Introduction

Functional approaches in precision oncology entail exposing tumor cells to therapeutic interventions to identify drug candidates and to rapidly assess the potential efficacy of drugs for individual patients and therapy selection1,2,3,4,5,6,7,8. While a small number of actionable mutations are known, the majority of newly-diagnosed tumors lack any currently actionable genomic alteration9. By directly measuring the effect of drugs on tissues or cells, functional precision medicine methods can inform on the therapeutic resistance and sensitivity landscape of tumors without requiring full knowledge of the underlying molecular vulnerabilities a priori2,3,4,10,11,12,13,14,15,16,17.

The broadly adopted model systems for screening assays each have limitations. Two-dimensional (2D) cell lines are relatively simple and inexpensive to culture but fail to represent the architecture, behavior, and drug response of native tissue18,19,20,21. Mouse models have additional complexity but carry inherent, species-specific variations that limit their translation to human patients22,23. Patient-derived xenograft (PDX) models aim to better recapitulate human cancers yet are still constrained by the large cost and time associated with their use which translates to the impracticality of performing large drug screenings1,24,25. Three-dimensional (3D) tumor organoids are promising models for precision medicine that can be established rapidly and effectively from a variety of cell lines26,27 and tissue sources28,29, and accurately mimic a patient’s response to therapy3,4,8,12,13,14,15,28,29,30,31,32. They are physiologically-relevant, personalized cancer models well-suited for drug development and clinical applications8,28,29,31. We previously developed a screening platform that takes advantage of patient-derived tumor organoids seeded in a mini-ring format to automate high throughput drug testing, with results available within one week from surgery3,16,33. The key outstanding limitations to the broad adoption of organoid-based screenings remain the time-intensive and operator-to-operator susceptibility of the cell seeding steps as well as destructive, population-level approaches required for subsequent organoid analysis34,35.

To overcome these limitations, we developed an organoid screening approach that combines automated cell seeding via bioprinting with high-speed live cell interferometry (HSLCI) for non-invasive, label-free, time-resolved imaging. Bioprinting, a technique for precise, reproducible deposition of cells in bioinks onto solid supports, is rapidly gaining traction in cancer biology36,37,38,39,40. Within the deposited bioink, embedded cells can interact with physiological microenvironment components34. We then use HSLCI, a type of quantitative phase imaging (QPI)41,42,43,44,45, to rapidly monitor changes in dry biomass and biomass distribution of single organoids over time. The HSLCI measures the phase shift of light transmitted through the sample which is used to calculate the mass density of each organoid46,47. Biomass is an important metric of organoid fitness as its dynamics are the direct result of biosynthetic and degradative processes within cells46,47. HSLCI uses a wavefront sensing camera to enable reconstruction of the phase shift of light as it passes through a cell and interacts with matter43,46,47. Due to the defined linear relationship between the mass density and refractive index of biomolecules in solution, which is invariant with respect to changes in cellular content48,49,50,51,52, measured phase shifts can be integrated across the area of an image and multiplied by a conversion factor to obtain the dry biomass density of imaged cells46,47. This conversion factor is the inverse of the specific refractive increment, defined as the slope of the relationship between the refractive index and mass concentration of a solution46,47,50. QPI measurements of biomass changes allowed resolution of drug-resistant and drug-sensitive cells in 2D cell culture models within hours of treatment41,42,45,53,54,55,56,57. HSLCI-measured response profiles matched drug sensitivity from PDX mouse models of breast cancer42. However, HSLCI has so far been applied exclusively to screening of cancer cell lines grown in 2D or single-cell suspensions of excised PDX tumors41,42,44,45.

Here, we introduce a comprehensive pipeline that combines bioprinting, HSLCI, and machine learning-based organoid tracking and classification. Using cell lines as a model for 3D tumor cell growth, we demonstrate that bioprinted cells deposited in uniform, flat layers of extracellular matrix allow label-free, time-resolved, non-destructive quantification of growth patterns and drug responses at single-organoid resolution.

