A High Throughput Cellular Screening Assay for Small Molecule Inhibitors and Activators of Cytoplasmic Dynein-1-Based Cargo Transport

Abstract

Cytoplasmic dynein-1 (hereafter dynein) is a six-subunit motor complex that transports a variety of cellular components and pathogens along microtubules. Dynein’s cellular functions are only partially understood, and potent and specific small molecule inhibitors and activators of this motor would be valuable for addressing this issue. It has also been hypothesised that an inhibitor of dynein-based transport could be used in anti-viral or anti-mitotic therapy, whereas an activator could alleviate age-related neurodegenerative diseases by enhancing microtubule-based transport in axons. Here, we present the first high-throughput screening (HTS) assay capable of identifying both activators and inhibitors of dynein-based transport. This project is also the first collaborative screening report from the Medical Research Council and AstraZeneca agreement to form the UK Centre for Lead Discovery. A cellular imaging assay was used, involving chemically-controlled recruitment of activated dynein complexes to peroxisomes^1,2. Such a system has the potential to identify molecules that affect multiple aspects of dynein biology in vivo. Following optimisation of key parameters, the assay was developed in a 384-well format with semi-automated liquid handling and image acquisition. Testing of over 500,000 compounds identified both inhibitors and activators of dynein-based transport in multiple chemical series. Additional analysis indicated that many of the identified compounds do not affect the integrity of the microtubule cytoskeleton and are therefore candidates to directly target the transport machinery.

Introduction

Cytoplasmic dynein-1 (dynein) is a 1.4 MDa motor complex that plays an essential role in cellular organisation. Dynein moves towards the minus ends of microtubules in association with a large variety of cellular constituents—including endosomes, lysosomes, protein aggregates, mitochondria and mRNAs—and thereby mediates their polarised trafficking³. Neurons are particularly dependent on efficient dynein-based cargo transport due to their long processes, as illustrated by the association of mutations in the dynein transport machinery with neurodevelopmental and neurodegenerative diseases⁴. Pathogenic viruses such as HIV and herpes virus have also evolved the ability to engage the motor complex for translocation through the cytoplasm^{3, 5, 6}. In addition, dynein organises microtubule arrays during interphase and mitosis, for example by generating pulling forces on microtubule ends at the cell cortex^7–9.

Dynein comprises six subunits, each present in two copies per complex³. Microtubule binding is mediated by the heavy chain, which also contains ATPase activity. Repeated conformation changes in this subunit, which are coupled to cycles of ATP hydrolysis, drive the motor complex towards the microtubule minus end. In mammalian cells, the motility of the dynein complex is activated by association with another multi-subunit complex, dynactin, and one of a number of specialised coiled-coil proteins bound to cargoes (termed ‘activating adaptors’). Dynein, dynactin and the activating adaptor form a mutually dependent triple complex that can move in vitro over a distance of many microns^{3, 10–12}.

The activities of dynein in cells are only partially understood, and many academic groups are working to address this issue. Because of dynein’s numerous roles, a key challenge is discriminating between direct and indirect effects when the function of the motor complex is manipulated using genetic tools. For this reason, the ability to acutely inhibit or activate dynein-based transport using small molecules is highly desirable. Dynein inhibitors may also be useful in anti-viral or anti-mitotic therapies, whereas activators could be used in animal models to explore the hypothesis that enhanced microtubule-based transport in axons can alleviate age-related neurodegenerative diseases^13–18. Small molecules that inhibit dynein have previously been reported but questions remain about their potency, selectivity or mode of action^19–26. To our knowledge, no molecules have been described that are specific activators of dynein-based transport.

Here, we present the first cellular high throughput screening (HTS) assay and associated analytical tools for the identification of candidate small molecule modulators of a model dynein-dynactin-activating adaptor complex. Such a system opens the possibility of finding molecules that affect specific aspects of dynein biology in vivo, including regulation by dynactin and activating adaptors, as well as by signal transduction pathways. This chemical equity would allow a more nuanced approach to manipulating dynein function in academic and applied settings.

The opportunity to develop such an assay has arisen from the innovative agreement in place between the Medical Research Council (MRC) and AstraZeneca, whereby academic groups in the UK can apply for funding from the MRC and gain access to the infrastructure within the UK Centre for Lead Discovery (UKC4LD). This has created a unique combination of biological knowledge, financial support and screening knowhow to progress novel targets in a collaborative setting^27–29.

