Abstract
Introduction
In this the third annual SBI2 conference, the afternoon of the opening day was dedicated to short presentations and discussions revolving around high-content screening (HCS)/high-content analysis (HCA) 3D imaging: Best practices and unmet needs. The growing interest in conducting assays in 3D is driven by the realization that there is often a disconnect between observations made with cells grown in 2D in a tissue culture plate and biology occurring in an organism that occurs in the context of other cell types and a complex extracellular matrix (ECM). This is apparent when you look at the high failure rate of earlier candidates in the drug discovery process. The push toward 3D assay systems is an attempt to increase clinical relevance, while at the same time preserving an acceptable level of throughput to evaluate potential clinical candidates. With growing scientific interest in 3D biology, imaging-based assays are emerging as an obvious choice to measure complex phenotypes and move toward the realization and quantification of the concept of an “organ in a dish.” 1
Our goal for the 3D colloquium was centered upon sharing experiences and best practices and bringing key opinion leaders to our meeting to communicate on all aspects of 3D imaging and the major challenges. The overwhelming interest and number of participants meant we had to quickly move to a larger room to hold all educational session instructors, scientific program presenters, attendees, and sponsors and vendors. 3D imaging was discussed in last years' colloquium 2 and common aspects were raised again this year. However, there have been significant leaps forward in the past 12 months, demonstrating the interest and maturation of the field and the sense that it is only going to continue to improve and be a large part of future meetings. The primary topics discussed in the colloquium were relevance and challenges of the 3D model and guidance on technical aspects of acquiring and analyzing 3D data. As the complexity of 3D models continues to develop in areas of bioprinting, nanoengineering, and clinical relevance, understanding how we best interrogate spheroid/organoid models to capture as much “content” and representation as possible is a major focus in the field. The way this is undertaken at present and for the future regarding data flow, architecture, and data analysis will help realize the expectations of 3D cellular and tissue models to transform drug discovery and development. 3
The colloquium sessions were broken down as follows: 1. 3D models 2. Applications of 3D models in research 3. CellProfiler 3D features mini update 4. 3D image acquisition 5. 3D image analysis
3D Models
Topic discussion leader: Carrie Lovitt (Griffith University, Australia).
Panel members: Judi Wardwell-Swanson (InSphero, Inc.), Mark Rothenberg (Corning Incorporated, Life Sciences), and Glauco Souza (Nano3D Biosciences).
Session Summary
There are a wide variety of innovative technologies that support the generation and maintenance of 3D models. 4,5 They include hanging drops, hydrogels, micropatterned surfaces, porous scaffolds, ultra-low adhesion surfaces, bioprinting, and microfluidic organ-on-a-chip. Along with these innovative 3D technologies, there are a growing number of high-content compatible 3D applications. A few of the 3D applications utilizing high-content end points that were highlighted during the 3D models session were cellular adhesion, invasion, and cell proliferation; targeting of proliferative or nonproliferative cells; and tissue-specific measurement of compound toxicity.
Three 3D technology providers gave presentations during the 3D models session. The presentations included InSphero's GravityPlus Hanging Drop/ultra-low attachment (ULA) platform; Corning's Matrigel, HydroGel, ULA plates, and permeable supports; and Nano3D's magnetic bioprinting platform. Each of the presenters provided an overview of their respective technology platform and initiated the discussion around the following topics:
Scaffold-free methods, such as InSphero's hanging drop/ULA platform, can eliminate assay interference and other unwanted artifacts associated with artificial and exogenous ECMs. Scaffold-free technologies also promote the formation of organotypic 3D models through cell adhesion-promoted assembly and through interactions with endogenous ECM, which can result in a more native tissue-like architecture and morphology. 1 The resulting “microtissues” have remarkable longevity and stable protein expression and function, typically upwards of 4 weeks. Some of the challenges associated with scaffold-free platforms are the potential for spheroid loss during downstream processing. And although scaffold-free 3D platforms are generally more compatible with imaging technologies than scaffold-based platforms, there can still be issues with focus, light penetration and scattering, segmentation, and the ability to locate the spheroid. Innovative plate designs, such as InSphero's GravityTRAP plate, are helping make scaffold-free 3D models compatible with automation and imaging by offering media exchange compartments and flat imaging quality plate bottoms. Tissue assembly and maturation are often slower in scaffold-free settings and some cell types may not be amenable forming spheroids in the absence of exogenous matrix.