Results

Bioprinting enables seeding cells in Matrigel in suitable geometries for quantitative imaging applications

To address current limitations3,16,33 and facilitate non-invasive, label-free imaging of 3D organoids by HSLCI, we created an automated cell printing protocol using an extrusion bioprinter. As a base, we used an organoid screening platform in which organoids are grown from cells seeded in mini-rings of Matrigel around the rim of 96-well plates, with the empty center allowing the use of automated liquid handlers, facilitating media exchanges and addition of perturbagens3,16,32,33. We retained the empty center architecture but altered the geometry to bioprint mini-squares of cells in Matrigel (Fig. 1a). Positioning the sides of the square in the HSLCI imaging path allows sampling of a larger area and limits imaging artifacts caused by uneven illumination at well edges43 (Fig. 1a). Our bioprinting protocol entails suspending cells in a bioink consisting of a 3:4 ratio of medium to Matrigel. We then transferred this material to a print cartridge, incubated at 17 °C for 30 minutes, and bioprinted into each well at a pressure between 7 and 15 kPa, resulting in <200 µm prints on standard glass-bottom plates (Fig. 1b).

Fig. 1: Bioprinting enables the seeding of Matrigel-encapsulated cells optimized for efficient HSLCI.

a Schematic of wells with mini-rings (top) and mini-squares (bottom) relative to HSLCI imaging path (blue arrows). The top views (left) demonstrate that transitioning from rings to squares increases the area of material in the HSLCI imaging path. The side views (right) show that organoids in the square geometry align to a single focal plane better than organoids in a ring. b Plasma treatment of the well plate prior to printing optimizes hydrogel construct geometry. Bioprinting Matrigel onto untreated glass (left) generates thick ( ~ 200 µm) constructs that decreases the efficiency of organoid tracking by increasing the number of clusters out of the focal plane. Whole well plasma treatment (middle) increases the hydrophilicity of all well surfaces causing the Matrigel to spread thin ( ~ 50 µm) over the surface; however, the increased hydrophilicity also draws bioink up the walls of the well. Plasma treatment with a well mask facilitates the selective treatment of a desired region of the well (right). This leads to optimal constructs with a uniform thickness of approximately 75 µm across the imaging path. c Individual organoids can be tracked over time across imaging modalities. Five representative HSLCI images are traced to the imaging path across a brightfield image. d Cell viability of printed versus manually seeded MCF-7 cells in a Matrigel-based bioink, 1 h after plating. Data are presented as mean values ± SD. A one-way ANOVA was performed (n = 4, p = 0.0605) with post-hoc Bonferroni’s multiple comparisons test to compare all bioprinted conditions against the manually seeded control. Adjusted p-values were 0.0253, 0.6087, >0.9999, 0.1499 for print pressures 10, 15, 20, 25 kPa, respectively. e H&E staining shows the development of multicellular organoids over time regardless of seeding method. The prevalence and size of multicellular structures increases with culture time. Ki-67/Caspase-3 staining demonstrates that most cells remain in a proliferative state throughout culture time. While some apoptotic cells were observed in organoids cultured for 72 hours, the majority of cells show strong Ki-67 positivity. Ki-67 is stained brown, and Caspase-3 is stained pink. Scale bar is 60 µm. Source data is provided as a Source Data file.

We next coupled bioprinted cell lines grown in 3D clusters to an HSLCI platform. HSLCI uses a wavefront sensing camera and a dynamic focus stabilization system to perform continuous, high throughput, label-free, QPI of biological samples, tracking their biomass changes over time41,42. However, efficient high throughput QPI of 3D organoids using HSLCI is hindered by geometry considerations; when an object of interest is out of focus, measured phase shifts cannot be assumed to maintain a direct relationship with mass density43. Thus, we attempted to generate thinner layers of Matrigel to yield a relatively greater number of organoids in focus that can be quantitatively assessed at any given time. To generate thinner (<100 µm) constructs amenable to efficient, label-free HSLCI imaging, we increased the hydrophilicity of the surface of 96-well glass-bottom plates by oxygen plasma treatment58. We developed 3D masks composed of BioMed Amber Resin (FormLabs) to selectively functionalize the region of interest (Supplementary Fig. 1A–C). Bioprinting post-plasma treatment generated uniform mini-squares with organoids closely aligned on a single focal plane at ~70 µm thickness (Fig. 1b). The printed structures are consistent (Supplementary Fig. 1D) and amenable to massively parallel QPI by HSLCI as we aligned the legs of the bioprinted mini-square construct with the HSLCI imaging path (Fig. 1c).

Lastly, we verified that the printing parameters used did not alter cell viability by directly comparing MCF-7 cells manually seeded according to our established protocol3,16,33 to cells printed through a 25 gauge needle (260 µm inner diameter) using extrusion pressures ranging from 10 to 25 kPa. We did not observe any reduction in cell viability as measured by ATP release assay (Fig. 1d). These results are consistent with the existing literature as reductions in cell viability are often associated with higher print pressures (50–300 kPa)59,60. Taken together, this describes a method for bioprinting layers suitable for HSLCI imaging without impacting cell viability, while supporting automated liquid-handling for high throughput applications3,16,33,61,62.

