Digging Deeper: Methodologies for High-Content Phenotyping and Knowledge-Abstraction in C. elegans

Abstract

Deep phenotyping is an emerging conceptual paradigm and experimental approach that seeks to measure many aspects of phenotypes and link them to understand the underlying biology. Successful deep phenotyping has mostly been applied in cultured cells, less so in multicellular organisms. Recently, however, it has been recognized that such an approach could lead to a better understanding of how genetics, the environment, and stochasticity affect development, physiology, and behavior of an organism. Over the last 50 years, the nematode Caenorhabditis elegans has become an invaluable model system for understanding the role of the genes underlying a phenotypic trait. Recent technological innovation has taken advantage of the worm’s physical attributes to increase the throughput and informational content of experiments. Coupling these technical advancements with computational/analytical tools has enabled a boom in deep phenotyping studies of C. elegans. In this review, we highlight how these new technologies and tools are digging into the biological origins of complex multidimensional phenotypes seen in the worm.

1. Introduction

One of the great drivers of biological research in the 20th century was the desire to understand how the information encoded in an organism’s genome gives rise to its physical and behavioral phenotypes, collectively referred to as the organism’s phenome¹. However, the phenotype an organism displays is not merely a reflection of its genes but the integrated product of its genotype coupled with the environmental conditions and stochastic effects that the individual experiences before observation^{2, 3}. This interplay means that an organism’s phenome contains a large multi-dimensional set of observable characteristics. Thus in a population of individuals, the resultant phenospace is vast. Understanding the specific contribution of an individual’s genome, its environment, and stochastic processes to its position in phenospace presents a formidable technical challenge. Fortunately, the pace of technological innovation in modern biology is beginning to provide the tools that are necessary for both quantifying these multi-dimensional characteristics and their underlying causes, opening up a new experimental paradigm known as deep phenotyping⁴. As a consequence of its relative newness, the term ‘deep phenotyping’ is used broadly within the scientific literature. In this review, we define deep phenotyping as the coupling of high-throughput experimental techniques with computational analyses to enable the generation, examination, and interpretation of high-dimensional biological data.

Caenorhabditis elegans has many attributes that make it an ideal model system for deep phenotyping studies. For example, the worm is a small poikilotherm with an adult body length of ~1mm that feeds on bacteria. It has a rapid life cycle, going from egg to reproductive adult in ~3 days at 20°C with a deterministic developmental lineage⁵. These features coupled with the fact that the worm primarily reproduces as a self-fertilizing hermaphrodite mean that it is possible to grow large near-isogenic populations in highly controlled environmental conditions. Also, at least ~38% of the protein-coding genes in the worm genome have a human ortholog⁶, which means that insights gained from studies in the worm can inform us about human biology. For example, genes involved in apoptosis⁷, axonal migration^{8, 9}, as well as microRNAs^{10, 11} and the phenomenon of RNA interference (RNAi)¹² were all initially identified in C. elegans and have since been shown to have roles in human disease^13–15.

In this review, we outline the technological and analytical developments that have enabled deep phenotyping studies in C. elegans (Section 2). We then examine how these tools have yielded greater insight into the biology of complex phenotypes using several different examples (Section 3). Finally, we offer a prospective outlook on the future of deep phenotyping experiments in C. elegans and other systems (Section 4).

2. Recent Development of Tools and Techniques that Enable Deep Phenotyping

Historically, phenotypic analysis of C. elegans relied on manual measurement of morphometric or behavioral features, such as alterations in body shape or defects in movement¹⁶. However, modern biological techniques have dramatically expanded the definition of phenotype¹. For example, access to RNAseq and mass spectrometry is now more widely available and less cost prohibitive, enabling transcriptomic and proteomic descriptions of individuals. Similarly, advancements in both hardware and computational tools have led to an increase in both the throughput of experiments and their informational content. In this section, we examine three specific areas that have enabled deep phenotyping studies of the worm.

2.1. Manipulating the genome

A sophisticated set of techniques allows manipulation of the worm genome to elicit phenotypes, including both forward and reverse genetic approaches⁵. RNAi of gene expression can be achieved by feeding worms bacteria expressing double-stranded RNA corresponding to a gene of interest^{17, 18}. This has enabled multiple reverse genetic screens to uncover the phenotypes associated with various gene inactivations^{19, 20}. More recently, several groups have introduced CRISPR/Cas9 methods that are optimized for manipulating the C. elegans genome^{21, 22}. These methods allow for the rapid and efficient knock out of genes or introduction of fluorescent markers of either gene or protein activity.

