Approaches to benchmark and characterize in vitro human model systems

Over the past ∼25 years, a remarkable amount of progress has been made in the ability to culture human embryonic stem cells (hESCs) or reprogram cells into an induced pluripotent state (hPSCs). These stem cells can be used to recreate aspects of embryogenesis and organogenesis ‘in a dish’ (Thomson et al., 1998; Takahashi and Yamanaka, 2006). hPSC-derived model systems such as gastruloids, which mimic human gastrulation (Simunovic and Brivanlou, 2017; Siggia and Warmflash, 2018; Heemskerk, 2020), and organoids, which model organ-specific development (Frum and Spence, 2021; Clevers, 2016; Huch et al., 2017), have become powerful tools for studying human development, which is challenging to investigate in vivo due to limited access to tissue and to ethical concerns (Fu et al., 2021). Human model systems can also be derived from patient samples (i.e. from a biopsy) and, collectively, both stem cell- and patient-derived tissues can be cultured in several formats, such as standard 2D culture, 3D culture and air-liquid interface cultures, and in various engineered systems, such as microfluidic ‘organ-on-a-chip’ systems (reviewed by Hofer and Lutolf, 2021) (Fig. 1).

Early studies that used hPSCs to differentiate embryonic cell lineages or organoids modeled a variety of developing tissues and stages, including early gastrulation, neural tissue, gut, lung, kidney and optic cup (Eiraku et al., 2011; Spence et al., 2011; Mou et al., 2012; Sasai et al., 2012; Hogberg et al., 2013; Takasato et al., 2014; Warmflash et al., 2014). These in vitro human models were often compared with developing mouse embryos because of the rich literature available and the relatively deep understanding of the developmental events, cell types, essential genes and well-characterized molecular markers during murine development, and also because of the lack of access to human fetal tissue as a reference (Frum and Spence, 2021; Fu et al., 2021). However, mice are an imperfect model for understanding human-specific aspects of development. Accurate benchmarks are essential because organoids and gastruloids are only useful as models to the extent that they accurately recapitulate human-specific development. The analysis of human models was also limited; generally, only a handful of cell type-specific markers were assayed, which showed that not all the relevant cell types within the organ were represented and complex in vivo structures were only partially recapitulated (Spence et al., 2011). There were also significant technological limitations at the time. Interrogating tissues was also generally limited to low-throughput assays, such as in situ hybridization and immunofluorescence, or to bulk-sequencing-based approaches that lacked single-cell resolution. Therefore, the amount of information gained from early benchmarks for organoid systems was relatively limited.

Recent technological capabilities in benchmarking have significantly improved (Box 1), and advances in single-cell genomics and spatial-imaging technologies, alongside new computational approaches, have led to an unprecedented capacity to interrogate cellular diversity at single-cell resolution (Macaulay and Voet, 2014). These technologies have enabled large-scale efforts to catalog cell types within the developing and adult human body, which in turn provides an important benchmark for organoid and gastruloid model systems that can be used to refine and improve these models. In addition, these atlas-building efforts have provided a foundational scientific resource to aid understanding of human development from fertilization to birth and through childhood (Taylor et al., 2019; Haniffa et al., 2021), and understanding of normal adult tissue homeostasis and disease states.

This Spotlight discusses in vitro human model systems alongside enabling technologies that allow deep molecular characterization at single-cell resolution. Focusing largely on organoids, although this discussion also applies to other model systems such as gastruloids, it aims to explain how cell and tissue atlases provide a roadmap by which to benchmark in vitro models of human development. Current limitations the field faces and how these limitations might be overcome in the future are also discussed.

How does one know if an in vitro system (i.e. organoid) accurately recapitulates the native organ it was developed to model? Theoretically, if an organoid model accurately mimics an organ, it will recapitulate all cellular and functional aspects of the organ itself. There is no simple way to holistically evaluate cellular heterogeneity and function, but with the current technologies widely available in the field, several features of the organoid can be assessed.

