Society of Biomolecular Imaging and Informatics High-Content Screening/High-Content Analysis Preclinical Translational Imaging Colloquium: How Emerging Technologies in Disease Models,High-Content Imaging,and Data Analytics Are Changing Our Approach to Drug Discovery

Introduction

For the fifth annual SBI ² conference, as in previous years, the afternoon session on the opening day was devoted to a colloquium reflecting current trends in the field. This year the focus was on the impact of emerging technologies in high-content analysis, cellular model systems, and data analytics, and their impact on drug discovery (DD). A panel of key opinion leaders reflected on their experiences in technology development, disease relevance, and high-dimensional data analytics, and offered perspectives on practical applications, translational approaches, and gaps in the state of the art.

The colloquium featured two main discussion topics: clinically relevant model systems and informatics. To address these topics, sessions were broken down as followed:

Biological relevance, cell profiling, and screening

Invited Key Opinion Leader Panelists were as follows:

Rochelle Prantil-Baun, Senior Staff Scientist, Wyss Institute, Harvard University, Boston, Massachusetts

Biological Relevance, Cell Profiling, and Screening

Key to the success of the technology improving DD is the ability to identify opportunities in a relevant (if not predictive) context. Pipelines have taken the reductionist approach to DD to quickly hone in on tractable options. Reducing the readouts to simple scalable systems comes at the cost of poor translation. Innovations in cell model systems (induced pluripotent stem cells [iPSCs], gene editing, etc.) and the ability to multiplex using high-content multiparameter approaches have enabled the field to capture needed biological relevance, improving the translatability of our screening hits.

The panel first addressed innovations around data acquisition, including model systems and modality (i.e., high-content analysis). Panel members offered that pharma has been working for years on physiologically relevant cells. Primary and patient-derived iPSCs have been the work horses for disease relevance. However, it has become evident that there are limitations to practical application of these cell models: (1) variability and scalability and (2) data storage. The resources required for full deck screen are difficult to procure for most groups using primary cells and to a lesser extent iPSC-derived cells. Two areas should be taken into consideration when evaluating these resources. First, researchers should adjust their expectations for model systems challenged by scale. Rather than a full deck (∼100K-2M interrogated wells), focused libraries might be used to inform biology around specific target classes, as well as offer starting points for chemistry. Second, more features from these cell models should be extracted, making high content the amenable platform for enriching cellular readouts. In addition, single-cell data should be considered for assay readouts, especially as complexity and heterogeneity in the cellular systems increase and the analytical methods to analyze these complex data sets matures and becomes more available.

Even with the challenges posed by limited scale, content-rich readouts can still burden the task of DD. The balance of cost/benefit for this approach depends on to what degree the DD scientist wants to inform biology or create diversity in compound series to follow up. A possible solution to bring both biological insight and compound diversity into the campaign could be a link between the image-based readouts and “omics” (e.g., transcriptomics, proteomics, and metabolomics). Another solution could be establishing phenome databases against which perturbations—compounds from a screen or genetic mutations—could be compared. Orthogonal “omics” readouts and a phenome reference set offer annotation for HTS to interrogate phenotypes, which could result in a more meaningful use of data. Furthermore, as the data density builds across focused and randomized campaigns, the researcher establishes the consequences of targeted perturbations using feature readouts as profiles.

Cell painting, an image-based technique of labeling cellular compartments and measuring hundreds of features using high-content analysis, offers a means to systematically assess gross cellular profiles across disease models and modes of perturbations. By taking a chemogenomic approach of cell painting, profiles can be generated as reference phenotypes to facilitate the elucidation of compound mechanism of action and pathways associated with disease risk. Furthermore, cell painting typically establishes large data sets to which multiple analytical techniques, such as those that use artificial intelligence, can be deployed as a means of maximizing the insight of the biological impact of a given experimental condition.

The panel admits that there is a growing interest in cell painting as a technology to probe biology, differentiate compounds, identify mechanism of action of novel compounds, and expose potential liabilities of leads through toxicity profiling. However, “ground truth” has yet to be established with this approach, and the robustness of cell painting remains unknown, which continues to feed skepticism in its deployment. Contributing to the skepticism is the uncertainty around the breadth of biology that can be explored with cell painting. In addition, the variety in cell type, growth conditions, imaging platforms, and other experimental design components makes it difficult to assess the value of the platform without a ground truth reference. The seminal work by the Carpenter Lab ¹ details a protocol in one cell line. One early litmus test could be the direct comparison of this data set to an in-house-generated data set. Pharmaceutical companies are actively pursuing this exercise through alliances, commercial and academic venues, and government institutes. The problem is that the results may not be shared among industry stakeholders, making the overall attempt to assess cell painting's robustness difficult. Panel members pointed out that cell painting consists of profiles generated by in-house groups with the assumption that the resulting data achieved high-quality metrics and robustness. Thus, how important is it to have a cross site comparison? Furthermore, the cell type used by each could (and does) vary, and it could be argued that this limits the utility of cross site reference. They added that it remains risky to deploy this technique on human primary cells, due to the inherent variability in the cell culture, making it difficult for a screen, especially when up to four to five replicates may be needed to interpret the heterogeneity of the primary cells themselves. However, cell painting could clarify the consequences of different hit clusters from a phenotypic screen.