Investigators continue to use technologies such as the Corning® Matrigel Matrix or permeable supports in new and inventive ways to promote 3D culture systems with excellent results. There are many advantages of using 3D systems in drug screening, the predominant reason is that these systems are known to stimulate in vivo cellular conditions more realistically than the standard 2D systems currently in use in screening protocols. There are challenges, including our limited knowledge on the nutritional requirements for growth, proliferation, and functionality in a 3D microenvironment.
3D spheroid assembly through bioprinting, such as Nano3D's magnetic bioprinting platform, represents yet another innovative approach to 3D model generation. Nano3D's technology promotes very rapid spheroid formation (often less than 15 min) by drawing nanoparticle-decorated cells to a single magnetic site in the microtiter plate. Once the cells are brought into proximity by the magnet, they are further induced to form spheroids through cell contact and cell–cell adhesion. Nano3D's magnetic technology, like other bioprinting technologies, allows the researcher to guide the tissue assembly and the resulting architecture. Nano3D showed a clear example of a vasoconstriction assay, which utilizes ring-shaped “vessels” comprising primary smooth muscle and endothelial cells. Another advantage of magnetic bioprinting technology is the retention of spheroids through the magnet during processing steps, thereby greatly facilitating automation and downstream processing. Finally, Nano3D technology is compatible with flat-bottomed plates and, therefore, compatible with most imagers and the magnet arrays are available up to 1,536 well format, enabling true high-throughput screening (HTS).
Beyond the development of new technologies for culturing and screening in a 3D model, the field needs development of robust and characterized 3D phenotypic models. And as we move toward the use of more relevant 3D cell models, there will be significant challenges related to the identification of reliable sources of high-quality primary cells, validation of individual donor lots to ensure reproducibility of the models, and the optimization of cocultures to achieve tissue-like morphology and function.
Other challenges include our limited knowledge of the nutritional requirements for growth, proliferation, and functionality in a 3D microenvironment. The 3D-specific assay methods and kits currently available are providing experimentally rich data, but more are needed in a variety of different assay formats. The ideal level of complexity is still very much up in the air and may be best answered by the individual investigator and the biological question(s) they are trying to address (in other words, the model should be fit for purpose and that will help dictate the required level of complexity). In this session, the complexity of the models ranged from 3D monocultures (in the case of tumor growth and invasion assays) to complex multicellular organoids (in the case of an adipose model).
As 3D assay models move into high-throughput systems, the ability to assess functionality using HCS will become highly relevant. Imaging of a limited number of real-time events and fixed end points in 3D is possible with existing technologies, but it is clear that there is still a significant gap with regard to 3D-compatible high-content assays, imaging-compatible 3D plastic ware, 3D-compatible imagers, and the capabilities of the image analysis software. Light sheet microscopy and novel clearing agents were mentioned as potential solutions for 3D imaging challenges.
Other required advances include compatible liquid handling automation for 3D assays so that these formats can be more readily moved into a high-throughput screening mode. As technologies evolve, which allow for the improved cellular growth, functionality, and ability to screen in 3D cellular formats, the ability to miniaturize and increase the assay footprints will expand as well. Although 3D HCS currently has been enabled to query selective questions regarding target function and cellular cytotoxicity, 6,7 the complexity of the models trends toward simple, practical formats. Complexity, then, in scaling the screening processes is a current boundary. Thus 3D formats such as 384 and 1,536 well and bioprinting formats will be required if 3D models are to be utilized in a high-throughput phenotypic screen setting.
In conclusion perhaps the key challenge is the retrospective benchmarking of each new 3D models against preclinical and clinical data to gain a better appreciation of their true predictive potential in early safety toxicity prediction and on-target functional therapeutic potential.
Applications of 3D Models in Research
Topic discussion leader: Serena Silver (Novartis, Cambridge, MA).
Panel members: Bonnie Fiedorek Sloane (Wayne State University School of Medicine, MI), Judi Wardwell-Swanson (InSphero, Inc.), and Glauco Souza (Nano3D Biosciences).
Session Summary
The push to develop 3D models of cell growth and tissue organization is generally driven by the desire to better represent in vivo processes and drug responses. This type of biology has a long history dating back to measurements of pharmacological and physiological responses of whole organs and tissues in “organ-bath” type assays to the present day, where we are seeing iPS-derived cells being engineered into functional systems in a 3D matrix. 8,9 Beyond the physiological significance of these 3D systems is the issue of how we measure them, and HCA using a variety of methods is clearly a viable option and was discussed in depth throughout the colloquium. The questions and discussions hereunder will help give the reader an idea of the viewpoints of experts within the room and likely sources of information for their own research.