2.2. Hardware employed in deep phenotyping

The hardware development that has enabled deep phenotyping studies in worms falls into two categories. The first is an improvement in the technology that allows handling of worms and the second is an improvement in the imaging technology that records the output of an experiment. Both categories have seen significant reductions in component costs and the accumulation of many designs and proof-of-principle experiments, which increases the accessibility of the hardware needed for deep phenotyping experiments. We highlight some technical aspects in this section and give specific examples in section 3.

Manual handling of worms is labor intensive and acts as a barrier to the scale and scope of an experiment. However, the worm’s small size makes it very easy to manipulate using microfluidics^{23, 24}. In the last decade or so, the microfluidic devices used by the C. elegans research community have been fabricated mostly from polydimethylsiloxane (PDMS). The fabrication of devices with this material is relatively cheap and straightforward, and PDMS is non-toxic to worms. It is also optically transparent, which makes it compatible with many forms of microscopy used to study the worm. Its elastic properties allow for simple on-chip valves that can control the flow of fluid containing the worms, which enables automation of animal handling leading to increases in sample throughput^{23, 25–27}. Microfluidics can also be used to tightly control the microenvironment surrounding the worm within the device, which is hard to do on an agar plate. Device/system designs that are widely used for worm experiments (Fig. 1) include arena or multi-chamber arrays^28–31, imaging and sorting devices^{24, 27, 32–37}, and devices enabling more complex manipulations^38–41, specific examples are described in section 3.

Figure 1.

Microfluidics enables high-throughput experimentation in C. elegans. Examples of microfluidic devices routinely used in worm deep phenotyping studies. A. Multi-chamber arrays for simultaneously studying large numbers of individual worms (adapted from Chung et al²⁹) B. Sorting devices used to rapidly isolate worms with specific characteristics (adapted from Chung et al²⁴) C. Arena devices for behavioral assays (adapted from Albrecht and Bargmann²⁸)

C. elegans has proven itself to be amenable to multiple imaging modalities. Being optically transparent, early in vivo studies of worm development^{42, 43} used Nomarski optics to provide sufficient contrast to view individual cell nuclei. Ultrastructural studies of C. elegans have relied on various forms of electron microscopy⁴⁴. Increasingly, high-throughput experiments rely on some form of automated imaging, mostly involving fluorescent reporters. The demonstration of fluorescent protein reporters in C. elegans⁴⁵ paved the way for new forms fluorescence microscopy that are compatible with high-throughput experimentation. Examples include light sheet and lattice light sheet microscopy (LLSM) platforms that have been used for studying embryogenesis^46–49, as well as protein dynamics in adult worms⁵⁰. Recently, a new form of LLSM with adaptive optics has been used to study cell movements during vulva development in stunning detail⁵¹. Several new microscopy platforms, including two-photon-based⁵² and light field^{53, 54} systems, allow fast volumetric ‘whole-brain’ imaging of calcium dynamics in C. elegans. A new form of volumetric flow cytometry using line excitation array detection was used to monitor protein aggregation in worm models of Huntington’s disease⁵⁵. Several super-resolution microscopy systems have also been developed used to study embryogenesis^{56, 57}, the structure of muscle⁵⁸, and the distribution of glutamate receptors⁵⁹ in C. elegans.

Unlike imaging of cellular structures or activities, many high-content behavioral studies utilize dark-field or transmission imaging to enhance the contrast of the transparent worms to allow longitudinal tracking and quantification of whole-animal phenotypes^{60, 61}. The hardware needed for such systems is becoming cheaper and more readily available^62–65. Behavioral tracking systems have also been integrated into microscopy systems that can use targeted illumination of specific cells in worms in optogenetic experiments^{66, 67}. Such systems have been used to determine how the mechanosensory system of the worm encodes spatial and temporal information about stimuli to ensure appropriate behavioral responses^{68, 69}. The increased availability of these systems has been central to the proliferation of behavioral deep phenotyping studies (see section 3.5).

2.3. Computational Techniques for Deciphering High-Content Phenotypic Information

Due to hardware-related gains in throughput, many experimentalists now face a dramatic increase in the amount of data produced from experiments. Making sense of these data can be beyond the analytical capability of human vision or comprehension. Therefore, extracting useful information from large volumes of data requires workflows that can parse the content efficiently and in an interpretable manner. For image-based data, the typical structure of a deep phenotyping workflow involves segmentation to identify objects of interest within each image^{70, 71}. This is then followed by feature extraction and classification, which provides a quantitative breakdown of the differences between images allowing the assignment of phenotypic profiles^{70, 71}. Open-source image analysis platforms such as ImageJ^{72, 73} and CellProfiler⁷⁴, both of which have C. elegans-specific plugins^75–77, conveniently offer many of the computational tools required to analyze high-content data.

