Evaluating cell growth and hypoxic regions of 3D spheroids via a machine learning approach

A549 (human lung cancer cell line, CCL-185), HepG2 (human liver cancer cell line, HB-8065), Hep3B (human liver cancer cell line, HB-8064), HCT116 (human colorectal carcinoma cell line, CCL-247), and HT-29 (human colon adenocarcinoma cell line, HTB-38) were obtained from ATCC (American Type Culture Collection). The BEAS-2B (human bronchial epithelial cell line) was acquired from Kyung Hee University. Before the experiment, cells were thawed and allowed to acclimate for 3 cycles. A serum-containing medium diluted with 500 ml of Dulbecco's Modified Eagle's Medium (DMEM, Gibco$^\textrm{TM}$, 11995-065), 10$\%$ fetal bovine serum (FBS, Gibco$^\textrm{TM}$, 26140079), and 1$\%$ penicillin-streptomycin (Gibco$^\textrm{TM}$, 15140-122). Cultured adherent cells were harvested in T75 flasks (Sigma Aldrich) containing 0.25$\%$ trypsin-ethylenediaminetetraacetic acid (EDTA) (1X). Cells were counted using 0.4$\%$ Trypan blue (Sigma Aldrich) and placed in a T75 flask containing 2 × 10⁶ cells ml⁻¹ in 15 ml of DMEM, until the cell confluence reached 80-90$\%$ after 2 days (37 ^∘C/5$\%$ CO₂). To reduce the effect of FBS fluorescence on the proteins, a serum-free medium was prepared by adding only 1$\%$ penicillin-streptomycin to 500 ml of DMEM.

After thawing, all cells were subjected to centrifuging at 1300 rpm to remove the stock solution. After centrifugation, 1 ml of culture medium was added for resuspension, and each cell was seeded and incubated in a T75 flask containing 15 ml of culture medium. (37 ^∘C/5$\%$ CO₂). To stabilize the cells, two sub-culturing steps were performed. The culture medium was removed, and the cells were washed once with phosphate-buffered saline (PBS). Subsequently, 3 ml of 0.25$\%$ trypsin-EDTA (1X) was added to detach the cells from the culture dish bottom and incubated for 2 min. An additional 7 ml of DMEM was added, followed by centrifugation at 1300 rpm for 3 min to harvest the cell pellet. Following this, the cell count was determined, and $2.0 \times 10^4$ cells/well were seeded in an ultra-low attachment (ULA) 96-well plate (S-bio, MS-9096UZ) for 3D cell culture. The cells were cultured for 10 days to form and observe rotational ellipsoids. We utilized the IncuCyte S3 system (Sartorius, 60049, Germany) for continuous cultivation and captured images every 6 h over 10 days to monitor the growth and size changes of the spheroids in real-time. Afterwards, we utilized the Image-iT$^\textrm{TM}$ Red Hypoxia Reagent (Invitrogen, H10498, 1 mg) as a hypoxia tracker to effectively distinguish hypoxic regions based on changes in the SoS. The spheroids were treated with 20 nM of this reagent and cultured in the plates in which they had formed. Subsequently, the spheroid images representing the hypoxic region were measured using 4× magnification at intervals of 12 h for 7 days, which was implemented in the IncuCyte S3 system (37 ^∘C/5$\%$ CO₂).

Additionally, the Image-iT$^{TM}$ Red Hypoxia Reagent (Invitrogen, H10498, 1 mg) was used to further distinguish hypoxic regions within the spheroid structure. This reagent functions as a hypoxia tracker, allowing the differentiation of hypoxic areas within the spheroids and assessing the influence of spheroid size changes. The spheroids were treated with 20 nM of the hypoxia reagent and cultured within plates where the spheroids were formed. Real-time 4× images of the hypoxic region were obtained within the spheroid at 12 h intervals to track temporal changes that reveal the dynamic nature of the hypoxia response within the spheroid. Consequently, the acquisition of hypoxia images allowed us to explore the relationship between the SoS changes and the occurrence of SoH.

