Organoids, a product of 3D biology, are cells grown to develop a similar structure to their source tissues, in turn mimicking the behavior and properties of that cell. Researchers often use fluorescence microscopy to observe the development and morphology of organoids. However, this manual analysis is time consuming and leaves a high margin for error. To address these issues, some researchers have leveraged machine learning (ML), a form of artificial intelligence (AI) that humans can “train” to recognize patterns and make more accurate predictions.
For example, a September 2021 study published in Development showcases MOrgAna (ML-based Organoid Analysis software), a ML model designed for analyzing the morphological and fluorescence characteristics of organoid models via image processing. Described as an “easy-to-use ML pipeline,” MOrgAna represents an accessible path to augmenting organoid analysis with ML. Rather than analyzing organoids manually, researchers can instead invest that time into training a ML model to perform the analyses autonomously.
Similarly, an October 2020 study published in Nature Communications details the use of ML to analyze 3D preclinical colorectal and bladder organoid models and identify robust drug biomarkers. The identified biomarkers accurately predicted patient responses to particular drugs used to treat colorectal and bladder cancer.
These studies and others illustrate that ML was the next step in organoid-centric research, and they have aided in developing more effective therapeutic treatments. However, conventional ML is not without its limitations. For instance, the accuracy of ML models improves when fed new data, which means that humans must intervene continually if they wish to improve its accuracy. Additionally, ML models can be time consuming and costly to generate. A multi-disciplinary expert is needed to design the model’s “feature extractor”, which enables the ML model to identify points of interest in data after being trained. In the case of 3D organoid analysis, the model’s feature extractor would serve to identify cell components, disease markers, and other characteristics in the 3D image. Implementing a ML model’s feature extraction ability can be long and difficult.
To address these hurdles, some researchers have begun using a variation of ML called deep learning (DL) to analyze organoids—and it’s possible that DL will mark a new era in organoid research.
Deep learning as the way forward
DL is a subset of ML that uses artificial neural networks—that is, digital counterparts to the human brain—to learn. Much like how biological brains are comprised of neurons, artificial neural networks are comprised of “nodes” that are stacked into “layers.” Simply speaking, the more layers a DL model has, the more advanced it is. The main advantage that DL models have over traditional ML models is that their artificial neural networks allow them to effectively learn the same way the human brain does. Consequently, DL models can self-improve, are highly flexible, and are capable of identifying complex patterns and correlations that ML models cannot perceive. In the context of organoid analysis, these traits could manifest as being able to identify a wider range of organoid types, identifying trends between organoid images more accurately, and effectively handling more complex datasets than ML models can. For instance, a DL-based organoid analysis solution called OrganoID was trained on images of pancreatic cancer organoids, but can accurately identify organoid models of lung, colon, and adenoid cystic carcinoma organoids as well. Traditional ML models likely would not be able to replicate those results because they require researchers to manually design unique feature extractors for different organoid types, rather than generating their own feature extractors and then applying that extractor to novel data. Moreover, researchers have demonstrated that DL models can outperform conventional ML models in image processing, which is the method by which organoid analyses are carried out.
While DL models require a much larger set of initial training data than traditional ML models, they can realize feature extraction autonomously—eliminating the time-consuming and costly process of designing a feature extractor for conventional ML models while yielding higher accuracy. Though traditional ML models are still powerful and well-suited to certain types of tasks, they cannot process the same volume or complexity of data that DL models can. This naturally limits their analytic capability.
The paper, “Organoids revealed: morphological analysis of the profound next generation in-vitro model with artificial intelligence,” was written by researchers from the State Key Laboratory of Bioelectronics at Southeast University, China, and published in January 2023. DL, the authors argue, is a fast-approaching future, with the body of literature around the use of DL in biomedical contexts already mature. On its current trajectory, the authors “anticipate more combinations of DL algorithms and high-content analysis.” Namely, they “expect the development of a ‘self-learning microscope’ that would enable automated, high-speed, high-throughput, highly repeatable analyses that could make organoid analysis more comprehensive and easier to execute. “Luckily,” the paper states, “more and more biological and medical researchers are now devoted to organoid exploration and study. We believe that readers will observe a significant increase in the scope of DL-based organoid analysis in the future.”
While the future looks promising, DL analysis still presents its own set of challenges. For instance, most scientific literature involving DL uses supervised learning, which requires manually annotated data (as opposed to unsupervised learning, which does not require manually annotated data but whose results are difficult to verify): “To achieve good accuracy, neural networks usually require many annotated samples to perform training tasks. Collecting annotated high-quality datasets...for supervised learning is usually a very difficult task, and manual annotation on these images is also very tedious and expensive,” the paper says. While DL models perform feature extraction autonomously, which is a significant advantage over traditional ML models because human intervention is not continually required, researchers must still invest significant time into annotating the training data. The authors propose the construction of standardized databases for organoid images that any lab could access to train their own models. Furthermore, DL demands very powerful computers, and most current hardware cannot keep up with those demands. As a consequence, the 3D organoid images are effectively downscaled, which allows them to be processed but can cut elements of interest from the image.
There is still work to be done, but the future is clear: the trend of leveraging DL to facilitate organoid analysis is gaining traction, and we are likely to see substantial growth in research around the topic. Ideally, this technology will make organoids more effective and consistent, allowing for therapeutic solutions to be developed more efficiently than ever.