A shared interest on the panel was the establishment of databases of phenotypically active compounds. Leveraging the phenome of bioactive compounds creates a valuable reference source for deconvolving mechanisms of action. According to our panel, industry stakeholders are leveraging such phenomes the same way as informatics. This screening strategy supports the use of morphological, functional, and protein localization examinations as key features to quantify. Moreover, public data sets could provide supplemental support to in-house annotation and answer: Can annotation in phenotypic screens support the insight from phenome databases? It is imperative that industry participants (pharma, biotech, etc.) embrace phenotypic screening. Recent commentaries have documented this trend. ^{2 –4}

Patient Phenotyping for Instructing Phenotyping Screens

Primary cells and patient-derived iPSCs have been used extensively in phenotypic screens. The value of these systems has escalated with the advent of gene editing. Without question, the panelists advocated for the use of isogenic lines to provide a clearer contribution of mutations to disease biology. Isogenic lines also offer a means to clearly differentiate compound effect on targets while maintaining high-content readouts. This provides a deeper dive into disease biology in the early phase of DD than does the targeted approach. These systems are also amenable for small-scale deep phenotyping and biologics DD.

A consensus from the pharma representatives on our panel cautioned that the cost of patient-derived cells, iPSCs, and complex model systems, such as the three-dimensional (3D) systems (discussed as follows), is relatively high. Although there are indications that some 3D model systems recapitulate disease biology, which fail to be modeled in two-dimensional (2D) cultures, it is not clear to what extent this is generally true. Most screens continue to be performed in 2D monolayer cultures emphasizing a target pathway or functional response (i.e., mitosis and neurite outgrowth). With the promise to represent disease biology with nascent genetic background, it remains to be seen how such cells direct DD toward success. These complex model systems are expensive and require significant infrastructure from tissue and cell culture to hardware through data acquisition and data storage.

The panel advocated for the continued effort toward biological relevance and cellular profiling for driving early-stage DD and supporting screens. Managing the expectations of appropriate deployment of such technology is imperative, especially when it comes to the use of precious patient-derived material, such as primary cells. Rather than attempting to achieve high-density plate formats for each case, the panel endorsed the use of 96-well plates and medium throughput. This format accommodates both discovery biology and “hit to lead” in vitro pharmacological studies. Also, utilization of multiple primary cell lines for each study was strongly suggested to mitigate the challenges of line-to-line/patient-to-patient variability. The balance between general and specific responses should be assessed through the use of such data sets. One must accept that scalability and biological relevance are inversely related; profiling fewer conditions can provide more insight into disease biology and/or compound effects. Understanding these aspects will enable the researcher to take a more impactful yet parsimonious approach to deploying biological relevance in a phenotypic screen.

Influence of Artificial Intelligence and Omics On Hca/hcs

Unlike the colloquium of 2017 where the panelists and audience voiced some exposure and knowledge with artificial intelligence and machine learning, this year's panelists and audience illustrated specific cases and experiences using these techniques. ^5,6 Although the general adaptation and integration within the current HCA/HCS workflows was less common, these proof-of-concept examples were provocative. Literature and recent commentaries confirm the DD interest in utilizing neural networks in the processing of cellular images. ⁷ Interest and discussion revolved around how to best implement the key components of such a workflow, what specific types of data sets might produce greater successes to predict meaningful interpretations, and how “ground truth information” is integrated within training sets. The consensus opinion was that with ample robust imaging data, omics, and detailed compound annotation information, AI is an important new pathway for DD successes. As the challenges continue in gathering the overhead in expertise, hardware, and sufficient large high-quality data sets, there remained an underlying question regarding what type of data is best to use.

Conventional feature extraction of cellular and tissue images is a well-established method in analyzing high-content screening images. Instrument-specific and instrument-agnostic software enable user interfaces that abolish the need for any algorithm development with minimal interventions for optimizing segmentation. Although the process can still be a time-consuming task and a moderate level of expertise with image analysis tools are required, the newer AI methods such as deep convolutional neural networks can also address these problems. These models can automatically extract informative features from raw pixel values and does so in less time with less user effort and likely bias, resulting in a richer phenotypic analysis of these images. Thus, AI analysis may help close the gap established within workflows as robust predictive tools to combine data sets and help generate stronger hypotheses within the context of DD.