Several different segmentation algorithms have been demonstrated in C. elegans image informatic workflows. These include algorithms to accurately segment fluorescently-labeled nuclei for cell identification and tracking, either during embryonic development^78–81, within the context of a digital atlas of the L1-stage larva^{82, 83}, or for tracking calcium transients across neurons in ‘whole-brain’ imaging data^{84, 85}. Algorithms also exist for segmenting fluorescently-labeled synapses to study synaptogenesis^{32, 86}, as well as automatically tracing fluorescently-labeled neurites⁸⁷ in the worm. Brightfield images can also be automatically segmented to calculate various size-related metrics of worms using the ImageJ plug-in WormSizer⁸⁸.

Machine-learning is increasingly used to automate segmentation, feature extraction, and classification of images in high content workflows^{70, 89}. DevStaR⁹⁰ is a machine learning-based platform that has been developed to score phenotypes in worms automatically. This system can segment brightfield images and quantify animals belonging to different developmental stages in a mixed population. Zhan et al. developed a modular image-processing pipeline that allows rapid development of custom supervised learning-based classifiers employing support vector machines (SVM)⁹¹. This pipeline has been used to train a classifier to identify the head of the worm in brightfield images⁹¹, as well as classifiers to identify and differentiate between the ASI and ASJ⁹¹ and ASI, ADF and NSM neurons⁹² in fluorescence microscopy images. WorMachine is another modular pipeline that can automatically score a wide-range of phenotypes using trained classifiers⁹³. This system can identify the sex of the animal, as well as quantify complex partially penetrant phenotypes or intracellular protein aggregation in a group of worms⁹³. The development of these image analysis pipelines allows researchers without machine vision expertise to design and perform deep phenotyping studies in C. elegans. Additional examples of machine learning in image informatics and behavioral classification are provided in sections 3.2 and 3.5, respectively.

3. Recent Applications of Deep Phenotyping

Given that the hardware and computational tools to perform deep phenotyping experiments in C. elegans now exist, we next examine how these technologies have been used and what they have taught us. In this section, we review recent studies covering multiple areas of worm biology.

3.1. Embryogenesis/lineage tracing

C. elegans is one of the few metazoans for which the entire somatic cell lineage can be traced from single-cell embryo to adult^{42, 43}. However, tracing cell lineage in the developing worm embryo is exceptionally challenging due to the rapidity of cell division and morphological similarity of the cells⁹⁴. The relative complexity of identifying phenotypic aberrations in the embryonic lineage compared to that of the post-embryonic cells has meant that the mechanisms governing embryonic cell division and differentiation have been harder to elucidate. The advent of 4-Dimensional imaging systems^{95, 96} removed the need for manual observation of embryogenesis but not curation of the resultant images into lineages. Deep phenotyping is still possible using manually curated data, for example, a study involving a whole genome RNAi screen identified 661 genes involved in the early embryogenesis⁹⁷. A subsequent study integrated transcriptional, protein-interaction and visual phenotypic data of these 661 genes to create predictive models of cellular events during embryogenesis⁹⁸ (Fig 2A). The use of transgenic worms that ubiquitously express fluorescently-tagged histone proteins has enabled the development of automated tracing algorithms that track the embryonic lineage through to the 350-cell stage^{47, 79, 80, 99–101} (Fig 2B).

Figure 2.

Deep phenotyping is a tool to produce informationally rich datasets. A. Integration of transcriptomic, proteomic and phenotypic data can be used to create models of cellular events in early embryogenesis. (adapted from Gunsalus et al⁹⁸). B. Use of two-color imaging can aid automated embryonic lineaging (upper panel) which can then be used to create spatiotemporal maps of gene expression (lower panel) (adapted from Du et al¹⁰⁸). C. Deep phenotyping can reveal a broader swathe of the phenotypic spectrum than traditional screens. Subtle mutants (labelled in pink) lie closest to wild-type (WT), while most previously identified mutants are the farthest from WT (adapted from San-Miguel et al⁸⁶).