Examples of complex 3D technology models discussed in this session included wound healing (n3dbio), cell migration (n3dbio), liver (InSphero), diabetes (InSphero), and MAME cocultures (mammary architecture and microenvironment engineering, B. Sloane). Glauco Souza indicated that the application of 3D models should generally depend on the question being asked, that is, relative to the cell/disease. There are traditional models for cancer, liver, toxicity, and stem cells. The N3D system provides rapid phenotypic models that bring cells together in a unique and rapid manner. Community concerns remain, however, to fully appreciate and understand how 3D in vitro systems reflect biological systems compared with 2D systems. The N3D system is being used in a personalized medicine context with comparisons with animal data alongside analysis of patient-derived cell models.
Judy Wardwell-Swanson from InSphero presented a compelling study of 3D model validation through partnership with pharmaceutical companies for characterization of model biomarkers; 3D naive liver microtissues were compared with liver tissue in both function and morphological architecture. The study compared 109 clinical compounds known to cause liver damage and showed that 3D models were more predictive of in vivo/clinical data with fewer false positives than 2D hepatocyte models. Other examples of simple metrics (cell count) and complex metrics (morphology and activity probes) were discussed comparing 2D and 3D results.
The term “being predictive” was mentioned throughout the discussion, but how predictive is it, actually? Very often when comparing results, we see things that are expected or that one is supposed to see. It was referred to as a problem of “Tautology”—the saying of the same thing twice in different words. In general though, at this point in time we lack sufficient quantitative comparative data between 2D and 3D models regarding predictive outcomes with respect to breath of functional activities as well as sufficient statistical number of examples. Although efforts continue to improve 3D and 4D (3D plus time) systems to highlight a representative number of compounds in comparison with other “predictive modeling” efforts, the sheer numbers are relatively low and much more needs to be done to overcome the questions such as are spheroids representative of the organ tissue? Similarly how to better model diffusion of nutrients, although not blood, that are functionally stable for multiday studies is challenging.
Bonnie Sloane, Wayne State University, spoke on spatiotemporal 3D models using live cell imaging and demonstrated the interactions of the tumor microenvironment with cellular and noncellular examples. Her discussions revolved around spatiotemporal modeling and the importance of time dependencies. This type of study provided end points for live cell assays to address changes and maturity of the cellular systems.
Audience participation commented on whether expansion microscopy and chemical clearing agents were being used in the 3D community. While discussion was pursued, it was mentioned that while using clearing agents in conjunction with light sheet microscopy to image spheroids has been employed, few have yet to employ it within the realm of HCS/3D spheroid studies. InSphero, however, report that they have employed clearing agents to image 150 to 200 μM into their spheroids.
CellProfiler 3D Features Miniupdate
Topic discussion leaders: Allen Goodman, Claire McQuin, and Anne Carpenter (Broad Institute of Harvard and MIT, MA).
An introduction of the newly released 3D features toolset incorporated into CellProfiler was described. Representative examples using a beta version were shown describing different cell types, 3D culture conditions, stains, and the current 3D capabilities.
A brief overview of future plans was presented, which includes adding capabilities to quantify 3D models at lower resolution, increase the throughput for organoids/tumor spheroids quantitation, add additional algorithms specific to linear structures such as myelin sheaths and metastatic-like stellate projections, and the creation of a 3D visualization and volume-rendering viewer.
3D Image Acquisition Section
Topic discussion leader: Daniel V. LaBarbera (University of Colorado, Anschutz Medical Campus, CO).
Panel members: Leslie Griner (Novartis, Cambridge, MA), Aaron Risinger (PerkinElmer), and Vipat Raksakulthai (Molecular Devices).
Session Summary
Over the past decade, there has been a boom in concerted multidisciplinary efforts by scientists and biomedical engineers with a vision of developing 3D ex vivo tissue models of human organ function and other models of human anatomy and disease. These 3D models are referred to as organoid, organotypic, or spheroid and are often used interchangeably within the collective literature. 10 For screening applications, a 3D model is only as useful as the information that can be extracted from it with appropriate high-throughput and efficiency. Dan LaBarbera led a discussion, joined by Leslie Griner, (Novartis), Aaron Risinger (PerkinElmer), and Vipat Raksakulthai (Molecular Devices), on the specific challenges of high-efficiency image acquisition of cells within 3D microtissues. Emphasis was primarily devoted to the new developments of automated high-content microscopy that address spheroid detection and “best imaging outcomes.” Optimizations such as U-bottomed and low-bind plates continue to enable high-resolution imaging techniques with maximal projections of spheroids (stacked z planes interposed for a 2D image) combined with high numerical aperture (NA) water immersion lenses able to resolve single cells. Other trends to utilize bright-field microscopy at low magnification before switching to higher magnification and resolution are being adopted to refine region of interest (ROI) locations in an automated manner. Although multiplane z stacks are typically produced, there are challenges with large spheroid/organoids because of the impact of limiting z ranges. Although inroads were described as to state-of-the-art practice, it was recognized that major challenges remain in image acquisition that limit the widespread HTS use of organoids for HCA/HCS drug discovery.