Sorafenib (SML2653-5MG, Sigma Aldrich), known for its potential to inhibit tumor growth and suppress angiogenesis, was procured to substantiate the feasibility of the predicted spheroid size application. Drug treatment was performed at a final concentration of 10 µM, guided by previous literature [24, 25]. The treatment was administered to HepG2 spheroids cultured and grown for three days, followed by observation for 48 h to determine alterations in SoS and SoH, which resulted by sorafenib treatment.

The automated cell size detection method begins by selecting a specific point within a cell. From this initial point, a region of interest (ROI) is defined that is large enough to contain the spheroid cells. The first step of the method is to calculate the average values for each RGB channel across the ROI, which are then used as thresholds to distinguish cell from non-cell areas. Starting from the central pixel of the ROI, the method compares the RGB values of each pixel within the ROI with these thresholds. Pixels having RGB values below the threshold are classified as part of the cell. This process of comparison and classification is systematically applied to each pixel within the ROI, allowing the occupied area of the cell to be determined by counting the classified pixels. The area represented by each pixel is defined as 7.974 µm² per pixel. To determine the SoS and SoH in each image, this per-pixel area measurement is multiplied by the total number of pixels identified as part of the cell.

We employed a two-layer one-dimensional CNN to predict SoS and SoH. Each convolutional layer in our model was equipped with a rectified linear unit (ReLU) activation function [26] and followed by a one-dimensional max-pooling layer with a stride and kernel size of two. Because the spheroid size is inherently a positive value, we configured the output of the last fully-connected (i.e. linear) layer to be treated as an absolute value. Our dataset was split 7:3 for training and testing. For SoS prediction, we used five early-stage sizes from a total of 41 images ( collected at 6 h intervals over 10 days) from a single well to predict the size for each day. Similarly, for SoH predictions, we used five early-stage sizes from 21 images (also collected at 6 h intervals over 10 days) from a single well for daily size predictions. The input format to the model was a two-dimensional matrix where the rows represented normalized SoS and SoH values by growth date and the columns represented the number of cells. To minimize the difference between the predicted and actual SoS/SoH values, we used the mean square error (MSE) loss function. We employed the Adam optimizer with a learning rate of 0.0002. The model was trained with a batch size of one over 500 epochs. All unmentioned hyperparameters followed the default configuration of PyTorch. The ML models were independently trained for each cell line using a single NVIDIA RTX 3090 Ti GPU. The details of the CNN model architecture and the reasons for the chosen hyperparameters are provided in figure S1.

The SoS and SoH are essential parameters for evaluating and designing studies of intercellular interactions in cell biology and biomedical research [27–29]. Specifically, accurately measuring and extracting the SoS from enormous amounts of cell images for high-throughput study are essential for various applications. The first two examples located in figures 1(a) and (c) demonstrate the ability to measure SoS accurately, while the others located in the figures 1(b) and (d) highlight the importance of high-throughput capacity for further investigations (figure 1).

Zoom In Zoom Out Reset image size

Accurate measurement of SoS is significant in microfluidic chip design. Complex structures necessitate optimized cell vessel widths, making precise spheroid measurement fundamental for efficient design [30, 31]. Moreover, precise measurement of SoS is a substantial parameter in 3D imaging. High-quality 3D spheroid imaging requires accurate consideration of changes in spheroid thickness and size, both laterally and perpendicularly [12, 32]. Issues such as focusing particularly arise for larger spheroids, in which case the light transmission decreases. As shown in figures 1(b) and (d), extracting sizes from several cell images is important for the highly reliable evaluation of drug efficacy and the toxicity of various materials (e.g. nanomaterials and micro-plastics). Many researchers use 3D spheroids of cell-ECM similar to the in vivo environment to evaluate drug efficacy and toxicity of various substances. For reliable evaluation of drug efficacy and toxicity, the high-throughput approach is required, which integrates the quantitative elements (e.g. SoS, SoH) into the qualitative analysis of intercellular interaction [33, 34].