With automated image collection and curation in place, it is now possible to examine different aspects of embryonic cell division. Several studies used automated lineage tracing to identify the precise cellular expression patterns of known embryonic genes^102–104. This has allowed for the construction of a single cell resolution atlas of gene expression revealing when and where transcription factors are expressed in the developing embryo^{103, 104}. It is now possible to ask how different genetic perturbations affect embryogenesis. Early examples of this include identifying the subtle roles of a single transcription factor¹⁰⁵ and the distinct roles of highly-related/recently duplicated genes¹⁰⁶ in defining different aspects of the embryonic lineage. More recent studies have demonstrated the roles of several hundred genes in cell fate choice^{107, 108}, and the regulation of asynchronous cell division¹⁰⁹. A specific screen of chromatin regulators has also revealed distinct roles for several chromatin-modifying complexes during embryogenesis¹¹⁰. Together these studies are revealing the genetic programs and molecular mechanisms that specify how a single fertilized oocyte becomes an embryo containing a complex collection of differentiated cell types.

In contrast to tools available for deep phenotyping of the embryonic development, the ability for high-throughput lineaging of post-embryonic worms has been lagging. However, recently Keil et al. have demonstrated a microfluidic device that allows the long-term culture of larvae coupled with the ability to routinely immobilize the animals in order to take high-resolution images with a variety of microscopy techniques¹¹¹. Using this platform, the authors demonstrated the ability to image animals from the L1 larval stage through to reproductive adulthood, while examining vulva precursor cell development, apoptosis during larval molting, neuron differentiation and neurite outgrowth. This type of platform coupled with future developments in post-embryonic phenotyping is likely to lead to a complete description of the biological events involved in the development of a multicellular animal.

3.2. Automated high throughput genetic screens

One of the most potent aspects of C. elegans as a model system has been the ability to carry out forward genetic screens in order to identify mutations that affect all possible elements of a gene. However, most genetic screens rely on visually identifiable phenotypic differences from the control, which inherently limits the ability to identify mutations that have weak or non-obvious phenotypes but still provide valuable information about a gene’s function. The power of deep phenotyping in identifying subtle mutants that may be missed by visual inspection was demonstrated in automated screens to find mutations affecting synaptogenesis^{32, 86, 112}. Using a microfluidic sorting system³² (Fig 1B), hundreds to thousands of worms were continuously imaged, processed, and sorted in real time. An online SVM-based image processing algorithm was developed to classify multidimensional features of fluorescently-labeled synapses on the fly^{32, 86}. Clustering of these multidimensional features of all the worms from the screen revealed the phenospace of the entire mutational spectrum^{32, 86} (Fig 2C). These clusters indicate where each new mutant resides in phenospace, allowing the inference of their potential genetic relationships. This study offers a powerful example of how deep phenotyping integrates high-throughput hardware with computational tools to provide mechanistic insights that would not have been possible for a human observer.

3.3. Measuring aging and age-related decline

Over the last 30 years, there has been a concerted effort among scientists to understand the biology of aging. C. elegans is the premier model organism for the study of longevity, due to its short lifespan and powerful genetics¹¹³. The majority of lifespan studies in worms are performed manually by periodically examining animals maintained on agar plates under a stereomicroscope for either spontaneous or stimulated movement¹¹⁴. These manual experiments place constraints on the types of phenotypic and scale of demographic data that can be obtained by a human observer. Deep phenotyping technologies offer a more efficient and cost-effective way of collecting high-throughput high-temporal resolution lifespan data.

The first push towards automating aging studies in worms used scanners to capture time-lapse data on worm movement on agar plates^{115, 116}. In these systems, death is defined as a persistent absence of movement, as there is no ability to stimulate the worms to test for induced movement. Stroustrup et al. used their automated system, termed the Lifespan Machine¹¹⁶, to examine the demographic features of large populations¹¹⁷. The authors found that at the population-level aging appeared to scale temporally between groups that had very different mortality rates due to a variety of different factors. This scaling implied a single effective rate constant of aging in C. elegans and that interventions that altered longevity did so by changing this rate¹¹⁷.

