-103248, MedChemExpress), and Docetaxel (HY-B0011, MedChemExpress) (Supplementary Table S2). During the treatment process, TNBC organoids are photographed to monitor their morphology and grayscale changes, and viability tests are performed on D4 (Fig. 4A). At that time, some cells inside the organoids turn dead and blackened. Therefore, the overall grayscale change could be used to reflect the degree of cell death within organoids. The grayscale deviation of the organoids is used to describe the level of unevenness in organoid appearance caused by cell death, which results from the degree of chemotherapy drug treatment (gray value is 0 = pure black, gray value is 255 = pure white)^26–28. The overall evaluation of the drug’s efficacy was also verified by the results of the endpoint cell viability assay. Spheroids’ size and grayscale information can be identified and obtained by automatic organoid boundary recognition using deep learning. Details of the correlation between size and grayscale, and viability of spheroids are shown in Supplementary Fig. S7.

Fig. 4. Drug sensitivity tests of (S)-10-hydroxy camptothecin on on-chip TNBC organoids.

A Molecular structure of the drug and the treatment schedule; B Bright-filed imaging, live/dead staining results, and size distribution of TNBC organoids cultured in the GDOP chip; Images (overall (D0) and detailed) and grayscale analysis results of TNBC organoids under drug treatment in 0 (C) 0.0005 (D) 0.005 (E) 0.05 (F) 0.5 (G) 5 (H) and 50 μM (I) Normalized size changes (J) grayscale (K) and grayscale SD (L) of organoids under different drug concentrations (Significance analysis: drug effects vs. control on D4, samples at each concentration n_chamber = 5); M Viability of TNBC organoids post-treatment (n_channel = 5). Scale bars = 200 μm.

Chemotherapy drug (S)-10-hydroxy camptothecin is a compound derived from the natural product camptothecin, which is a potent topoisomerase I inhibitor with anticancer activity and isolated initially from Camptotheca acuminata²⁹. In several research papers, it has been shown that the drug has a positive effect in breast cancer therapy^30–32. The drug treatment process, bright-field and fluorescence monitoring, and endpoint viability measurement are proven to be convenient and feasible during drug screening on GDOP chip. We cultured TNBC organoids and performed concentration gradient drug sensitivity tests on standard well-plates and GDOP chips, respectively. For TNBC organoids formed and cultured in the chip for 4 days, the chemotherapy drugs are added by gravity-driven perfusion twice from D0 of treatment to D4 daily. The on-chip test results of Docetaxel and Toyocamycin are presented in Supplementary Figs. S8 and S9. The on-chip experimental results (Fig. 4) are comparable to those of the well-plates (Supplementary Fig. S10) with (S)-10-hydroxy camptothecin. The loose morphology of organoids formed in well-plate hinders bright-field monitoring and data analysis, whereas their final viability test results are consistent with the organoids formed on-chip.

Figure 4B shows the live/dead staining of TNBC organoids in the chip and the size statistics of organoids at different chamber positions before the drug experiment. Each chamber only has one main TNBC organoid, and few dead cells can be noticed. The organoids’ size distribution pattern is consistent with T47D spheroids, mostly between 270 μm and 370 μm. This state is suitable for drug experiments. Figure 4C–I shows the bright-field images on D0, D2, and D4 (left) together with the mean value (MEAN) and the standard deviation (SD) of TNBC organoids grayscale over time with different drug concentrations (right). To ensure grayscale consistency in the image background, the obtained bright-field images were normalized regarding grayscale. As the drug treatment continues, the number and size of accessory spheres and dead cells near the main spheroid gradually increase.

The MEANs and SDs of grayscale between the control group and the 0.0005 μM group decrease slowly. The MEANs and SDs of grayscale in 0.005, 0.05, and 0.5 μM groups decrease linearly. This is a result of the relatively mild drug effect and gradually increasing dead cells inside the organoids over days. In the groups with high chemotherapy drug concentrations (5 and 50 μM), the MEANs and SDs of grayscale decrease rapidly after drug treatment and then reach equilibrium. This may be due to the direct toxicity of high drug concentrations. As shown in Fig. 4J, the size of organoids will still increase when no drug is added, or the drug concentration is low, whereas the organoids will shrink or even lose their clear boundary when the drug concentration is high. The figure does not present organoids in 5 μM and 50 μM groups because they are broken on D1. Figure 4K, L show that the MEANs and SDs of organoids grayscale decrease faster when the drug concentration increases. The final viability test shows that this chemotherapy drug’s half maximal inhibitory concentration (IC₅₀) on TNBC organoids is 0.1281 μM (Fig. 4M).