Recognizing these challenges, our approach focuses on leveraging high-throughput methodologies to accurately measure the SoS and SoH. This is critical for accurate measurement and processing of SoS and sets the stage for subsequent analysis such as cell viability and drug efficacy. For this reason, we initially developed an automated detection method to accurately measure the SoS and SoH. Afterwards, we utilized the ML models that predict the SoS and SoH with high precision and suggested that the predicted sizes can be a standard for evaluating drug efficacy.

Spheroids grow in a spherical shape; therefore, the size measured in a two-dimensional image can represent the overall SoS and is an important parameter for various studies. The variation of SoS is one of the measures used to evaluate toxicity, and drug efficacy in spheroids. Since each zone of the spheroids (i.e. proliferating, quiescent, and necrotic) reacts with drugs and nanoparticles, causing changes in SoS, it is possible to easily assess the degree of toxicity and drug efficacy by measuring and comparing their sizes [8, 35]. Recently, image segmentation was applied to measure the SoS quickly and accurately. However, the segmentation method requires a laborious preliminary task of labeling to train the model and sufficient computer resources to process many cell images. For this reason, we developed a versatile, fast, accurate, and label-free automated detection method to measure spheroid size.

As shown in figures 2(a) and (b), the spheroids cells of HepG2 and A549 continuously grew for up to 10 days as they interacted with their environment. By utilizing the stably grown cell images, we scanned the cell images using our automated detection method and confirmed that our method successfully detected each area of cells. Specifically, we detected the exact area of the cell by preventing debris from being detected. We compared the regions detected using the IncuCyte S3 and our automated method to evaluate the accuracy of our automated detection method. The regions detected using our method were slightly larger than those detected using the IncuCyte S3 method over 10 days (figures 2(c) and (d)). On the other hand, this discrepancy suggests that our automated detection method detects cellular regions with greater accuracy than the IncuCyte S3 method. As shown in figure S2, the IncuCyte S3 system often misses cell edges and occasionally misidentifies non-cellular areas (e.g. scratch) as cellular regions. The overall consistent difference of approximately 2%–3% indicates that our automated detection method detected areas with high precision. In addition, our automated detection method scanned the images at high speed (i.e. 10 000 images per hour) by utilizing a lightweight algorithm for the large amounts of cell images. Taking these advantages into account, our accurate and fast automated detection method enables precise measurement of SoS changes in real time, comparable to conventional image detection methods such as ImageJ and IncuCyte S3 method.

Zoom In Zoom Out Reset image size

The advantage of predicting cell growth is that the desired SoS for various applications (e.g. evaluating drug efficacy or toxicity) can be determined without preliminary experiments. Additionally, prior knowledge of the growth curve of a particular cell line can be used to exclude unexpected outliers during real-time experiments, resulting in a reduced waste of resources. For this reason, we utilized the sizes of five cancer cell lines (HepG2, Hep3B, A549, HCT116, and HT-29) and one normal cell line (BEAS-2B) to train a CNN models on cell growth. By applying our automated detection method, the SoS of each cell line was obtained, and the one-dimensional CNN was trained to investigate the variation of cell growth in response to time (detailed model information is available in section 2.5). We intended the model to predict the SoS during the whole stages (0–10 days) when the early stages of cell growth (5 images, around 1 day) were inputted to take advantage of the aforementioned cell growth prediction. Despite these relatively constrained conditions, the cell growth of each of the six spheroid types is effectively trained into the CNN models with high accuracy, as shown in figure 3(a). To comprehensively analyze the model performance, we first froze each ML model for six cell lines, evaluated error metrics such as mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE), and summarized these results in table S1. From the statistical analysis, among the six cell lines, Hep3B has the highest error values with an MAE of 0.1098, an MSE of 0.0250, and an RMSE of 0.1581 (units: $10^5~\mu$m). Despite these error values, they are still quite small compared to the actual cell sizes (∼1–6, units of $10^5\,\mu$m), indicating that the ML models are significant in predicting the SoS. As shown in figure 3(b), the growth rate of some cancer cell lines (i.e. A549, HepG2, and Hep3B) significantly changed from day 5 to day 6, while that of other cancer cell lines (e.g. HT-29 and HCT116) progressively increase after the cells initially became smaller owing to transient aggregation. On the other hand, BEAS-2B spheroids are a normal cell line; therefore, we confirmed that it has limited growth, unlike the aforementioned cancer cell line. Our CNN-based ML models reliably predicted the variable cell growth for each of these cells. Therefore, our trained model effectively evaluated the cell growth and serves as a criterion to assess the status of cells.