These scanner-based lifespan studies are unable to gather data at the level of the individual, as their temporal resolution does not allow the tracking of each worm until it dies¹¹⁶. As a result, these systems cannot assess the level of inter-individual variation in lifespan within a population, which can be dramatic even for isogenic worms raised in identical conditions¹¹⁸. Two groups have recently demonstrated robust high-throughput automated acquisition of lifespan data from individual animals^{30, 119}. Pittman et al. use a polyethylene glycol hydrogel (PEG) substrate that is seeded with spots of E. coli as a food source for developing embryos that are individually placed in a single spot, once the embryos are deposited the substrate is sealed using a layer of PDMS³¹. The PDMS acts as a barrier confining each animal upon hatching to its local spot of E. coli. The PEG substrate can then be imaged using a wide array of microscopy platforms, allowing the tracking of multiple features of the development and lifespan of each worm individually. Zhang et al. used this system to show that when tracked at the level of the individual, aging does not display temporal scaling, with long-lived animals displaying an extended decrepitude phase the authors termed ‘twilight’ compared to short-lived individuals¹¹⁹. The WorMotel platform^{30, 62} can record both spontaneous and stimulated movement, which is evoked by the use of a brief pulse of blue light, to assess whether a worm is alive or not. Using this system Churgin et al. also demonstrate that short-lived and long-lived individual lifespans from an isogenic population do not temporally scale³⁰. The differences seen between a population-level measure of longevity and those that account for inter-individual variability demonstrate the increasing power of deep phenotyping technologies.

Dietary restriction (DR) is one of the most robust interventions that has been shown to extend lifespan in an evolutionarily conserved manner¹²⁰. There is considerable interest in finding drugs that can induce the phenotypic effects of DR without the concomitant need for drastic diet alteration in humans. Using the lifespan machine¹¹⁶, Lucanic et al. designed a high-throughput screen designed to identify potential DR mimetics in C. elegans¹²¹. The authors screened through a library of 30,000 small molecules, identifying 57 compounds that repeatedly extended the lifespan of treated animals versus control worms. Several of these compounds were structurally related, containing a nitrophenyl piperazine moiety, and further analysis of the most potent of these, NP1, suggested that it extended longevity by inducing DR-like effects¹²¹.

One of the reasons for the rise in interest in the study of aging is that the incidence of both cognitive impairment and neurodegenerative disease increases with old age, which imposes a significant societal cost as the proportion of elderly individuals relative to those of working age in the general population continues to grow¹²². There is a strong desire within the research community to develop treatments that can counteract the debilitating neurological effects of aging. C. elegans also displays age-related declines in cognitive ability and the morphological structure of its nervous system¹²³. Bazopoulou et al. have developed a microfluidic platform to monitor the calcium responses of a specific polymodal neuron, ASH, as it ages¹²⁴. They then used this system to conduct a pilot screen to identify small molecules that could delay the age-related decline of ASH activity. Several molecules from a panel of 107 FDA-approved compounds delayed the decline of calcium responses in ASH during aging, the most potent of these were Tiagabine and Honokiol¹²⁴. This study demonstrates the power of deep phenotyping technologies to acquire dynamic readouts, such as neural activity, in a high-content manner.

3.4. Drug discovery and small molecule biology

Deep phenotyping is increasingly being embraced by researchers in the field of drug discovery as it allows rapid screening of the pleiotropic effects of different molecules. Given its ease of culture and amenability to high-throughput experimentation, the worm has been used to model many different human diseases^{125, 126}. While many researchers are developing methodologies to use these worm models of disease to speed up drug discovery, these platforms do not fit our definition of deep phenotyping as they tend to analyze single and relatively simple phenotypes. However, it is worth noting that as they serve as the as signposts for where we hope newer deep phenotyping platforms will be able to take drug discovery screens in model organisms.

In a worm model of cirrhosis, a mutant human α1-antitrypsin (ATZ) fused to GFP aggregates within the endoplasmic reticulum of intestinal cells replicating the phenotype seen in diseased hepatocytes¹²⁶. A plate reader-based screen of a commercially available compound library yielded 33 compounds that decreased the rate of GFP-aggregation within the worms’ intestinal cells¹²⁶. More recently, a high-throughput genome-scale RNAi screen of ATZ model worms was performed to find gene inactivations that alter the intestinal GFP-aggregation¹²⁷. RNAi of 100 genes led to decreased levels of GFP-aggregation in the ATZ worms. An in silico approach then identified drugs that are known to target the mammalian orthologs of the worm genes and tested them for the rescue of the GFP-aggregation phenotype.

Poly-glutamine expansion (polyQ) diseases, such as Huntington’s, are another example of a protein aggregation disorder that has also been successfully modeled in worms¹²⁸. Recently, a high-resolution microfluidic system was unveiled for high-throughput drug discovery using a worm polyQ model as a demonstration of the potential of the platform¹²⁹. Using this platform, Mondal et al. rapidly screened ~100,000 worms through their device testing 983 FDA-approved compounds for their ability to reduce YFP aggregation in a high-polyQ strain¹²⁹. The screen resulted in four compounds that had a statistically significant reduction in protein aggregate formation. Another high-resolution microfluidic system has been developed to identify compounds that affect axon regeneration in C. elegans¹³⁰. This platform is capable of performing high-throughput laser microsurgery specifically on the axons of the PLM neuron and then moving the axotomized worms into media containing different compounds to test for their effects on neurite regrowth. This screen revealed that the compounds staurosporine and prostratin, respectively, inhibit and promote axonal regeneration¹³⁰.