Ideally, a human organoid system will possess all the specific cell types found in the organ of interest. These cells should include all those that facilitate the highly specific functions of that organ, such as hepatocytes in the liver (MacParland et al., 2018), alveolar epithelial cells of the lung (Travaglini et al., 2020) or enterocytes of the intestine (Elmentaite et al., 2020). Furthermore, the organoid requires other components of the organ, such as nerves, blood vessels, immune cells, etc., to fully mimic its native counterpart. Currently, cell types are largely characterized based on their expression of known marker genes or proteins and identified using methods such as single-cell or single-nuclei RNA sequencing to assess the transcriptome of specific cells. Other methods, such as single-cell ATAC sequencing, ChIP-seq and/or CUT&RUN are used to assess how closely in vitro cells mimic the in vivo epigenome. There have also been recent developments combining multiple techniques for a more holistic ‘multi-omic’ analysis where multiple types of analysis can be run on a single cell in a sample (i.e. dual scRNA-seq and scATAC-seq).

Organoid models are ideally organized in the proper spatial pattern with the higher-order structures found in an organ. For example, sophisticated intestinal organoids should contain epithelial cells organized into villi with crypts containing stem cells, with stroma, muscle, vasculature, neurons and immune cells in a highly organized structure (Elmentaite et al., 2020). Recent approaches have begun to develop and deploy high content image-based technologies to probe the spatial organization of in vitro models, and include immunohistochemistry-based approaches, such as 4i (Gut et al., 2018), and spatial transcriptomic approaches (Cho et al., 2021) (Box 1).

Finally, and perhaps most importantly, it is crucial to assess whether organoid function mimics that of its in vivo counterpart. Most organs perform a variety of specialized functions and, ideally, in vitro models would recapitulate all of these functions. For example, human intestinal organoids that perfectly recapitulate all functions of the in vivo gut would absorb nutrients, undergo peristaltic contractions, secrete mucus, have full immunological function, contain a functional enteroendocrine system and maintain a healthy microbiome (Kastl et al., 2020; Volk and Lacy, 2017; Quadrato et al., 2017; Eicher et al., 2022). As most in vitro organoid models lack the full complement of organ-level functions, such as vascular and neuronal networks, which are required for functionality, functional analysis of organoids often takes place at the cellular level (i.e. nutrient absorption/uptake). For example, gut organoids can recapitulate many individual functional aspects of the gut, including hosting a microbiome (Hill et al., 2017) and absorbing nutrients (Spence et al., 2011); however, their ability to simultaneously perform all functions of the gut is limited (Zachos et al., 2016).

A significant advancement in the field that has greatly enhanced the ability to fully characterize and benchmark organoids has been the explosion of publicly available benchmarking datasets from many developing human organ systems, driven by advancements in enabling single-cell technologies (Box 1). Large atlases of the adult (Jones et al., 2022) and developing human (Cao et al., 2020; Domcke et al., 2020; Yu et al., 2021) have been published, both as large cohorts of sequenced tissues across the human body, as well as more nuanced studies of separate organs (Miller et al., 2020; Holloway et al., 2021; Hein et al., 2022a; Howden et al., 2021) (Table 1). Large collaborative efforts, such as the Human Cell Atlas and the Human Developmental Cell Atlas aim to generate comprehensive reference maps of all cell types in the human body. For example, Tabula Sapiens is a collaborative effort to provide a cell atlas for adult humans and contains data from cells of 24 organs from 15 human subjects. These data include single-cell transcriptomics, histological sections and microbiome samples, and many samples were collected from the same donor, controlling for differences among individuals. The effort has identified over 400 cell types, and characterized organ-specific differences among cell types (Jones et al., 2022). There are similar efforts under way to generate developmental and pediatric human atlases. Published atlases generally include transcriptomics data (Box 1), but some also include information about the epigenome (Domcke et al., 2020).

Organoids and gastruloids often represent developing stages of organs, rather than adult tissue. This is especially true of models derived from hPSCs, which are relatively immature and model early development. Benchmarking data of human development is, therefore, necessary to characterize how well the models recapitulate human organs. Efforts like the Human Developmental Cell Atlas are essential to generate reference data of human development. Whole-embryo reference datasets of the early stages of development, including pre-implantation and gastrulation embryos have been published (Molè et al., 2021; Blakeley et al., 2015; Tyser et al., 2021; Yan et al., 2013). Furthermore, developmental atlases exist for a variety of organs across developmental time points, including brain, gut, lung, heart, liver, kidney, placenta and thymus (Haniffa et al., 2021; Sountoulidis et al., 2022 preprint; He et al., 2022 preprint).