Zoom In Zoom Out Reset image size

In terms of toxicology, it is difficult to distinguish between the toxicity of the target substance and the toxicity caused by hypoxia; subsequently, spheroids without hypoxic regions are required. On the other hand, in terms of pharmacology, the mechanisms underlying the activity of drug receptors are different for each zone of the spheroid; for this reason, it is necessary to have spheroids with enough hypoxic regions. Therefore, detecting the SoH with high accuracy and predicting its growth of spheroids are required for both toxicology and pharmacology [25]. For this reason, we effectively detected and predicted the SoH according to the growth date.

Figures 4(a) and (b) illustrate the detection of the hypoxic region of BEAS-2B and HepG2 spheroids. The first and second lines are the bright field (BF) and dark field (DF) images filtered in red, respectively, where the red-colored areas represent hypoxic regions. These different cell line-dependent hypoxic regions correspond well with the regions detected via the computer in the third row of figures 4(a) and (b). Notably, hypoxic regions in BEAS-2B are detected on days 3–4 and gradually increase in size, whereas hypoxic regions in HepG2 are detected on day 5 and significantly increase in size (figure S3). Subsequently, we trained a CNN model to predict the SoH for BEAS-2B and HepG2 cells. As shown in figures 4(c) and (d), our ML models learned the growth of hypoxia with high performance (R²: 0.9949 and 0.9953 for BEAS-2B and HepG2 spheroids, respectively). The scatterplot of HepG2 spheroids is discrete in the early stages, which is caused by the rapid growth of hypoxic regions in HepG2 spheroids. The previous high R² values indicate that the ML model has accurately learned the changes in SoS and SoH, which suggests that researchers can easily determine the SoS and SoH for a given day under the same experimental conditions using an early stage of few images.

Zoom In Zoom Out Reset image size

Sorafenib is a widely used drug in the treatment of liver cancer that functions by primarily inhibiting tumor growth through various mechanisms. It inhibits tumor angiogenesis, thereby cutting off the access of the tumor to essential nutrients and oxygen [36]. Additionally, sorafenib inhibits the activity of Raf kinase, an enzyme involved in signaling pathways that control cell proliferation and survival. By inhibiting Raf kinase, sorafenib effectively controls the proliferation of cancer cells and reduces tumor growth. This induces programmed apoptosis in cancer cells, thereby controlling uncontrolled tumor cell proliferation, and contributes to tumor shrinkage via the elimination of damaged cells [37, 38]. For these reasons, we treated spheroids of HepG2 cells, as a representative liver cancer cell line, with sorafenib for an effective evaluation of the drug efficacy. The efficiency of drug treatment was evaluated by comparing the variance of the SoS and SoH.

To demonstrate the applicability of ML models for predicting the SoS and SoH, we replaced the criterion for evaluating drug efficacy (i.e. control) with the predicted value from ML models. Figure S4 demonstrates that we used 10 early-stage SoS measurements, which correspond to the initial 2.5 days, as input to the pre-trained ML model to generate the SoS for a 5 day cell growth period. The pre-trained ML model delivers control values with high precision (R² = 0.9713) when given an unseen dataset. This high level of accuracy shows that the model has effectively learned the growth patterns of spheroids cells, indicating the pre-trained model is not overfitted. As shown in figures 5(a) and (c), the cellular growth of spheroids was inhibited from 3 days after drug treatment; this tendency is consistent with previous Wang et al's research showing that sorafenib worked effectively beyond 3 days after treatment [39]. In addition, the inhibition of the growth of spheroids was confirmed by the detection of hypoxic regions; the spheroid growth inhibition increased significantly after 3 days of drug treatment, indicating the inhibition of growth and hypoxia-induced apoptosis in the HepG2 spheroids (figures 5(b) and (d)).

Zoom In Zoom Out Reset image size