Application of deep phenotyping technologies may also enable the cost-effective study of widespread but often neglected diseases. Parasitic nematodes are thought to infect a billion people worldwide and are also a significant source of infection in many animals and plants that humans are dependent on for food and their livelihoods. However, development of anthelmintic drugs has not kept pace with the acquisition of drug-resistance by these nematodes, and so there is an urgent need for new therapeutics (see references in ^{131, 132}). Parasitic nematodes are difficult to work with directly due to the need to grow them in a host system, so C. elegans has become a model system for anthelmintic toxicology. The WormScan platform can simultaneously measure the mobility, brood size, body size and lifespan of worms either on agar plates or liquid culture^{115, 131}. This system was used to screen through 26,000 compounds resulting in the identification of 14 potential anthelmintic compounds. The INVAPP/Paragon assay also utilizes a 96-well plate format for culturing worms in liquid. However, the automated image capture relies on a high-frame-rate camera¹³² instead of a scanner, giving this system increased sensitivity in the detection of drug-induced motility defects. A proof-of-concept drug screen using this system with a 400-compound library identified 14 molecules that impaired worm growth¹³². A separate compound library screen against the parasitic nematode Trichuris muris using INVAPP/Paragon uncovered an entirely new class of promising anthelmintics¹³³. These deep phenotyping-based studies of C. elegans allow researchers to rapidly identify potential anthelminthic drugs and their mode of action without the need to perform the complex in-host assays required to study parasitic worms.

Small molecule screens are not only a tool to identify therapeutic chemicals, but they can also be used to study the biology of genetic pathways. The gene skn-1 is a transcription factor that regulates the worm’s response to oxidative and xenobiotic stress¹³⁴. Leung et al. performed a plate reader-based screen for small molecule activators of the SKN-1 protein by measuring the induction of a GFP-based transcriptional reporter of the skn-1 target gene gst-4¹³⁵. These authors then subsequently demonstrated the ability of their system to perform an ultra-high-throughput screen for inhibitors of SKN-1 amongst a compound library containing over 364,000 small molecules¹³⁶. This screen resulted in 125 that specifically lowered the fluorescent signal via inhibition of the gst-4::gfp reporter, suggesting that these molecules were SKN-1 inhibitors¹³⁶. A similar strategy has been used to study the male-specific linker cell, which undergoes cell death just after the molt between the fourth larval stage and adulthood¹³⁷. The death of this cell is non-apoptotic as it is independent of caspases; however, its death shares many morphological features with other non-apoptotic cell death observed in vertebrate development¹³⁷. Schwendeman and Shaham performed a proof-of-concept small molecule screen to identify potential inhibitors that may shed further light on the biology of this phenomenon¹³⁸. A screen of 23,797 compounds using a laser scanning cytometer resulted in the identification of six compounds that caused persistence of the linker cell by inducing some form of global developmental delay in the worms that was rescuable upon removal of the worms from the drug¹³⁸. Together these studies demonstrate how the use of deep phenotyping technologies could enable the quantitative measurement of multiple morphometric traits to gain new insight into biology.

3.5. Behavioral analyses of freely moving worms

One of the original motivations that drove the development of C. elegans as a model system was the desire to understand how the nervous system of an animal gives rise to all the behaviors it elicits¹⁶. A recent extensive review by McDiarmid et al.⁶¹ lays out the history and biological significance of behavioral studies of C. elegans. Here we discuss briefly recent improvements in worm tracking hardware and software and show how deep phenotyping studies could reveal mechanisms underlying emergent behaviors.

The rapidity of changes of behavior in freely moving worms makes it nearly impossible for a human observer to record all events in real time. Thus, the majority of behavioral studies employ some form of automated imaging. The earliest systems could record the spatial position, speed, and turning rate of individual worms¹³⁹. However, newer tracking platforms offer more comprehensive descriptions of worm behavior. Worm tracking systems fall into two categories: single-worm trackers for high-resolution analysis of individual behavior^140–143 and multi-worm systems for population-level studies^{65, 144–149} (Fig 3). These systems have elucidated the behavioral genetics of several sensory modalities. Explicit analyses of different behavioral features of thousands of animals from 239 genotypes revealed uncovered 87 genes, including components of the Gαq signaling pathway, involved in locomotion and predicted 370 specific genetic interactions among them¹⁵⁰. Similar studies have identified genes involved in thermotaxis¹⁵¹, chemotaxis^{144, 152} and mechanosensation¹⁴⁴. Most of these studies were performed on agar plates. However, the development of microfluidic arenas allows for precise spatiotemporal control of the chemical environment revealing behaviors are not observable in plate-based experiments²⁸.

