A protocol described in Nature Protocols combines 3D bioprinting with label-free quantitative phase imaging (QPI) using high-speed live-cell interferometry (HSLCI) and machine learning (ML) analyses to profile drug responses in bioprinted tumor organoids at single-organoid resolution.

“We demonstrate that the protocol can be leveraged to automatically generate plates containing thousands of organoids for high-throughput imaging and drug screening experiments, quantify growth and drug response heterogeneity, resolve rare phenotypes, and identify predictive features of drug response profiles for fundamental studies and therapeutic decision-making,” wrote Michael A. Teitell, MD, PhD, of the University of California, Los Angeles (UCLA), and colleagues.

The protocol can be completed in 2 weeks or less and, according to the investigators, generates organoid growth and drug response data within approximately 1 week of organoid establishment—a timeframe they noted is compatible with therapeutic decision-making and functional precision medicine efforts. It is also adaptable to organoids derived from a variety of cell sources and alternative screening paradigms.

Bioprinting, QPI, and ML

The protocol consists of three stages:establishing and culturing bioprinted organoids; high-throughput drug screening; and data processing and analysis.

The workflow begins by bioprinting cells within a temperature-sensitive extracellular matrix (ECM) into thin (~70 μm), flat, square-shaped constructs in 96-well plates. The investigators developed this geometry to improve imaging efficiency compared with conventional 3D seeding approaches, which generate organoid models in thick domes or droplets of an ECM, causing structures to grow at random positions throughout the full thickness of the matrix and increasing light scattering.

After the organoids are established and exposed to drug treatment, they are imaged using HSLCI, a high-throughput, label-free QPI platform that measures the phase shift of light passing through the organoids to calculate dry biomass. According to the investigators, biomass measurements over time provide direct indicators of cell fitness, making them especially useful for functional precision medicine. A typical experiment images approximately 25 unique fields of view across each of the interior 60 wells of a 96-well plate, for a total of about 1,500 unique fields of view, with an imaging interval of approximately 10 minutes between successive frames. The investigators reported a lower limit of approximately 2.4% for the coefficient of variation of HSLCI biomass measurements.

The downstream analysis pipeline applies ML at multiple stages. A U-Net convolutional neural network segments organoids in phase images, the open-source TrackMate software tracks organoids across time-course images, and an XGBoost classifier excludes tracks from out-of-focus organoids before biomass and other quantitative metrics are calculated.

Applications and Limitations

The investigators demonstrated the workflow using organoids generated from the MCF-7 and BT-474 breast cancer cell lines and a surgically resected metastatic leiomyosarcoma tumor sample; "[these serve as examples that] our protocols are compatible with tumor organoid models derived from a variety of cell and tissue sources,” they wrote.

Despite its broad applicability, the investigators cautioned that the current protocol has several limitations across its stages. Among them, they noted that the 2D phase-shift maps of the HSLCI platform can provide accurate biomass information only for organoids that remain in focus. This could limit platform efficiency for imaging 3D-cultured tumor organoids capable of moving orthogonally to the focal plane, although development of an automated focusing module minimized this issue. In addition, a typical experiment can produce more than 10 TB of interferograms and more than 2 TB of downstream analysis files—accounting for “very large amounts of data storage space,” the investigators wrote. They also stated that the ML models in the analysis pipeline may require retraining for organoid phenotypes that differ from those of the MCF-7 and BT-474 cell line–derived organoids used in the training data sets.

“The methods our protocol describe are of broad interest to researchers using 3D models for high-throughput screening and are applicable well beyond the cancer field. They are adaptable to a range of models, including patient-derived tumor organoids, and can conceivably be applied to performing functional precision medicine studies, identifying and assessing drug candidates and answering a variety of developmental or biological questions using these 3D models,” the investigators wrote.

DISCLOSURES: The work was supported in part by grants from the Air Force Office of Scientific Research and the Department of Defense to Dr. Teitell; a National Science Foundation Graduate Research Fellowship grant and a Eugene V. Cota Robles Fellowship to Dr. Wang; a Jonsson Comprehensive Cancer Center Fellowship to Dr. Tebon; a Virginia Commonwealth University Wright Center for Clinical and Translational Research grant to Dr. Reed; a UCLA David Geffen School of Medicine Seed Award to Dr. Soragni; and multiple National Institutes of Health and National Cancer Institute grants to multiple authors. Drs. Wang, Tebon, Soragni, and Teitell are inventors on a patent application related to aspects of the bioprinting process. The other authors declared no competing interests. For data and code availability, visit nature.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.