A multi-dimensional, time-lapse, high content screening platform for the flatworm pathogen causing schistosomiasis

Approximately ten percent of the world’s population is at risk of schistosomiasis, a neglected, parasitic disease of poverty caused by the Schistosoma flatworm. To facilitate drug discovery for this complex organism, we developed an automated, time-lapsed highcontent (HC) screen to quantify the multi-dimensional responses of Schistosoma mansoni post-infective larvae (somules) to chemical insult. We describe an integrated platform to dispense and process worms at scale, collect time-lapsed, bright-field images, segment highly variable and touching worms, and then store, visualize, and interrogate complex and dynamic phenotypes. To demonstrate the method’s power, we treat somules with seven drugs that generate diverse responses and evaluate forty-five static and kinetic response descriptors as a function of concentration and time. For compound screening, we use the Mahalanobis distance (dM) to compare multidimensional phenotypic effects induced by a library of 1,323 approved drugs. We characterize both known antischistosomals as well as identify new bioactives. In addition to facilitating drug discovery, the multidimensional quantification provided by this platform will allow mapping of chemistry to phenotype.


Introduction
The Schistosoma blood fluke (helminth) causes schistosomiasis, a neglected tropical disease (NTD) [1][2][3] that infects over 200 million people and puts more than 700 million people at risk of infection in 78 countries [4][5][6]. Parasite eggs cause chronic inflammatory and fibrotic responses that impair visceral and/or urogenital organ function; co-morbidities include increased risks for bladder cancer and HIV [7,8]. Praziquantel (PZQ) is the only available drug for schistosomiasis.
Although reasonably active against mature schistosomes, PZQ displays little to no efficacy against developing parasites [9,10]. Also, increased utilization of PZQ raises concerns that drug resistance will emerge. Thus, new drugs are needed [11].
Anthelmintic drug discovery has traditionally relied on phenotypic screens using parasites in culture or in small animal models [12]. Primary screening of cultured schistosomes has often used post-infective larvae (called schistosomula or somules) that can be obtained in their thousands to tens of thousands from vector snails for relatively little effort and cost, in contrast to adult worms that can only be harvested in low numbers (hundreds) from small mammals. Single-metric assays, in which somules are scored as alive or dead, have been reported. (e.g., [13] for review). However, single-metric approaches suffer a number of drawbacks; in some cases, even the clinically used drugs do not score as active in these assays. High-content imaging of live somules offers the potential to visualize complex and non-lethal (but potentially therapeutically relevant) phenotypic responses to drug treatment.
As part of our research program to develop novel methods for anti-schistosomal drug discovery [14][15][16], we report a fully integrated, automated and multiparametric imageanalysis platform for high throughput phenotyping of living parasites. Starting with a set of seven drugs known to induce changes in shape and motion [14,17], we describe a series of protocols to quantify those changes as a function of time and concentration. We then demonstrate the utility of the method for high-throughput screening using a set of 1,323 approved drugs. Our approach offers key advances in method integration, including several of general utility to the drug screening/imaging community: a) automated liquid handling of 100 µm-sized organisms, b) manipulation of the focal plane to facilitate identification of low-contrast, variable and touching objects, c) time-lapsed tracking to define frequencies and rates of motion, d) a public system for storage, visualization and querying of the complex phenotypic data, and e) use of a statistical metric (Mahalanobis distance, dM) to compare multidimensional phenotypes for high-throughput screening.
Sunitinib (S-8803 as the Malate Salt) and staurosporine (S-9300 as the Free Base) were purchased from LC laboratories. The library of approved drugs included 1,129 compounds from the Microsource Drug Collection and an additional 194 compounds donated by Iconix or purchased from commercial vendors. The set included drugs approved by the FDA (85%) and by the analogous European and Japanese agencies (15%).

High-throughput sample handling for S. mansoni somules
We selected somules for primary assays as we can obtain 10 4 -10 5 somules/week from freshwater vector snails. As somules (~300 x 150 µm) rapidly settle out of solution, we used a magnetic tumble stirrer containing eight stirring paddles to maintain worms in suspension ( Fig. 1) and allow their transfer using 96-or 384-channel pipets. Plate geometry and number-of-somules/well were optimized for accurate imaging and counting ( Supplementary Fig. 1a). In particular, U-bottom 96w plates concentrated the parasite as a monolayer into one central field of view, thus facilitating automated imaging. Routine robotic protocols then dispensed compound and shuttled plates between an incubator (Cytomat 2C) and a high-content imager (GE IN Cell Analyzer 2000) for data collection.