Bioprinted tumor cells maintain histological features of manually seeded 3D cultures

To verify that bioprinting did not perturb tumor biology, we directly compared the histology and immunohistochemical profiles of bioprinted and hand-seeded cells from two breast cancer cell lines, BT-474 and MCF-7. These lines were selected for their differing molecular features such as their human epidermal growth factor receptor 2 (HER2) and estrogen receptor (ER) status63,64. We grew organoids from cells seeded in maxi-rings (1 × 105 cells/ring) to obtain sufficient material for downstream characterization. Cells were either manually seeded into 24 well plates3,16,33 or bioprinted into 8-well plates at an extrusion pressure of 15 kPa. The bioprinted cells and cell clusters were morphologically indistinguishable from manually seeded ones in brightfield images and hematoxylin and eosin (H&E)-stained sections taken 1, 24, and 72 hours after cell seeding (Fig. 1e). Both bioprinted and manually seeded cell clusters grew in size over time and bioprinting did not alter their proliferation rates (Ki-67 staining) or apoptosis (cleaved caspase-3; Fig. 1e). Hormone receptor status was in agreement with literature reports for both cell types64,65,66,67 and unaltered between bioprinted and manually seeded cells as shown by immunohistochemistry (IHC) for HER2 and ER (Supplementary Fig. 2). Chaperone protein expression also remained consistent with manually printed cells (Supplementary Fig. 3). Thus, bioprinting did not influence histology for the tested cell types.

Bioprinted and manually seeded cells are molecularly indistinguishable

While bioprinted cells are histologically indistinguishable from manually seeded ones, this does not preclude molecular changes caused by the printing process. We therefore performed a detailed analysis of the transcriptomes of manually seeded and bioprinted cells 1, 24 and 72 hours post-seeding. By pooling RNA sequencing (RNAseq) data across two independent experiments, we assessed the distributions of 30,544 RNA transcripts and found no significant differences between seeding approaches (Fig. 2a). The overall transcriptomes of manually seeded and bioprinted cells were highly correlated (Fig. 2b), with no individual transcripts differing significantly in abundance for either cell line (FDR < 0.05, t-test; Fig. 2c).

Fig. 2: Bioprinting does not significantly alter transcriptomes.

a Distributions of total number of RNAs detected (above) and RNA abundance (below) measured as transcripts per million (TPM) are similar between manually seeded (left) and bioprinted (right) cells. n = 30,544 transcripts were assessed across two independent experiments. Boxplot centers represent each median, edges of the boxes represent the 25th and 75th quartiles, whiskers represent the range of the data, and individually plotted points are outliers greater than 1.5 times the interquartile range from the median. b Spearman’s rank correlation of RNA abundance of manually seeded and bioprinted cell line 3D models at three time points (t = 1, 24, and 72 hours post-seeding). We found strong correlations between RNA abundance in manually seeded and bioprinted cells for both cell lines. p-values are derived from a two-tailed test of correlation between paired samples. c Volcano plots of paired, two-tailed t-test results comparing the RNA abundance of manually seeded and bioprinted MCF-7 and BT-474 with unadjusted p-values (left) and false discovery rate (FDR) adjusted p-values (right). No transcripts were preferentially expressed based upon the seeding method for either cell line (n = 0 out of 30,544 genes, FDR < 0.05, t-test). d Distribution of percent spliced in (PSI) exons are similarly distributed among BT-474 (top) and MCF-7 (bottom). The distribution of PSI is similar between manually seeded (left) and bioprinted (right) cells. PSI of 1 indicates that the exon is exclusively included, while a PSI of 0 indicates that the exon is exclusively excluded. Source data is provided as a Source Data file.

We next examined pre-mRNA alternative splicing events since these can induce functional changes even in the absence of variations in mRNA levels68,69,70. We found that the density of exon-inclusion and exon-skipping isoforms was unchanged (Fig. 2d). Similarly, we found no association between the seeding method and the number of fusion transcripts (p = 0.0519, Mann-Whitney U-test), although a fraction of samples had large numbers of RNA fusions, reflecting the wide-spread trans-splicing and genomic instability of immortalized cell lines71 (Supplementary Fig. 4A). Finally, there were no significant differences in the number or nature of RNA editing sites between printed and manually seeded samples (p = 0.977, Mann-Whitney U-test; Supplementary Fig. 4B). These findings demonstrate that the bioprinting protocol does not significantly impact the molecular characteristics of tumor cells as measured by bulk RNA sequencing.