The HTS assay is based on a previously described strategy in which active dynein motors are inducibly recruited to fluorescent peroxisomes in order to elicit their relocalisation within the cell¹ (Fig. 1a). The N-terminal region of the activating adaptor BicD2 (BicD2N)³⁰ is recruited in a rapamycin-dependent manner to the surface of peroxisomes using the FRB-FKBP heterodimerisation system³¹. This leads to the association of these organelles with dynein and dynactin complexes and their translocation towards microtubule minus ends, which are concentrated at the microtubule organiser centre (MTOC) in the perinuclear region^1,2. The ability of the peroxisome relocalisation assay to report on dynein-mediated transport has been further corroborated by live imaging of cells deficient for the dynein activator Lissencephaly-1³².

Figure 1. Dynein-Dynactin-BicD2N-Peroxisome Trafficking Assay.

(a) Cartoon showing principle of the assay. Minus end-directed movement of peroxisomes along microtubules is stimulated by adding rapamycin to cells that express both GFP-BicD2N-FRB and a peroxisome targeting sequence (PTS) fused to RFP and FKBP (PTS-RFP-FKBP). Rapamycin promotes the association of peroxisomes with activated dynein-dynactin-BicD2N complexes via the FRB-FKBP heterodimerisation system. The assay uses fluorescence microscopy to monitor the translocation of GFP-tagged BicDN and/or RFP-tagged peroxisomes to the perinuclear microtubule organising centre (MTOC), where minus ends of microtubules are nucleated. Diagram not to scale. (b, c) Representative high magnification confocal images (40x oil objective, 1.3 NA) of GFP-BicD2N-FRB/PTS-RFP-FKBP U-2 OS cells illustrating rapamycin-mediated (b) and microtubule-dependent (i.e. disrupted by nocodazole) (c) relocalisation of RFP and GFP. Examples of clustering of the fluorescent proteins in a discrete spot in the perinuclear region are shown with cyan arrowheads. As expected, these clusters are located at the MTOC (Supplemental Fig. 2). Consistent with previous observations², GFP-BicD2N-FRB does not accumulate at the MTOC in the absence of rapamycin (b). This finding suggests that active transport in vivo is dependent on the concerted action of multiple motors on the peroxisome surface. (d) Quantitative analysis (mean intensity of GFP spots per cell) of an independent experimental series confirms rapamycin (Rap.)-mediated and nocodazole (Nocod.)-sensitive relocalisation. Mean GFP spot intensity values are normalised to the median value for rapamycin only. Number of cells analysed for each condition are shown in italics. Boxes show interquartile range (25^th-75^th percentile of values) and horizontal line is the median. Vertical lines illustrate 1.5x the interquartile range with outliers shown with circles. Statistical significance (compared to the rapamycin only sample) was evaluated using a pairwise t-test with p-values adjusted for multiple comparisons using the false discovery rate correction (****, p <0.0001). Note that nocodazole plus rapamycin results in higher mean GFP spot intensity per cell than the condition in which rapamycin and nocodazole are absent (DMSO only) (p<0.0001) because rapamycin induces concentration of GFP on peroxisomes that, in the presence of nocodazole, are dispersed in the cytoplasm (c). In (b), cells were fixed 150 min after treatment with vehicle (DMSO) or 2 nM rapamycin. In (c and d), cells were treated with 10 μM nocodazole or vehicle (DMSO) for 180 min before fixation, with 2 nM rapamycin also present for the last 150 min.

We reasoned that because kinetics of peroxisome relocalisation can be tuned by varying the assay endpoint and concentration of the chemical inducer of FRB and FKBP association¹, it may be possible to design an assay that identifies agonists and antagonists of dynein-based transport in a single screen. Previously, the peroxisome relocalisation assay was carried out manually using high magnification objectives in order to image a small number of cells over time in great detail^1,2. This approach was not practical for the compound numbers necessary to screen the chemical diversity in the AstraZeneca compound collection. A 384-well plate-based assay format, with a low magnification objective and semi-automated liquid handling and microscopy, together with optimised image analysis procedures, allowed a suitably high throughput to be achieved. Testing of over 500,000 compounds identified multiple inhibitor and activator series that can be analysed in the future in assays that deconvolute their specific modes of action. We demonstrate this point by categorising hits that affect the integrity of microtubules and are therefore candidates to modulate dynein-based transport indirectly, and by identifying undesirable activator hits that function in the absence of rapamycin or are inherently fluorescent.

Materials and Methods

High-throughput screening

505,500 compounds from the AstraZeneca (AZ) collection were screened using the high-throughput inducible peroxisome relocalisation assay. This involved a combination of liquid-handling work-cells and manual manipulation, with a three-channel fluorescence endpoint read on an automated dual CX5 CellInsight setup: a Cart-to-Cart system involving two CellInsight CX5 High Content Screening Systems (Thermo Fisher Scientific) fed from a Steristore incubator via a CoLAB Flex A-Cell robot (HighRes Biosolutions) (Supplemental Fig. 1).