Imaging schistosomes by automated, bright-field image analysis
Bright-field, time-lapsed images were generated for control and drug-treated somules using a 10x objective. Every 24 h for 3 days, images were collected at 1.66 Hz (the maximum frame rate for the IN Cell Analyzer 2000) to generate 20 s video recordings. We employed bright-field imaging as it is mechanism-agnostic, non-invasive and fast, and because the schistosome is not yet amenable to the transgenic incorporation of fluorescent proteins. However, in bright-field, somules do not present a high-contrast edge relative to background, thus limiting object segmentation (detection of the object's outline). We, therefore, lowered the focal plane 40 µm below the bottom of the well in order to artificially generate a dark edge that facilitated segmentation without a significant loss of interior density features (texture; Supplementary Fig. 1b).
We observed two strong distributions of somules during our studies. Initially, parasites had a translucent body with a discernable outline. However, under the influence of toxic compounds, worms can become progressively opaque, such that the worm outline is indistinguishable from the interior of the worm. This opacity is associated with degenerating/dying parasites with the transition from "clear" to "opaque" being irreversible.
To accurately identify both classes of worms, we segmented the somules using three customized protocols, optimized to detect somules independently by considering the entire worm body, only the worm outline or the worm interior (Fig. 2). The largest segmented area obtained from these protocols was selected as the true somule. Each somule was next described using 15 (Supplementary Fig. 1). In-depth protocols are provided in the Detailed Segmentation Methods.
Once worms were classified as clear or degenerate, the 15 calculated features were evaluated across three modesstatic, rate and frequency. In the static mode, we considered feature measurements in each frame. In the rate mode, we determined the mean change in a feature between consecutive frames, e.g., rate-of-change in length. In the frequency mode, we analyzed the number of times the value of a feature increased or decreased, e.g., the number of length contractions/second. As somules showed little translational movement in the u-bottom wells, we did not record their displacement. When combined, the static, rate and frequency modes for each of the 15 features yielded 45 measurements for each somule (see next section).
Two statistical approaches were used to evaluate the significance of changes in worm phenotypes in response to drug treatment. First, the mean and standard deviation for each descriptor for all somules within a well were computed and normalized to determine the Effect Size (ES): where x is the mean descriptor due to drug exposure, μ is the DMSO mean and σ is the standard deviation of the descriptor for parasites in DMSO [20,21]. Being dimensionless, ES can be useful in comparing effects across different features where an ES >2 (p <0.05) is generally considered significant. In addition to evaluating individual descriptors, we compared parasites in this descriptor space using the Mahalanobis distance (dM) [22]. dM measures the multi-dimensional, scale invariant distance between a test well and a standard condition, e.g., DMSO-treated somules, by taking the result of quadratically multiplying the mean difference and the inverse of the pooled covariance matrix. dM and ES are similar in that they both measure a distance from the DMSO reference; however, ES measures the distance for just one feature, whereas dM measures the distance for a group of variables. dM is not dependent on the measurement unit and can identify test wells that have one large difference or multiple small differences compared to DMSOtreated controls.

SchistoView: query-visualization of phenotypic screening data
We developed SchistoView (Fig. 3, http://haddock9.sfsu.edu/schistoview/home) which comprises a graphical user interface supported by a MySQL database. SchistoView allows users to visualize and query concentration-and time-dependent somule response data, from computed statistics for a given well to features for individual somules. Figure 3 shows a screenshot of Schistoview and describes the features of the graphical user interface.