Machine learning-based image segmentation and classification enables single-organoid analysis

Our complete pipeline includes cell bioprinting (day 0), organoid establishment (day 0-3), and full media replacement (day 3, Fig. 3a) followed by transfer to the HSLCI incubator. Within 6 hours of media exchange, we started to continuously image the plates through 72 hours post-treatment. At the end of the imaging period, we performed an endpoint ATP release assay to assess cell viability (Fig. 3a). For downstream data processing, we first converted interferograms collected by HSLCI to phase shift maps using the SID4 software development kit (GPU version, v741)43. Then we performed image segmentation and single-organoid level analyses using two types of machine learning algorithms (Fig. 3a).

Fig. 3: Bioprinting enables single-organoid tracking with high-speed live cell interferometry.
figure 3

a General schematic of the pipeline. Extrusion-based bioprinting is used to deposit single-layer Matrigel constructs into a 96-well plate (Day 0). Organoid model establishment and growth (Day 0–3) can be monitored through brightfield imaging. After treatment (Day 3), the well plate is transferred to the high-speed live cell interferometer for phase imaging (Day 3–6). Coherent light illuminates the bioprinted construct and a phase image is obtained. Organoids are tracked up to three days using the HSLCI and changes in organoid mass are measured to observe response to treatment. b Total number of tracks (left) and mean number ± SD of tracks per well (right) for cell clusters from each cell line at four time points (t = 6, 24, 48, and 72 hours after treatment). The total number of tracks across replicate wells treated with vehicle (1% DMSO) was 921 for MCF-7 (n = 12 wells) and 438 for BT-474 (n = 12 wells) c Mass distribution at four time points (t = 6, 24, 48, and 72 hours after treatment). Black bars represent the mean with error bars representing the standard deviation. Mean and standard deviation calculated based on n = 804, 855, 859, 803 for MCF-7 and n = 421, 420, 423, 402 for BT-474 cell clusters at t = 6, 24, 48, and 72 hours, respectively. d Hourly growth rate (percent mass change per hour) of tracked MCF-7 (left) and BT-474 (right) cultured in 1% DMSO. Data presented as mean ± SEM for each hour calculated based on growth rate data for n = 921 MCF-7 and n = 438 BT-474 tracked clusters. e Representative HSLCI-acquired phase images at four time points (t = 6, 24, 48, and 72 hours after treatment). Brightfield images taken immediately before treatment are shown on the left. f Calculated mass of each representative organoid over time. Source data is provided as a Source Data file.

To reliably identify unique cell clusters within each imaging frame despite the presence of background noise, debris, and out-of-focus organoids, we performed image segmentation using a U-Net architecture72 with ResNet-3416,73 as the backbone. U-Net, a type of convolutional neural network (CNN), consists of an encoder that extracts rich feature maps from an input image and a decoder that expands the resolution of the feature maps back to the image’s original size. The long skip connections between the encoder and decoder propagate pixel-level contextual information into the segmented organoid masks. The resulting segmentation images are very detailed even when provided small training datasets. We used a training dataset consisting of manually labeled organoids in 100 randomly selected imaging frames. This model created binary masks indicating whether each pixel of the image belonged to an organoid or the background with a mean Jaccard Index of 0.897 ± 0.109 at the 95% confidence level. The CNN reliably created masks omitting phase artifacts resulting from aberrant background or out-of-focus organoids (Supplementary Fig. 5).

Next, we determined the mass of each cluster in segmented masks by integrating the phase shift over the organoid area and multiplying by the experimentally determined specific refractive increment48,50,51,54,74. We assembled data from discrete segmented cell clusters and organoids into time-coherent tracks using TrackMate75,76 and filtered mass versus time data for each tracked object using a XGBoost classifier77 trained to exclude organoids moving in and out of focus, frequently overlapping and/or separating from neighboring clusters, or incorporating debris (Supplementary Table 1). We validated the model via cross-validation with 3-fold resampling of the sample population. The 3-fold resampling cross-validation score of the classifier was 93.1 ± 1.3% with a 0.966 ± 0.020 AUC (Supplementary Table 2). We also observed trends in the features used to classify each track, with excluded tracks typically having an increased number of missing frames as well as smaller interquartile ranges, and smaller initial mass (Supplementary Figure 6).