Importantly, these data resources can be used by research groups without access to human tissue, providing an essential community resource that is particularly important to advance in vitro human model systems. Tools such as online interactive platforms for biologists, which are less computationally intensive, are important to make these datasets accessible and user-friendly. Platforms have been developed, such as the UCSC Cell Browser, CellxGene (Cell-by-Gene; https://github.com/chanzuckerberg/cellxgene) and the EMBL-EBI Single Cell Expression Atlas. There has also been the development of other interactive tools that house benchmarking data beyond scRNA-seq, including the Allen Institute Cell Explore, which includes 3D imaging datasets, genomics and more. These tools are limited by the number of datasets available, and the availability of new online datasets, but as these resources continue to expand, they will become integral to the research community.

Although there have been significant efforts in the field to develop developmental atlases, there is more work to be done. Atlases of individual organs are limited because they do not enable comparison across organs from a single donor. Moreover, many single-cell atlases of developing organs do not take a nuanced approach to characterize different spatial domains within a tissue (i.e. trachea versus small bronchi/airway in the lung). In addition, limited time points across development do not represent the full spectrum of human development during fetal and postnatal life, and they often do not represent the full racial and ethnic diversity of the human population. The Human Developmental Cell Atlas aims to develop atlases with multiple replicates at each time point in development, which will enable the study of developmental trajectories (Haniffa et al., 2021). Better understanding the full developmental trajectory for each cell in every organ will be essential for continued improvement of model systems. Furthermore, the lack of ancestrally and/or racially diverse atlases is a major limitation; without diverse samples, the field cannot be confident that it is creating in vitro models that are relevant to the whole population.

Datasets themselves are also not without limitations. Currently, the field has mostly focused on creating scRNA-seq datasets and is starting to emphasize curation of scATAC-seq datasets that will fill major gaps in our current benchmarking capabilities, and will improve our understanding of regulatory networks and genomic architecture that underlie gene expression. Moreover, there are far fewer benchmarking efforts using spatial transcriptomic or proteomic technologies. This is likely due, in part, to the limited commercial availability of these technologies; however, as the advancement of in vitro models accelerates, the need for benchmarks that go beyond the transcriptome for a more comprehensive view of organogenesis and development will become increasingly important. Fortunately, such technologies are already moving forward. For example, recent studies have begun to interrogate retina organoid development using quantitative spatiotemporal measurements across different molecular modalities, including transcriptomics and proteomics (Wahle et al., 2022 preprint). Furthermore, even the most widely used technology, scRNA-seq, still comes with limitations and lack of standardized procedures that make it non-trivial to compare datasets created by different labs. Some of the most pressing limitations for single-cell benchmarking efforts include the variability in quality/read depth/resolution varying in datasets made from different protocols at different labs, accuracy of mapping to reference genomes, and integration of multiple data types across sample, tissue and developmental time (Lähnemann et al., 2020).

New knowledge gained from cell atlases has not only shed light on the complexities of human development and organogenesis but has also provided a roadmap to improve and refine all current in vitro model systems themselves. These improved in vitro models can then be assessed by benchmarking techniques discussed above to validate their similarities (or differences) from the native tissue.

Armed with the data obtained from atlas-building and spatial studies, it is now possible to compare the cell types and spatial organization, and the transcriptional and epigenomic make-up of organoids with that of in vivo human organs. This work has led to the discovery of new cell types (Basil et al., 2022; Kadur et al., 2022; Han et al., 2020; Pollen et al., 2015) and the improvement of organoid differentiation protocols and culture conditions (Holloway et al., 2020; Howden et al., 2021; Yu et al., 2021; Childs et al., 2022 preprint; Hein et al., 2022a). hPSC-derived organoid models are derived using a technique called directed differentiation, which involves mimicking the signaling that occurs during development to generate the specific organ of interest (Frum and Spence, 2021). Therefore, in addition to their utility in benchmarking, cell atlases provide researchers with new information about cell-cell signaling and cell-cell interactions in vivo, providing a framework to refine and improve directed differentiation protocols used to create in vitro systems, ultimately creating models that better recapitulate human development. This approach has been used to improve models of the gut, kidney and brain (Childs et al., 2022 preprint; Hawkins et al., 2017; Quadrato et al., 2017; Combes et al., 2019). For example, a method to create 3D human lung organoids from hPSCs (Dye et al., 2015) has recently been refined based on benchmarking to human data, leading to a more robust model with fewer off-target lineages. By using insights from single-cell genomics data from fetal lung tissue, the authors developed a method that improved definitive endoderm differentiation followed by lung epithelium specification, leading to 100-fold higher expression of lung-specific markers, ultimately leading to a purified population of stable lung organoids composed of lung progenitor cells that could be maintained in long-term culture. The lung organoids were then benchmarked against the in vivo lung using a variety of single-cell computational approaches, and were shown to have a higher degree of transcriptional similarity compared with the previous differentiation methods (Hein et al., 2022b). Single-cell datasets are extremely powerful as they act both as a reference to confirm the quality of organoid models, and as a reference to develop and improve in vitro culture systems.