The results of organoid dimension, grayscale MEAN, and grayscale SD obtained by on-chip bright-field process monitoring are consistent with the viability test results. The drug efficacy and the drug concentration interval where the IC₅₀ value is located can be obtained during the treatment process, instead of merely on the endpoint. Therefore, the efficiency and accuracy of drug screening can be improved. Subsequent drug screening can be performed by tracking the morphological changes of organoids through process monitoring. To validate the reproducibility and reliability of the nested experimental design, we utilized 840 raw datasets (Supplementary Data 1) obtained from the aforementioned TNBC organoid drug sensitivity assays of (S)-10-hydroxy camptothecin on GDOP platform. We systematically assessed the consistency of parallel channel and technical chamber replicates, and further demonstrated the reliability and feasibility of this design via variance component analysis and linear mixed modeling (LMM). Detailed experimental methods and results are provided in the Supplementary Note S4.

Drug toxicity tests on breast cell line spheroids

Drug screening includes experiments on drug sensitivity tests on tumors and drug toxicity tests on other healthy organs. Toxicity evaluation of chemotherapy drugs on regular organoids is critical because it enables accurate drug efficacy and toxicity predictions in a 3D cellular environment close to an in vivo environment. Thus, the evaluation can help optimize personalized administration plans, reduce side effects, accelerate the process of new drug development, and lower the risk and cost. MCF10A cells show limited proliferative capacity and can form 3D structures under appropriate growth conditions. They can be used to assess the efficacy of drugs on normal mammary epithelial cells. Comparing the assessment results of MCF10A with the results of TNBC organoids can help determine the selectivity and safety of the drug. MCF10A cell suspensions in 0.5 × 10⁶ cells/mL are seeded into GDOP chips, and drug concentration tests are performed after four days of on-chip culture. We have tested all the 3 above mentioned drugs, and the drug toxicity results of Docetaxel and Toyocamycin are presented in Supplementary Figs. S11 and S12. Followings are detailed analysis of the drug sensitivity test of (S)-10-hydroxy camptothecin on MCF10A. Excluding the concentrations with poor efficacy in the TNBC organoid drug sensibility tests described above, 0.05, 0.5, and 5 μM drug concentrations are selected for drug toxicity experiments with MCF10A 3D spheroids. The drug toxicity results of adherent MCF10A cells can be found in Supplementary Fig. S13.

Figure 5A shows the bright-field photos taken on D0, D2, and D4. Combined with the statistical results of the normalized spheroids size, grayscale MEAN, and grayscale SD shown in Fig. 5B–D, it is concluded that the size of MCF10A 3D cell spheroids increases at low drug concentration and decreases at high drug concentration, grayscale MEAN and SD decreases with the increase of drug concentration and spheroids lost their boundary integrity and smoothness when undergoing the highest drug concentration. After measuring the endpoint viability of MCF10A spheroids, the results show that the drug under the concentration of 0.05 and 0.5 μM has a distinct ability to kill the TNBC organoids and at the same time gives less killing effect on MCF10A cells (Fig. 5E). It can be concluded that 0.05–0.5 μM may be a suggesting concentration range for TNBC chemotherapy with (S)-10-hydroxy camptothecin.

Fig. 5. Drug toxicity tests of (S)-10-hydroxy camptothecin on on-chip MCF10A spheroids.

A Bright-field process monitoring during drug toxicity tests; changes of normalized MCF10A spheroids dimension (B), grayscale MEAN (C), grayscale SD (D) under different drug concentrations (samples at each concentration n_chamber = 5) over time; E Comparison of TNBC organoid sensitivity and MCF10A spheroids toxicity results with chemotherapy drug treated under different concentrations, n_channel = 3. Scale bars = 300 μm.