Trends in mass accumulation can be quantified by HSLCI with single-organoid resolution

HSLCI-based imaging allowed continuous tracking of n = 921 MCF-7 cell line-derived organoids in 12 replicate wells (median: 78.5/well) and n = 438 BT-474 in 12 replicate wells (median: 36/well, Fig. 3b). Due to cells moving in and out of the field-of-view (FOV), the number of tracked objects at each time point varied slightly. Overall, we tracked an average of 821 ± 28 MCF-7 and 412 ± 9 BT-474 cell clusters at any given time throughout imaging (Fig. 3b).

Unlike chemical endpoint assays or other live imaging modalities, HSLCI-based imaging facilitates parallel mass measurements of individual cells and organoids. The initial average measured mass at the start of imaging, corresponding to approximately 78 h total in culture, was larger for MCF-7 (1.36 ± 0.84 ng) than BT-474 (1.12 ± 0.61 ng, Fig. 3c). Considering a single cell mass of 200–300 pg55, clusters contained ~4–7 cells after 3 days in culture as measured by HSLCI, consistent with the observed histopathology (Fig. 1e). The difference in size between MCF-7 and BT-474 persisted throughout the entire imaging duration (Supplementary Table 3). BT-474 grew at a rate of 0.80 ± 6.07% per hour while MCF-7 demonstrated slower average hourly growth rates (0.33 ± 4.94% per hour, Fig. 3d). The growth rate of the BT-474 in 3D was slightly slower than BT-474 cells observed after 6 hours in 2D culture ( ~ 1.3%), while the MCF-7 in 3D showed a much lower growth rate than previously reported 2D cultures ( ~ 1.7%)55. We also observed positive associations between initial mass and growth rate in organoids originating from both lines; however, the association between these factors is stronger for MCF-7 cells (Supplementary Fig. 7). The data presented provides evidence that cell-line specific 3D growth characteristics can be quantified by HSLCI.

Drug responses in 3D cultures can be quantified by HSCLI

We then tested the utility of our platform in detecting drug responses in high throughput 3D organoid screenings (Fig. 3a). As proof-of-principle we tested staurosporine, a non-selective protein kinase inhibitor with broad cytotoxicity78, neratinib, an irreversible tyrosine kinase inhibitor targeting EGFR and HER279, and lapatinib, a reversible tyrosine kinase inhibitor also targeting EGFR and HER280. We tested staurosporine at 0.1, 1, and 10 µM, while neratinib and lapatinib were screened at 0.1, 1, 10, and 50 µM (Fig. 4 and S8). The concentration ranges tested include and extend beyond the maximum plasma concentrations reported for both lapatinib (4.2 μM)81 and neratinib (0.15 µM)82. We excluded wells treated with 50 µM neratinib from image analysis because drug precipitation at this concentration resulted in optical opacity. Precipitates and other opaque materials limit data acquisition by causing uninterpretable phase images.

Fig. 4: HSLCI enables high throughput, longitudinal drug response profiling of 3D models of cancer.

a Representative HSLCI-acquired phase images of MCF-7 and BT-474 grown in 3D and treated with 10 µM staurosporine, 10 µM neratinib, and 10 µM lapatinib. b Each bar represents the mass distribution at 6-, 24-, 48-, and 72 hours post-treatment (left to right). Black horizontal bars represent the median with error bars representing the interquartile range of the distribution. Sample sizes, summary statistics, and exact p-values for each condition are available in Supplementary Table 3. Statistical significance was assessed using Kruskal-Wallis tests. For samples with p-values lower than 0.05, we performed two-tailed Mann-Whitney U-tests against the vehicle control at the respective time points. p < 0.05 is denoted by *, p < 0.01 is denoted by **, and p < 0.001 is denoted by ***. c Hourly growth rates, calculated as percent mass change per hour, are compared between organoids treated with 10 µM staurosporine and vehicle, 10 µM neratinib and vehicle, and 10 µM lapatinib and vehicle. Data are presented as mean growth rate ± SEM. Growth rate data is derived from n = 921, 592, 249, and 292 MCF-7 cell clusters and n = 438, 299, 110, and 127 BT-474 clusters treated with the vehicle, 10 µM staurosporine, 10 µM neratinib, and 10 µM lapatinib, respectively. Source data is provided as a Source Data file.