One panelist described the clinical challenges to collate the overwhelming individual patient data, deliver it to the AI teams, and await an output that could be utilized for patient decision making. As a key issue for medical successes is a timely response in prescribing an action plan, he described that the observable dominant features are the stable keys to the pathology of the system such that tailoring for those alongside the developing AI processes is current practice. Further discussion focused upon how to make sense of the vast amount of patient-generated data (electronic health records, biopsy analysis, primary human tumor analysis, and patient-derived tumor xenographs) and where AI pathways can help close not only the gap of confident predictive outcome but also the timeliness of those results.

Higher Dimensional Models for Dd: Ooc and 4d

The final points of discussion for this year's colloquium focused on 3D/4D biological model systems for DD. As alluded to earlier, typically the higher complexity in model systems—at attempts to recapitulating disease biology—is inversely proportional with scalability. However, if the model is more predictive, its earlier deployment in the DD pipeline could facilitate higher efficiency and success rates to identify new and better drugs. Successes of in vitro and in vivo correlations with OOC were highlighted as better models for complex diseases, especially those where multiple tissues and systems are affected (e.g., neurodegenerative diseases and metabolic diseases). In addition, these systems offer insight into compound distribution and bioavailability.

It is not impossible to think that OOC could help predict PK/PD in vivo. Could the OOC be the “killer App” for DD, although not likely for the primary screen? It is possible that OOC for specific disease niches could serve as practical and correlative models before investing in either animal studies. By deploying such platforms just upstream (target validation) and downstream (hit validation) from the primary screen, it is likely that the OOC will facilitate deeper phenotyping than on conventional 2D substrates. Our panelists echoed that the OOC would likely enable more disease-relevant biology than other cell-based systems, particularly when it comes to the all-important microenvironment—a source of novel biomarkers and putative targets.

Nonetheless, the challenges are substantial for developing these systems, and it is essential to understand their limitations and fit for purpose. Moreover, these systems may not be suitable for many of the high-content and other image-based systems, requiring customized systems. The panel suggested two areas that could facilitate the adoption of the OOC model for DD: (1) use these systems to build better predictive models against which hits could be triaged and (2) develop models that the FDA could find supportive of the disease modulation, such as toxicity prediction. OOC offers the chance to examine multiple types of readouts concomitantly, such as morphology (image based) and secreted protein and lipid factors (biochemistry). The real advantage of OOC is the noncell autonomous biology that can be explored. Panelists strongly suggested that one take a very specific (if not targeted) approach to validate the readouts with caution as general property trends may invalidate the model. Specifically, the key to the utilization of OOC systems is to identify the dominant features recapitulated in the disease biology.

Two challenges were mentioned for these 3D/4D systems: (1) The biology remains limited within a specific microenvironment system bridging the gap from micrometers to millimeters. These systems are not organisms and still represent a small cross section of disease biology. How one then identifies and defines the appropriate metrics representing the disease biology robustly becomes the task. (2) These systems are prone to variability. More insight using omics level could be used to normalize the data or at least help manage expectations. However, one must approach the idea of heterogeneity differently—if not embrace it. Could single-cell analyses capture value in these systems rather than mean readouts? Could capturing kinetic measurements be necessary to identify the appropriate point of intervention? Opportunities such as these are the present and future arenas of deep investigations.

Perhaps the most provocative use for OOC in DD brought up by our panelists was in precision medicine. In contrast to expanding systems into general models, the OOC could offer a system to deploy patient-derived cells/tissues to better understand disease biology of individuals, medical interventions, and enable a stratification process of responding populations. The challenge, as several panelists pointed out, is the reliability of such systems in predicting outcome, which must be empirically determined. However, the combination of multiple platform readouts, the integration of multiple cellular/organ systems, and the opportunity to deploy gene editing, pharmacology, and other perturbations to OOC offers new avenues to explore disease biology and predict therapeutic efficacy—all in a human system.

Summary

Our goal for the colloquium was to address to what degree advances in high-content analysis, cellular models of disease, and data analytics are improving DD. In addition, we challenged aspiration with practical feasibility in an honest discussion around “biological relevance, cell profiling, and analytics” to identify gaps. What emerged was an enlightening discussion around the balance of the potential of enabling technologies offering new opportunities and a healthy skepticism of tractability, given the resources needed to deploy these technologies. As expressed by many of our key opinion leaders, there have been many recent advancements in the development of cellular models, computational data analysis techniques, all combined with accessible omics and imaging data sets. With that, there exist the next stages of new opportunities to advance discovery outcomes by combining biologically relevant and cellular phenotypic profiling data to drive more successful and translational early-stage DD processes and results.