In-situ hiPSCs self-assembly for on-chip brain organoid establishment

The main organs that need to consider the toxic effects of chemotherapy drugs include brain, liver, and heart³³. Utilizing corresponding organoids derived from hiPSCs can achieve high-throughput, authentic, convenient, in-vitro drug toxicity tests. The brain is the central regulator of all vital functions, and its dysfunction or damage can lead to severe consequences. Studying the neurotoxicity of drugs is highly significant, which is why brain is a commonly used model in drug toxicity studies. Furthermore, brain organoids are derived from hiPSCs, and their differentiation process and resulting structures are relatively complex and time-consuming. If this process can be achieved in our chip platform, the platform’s adaptability could be effectively demonstrated. Therefore, to further verify the compatibility of the constructed GDOP chips on different types of organoids, we used the on-chip inducing and differentiation process of brain organoids from hiPSCs as an example.

Unlike the conventional manual embedding cell spheroids into Matrigel³⁴, the GDOP chip does not need Matrigel embedding or centrifugation, and the spontaneous continuous perfusion can significantly promote the diffusion of nutrients in the culture chamber, making the nutrients distribute evenly inside the chip. As shown in Fig. 6A, hiPSC cells are seeded into the chip on D0, and a large number of EBs are formed in the chambers thereafter. The main effect of the M1–M4 medium is to maintain cell stemness, avoid premature differentiation, regulate the internal structure of brain organoids, and contribute to the formation of different brain regions. Changing the M5 medium daily after D15 can gradually promote the differentiation and growth of neuroectoderm into brain organoids.

Fig. 6. Establishment of on-chip derived brain organoids.

A Differentiation process of brain organoid from hiPSCs; B Brain organoids on the GDOP platform; C Calcein-AM/PI images of brain organoids; chip and bright-field process monitoring. D Bright-field images of brain organoids on D3, D7, D15, and D30; E Viability heat map of all brain organoids in one GDOP chip on D30 (values from CTG viability assay); F Distribution of viability values for different channels at D30; Size distribution G and grayscale distribution H of brain organoids in different channels on D3, D4, D7, D12, D15, and D30.

Figure 6B shows the overall view of the chips on D30. GDOP chip contains 48 brain organoids, each located in one of the 48 chambers. Figure 6C is a z-stack Calcein AM/PI fluorescent image (BZ-X800E integrated fluorescence imaging system, Keyence Corporation) of one brain organoid on D30. It shows optimal organoid viability rate with minimal dead cells. Figure 6D shows different stages of brain organoids differentiated from hiPSCs in one of the channels in a GDOP chip. 3D viability assessment at D30, Fig. 6E, shows the fluorescence values of all brain organoids on a single chip (higher values indicate better viability). The values ranged from 27,350 to 37,750 and mostly clustered around 30,000. Figure 6F is the distribution of brain organoid viability in different channels, and the values are close. Together with the staining results in Fig. 6C, it indicates that the brain organoids in the chip exhibited optimal conditions and minimal variability in viability metrics. The size dimension and grayscale information of the organoids during differentiation process were recorded and statistically analyzed. As shown in Fig. 6G, H, the organoids grew larger over time, with progressively reduced transparency and grayscale values. It is clearly shown in the scatter plot that points representing eight organoids from the same channel and those from three different channels have close grayscale values and size at the same recording time. These results preliminarily demonstrate that the GDOP system enables high-throughput culture of brain organoids, which grow healthily and exhibit good consistency in this system.

We performed three identification methods on both the brain organoids in well-plates and on GDOP system: QPCR, IF, and Calcium imaging. To identify the neural differentiation in brain organoids, qPCR detection have been performed concerning eight markers representing pluripotency and nerve development as shown in Fig. 7A. Compares to the control group (D0, hiPSCs), the expression levels of pluripotency markers OCT4 and c-MYC of brain organoids on D30 decrease greatly, the expression level of nerve-related marker PAX6 increases enormously, indicating the differentiation of neuroectoderm in the brain organoids. The immunofluorescence (IF) identification results also reveal high expression levels of neural precursor cell markers SOX2 and NESTIN (Fig. 7B), indicating efficient neural differentiation in the brain organoids. The IF result of marker PAX2, relating to brain regions (early developing forebrain, midbrain, and hindbrain)³⁵, also shows a high expression level. In addition, we studied the layered cortical architecture and detected the preplate marker TBR1 and deep layer marker CTIP2 by qPCR³⁶. These cortical-related gene expressions increase significantly. Then, IF analysis shows regions with positive TBR1 and CTIP2 stains, indicating preliminary formation of the layered cortical architecture in the cerebral cortex. Calcium imaging experiments on brain organoids in 96-well plates (Fig. 7C and Video S1) and on-GDOP platform (Fig. 7D and Video S2) were also conducted, respectively, at D50 to verify their electrophysiology function. Both groups exhibited at least three spontaneous fluorescence fluctuations at similar frequencies. The peak amplitude (ΔF/F) of the brain organoid in GDOP chip was slightly higher than that in well-plate, indicating that organoids on GDOP system have relatively higher neuroactivity.