Exploring the parasite's multivariate responses using known anti-schistosomal compounds
We tested the time-lapsed imaging platform with seven compounds that induce diverse changes in the parasite [14,17]. Somules were exposed to an 11-point, 2.5-fold dilution series (from 2 nM to 20 M) of compounds in quadruplicate and images were collected after 2, 24 and 48 h. Raw images collected after 24 h of the first frame in each well are shown in Supplementary Fig. 2. Images were segmented and data extracted as described above. The results highlight important features of the imaging methodology, the depth of analysis offered and the underlying biology of the schistosome parasite.
The time-and concentration-dependent effects of drugs on worm behavior are visualized using heat maps extracted from SchistoView ( Fig. 4; Supplementary Fig. 3). The absence of cidal activity for these two drugs is consistent with their primary activity as paralytics (see below). Thus, the ability to capture phenotypic changes by dM [23] (i) affords a rapid and deep assessment of anti-schistosomal activity that is independent of degeneracy and (ii) adds essential value by identifying highly relevant anti-schistosomals that do not induce degeneracy, including the two clinically used drugs, PZQ and metrifonate. Finally, combining static and dynamic descriptors into the dM provides a more sensitive readout of phenotypic change than either modality alone (Supplementary Fig.   3).
The imaging platform quantifies drug-induced increases and decreases in parasite motility. Previously, L-imipramine was visually assessed to induce hypermotility [14]. We confirm this finding and, for the first time, quantify the response. As shown in Fig. 5a, imipramine induces a concentration-dependent increase in movement after 2 h between 10 nM and 1 µM (EC50 = 100 nM) followed by decreased motility at higher concentrations.  [24].
Despite the centrality of PZQ for the treatment of schistosomiasis, its mechanism of action is not well defined and is likely complex, leading to Ca ++ influx and tetanic paralysis [9]. This complexity is reflected in our phenotypic analysis (Fig. 5c) The time-dependence of the shortening effect by PZQ is also different from that of the shivering phenotype. Although both are observed at 2 h (see above), the shortening effect disappears by the 48 h time point (Supplementary Fig. 4f), whereas shivering remains unchanged (Fig. 3f). The rapid and small changes in length during shivering are most clearly quantified by frequency ( Fig. 3b; magenta square at the intersection of length and frequency). These data highlight the ephemeral or transient nature of some phenotypic responses, a concept that has not yet been considered in anthelmintic screening. Overall, the imaging methodology, as interrogated through SchistoView, allows for the orthogonal identification and quantification of individual concentration-and time-dependent changes.
For the other four members of the seven-drug test set, phenotypic effects at 2 h preceed degeneracy seen at 48 h (Fig. 4). For example, two known anti-schistosomal agents, simvastatin [17] and K11777 [18], induce gradual increases in degeneracy (simvastatin EC50 = 1 µM at 24 h; K11777 EC50 = 20 μM at 48 h). Degeneracy caused by staurosporine is apparent by 24 h and extends across the entire concentration range (Fig.   4), consistent with this inhibitor's high affinity for multiple kinases. Other changes include increased area (65% larger than clear worms in DMSO, 13% larger than degenerate worms in DMSO) and increased median diameter (51% larger than clear worms in DMSO, 20% larger than degenerate worms in DMSO; data not shown). Interestingly, within 2 h, sunitinib produces a gray to jet-black phenotype that is much darker than control (DMSOtreated) degenerate worms. The effective concentration for this color change (EC50 = 2.94 μM) follows a different concentration-response curve than changes in size, which are elevated at all concentrations. As with PZQ, the complexity of these changes may reflect the time-and concentration-dependent engagement of different targets [25].

Using multi-dimensional features for primary screening
In addition to phenotyping based on inspection of individual descriptors, the experimental platform and SchistoView software are applicable to high-throughput screens using dM.
We prepared 20 x 96-well plates with 40 somules/well that were incubated with DMSO (0.1%) or 10 µM compound from an in-house collection of 1,323 drugs approved for human use. Using the same sample preparation and imaging conditions as for the sevendrug set, plates were robotically handled without manual intervention. Screening proceeded at a maximum rate of one plate/37 min for four scan cycles. Images were automatically processed and analyzed, and data entered into the MySQL database; the screen generated 59,867,820 results for 553,492 segmented worms. The dM values (calculated from combined static, rate and frequency data), were then extracted and plotted in Fig. 6a.
Because dM does not have an inherent maximum value, the typical screening metric Z' cannot be calculated. However, 'hits' are usually picked based on standard deviations from the mean of the negative controls. For the data in Fig. 6a (combining static, rate and frequency), a dM value of 2.47 represents 3 SD from the mean of the DMSO-treated controls. Using this cutoff value, a total of 237, 263, 326 and 309 hits were identified for cycles 1-4, respectively. Notably, three drugs (PZQ, simvastatin and sunitinib) from the seven-drug test were also present in this screening set, and were identified as active, with dM value greater than 2.47.
The individual contributions of the static and frequency measurements to the identification of hits were illustrated by calculating a dM for each measurement in scan cycle 1 (Fig. 6b) Fig. 5). Thus, 67% of active compounds differed from DMSO-treated worms based on changes in rate but were not significantly different from DMSO based on static features. Hence, quantifying changes in motion was critical to identify the majority of the active compounds in the drug set.
Phenotypes among the 1,323 drug set were remarkably varied. For example, somules exposed to the hormone analog altrenogest displayed static features similar to those of DMSO controls (Fig. 6b), but exhibited a 144% increase in frequency of movement by length and a 147% increase in frequency of movement by area. Hence, altrenogest was 'active' based on the dM for frequency. By contrast, the antifungal, piroctone, only yielded changes in static features with a decreased variation in internal texture (63%) and an increase in form factor (110%; yielding a more rounded phenotype) relative to DMSO controls. Somules treated with the adrenergic agonist xylometazoline were altered in both static and dynamic descriptors, i.e., a greater mean length of 211 µm vs. 130 µm for controls and a 160% increase in frequency of movement by length, respectively.
The data obtained for the 1,323 drug set agree well with results from another drug screen (1,600 drugs tested at 10 µM from the FDA Pharmakon and Microsource Discovery collections) that employed an observation-based scoring system (Supplementary Table   2) [26]. Of the 236 compounds for which data were declared in 25 , 144 were also screened by us, and for these there was a 67% concordance in the number of actives (dM(static-rate- Somules are difficult to image due to their (variability in) movement and because they have a low contrast in bright field. We solved the image-collection challenge using round-bottom wells to constrain worms into one visual field and thus limit translation. We then addressed their low contrast by focusing slightly below the worm to enhance its outline. From there, we observed two basic classes of wormsclear and opaqueand optimized segmentation protocols for each. The resulting segmentation accuracy (precision of 90% and a recall of 95%, Supplementary Fig 1a) is an improvement on a previous report that employed bright field analysis (24.5±7%) where touching somules could not be evaluated [27]. Also, our imaging approach economizes on the number of parasites needed by 3-4 fold and measures how worms move rather than a simple classification of whether movement has occurred [27]. Finally, our methodology provides a solution to the critical issue of segmentation of touching objects in the analysis of bright field images generally [28][29][30].
The live imaging platform described here can be incorporated into a drug discovery pipeline upstream of ex vivo phenotypic screens of the adult schistosomes and rodent models of infection [14]. Recent advances in the image-based quantification of adult schistosome motility [31,32] have demonstrated the ability to quantify motion and could mesh seamlessly with the workflow described here. The platform will complement other advances relating to schistosome biology, including gene expression profiling [33], metabolomics [34], and CRISPR/Cas9 [35], that together will improve our ability to holistically quantify this globally important parasite's response to a range of drug-induced, environmental and developmental phenotypes. To facilitate such discoveries, the database and SchistoView interface are available online. Compounds were added using a 96-channel pin tool. Plates were maintained at 5% CO2 and 37 °C. At specified time points, a 6-axis robotic arm transferred the plates to the highcontent imager.