It was noted that many organoids are not spherical, 11 and the different formats have varying amenability to HTS/HCS, largely because of differences in either organoid shape or positioning within a larger volume. Liquid overlay, ECM embedded, mechanically supported culture systems, microfluidics, and microchip technologies were mentioned briefly. Panelist Vipat Raksakulthai addressed the challenge of efficiently locating rare events in a 3D space and described Molecular Devices' targeted imaging approach, in which a low-resolution objective is used for feature identification before centroids are logged, ROI drawn, and then higher resolution imaging performed on the smaller volume. In the example shown, the ROI was selected manually on a wide field platform, and the importance of correct parfocal alignment was highlighted.
Questions arose about the pros and cons of locating and scanning using image-based versus plate-based focusing technologies. PerkinElmer has traditionally offered plate-based solutions because they are faster and less invasive to the sample, although it was acknowledged that customer requests for image-based focusing are increasing. A brief discussion of ways to determine the most relevant focal plane followed, and ways in which even a fast range-finding feature—one capable of finding a relatively small sample in a relatively thick matrix, which could then be followed by higher resolution imaging—would provide significant time savings. There was a brief discussion of the challenges associated with the presence of multiple organoids within single images and the factors (e.g., step size, NA, and biology) that can play into object recognition. Although many of these were relatively new software packages, there was recognition that further developments in this area are needed.
Panelist Aaron Risinger focused on the challenges of identifying the z plane of interest and subsequently extracting meaningful data for analysis. Aaron described PerkinElmer's integration of imaging and analysis tools that facilitate this process, including in their Harmony software, tools for the identification and segmentation of 3D images for volumetric segmentation, and use of maximal projections. Aaron recommended microlens-enhanced spinning-disk confocal for 3D tissue resolution, and described work in progress to address other 3D-specific challenges, such as using texture analysis to resolve interior structures when photon budget may not allow more straightforwardly intuitive analyses. Although the standard identification of “best image” appears to be observer dependent, these approaches depend upon multiple types of data acquisition (single cell heterogeneity within a spheroid, subcellular signaling responses within the spheroid) with the trend toward merging the “best image” with the “best quantification and analysis.”
There was a brief discussion of ways to determine the most relevant focal plane and ways in which even a fast range-finding feature (one capable of finding a relatively small sample in a relatively thick matrix) that could then be followed by higher resolution imaging would provide significant time savings. Discussion and questions followed regarding whether analysis of image bleed-through issues was best performed on single planes, maximal projections, or in volumetric space. The challenges associated with the presence of multiple organoids within single images and the factors (e.g., step size, NA, and biology) that can play into object recognition resulting in skewed intensity results were recognized and discussed with no specific consensus as to how to generally proceed. Suggestions appeared to be research specific. The session ended with a discussion of the efficiency that might be gained from the minimization of differences in organoid positioning within a volume in any dimension, including in the z plane within a 3D space but also with respect to similarity in the x and y localization of objects within the individual wells of a plate (or growth positions of some other substrate).
3D Image Analysis Session
Topic discussion leader: Peter Horvath (Hungarian Academy of Sciences, Hungary).
Panel members: Scott Keefer (Thermo Scientific), Vipat Raksakulthai (Molecular Devices), Santosh Hariharan (Sunnybrook Research Institute, University of Toronto, Canada).
Session Summary
Peter Horvath introduced some of the major challenges currently faced for 3D HCA and in designing computational methods to analyze 3D volumetric data.
Image formation
With confocal and several light sheet microscopes, the axial (x–y) and lateral (z) resolutions differ, creating an additional computational correction step. On top of this, because of the axial–lateral asymmetry of the point-spread function, the interpretation of pixel information is nontrivial. These issues make all the following steps highly complex. Light sheet microscopy holds great promise in this area.
Image preprocessing
In addition to image formation problems, confocal microscopes usually have very high light exposure, resulting in photo bleaching and uneven exposure across the field of view (vignetting). There are now several advanced methods to correct these effects; however, none account for 3D, but instead treat volumetric data as a set of 2D sections. Another problem with 3D measurement and analysis is light penetration into thick samples. This decrease in illumination with thick samples is nonlinear and is crucial to correct.