Fig. 7. Identification of hiPSCs derived brain organoids.

A Real-time quantitative PCR detection and comparison of the gene expression in brain organoids (Control: result of D0). B IF identification and comparison of brain organoids, scale bars = 25 μm; Calcium datasets of brain organoid on D50 with bright-field images, calcium snapshots, and signals recorded at marked positions in well-plate C and on GDOP platform (D).

All the above experimental results are consistent with the findings of previous studies on brain organoid characterization^34,36, as they demonstrated that brain organoids established on the GDOP chip have the key features of the early human fetal brain, including neurogenesis, brain regionalization, cortical organization formation, and electrophysiological activity. Furthermore, the expression levels of various genes in brain organoids were consistent across different microchannels of the GDOP chip, and the expression of genes associated with neural and brain regional development in GDOP chip-derived brain organoids was comparable to or even higher than those of organoids cultured in 96-well plates. Taken together, these results confirmed that brain organoids underwent uniform differentiation and proliferation in the GDOP chips, the system’s suitability for large-scale screening applications could be validated.

Discussion

In this study, a gravity-driven organoid perfusion chip was proposed. Finite element simulations were performed to optimize the operation parameters of the chip and verify the effect of unidirectional perfusion. The size of cell spheroids inside the chip and the practicability of the drug treatment process were first validated by the T47D cell line. Then, the on-chip TNBC organoid model was established, and its IF identification results were consistent with the clinical histopathological test results of the parental tissue sample. The gradient concentration drug sensitivity tests were conducted on chip-cultured organoids and standard 96-well-plate-cultured organoids simultaneously with three commonly used chemotherapeutic drugs, among which the results of (S)-10-hydroxy camptothecin treatment have been thoroughly analyzed. The changes in organoid morphology and grayscale were monitored during the process. After the drug treatment tests, the relationship curve between the organoid viability and drug concentration was given by luminescent cell viability tests, and the IC₅₀ value in this circumstance was calculated as 0.1281 μM. Subsequently, drug toxicity experiments on on-chip MCF10A spheroids were performed. Combined with the result of the tumor drug sensitivity experiments, it could be suggested that the (S)-10-hydroxy camptothecin concentration for breast cancer chemotherapy would located between 0.05 μM and 0.5 μM. For the choice of patient-specific treatment plan, it should be further discussed in conjunction with clinical judgment. In addition, hiPSCs were seeded, differentiated into brain organoids, and cultured in the chip for 50 days. After qPCR identification (D30), IF identification(D30), and Calcium imaging (D50), the on-chip establishment of early-stage brain organoids was verified.

In summary, a novel microfluidic organoid culture platform for patient-specific drug response tests was described. (1) The organoid perfusion platform based on gravity-driven media replacement can realize in-situ monitoring, requires no external tubing and equipment, has high throughput, high efficiency, and a simple operation process. (2) The fluid dynamics simulation was used to aid the microfluidic chip design, optimize the operation parameters, and reduce cell experiments. (3) With systematical experiments, the diameter and grayscale analysis of organoids cultured in a GDOP chip were proven to be a promising treatment evaluation method. (4) Regular organoids and tumor organoids could be built on the same chip, and a complete drug screening process could be performed by experiments operated on a single chip. The platform can easily realize the whole process of drug sensitivity and toxicity evaluations and is expected to become an efficient means for cancer drug screening and personalized treatment development.

Despite comprehensive exploration of the GDOP system for anti-tumor drug screening, a critical aspect merits further investigation to enhance the clinical translation potential. Fluid dynamic simulation of the chip reveals that at 150 μm above the bottom, the flow velocity is 1.21 μm/s, generating a shear stress (≈0.006 dyn/cm²) around organoids during medium exchange, which is consistent with the physiological tumor interstitial flow range (0.3–3 μm/s) [ref. ³⁷]. Given that biomechanical cues are known to regulate tumor cell migration and invasion, future studies should explore the potential effect of shear stress control on tumor organoids developing and drug sensitivity.