Area
Area of a target μm^2

Median Diameter
The median internal distance perpendicular to the maximum curved chord. μm Length Maximum distance across a target. Boundaries may be crossed. μm

Form Factor
Estimate of circularity, expressed as a value between 0 and 1 (1 equals a perfect circle). value Perimeter Distance around a target. μm

Straight Chord
The maximum straight-line distance across a target without crossing a boundary. μm

Curved Chord
Maximum center line through target. μm

Bend
Bend = (max curved chord / max straight chord). ratio Pinch Pinch = (median diameter * length) / area) is the estimated area divided by the actual area. ratio Wave Wave = Perimeter / (2 * Area^0 .5 ) is the actual perimeter divided by the estimated perimeter. ratio

Mass
The sum of all pixel values in the shape. sum WMOI (Weighted Moment of Inertia) Index of the homogeneity of gray levels within a circular target. A value of 1 indicates the target is relatively homogeneous. If >1, the target has a higher proportion of bright pixels in its center. If <1, the target has a higher proportion of bright pixels around its perimeter. index

Density -Levels
Mean gray level value of the pixels contained within the target outline. Gray levels is an intensity scale, where black = 0 and white = 4095 (12-bit image).

SD -Levels
A standard deviation (SD) of pixel densities, which measures the pixel density variation within the target. SD values are available for any density unit. degrees

Rate
The average amount of changes between sequential time-lapse images. "units"/frame Frequency The number of times a feature changes direction. Hz Effect Size (ES) ES = (x-μ)/σ where x is the drug mean, μ is the DMSO mean, and σ is the standard deviation of DMSO. No unit, but the magnitude of the measurement can be thought of as the number of standard deviations from the DMSO control group for the selected feature.

Mahalanobis Distance (dM)
dM is a multi-dimensional effect size which measures the distance of a test point from a reference distribution. No unit, but the magnitude of the measurement can be thought of as the number of standard deviations from the DMSO control group. Figure 1. Optimizing parasite handling and segmentation. (a) Comparison of actual worms (counted by visual inspection) and the number of worm objects identified by the object classifier algorithm 'computer count'. Computational inclusion of non-worm objects with worm-like features leads to a systematic 10% increase in object count. (b) Images from a single sample well imaged at three focal planes (0, -40, and -80 µm from the outside bottom of the well). Lowering the focal plane improves the contrast of the somule outline ('edge'). Minus 40 microns improves the appearance of the outline while preserving some of the internal texture detail.