Figure 3.

Behavioral deep phenotyping of worms using automated trackers. A. Behavioral repertoire of C. elegans that can be automatically tracked. B. Example of a single worm’s behavior during an experiment. C. A multi-worm tracking system developed for tracking behaviors in a group of worms. (A. and C. are adapted from Albrecht and Bargmann²⁸, B. is adapted from Yemini et al¹⁴¹)

In addition to the improvement in worm tracker hardware, development of new algorithms has led to a more comprehensive description of worm behaviors. For example, PCA has been used to decompose the postural space of worms into eigenvectors referred to as “eigenworms”¹⁵³. Surprisingly, the postural space of locomoting worms on agar plates is low dimensional, with superposition of just four of eigenworms sufficient to describe the majority of the worm’s locomotory postures. This analysis, which dramatically reduces the complexity of quantifying behavioral patterns, has been built into many worm tracking systems^{141, 144, 153, 154}. This approach has enabled deep phenotyping studies of behavioral dynamics, such as the creation of a dictionary of behavioral motifs curated from both wild-type and 307 mutant strains¹⁵⁴. Similarly, an extensive phenotypic database of locomotory behaviors for a large number of strains was compiled through the comprehensive behavioral recording of multiple alleles of the same gene as well as some double and triple mutants¹⁴¹. These dictionaries¹⁵⁴ and databases¹⁴¹ uncover subtle behavioral phenotypes for mutants that cannot be discerned by manual observation, underscoring the importance of deep phenotyping pipelines.

Understanding how behaviors emerge from the integration of information about the external environment and the worm’s own internal state remains limited. This problem has been studied extensively in food-related behaviors¹⁵⁵. The behavioral state transitions a worm undergoes, as well as the informational value of the food it encounters while foraging can be modeled mathmatically^156–158. These models can be used to predict the worm’s response to food encountered while foraging^{156, 158}. We anticipate that when such predictive models are integrated into deep phenotyping studies, it will lead to a greater mechanistic insight into the genetics of foraging behavior. In addition, increasing experimental throughput is essential for deep phenotyping foraging behavior. WorMotel^{30, 62}, allows for highly parallelized monitoring of individual worms under uniformly controlled environmental conditions This platform has been used to examine the relationship between the roaming (active) or dwelling (sedentary) behaviors and food abundance^{159, 160}. These studies reveal the biological complexity of foraging behaviors. For example, serotonin produced by the ADF neurons promotes roaming while serotonin produced by the NSM neurons promotes dwelling¹⁶⁰. Other biological amines such as dopamine promote dwelling¹⁵⁹, while octopamine is involved in roaming behaviors¹⁶⁰.

Similar experimental parallelization has been demonstrated on plate-based foraging assays through the use of multiple cheap cameras¹⁶¹. Stern et al. examined roaming and dwelling behavior in individual worms continuously across their development¹⁶¹. The pattern of exhibited behaviors varied across the different developmental stages as well as between the onset and exit phases of each larval stage in a reproducible way. A suite of neuromodulators is responsible for the regulation of these behavioral patterns¹⁶¹. Interestingly, by tracking individual worms throughout their development, the authors also found that there was significant inter-individual variation in roaming behavior. Even though the worms were derived from isogenic populations, some animals consistently roamed less across all developmental stages while others consistently roamed more¹⁶¹. Quantifying this type of stochastic variation in any phenotype and understanding it biological origins is challenging, given the large numbers of individuals that need to be surveyed. Deep phenotyping offers an integrated way of achieving high-throughput experimentation with comprehensive behavioral analysis.