Materials and methods

Fabrication and surface treatment of the GDOP chip

The GDOP chip comprises a thick reservoir layer and a thin culture chamber layer. Polydimethylsiloxane (PDMS, Sylgard TM 184, Dow) prepolymer was poured into 2 customized metal molds for the 2 layers and cured at 80 °C. Then, the 2 PDMS layers were irreversibly bonded together by O₂ plasma treatment (60 s, 80 W). The assembled chip was autoclaved at 140 °C for at least 1 h. 1% (w/v) Pluronic F127 in 1× PBS was added into the microfluidic channel at room temperature after autoclaving, and then the chip was stored at 4 °C for 12 h¹⁵. The PF127 solution should be removed completely from the chip by high-pressure clean gas after the low attachment surface treatment (See details in Fig. S1). To prevent air bubbles in the culture chambers, the PDMS chip should be vacuumed (20 kPa, 0.5 hr) before being seeded with cells.

Operation of the GDOP microfluidic chip

The operation methods of the GDOP chip include: (1) Seeding: The cell suspension was injected from the bottom of the reservoir at a constant rate of 30 μL/s using pipetting equipment. (2) Medium exchange: 100 μL of fresh medium was added into the reservoir, and then 100 μL medium was moved away from the waste container. (3) Culture: The chip was placed into a petri dish with a small amount of pure water around the chip to maintain humidity, and then the entire dish was transferred into an incubator(37 °C, 5% CO₂). (4) Process monitoring: The culture dish was placed on a microscope or monitoring devices for image recording. (5) Viability measurement: CTG reagent was injected into the chip slowly through the bottom of the reservoir. Then, the chip was shaken on a shaker under light protection for 30 min. Then all the liquid in the channel (approximately 40 μL) should be aspirated through the bottom of the reservoir and transferred to a black well-plate accordingly (a channel per well). After collecting all the liquid from all chip channels, fluorescence detection was performed using a microplate reader to obtain the viability values. (6) Organoid harvesting: After aspirating all the liquid from the reservoir and waste liquid container, the upper and lower layers of the chip were separated along the bonding surfaces, with the organoids remaining in the chambers of the lower layer. The organoids then could be slowly aspirated and removed to next position using a 100 μL pipette.

Cells dynamics modeling

The microfluidic culture channel in the GDOP chip linearly connects multiple microchambers, as shown in Fig. 1B. When cells are seeded, the cell suspension flows from left to right. We need to pay attention to the distribution and flow of cells in the culture channel. The maximum value of the Reynolds number in the channel is 40, as shown in Eq2, and the flow field state is laminar.

$$Re=\frac{\rho UL}{\mu }$$ 2

ρ is the culture media density, μ is the dynamic viscosity of the culture media, U is the flow velocity (here is the corresponding cross-sectional velocity while the inlet flow rate is 50 μL/s). The “Two phases-flow, phase field” module and “Particle tracing” module in COMSOL Multiphysics 5.6 were coupled to study the detailed cell distribution inside the micro culture channels. The governing equations include:

$$\rho \frac{\partial \mathit{u}}{\partial t}+\rho \left(u\mathrm{\cdot }\nabla \right)u=\nabla \mathrm{\cdot }\left[-p\mathrm{I}+\mathrm{K}\right]+\mathrm{F}+\rho g$$ 3

$$\rho \nabla \mathrm{\cdot }u=0$$ 4

Equation 3 is the Navier-Stokes equation with inertial term on the left, pressure gradient term, viscous term, body force term, and gravity term on the right. Equation 4 is the continuity equation for incompressible fluids. The chip is initially filled with air. While cells are seeding, the boundary of phases changes. The governing equation for interface tracking is derived from the Cahn-Hilliard equation and the chemical potential equation:

$$\frac{\partial \phi }{\partial t}+u\mathrm{\cdot }\nabla \phi =\nabla \mathrm{\cdot }\left[\gamma \nabla \left(\frac{\partial \mathrm{E}}{\partial \phi }\right)\right]$$ 5

where, $\phi$ is the phase variable function, t is time, u is the velocity vector, $\gamma$ is the migration rate constant, and $\mathbf{E}$ is the total free energy, including bulk distortion energy, anchoring energy, and mixing energy¹⁶. The change in cell momentum over time after the cell suspension entered the micro culture channels is equal to the force exerted (${\mathbf{F}}_t$) on the cell, as described below:

$$\frac{\mathrm{d}\left(mv\right)}{\mathrm{d}t}={\mathrm{F}}_t$$ 6

where, $m\mathit{v}$ is the particle (cell) momentum. Equations 2–6 are coupled to build the cell's dynamics model. However, for convenience of calculation, we have only performed unidirectional coupling for Eq. 5. Since the particle size is small and the influence of particles on the flow field can be ignored, the unidirectional coupling can hardly affect the calculation results, yet it can save a lot of calculation time.