Representative HSLCI images demonstrate a range of responses to treatment (Fig. 4a). Six hours post-treatment, we detected a significant decrease in the average mass of BT-474 clusters treated with 10 µM neratinib relative to the control. The average masses at the start of the imaging window (6 hours post-treatment) did not significantly differ from the vehicle control for all other conditions (Fig. 4b, Supplementary Table 3). After 24, 48, and 72 hours, we observed significant differences in a number of treated samples (Supplementary Table 3). After 24 hours, control MCF-7 cell clusters averaged 1.56 ± 1.05 ng, while those treated with 1 μM and 10 μM staurosporine showed significant reductions in average mass to 1.18 ± 0.77 ng (p = 1.93 × 10−9, Mann-Whitney U-test) and 1.11 ± 0.69 ng (p = 1.49 × 10−13, Mann-Whitney U-test), respectively. BT-474 showed a similar pattern after 24 hours with control organoids averaging masses of 1.27 ± 0.69 ng while staurosporine-treated clusters averaged 0.76 ± 0.39 ng (1 μM, p = 2.27 × 10−32 Mann-Whitney U-test) and 0.79 ± 0.44 ng (10 μM, p = 1.59 × 10−28, Mann–Whitney U-test). Both MCF-7 and BT-474 rapidly responded to treatment with 1 μM staurosporine, as visualized by plotting normalized growth curves (Supplementary Fig. 9) and by the reduction in the average growth rate (Fig. 4c).

Responses to lapatinib and neratinib reflected cell line-specific trends consistent with HER2 expression (Supplementary Fig. 2). BT-474 quickly showed sensitivity to both neratinib and lapatinib, while MCF-7 only exhibited sensitivity to 10 µM lapatinib and neratinib (Fig. 4b, Supplementary Table 3). After 24 hours, the mean mass of BT-474 clusters treated with 0.1 µM neratinib decreased to 0.97 ± 0.44 ng from 1.27 ± 0.69 ng (p = 4.86 × 10−6, Mann-Whitney U-test) and those treated with 1 µM lapatinib decreased to 1.00 ± 0.49 (p = 3.37 × 10−5, Mann–Whitney U-test). When comparing EC50s of 2D cultures, manually seeded, and bioprinted 3D cultures, we found no significant change, albeit we observed a trend toward lower EC50 for bioprinted MCF-7 treated with neratinib (Supplementary Fig. 10, Supplementary Table 4). Results were consistent across independent experiments (Supplementary Fig. 11).

Intra-sample heterogeneity of drug responses

Our combination of HSLCI with machine learning-based tools provides mass tracking of individual cell clusters and organoids, allowing quantitation of intra-sample heterogeneity (Fig. 4b, Supplementary Movies 1 and 2). We assessed the ratio of imaged 3D clusters that gained, lost, and maintained mass over 12, 24, 48, and 72 hours for both control and treated samples (Fig. 5a, Supplementary Table 5). In the absence of drug treatment, 11.9% of BT-474 3D clusters lost more than 10% of their initial mass and 80.9% gained more than 10% of their initial mass over 72 hours. In contrast, only 50.8% of MCF-7 gained mass, and 32.6% lost mass. This heterogeneity in populations increases over time, with 23.2% of MCF-7 gaining more than 10% mass within 12 hours. This proportion nearly doubles to 44.8% after 24 hours but remains consistent at 48.6% and 50.8% after 48 and 72 hours, respectively. This pattern differs from BT-474 as the population of organoids that gained mass continually increases over the first 48 hours before plateauing between 48 and 72 hours. BT-474 that gained >10% mass increased from 30.1% after 12 hours, to 62.4% after 24 hours, and 80.9% after 72 hours.

Fig. 5: HSLCI enables identification of resistant and sensitive 3D cluster subpopulations and discerns response to treatment earlier than endpoint assays.
figure 5

a Plots showing the percentage of tracked clusters in each condition that gain (green) or lose (black) more than 10% of their initial mass 12, 24, 48, 72 hours after treatment. b Relative viability of treated wells in HSLCI-imaged plates of bioprinted MCF-7 and BT-474, determined by endpoint ATP release assays. Bars represent the mean and error bars represent the standard deviation. Each point represents the normalized luminescence signal from independent replicates. n = 12 and n = 11 wells were treated with the vehicle control for MCF-7 and BT-474, respectively. For treated wells in both experiments, N = 8 replicates were screened for each concentration of staurosporine and N = 3 replicate wells were screened for each concentration of both lapatinib and neratinib. Statistical significance was assessed using a two-tailed, unpaired t-test with Welch’s correction. p < 0.05 is denoted by *, p < 0.01 is denoted by **, and p < 0.001 is denoted by ***. Exact p-values are reported in Supplementary Table 6. Source data is provided as a Source Data file.