3.6. ‘Whole-brain’ imaging: the next frontier in deep phenotyping

How the nervous system encodes external environmental information in order to modulate animal behavior is the subject of intense examination in biology. This question becomes even more challenging to answer when one considers that in natural habitats the environment contains a myriad of conflicting cues that an animal must successfully integrate in order to perform behaviors that maximize its chances at survival. In C. elegans, the nervous system is essentially divided into three layers: sensory neurons that detect external stimuli, interneurons that integrate information from sensory neurons, and motor neurons that control the behavior of the worm, with a few polymodal neurons that perform several of these activities¹⁶². Understanding how these different layers receive and process information and communicate with one another to ensure the appropriate outcome is a complex problem. For example, the aversive chemotactic response to isoamyl alcohol exposure is stochastic, even though the chemosensory neuron that detects the stimulant depolarises in a deterministic way¹⁶³. This behavioral stochasticity is generated at level of the interneurons in this circuit via their collective activity¹⁶³. The development of genetically encoded calcium indicators has allowed the imaging of neuronal activity, ranging from a single neuron all the way up to the entire ‘brain’ in the worm’s nervous system (for a summary, see Cho et al¹⁶⁴). This, in turn, has paved the way for the development of ‘whole-brain’ imaging platforms that allow long-term observation and analysis of a large number of neurons. The goal of these systems is to dissect neuronal activity and identify not only neural circuits associated with specific behaviors but also global states that reflect the functional architecture of the brain.

Several studies have successfully characterized whole-brain dynamics under various experimental conditions. Kato et al.¹⁶⁵ used a microfluidic imaging platform⁵³ to carry out whole-brain recordings from immobilized worms, which revealed that the evolution of network dynamics among neurons are directional and cyclical. Different phases in this cyclical activity regulate motor commands that drive certain locomotory behaviors. Using a spinning disk confocal system, Venkatachalam et al.¹⁶⁶ developed a whole-brain imaging platform and studied representations of sensory input and motor output of individual neurons upon thermosensory stimulus in freely moving worms. Similarly, using a simultaneous worm tracking and whole-brain imaging system, Nguyen et al.^{85, 167} recorded whole-brain activities of freely moving without stimulation; in a related effort Nguyen et al.⁸⁵ showed a machine-learning approach to track neurons in the freely moving heads, which is an important step towards robustly and automatically analyzing such large sets of dynamical data. Most recently, Nichols et al.¹⁶⁸ investigated global brain dynamics during lethargus, a sleep-like state in the worm. Global brain activity becomes quiescent during lethargus. However, specific neurons remain active as these cells promote the establishment of the quiescence. By examining whole-brain dynamics, this work demonstrated that the transition to the ‘wakeful’ state is carried out by the re-establishment of activity in specific neurons that then drive global brain dynamics back to the aroused state¹⁶⁸.

From the above examples, it is clear that there is much to be gained by studying the entire nervous system, rather than analyzing specific subsets of neurons, as not all of the neurons nor their functional connections in the circuits governing various sensorimotor behaviors have been identified. These examples also point to opportunities for future technological and theoretical developments for analyzing and understanding such complex and dynamical systems. For instance, better imaging systems that allow coupling of other experimental techniques, such as optogenetics, and better/faster tracking algorithms along with automatic identification of neurons are still needed. Further, better theories may be needed in the future for interpreting these large volumes of curated data to make sense of how the brain processes information and makes decisions.

4. Future outlook

In this review, we have highlighted many recent conceptual and methodological developments for deep phenotyping, specifically using C. elegans as a model system. From the success of these studies, it is clear that by measuring many aspects of the morphology, functional output, or behavior of cells, circuits, tissues, and individual animals, we can expand the scope of biological studies. Furthermore, by using appropriate mathematical and statistical tools, we can better understand their underlying biological mechanisms. We believe that these approaches will become more ubiquitous as improved microscopy and other experimental tools and analytical pipelines using advanced computational and theoretical techniques become more accessible.

While many of these tools discussed above were used in a C. elegans-specific context, their general utility is applicable to a broad range of biological systems. For example, microfluidics is widely used in the analysis of single cells, with platforms such as DropMap¹⁶⁹ providing high-content single cell-resolution analysis of IgG-secreting cells. There has also been significant progress in developing automated animal handling and tracking systems that can work with diverse sets of organisms. The MAPLE robotic system and its derivatives can perform deep phenotyping experiments on Saccharomyces cerevisiae, Physarum polycephalum, Bombus impatiens, and Drosophila melanogaster, in addition to C. elegans^{65, 170}. Several studies have also recently demonstrated deep phenotyping in the vertebrate model system Danio rerio. High-throughput optical projection tomography has also been used to gene expression changes across the brain of different mutants of zebrafish¹⁷¹. Similarly, micron-scale tomography¹⁷² has been used to study the determinants of skeletal development in zebrafish¹⁷³. We predict that with future integration of efforts in different disciplines (e.g., biology, engineering, and computational sciences), the ability to link phenotypes to genotypes, environmental conditions, and stochasticity will significantly accelerate.