The evaluation of organoids using benchmarking data has also underscored the many limitations of even the most optimized in vitro model systems of human development. The most notable disparities between native tissue and in vitro models include the lack of a complete repertoire of all major cell lineages, increased plasticity in vitro, off-target patterning/cell types and transcriptional differences resulting from the culture environment.

Many organoids lack important cell types such as a blood/vascular network, peripheral nervous system and immune (Holloway et al., 2019). Without these major cell types and structures, model systems fail to recapitulate many functional aspects of organs in vivo, and experience additional constraints, such as being limited in size due to restricted nutrient availability in culture conditions. Furthermore, organoids, especially when derived from iPSCs, also appear to represent a very immature transcriptional state during development (Yu et al., 2021), leading to the lack of representation of mature cell types seen in the adult organ. The ability to create the entire array of cells that are needed to complete all the functions of the organ coupled with the ability to cause maturation in the transcriptional profile of these organoids is a major challenge in the field. Without this capability, it is not possible to study/model later stages of human development, leaving large gaps in the knowledge of how important events during organogenesis occur. More access to developmental stages and continued improvements in key technologies (Box 1) will be crucial for this progress to continue.

Plasticity and off-target patterning are also major issues, particularly in hPSC-derived models, as primary tissue-derived cells tend to be more committed to their fate. Even cells that appear committed have the ability to lose their identity and change lineages (Hurley et al., 2020). Although in vitro models might have cells that share a similar transcriptional signature to most of their in vivo counterparts, there are still often off-target lineages found within cultures, which inherently limit their functional applications. For example, lung organoids often contain small areas of gut tissue (Dye et al., 2015; Hein et al., 2022b).

There are also major differences between organoids and native tissue caused by the in vitro culture environment. Single-cell sequencing data have shown the impact of in vitro culture conditions on gene expression. For example, lung organoids derived from primary tissue grown in vitro show distinct transcriptional signatures when compared with the same cell types in the native lung (Miller et al., 2020; Hein et al., 2022b). The very nature of being generated in vitro introduces a transcriptional signature driven by the growth factors in the culture media that cannot be avoided. Better understanding growth factors, nutrients and metabolites, and the necessity of supportive cell types and scaffolds that better reflect the in vivo environment will be needed to begin to address these fundamental differences.

These limitations further highlight the fact that consistency (and inconsistency) across individual laboratories may be inherent to these systems despite having robust protocols and high reproducibility. Although there is great potential for understanding development and stem cell biology, as well as translational applications, further utility of organoid models would be enhanced if these limitations are addressed in the future.

Where does the field go from here? How can new knowledge from human cell atlases be applied to improve in vitro systems? The field needs to leverage new enabling technologies (Box 1) and combine these with current and future bioengineering approaches to better simulate the in vivo environment. Ideally, the creation of multiomic datasets in concert with higher throughput image-based analysis of different stages of human development could be used as a holistic reference for matched organoid datasets to obtain a more comprehensive picture of how well the models mimic development. In addition to RNA/DNA sequencing-based advances, high-resolution proteomics and the ability to better understand the physiological conditions of the native tissue (i.e. the metabolic and nutritional environment) will also greatly help advance in vitro modeling. Beyond these technical advances, adding poorly studied data from across the lifespan (i.e. pediatric stages) and data from both healthy and diseased states will round out the reference data sets needed to properly benchmark new and advanced in vitro models. More advanced co-culture approaches and creative bioengineering approaches will allow more complex organoids with increased diversity of cell type, spatial organization and function. These steps will enable researchers to create data-driven hypotheses to refine protocols to create improved in vitro systems that more accurately mimic human development and organogenesis.

We thank Dr Kate D. Walton, Dr Tristan Frum, Dr Lindy K. Brastrom and Ansley Semack Conchola for their constructive comments on the manuscript.