Upon treatment, we observed both inter-sample (MCF-7 vs BT-474) and intra-sample heterogeneity. In the presence of the HER2-targeting lapatinib (10 µm), 37.7% of MCF-7 continued to grow and an additional 10.0% maintained their mass after 72 hours of treatment (Fig. 5a, Supplementary Table 5). When treated with 10 µM neratinib, only 4.5% of MCF-7 clusters gained mass, while 18.1% remained stable. In contrast, BT-474 showed greater sensitivity to both drugs, with 11.4% of cell clusters growing and 73.7% losing mass with 10 µM lapatinib treatment (vs 11.9% for controls), and no cell cluster growing after exposure to 10 µM neratinib for 72 hours (Fig. 5a, Supplementary Table 5). A subset of BT-474 cells showed high sensitivity to 0.1 µM of both lapatinib and neratinib. In response to lapatinib, 12.8% of BT-474 clusters lost mass, while 17.9% maintained stable mass. When treated with neratinib, 49.1% lost mass and 22.4% maintained stable mass. Both responses contrasted with organoids treated with vehicle, of which 11.9% lost and 7.2% maintained mass. The heightened sensitivity of BT-474 cells to lapatinib and neratinib is expected given the higher expression of HER2 found in these cells63 (Supplementary Fig. 3).

A fraction of cells in all treatment-cell combinations were unresponsive to the drugs tested (Fig. 5a, Supplementary Fig. 9, Supplementary Table 5). These grew at similar rates to their vehicle-treated counterparts and comprised between 7.8 and 87.1% of all organoids depending on the cell line and drug. For example, 37.7% of MCF-7 cell clusters treated with 10 µM lapatinib grew after 72 hours, while an additional 10% maintained stable mass (Supplementary Table 5). Similarly, when treated with 10 µM neratinib for 72 hours, nearly 8% of BT-474 maintained their mass, and when exposed to 10 µM lapatinib for 72 hours, the proportion increased to over 25% (Supplementary Table 5). Our findings are indicative of a resistant population of organoids that can be rapidly identified by HSLCI imaging. These persisters may provide a unique model for understanding de novo and acquired treatment resistance.

Lastly, to validate the responses measured by HSLCI, we performed an endpoint ATP release assay on the same plates used for HSLCI imaging and assessed viability at the end of the 72-hour treatment (Fig. 5b). The ATP assays confirmed that both cell lines are highly sensitive to staurosporine, with near-zero viability at the 1 and 10 µM concentrations. Additionally, BT-474 showed significant reductions in viability when treated with 0.1 µM lapatinib and 0.1 µM neratinib for 72 hours (Supplementary Table 6). Overall, the results of the cell viability assays after 72 hours confirmed the trends observed in as little as 6 hours by HSLCI but fail to capture intra-sample variability.

Discussion

Cancer therapy is increasingly moving towards treatments tailored to each patient’s unique and heterogeneous disease83,84. Molecular precision medicine requires knowledge of the associations between molecular features and drug response;1,85 however, many of these relationships have yet to be established86. Functional precision medicine bypasses the need to have prior knowledge of these associations and several studies have related in vitro response with patient outcomes4,12,13,32. Key limitations towards the broad adoption of functional precision medicine have been the creation of physiological culture models, the development of high throughput systems, and the difficulty in measuring organoid heterogeneity7,31,34,35,87. Our pipeline overcomes each of these barriers by incorporating a robust 3D organoid bioprinting protocol and an imaging approach that facilitates single-organoid analysis of response to treatment.

We introduced bioprinting to enhance the throughput and consistency of our previously published organoid screening approaches3,16,33. We opted to print a Matrigel-based bioink due to its ability to preserve tumor characteristics ex vivo3,21. Matrigel behaves as a liquid in a narrow temperature range under its gelation temperature of 6–8 °C. Above these values, it shows time- and temperature-dependent changes in viscosity while it thermally crosslinks88,89. Without strict control of temperature and bioprinting preparation procedures, bioink can vary in viscosity. As a result, extrusion pressure may need to be adjusted with each print. Although these physical properties complicate its suitability for bioprinting, we show that bioprinting simple, single-layer Matrigel structures is attainable with strict temperature regulation. We further enhanced the quality of the Matrigel deposition by selectively modifying the print substrate with oxygen plasma treatment. The introduction of 3D plasma masks (Supplementary Figure 1) facilitated the selective treatment of a square region in each well. The increased hydrophilicity of the substrate in the exposed region guides the spreading of the material to maximize consistency in deposition volume and construct thickness while preventing obstruction of the center of the well. Bioprinting allowed us to finely control the size and shape of the deposited gel constructs, facilitating the use of HSLCI for downstream analysis (Fig. 1).