Fluid dynamics simulation

In order to characterize the flow inside the micro culture chambers, fluid dynamics simulations were used to study the flow field distribution, the nutrient transport mode, and the optimal cell seeding flow rate. The “Two phases-flow, phase field” model in COMSOL Multiphysics 5.6 was used to obtain the flow field distribution in every microchamber. The inlet velocity was set at 0.08 μL/s (the steady speed obtained from previous experiments), the outlet pressure was set as atmospheric pressure, gravity was applied, and the process was solved based on the Navier-Stokes equation. By unidirectional coupling the “Particle tracking” model with the “Two phases-flow, phase field” model with particle properties consistent with the cells, the CV in each culture chamber was simulated and counted to find the optimal cell seeding speed.

Sample collection

The patient-derived tissue used in this study was de-identified residual clinical material, which has been reported in our previous work³⁸. All procedures were in accordance with the ethical standards of the research ethics committee in the First Affiliated Hospital of Soochow University, and approval was obtained from the committee responsible. Detailed clinical information of the samples is provided in Supplementary Table S3.

Histology processing and IHC staining

Tumor tissues and organoids were fixed in 4% paraformaldehyde (P0099, Beyotime) for 24 h following standard protocols¹⁹. Paraffin-embedded samples were sectioned into 5 μm slices for hematoxylin and eosin (H&E) staining and IHC. Before staining, paraffin-embedded sections undergo deparaffinization and rehydration, followed by antigen retrieval methods to expose hidden epitopes. Subsequently, tissue sections were incubated with the primary antibody (mouse anti-Ki-67 mAb (9449S, CST, 1:800), rabbit anti-ER mAb (ab16660, Abcam, 1:200), rabbit anti-PR mAb (ab101688, Abcam, 1:200) and rabbit anti-HER2 mAb (2165, CST, 1:200)) for 30 min at ambient temperature followed by a HRP-labeled goat anti-rabbit IgG or goat anti-mouse IgG coupled with alkaline phosphatase (30 min). Fast Red was used as the chromogen (15 min), and counterstaining was performed with hematoxylin for 5 min. Slides were dried in a 60 °C oven for 30 min and mounted with a permanent mounting medium (Micromount, Leica Biosystems).

Culture, monitoring, and drug sensitivity tests on TNBC organoids

(S)-10-Hydroxy camptothecin, Docetaxel, and Toyocamycin were used to conduct drug sensitivity tests on breast cancer organoids (adsorption analysis of three drugs on GDOP platform see Supplementary Fig. S14). Tumor cell suspension (0.75 × 10⁶/mL) was seeded into the GDOP chip and cultured for 4 days until the tumor organoids formed and became morphologically stable. 0, 0.0005, 0.005, 0.05, 0.5, 5, and 50 μM of the three drugs diluted in culture media were then added to the chip. Cell viability was measured by luminescent cell viability test (CellTiter-Glo® Luminescent Cell Viability Assay (CTG), Promega (Beijing) Biotech Co., Ltd) at the endpoint, and the IC₅₀ value was calculated. The tumor organoid cultivation and drug treatment process was monitored by bright-field (lab developed live-cell imaging system) and fluorescent inverted microscopy (SOPTOP ICX4I, Sunny Optical Technology (Group) Company Limited).

Induced pluripotent stem cell cultures

HiPSCs were acquired from the Institute of Biochemistry and Cell Biology at the Chinese Academy of Sciences in Shanghai, China. The cells were cultured on Matrigel-coated plates under feeder-free conditions using mTeSR Plus medium. The culture media was changed every 24 h to ensure optimal cell growth in a 37 °C, 5% CO₂ incubator.