Segmentation
This is probably the biggest challenge in 3D analysis. If the analysis is more complex than simple thresholding or connected component search, then 2D algorithms require major changes. The simplest region-growing algorithms only require extensions to march toward more directions. With more complex models such as active contours, length becomes surface, area becomes volume, and normal vectors are calculated in 3D. Recently, several novel algorithms were proposed that incorporate different shape and regularization. 3D image analysis on the scale needed for HCA is very resource intensive. If this analysis is to be carried out locally, graphic processing unit (GPU) processing shows very promising results, because of its ability to rapidly handle large number of simple calculations very quickly in a massively parallel manner. For organizations with greater available resources, large servers and cloud infrastructure might be the best solution, but the trade-off is the increased time needed to transfer of GB or TB of data.
Feature extraction
In common with segmentation challenges faced in going from 2D to 3D, feature extraction may also require different interpretations and analysis to cope with increased complexity. A simple example is the use of Haralick features, 12 which are one of the most frequently used morphology measurement methods. Co-occurrence matrices used with Haralick features became 3D, resulting in a large increase in complexity. The most commonly used features have already been worked out for 3D because of 3D medical imaging (CT and MRI).
A common method suggested by the panel for viewing 3D images was to use orthogonal and oblique slices through identified objects and image stacks. Structures can be identified and more easily visualized by using volume and surface rendering as well as false color mapping of intensities. Although not 3D, uses might also find it easier to view images as individual planes. Another common visualization method is to view all sections as a maximum or mean projection onto a 2D plane. To compose and view all planes as a 3D representation, the surface can be seen through surface rendering and viewed as solid or through alpha blending, a technique to make pixels transparent or semitransparent to see foreground and background simultaneously.
A full review of the processing steps for 3D to 2D projections is beyond the scope of this commentary; however, there are sets of tools that are commonly available in image processing and analysis packages capable of the following transformations to generate a projection of 3D objects into a 2D plane: min-max projection, standard deviation projection, and nearest neighbor deblur. Some further discussion on how these can be handled in Image J can be found in (
The conclusion of the panel and most of the audience to this question was, yes, it is possible to analyze 3D images as 3D objects in a HCS environment. However, as has already been mentioned, this is computationally costly, there is a lack of available tools, and those available do not always improve the quality of data over a 2D analysis. Volumes can be rendered, and then measured in 3D space, an example in which this might be most valuable is with neurite tracing where it would be critical to trace the neuronal network in 3D through a biomatrix or an ex vivo tissue section. Automated image analysis is currently an area of active research and is likely to gain traction and an increase in users in the near term. Feature extraction and segmentation tools can be expanded into 3D but with a great increase in complexity. Although each pixel in a 2D plane has 8 neighbors, this increases to 26 in three dimensions.
The two main bottlenecks for 3D analysis of high-content images discussed by the panel and audience were speed and software. Computation speed and cost to the user are closely correlated and it is anticipated that this cost will continue to fall. There needs to be further development and acceptance of software capable of dealing with scale and complexity and still be understood by the vast majority of users in the screening community. Future developments are likely to include increased use of GPU computing, and expansion of 3D techniques to include 3D plus time (4D) and incorporation of orthogonal measures.
Session Conclusion
A major problem in automated image analysis and visualization of large-scale 3D imaging data is the increase in complexity and scale compared with 2D image sets. There is also a lack of 3D algorithms and visualization tools, likely because of several factors. 13 Thus 3D problems are often simplified to 2D, using methods such as using maximum intensity projection, often leading to satisfactory results in a significantly reduced time. 14
3D HCS analysis needs a significant increase in computational resources compared with 2D analysis, typically requiring orders of magnitude more CPU time and storage. 3 This computational load might be more than can be easily handled by PCs or small servers, but probably not large enough to justify the use of supercomputing. Most laboratories currently make use of the former methods, whereas a few lucky imaging centers are able to take advantage of high-performance computational resources with sufficient data transfer and computational capacity (such as Max Planck Institute, Dresden).
There is significant effort being put into developing effective and impactful 3D image analysis tools, which can be used on the scale needed for HCS. There are still computational and experimental bottlenecks hindering the adoption of 3D HCS, and we look forward to usable and useful solutions, especially open-source tools, which would reduce the barriers for entry for all researchers.
For those interested in learning more about 3D image analysis, the following references provide an excellent starting point. 3,14,15