We have optimized a low pressure, temperature-controlled extrusion-based bioprinting protocol to avoid altering tumor cells. Though it is possible that a subset of cells may be influenced by the bioprinting process, our bulk RNAseq analysis suggests this would be limited to a very small subpopulation. Overall, the sum of our orthogonal findings supports the conclusion that the cells we tested were largely unaffected by the mild bioprinting conditions used in our studies. This applies to the specific combination of bioprinting parameters and cell models included in this work. It will be important to continue evaluating the biological effects of bioprinting in the context of future studies that either use alternative printing protocols or are focused on different patient-derived samples.

Previous studies have used interferometry to quantify the mass of individuals cells cultured on 2D substrates to study cell division90, cytoskeletal remodeling91, mechanical properties92, and response to treatment41,42,54. Tomographic QPI has also been used to obtain high-resolution images of 3D objects such as cerebral organoids93 and cancer organoids94. An enduring challenge of adapting high throughput quantitative imaging techniques to organoid models of cancer is efficient visualization of 3D clusters35,95. HSLCI imaging records 2D projections of the phase shift of incident light caused by cultured cells or organoids at a single focal plane relative to the plate surface41. 2D phase shift maps only provide accurate biomass measurements for objects in focus43,96. When using HSLCI to measure biomass, efficiency can be reduced by organoids moving orthogonally to the focal plane over the duration of imaging. We mitigated this challenge by using bioprinting to generate uniform, thin constructs that maximized the number of organoids that could be tracked in parallel, coupled with a machine learning-based organoid classifier that excludes out-of-focus objects from the analysis. Alternatively, an imaging system capable of capturing multiple z-planes across the plate could address this issue. Such an approach, however, would decrease the frequency of organoid imaging and increase the size of the resulting dataset, which poses different challenges for timely and efficient data analysis.

By introducing region-specific reference images and machine learning-based methods for image segmentation and track filtering, we have been able to increase the number of organoids tracked approximately 15-fold from initial analyses in which only approximately 1% of organoids analyzed were retained using standard approaches (Supplementary Fig. 11). Further improvements will include shortening the 6-hour delay between drug treatment and imaging start, which will allow us to capture highly sensitive organoids undergoing cell death within that timeframe. Lastly, due to the large amount of data generated using HSLCI (approximately 250 GB per plate/day), data analysis remains a time-limiting factor.

Our work uses bioprinting in combination with time-resolved HSLCI for high throughput, label-free biomass profiling of 3D models of cancer. While other groups have previously bioprinted organoid models97,98,99,100 to assess drug treatments98,99,100, most of these systems, like other screening approaches, used endpoint chemosensitivity assays that return well-level average measurements at a single time point95. These endpoint assays are limited in their ability to uncover transient responses and heterogeneity in growth and drug sensitivity phenotypes95. In contrast, the time-resolved, organoid-level information returned by HSLCI enables the characterization of rare, drug-resistant cell subpopulations, providing additional insight that cannot be resolved by population-based assays. Genomic characterization of tumors has demonstrated that these malignancies are collections of evolutionarily-related subclones, rather than homogeneous populations101,102,103,104. This genetic diversity is one of the several factors that contributes to differential response to treatment. Endpoint assays, such as live-dead staining or ATP release quantification, characterize the average response to treatment. Though they may be useful for identifying drug sensitivity in majority cell populations, they fail to account for the response of resistant clones that may also be present. In the clinical setting, failure to treat the resistant populations may cause recurrence and long-term disease progression after initial responses105,106,107. Despite the development of 3D cancer models with varying extents of complexity and scalability, functional screening assays have been hindered by their inability to consider this heterogeneity of tumor responses. HSLCI allows non-invasive, label-free tracking of various features of bioprinted organoids over time, including size, motility, and mass density at single-organoid resolution. Because of the ability to quantitatively measure time-resolved, individual mass changes in response to treatment, it is possible to identify and isolate responsive and resistant subpopulations of cells which can lead to more informed clinical decision making when selecting a treatment approach45. The combination of throughput, time resolution, and number of organoids sampled94,108,109 paired with our short experimental time frame from seeding to drug sensitivity results3,16,33,110, make our HSLCI-based method valuable for screening tumor organoid models for research and, in the future, for possible clinical applications.