Establishment and differentiation of brain organoids

HiPSCs can be differentiated into brain organoids through carefully controlled culture conditions. On D0, suspension hiPSCs at 0.5 × 10⁶/mL were prepared and seeded into the chip. The culture media for this suspension consists of mTeSR1 supplemented with Y27632 and bis-antibody. The chip was incubated overnight at 37 °C in a 5% CO₂ atmosphere to facilitate the formation of embryoid bodies (EBs). On D1, the culture medium should be removed from the waste liquid container and replaced with fresh mTeSR1 supplemented with Y27632. This procedure was repeated on D2. The chip was perfused with medium M1 (80% DMEM/F12, 10 μM Y27632, 4 ng/mL bFGF, and 20% KSR) On D3. The chip was perfused on D4, D5, and D6 with medium M2 (80% DMEM/F12, 100 ng/mL SB431542, 100 ng/mL dorsomorphin, and 20% KSR). Starting from D7, the chip was perfused with medium M3 (80% DMEM/F12, 100 ng/mL bFGF, and 20% KSR) for five days. Starting from D12, on-chip perfusion culture was continued with medium M4 (80% DMEM/F12 and 20% KSR) for three days. Finally, from D15 to D17, the chip was perfused with medium M5 (DMEM/F12 and Neural Basal (1:1), supplemented with 1% B27, 1% NEAA, 1% GlutiMAX, 1 μg/mL heparin, and 1% P/S). Identification of the established brain organoids occurred on D30 [ref. ³⁴].

Organoid analysis based on deep learning

In order to accurately analyze the status of organoids in the chips, a deep learning approach based on YOLOv8 was employed for organoid detection and segmentation. The model was trained on annotated 1440 × 1440 pixels organoid images (via Labelme v5.4.0; Fig. 8A) to learn morphological features. Data augmentation (rotation, scaling, translation; Fig. 8B) and 50 training epochs optimized the model’s generalization ability. During training (Fig. 8C), final training loss reached 1.2623 and segmentation loss 0.82746 (Fig. 8D), indicating strong dataset fitting. Key metrics included 0.9429 training accuracy, mAP50 of 0.99239, and mAP50-95 of 0.93176—demonstrating high prediction accuracy, target detection performance, and robustness across object scales/shapes. For prediction, RGB images (Fig. 8E) underwent histogram-matched normalization and grayscale processing (Fig. 8F) to mitigate acquisition noise. Inference yielded precise organoid identification, bounding boxes, and segmentation results. Quantitative analysis (e.g., example organoid: 29,740-pixel area, mean grayscale 84.42, SD 28.36; Fig. 8G) was performed using segmentation data. Detailed training configurations and validation strategies were shown in Supplementary Note S5.

Fig. 8. Process of organoid recognition based on deep learning.

A Dataset annotation; B Data augmentation; C Training process; D Model structure; E Original image; F Normalized grayscale image; G Predicted results. Blue arrows indicate the model training process, and red arrows indicate the predicting process.

This method achieved high-precision organoid identification and high-throughput analysis (1000 images/min), significantly enhancing efficiency, automation, and accuracy in organoid research and drug screening. Training results confirmed the model’s efficiency, accuracy, and lack of overfitting on training/validation datasets, validating its feasibility and superiority in practical applications.

Real-time reverse transcription polymerase chain reaction (qPCR)

Gene expression of brain organoids cultured for 30 days was investigated using real-time quantitative PCR (qPCR), as previously described. Total RNA was isolated from brain organoids using an RNA extraction kit (Vazyme). Total RNA was then reverse transcribed into cDNA using the Takara PrimeScript II first-strand cDNA synthesis kit (# 6110 A, Takara Bio LNC.). The synthesized cDNA was used for qPCR with Taqman Fast Universal PCR Mastermix (# 4366073, Thermo Fisher Scientific). Gene expression quantification was performed for the targets listed in Supplementary Table S4. The relative mRNA abundance was normalized to the expression of glyceraldehyde 3-phosphate dehydrogenase (GAPDH) and evaluated by the 2^-ΔΔCT method. The target gene expression was quantified by C_T value comparison and normalization using glyceraldehyde 3-phosphate dehydrogenase (GAPDH) as an endogenous reference.

Immunofluorescence (IF) assay

Brain organoids were fixed using 4% paraformaldehyde (PFA) for 15 min at room temperature and washed three times with 1× PBS buffer, after which tissues were embedded in O.C.T compound. A cryotome was used to prepare 10 μm-thick slices. Before performing IF (Antibodies see Supplementary Note S6), slides preserved with sample slice were washed in 1× PBS buffer followed by cell disruption treatment with 0.25% Triton X-100 for 5 min at room temperature. After washing with 1× PBS buffer, the slides were blocked with 10% goat serum for 1 h at room temperature. After washing again in 1× PBS buffer, the slides were added with primary antibodies (Rabbit anti-OCT4A mAb, Rabbit anti-TBR1 mAb, Rabbit anti-SOX2 mAb, Rabbit anti-PAX2 mAb, Rabbit anti-PAX6 mAb and mouse anti-NESTIN mAb) and incubated overnight at 4 °C. After washing with 1× PBS buffer, the slides were added with Alexa Fluor 488-labeled Goat Anti-Rabbit IgG (H + L) and Alexa Fluor 555-labeled Donkey Anti-Mouse IgG (H + L) antibodies and incubated for 1 h at room temperature. The nuclei were then counterstained with 0.5 μg/mL 4′, 6-diamidino-2-phenylindole (DAPI) for 5 min at room temperature. After washing with 1× PBS buffer, the slides were sealed with 50% glycerol and readied for observation. All samples were observed using con-focal microscopy (LSCM 2000, SIBET, CAS).

Calcium imaging of brain organoids

On D50 of brain organoid culture, organoids were transferred to glass-bottom dishes, with the medium removed by washing. Sufficient pre-prepared Fluo-4 Direct Working Solution was added to completely submerge the organoids. The reagent was prepared by diluting an appropriate volume of 1 mM Fluo-4 Direct stock solution (F10471/F10472, Invitrogen) at a ratio of 1:500 with pre-warmed (37 °C) HBSS (60147ES, Yeasen), with the addition of 1×PowerLoad™ Concentrate (F10471/F10472, Invitrogen). The dishes were then incubated for 60–90 min at 37 °C in a 5% CO₂ incubator protected from light. After incubation, the dishes were directly mounted onto the environmental control chamber of a microscope stage, followed by re-equilibration at 37 °C for 10–15 min to further reduce fluorescent background and stabilize signals³⁹.

For image acquisition and analysis, a con-focal microscope (A1R HD25, Nikon) was used for continuous imaging at a frame rate of approximately 1 fps. ImageJ (FIJI, National Institutes of Health) software was employed to measure the baseline fluorescence intensity (F₀) in distinct regions of the organoids. Regions with fluorescent transients were selected to determine the average fluorescence intensity (F) at different time points, and the relative fluorescence change ΔF/F was calculated by ref. ⁴⁰:

$$\mathrm{\Delta }F/F=\left(F-{\mathrm{F}}_0\right)/{\mathrm{F}}_0$$ 7

Calcium transient curves were generated by plotting ΔF/F₀ against time.

Statistical analysis

All experiments were repeated at least three times. GraphPad Prism 7 software was used to calculate the p-value using Student’s t-test for between-group comparisons and one-way ANOVA or two-way ANOVA for multiple comparisons. Data are expressed as mean ± standard deviation (SD), and statistical relevance was evaluated using the following p values: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Ethics statement

The ethical approval for sample collection and use was obtained from the Ethics Committee of the First Affiliated Hospital of Soochow University and remains valid for the current study (Approval number: 2024541, December 19, 2024). All ethical regulations relevant to human research participants were followed. Informed consent for the publication of personally identifiable information has been obtained from all participants.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

S.W. and J.W.—Conceptualization; S.W., H.M., and J.W.—Methodology; S.W., J.W.—Investigation; S.W., H.J., X.C., T.Y., and Y.B.—Validation; S.W. and H.J.—Visualization; S.W., T.Y., and H.M.—Formal analysis; S.W. and T.Y.—Data curation; S.W. and X.Z.—Software; T.Y., J.W., and G.L.—Funding acquisition; J.W. and G.L.—Resources; L.L., J.W., and G.L.—Project administration; J.W., T.Y., and G.L.—Supervision; S.W., H.M., and X.Z.—Writing—original draft; X.C., T.Y., J.W., and G.L.—Writing—review and editing; All authors have read and agreed to the published version of the manuscript.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Dr Shirley Tang and Dr Ophelia Bu.

Data availability

All the data in this study are available upon reasonable request from the corresponding author. Source data for 840 sample datasets of organoid dimension and grayscale obtained by image recognition are available in Supplementary Data 1. Source data underlying graphs can be obtained from Supplementary Data 2.

Code availability

The code and training results of the developed model have been deposited to GitHub repository (https://github.com/yesterdayzxl/organoidRecognizatrion).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-026-09973-5.