A U-2 OS human bone osteosarcoma cell line stably expressing GFP-BicD2N-FRB and peroxisome targeting sequence (PTS)-RFP-FKBP (see Results) was grown in continuous culture in selective antibiotics prior to use for screening. PTS-RFP-FKBP, which was under the control of the cytomegalovirus promoter, was selected with Geneticin, whereas GFP-BicD2N-FRB, which was under the control of the chick β-actin promoter, was selected with Hygromycin. Cells were grown in bulk in multi-layer tissue culture T175 flasks and plated manually into black 384-well cell culture plates (Greiner Bio-One #781090) using a Multidrop Combi (Thermo Fisher Scientific). Cell culture conditions are provided in Supplemental Table 1. After 24 h of incubation at 37°C/5% v/v CO₂, cells were serum-starved by removal of media (Biotek EL406) and replacement with fresh serum-free media (Thermo Fisher Scientific Multidrop Combi). 24 h later, compounds diluted in serum-free media were transferred from 384-well stock plates (Labcyte) to cell plates using an automated BioCel system comprising a number of V-Stacks, a V-Code and a V-Prep (Agilent). Compound-dosed plates were pre-incubated for 30 min at 37°C/5% v/v CO₂ to allow slow binding compounds to engage the target before rapamycin (in serum-free media) was added manually (to a final concentration of 2 nM) to the compound-containing medium using a Multidrop Combi. After incubation for a further 2.5 h at 37°C/5% v/v CO₂, plates were manually fixed and stained for 30 min with formaldehyde solution (In-house supply; Multidrop Combi) containing the membrane permeable DNA dye Hoechst 33342 (Sigma). Fixative was removed, and plates washed three times in PBS using a Biotek EL460 with automated Biostack before being sealed with black plate seals and stored at 4°C until required. Plates were equilibrated to room temperature before reading.

Plates were read on an automated system comprising two CellInsight CX5 machines (Thermo Fisher Scientific), using a three-channel image acquisition (for GFP, RFP and Hoescht signals) and an analysis method described below. Each 384-well assay plate included 16 neutral controls (DMSO only) and 16 inhibitor controls (10 μM nocodazole (a known microtubule targeting agent) in DMSO). The median and robust standard deviation of these controls was used to generate robust Z’ (RZ’) statistics to determine plate variability. To calculate the percent inhibition of test compounds in the inhibitor assay, a normalisation using the compound wells as neutral controls and nocodazole-containing wells as inhibitor controls was used. This approach assumed that, due to the large number of inactive compounds being tested, the compound wells would be a more robust set of neutral controls than the limited number (16) of DMSO only wells. For the activator assay, a single point normalisation based around the compound controls only was used due to the lack of an activating control. The primary screen was carried out at a compound concentration of 10 μM, with a robust z-score determination (see below) for each compound used to select primary screening hits with robust z-score cut-offs of -5 (inhibitors) and +5 (activators).

Genedata Screener software (Genedata Version 14.0.7-Standard) was used for data analyses following CellInsight image acquisition. Primary hits were plotted in a 10-point XC₅₀ (where X equals either I (Inhibitory) or E (Effective) concentration) dose response from 1 nM to 30 μM in half-log intervals. A near-neighbour compound set (i.e. with related chemical structures) was generated from XC₅₀ screening hits and tested in dose response format to generate structure activity relationship (SAR) data. Data were fitted to a four-parameter logistic model to obtain XC₅₀, hill slope, minimum inhibition (S₀), and maximum inhibition (S_inf). Curves were categorised based on statistical measures to highlight those with the best fit. In cases where fitting failed, the S_inf and/or S₀ parameters were fixed, resulting in a three- or two-parameter curve fit.

All XC₅₀ plates were restained using a rat anti-α-tubulin primary antibody (Abcam Ab6160, diluted 1:2000) and an Alexa Fluor 680-conjugated goat anti-rat IgG secondary antibody (Invitrogen A21096, diluted 1:1000) and read on the CellInsight CX5 systems (Thermo Fisher Scientific) in the far-red channel. Images were analysed using Columbus software (Perkin Elmer) to assess the degree of microtubule disruption caused by the tested compounds. The user-developed Columbus algorithm calculated a number of features for each identified cell, which were based around intensity and textural appearance (ridges, spots, etc) and the distribution of these textures within the cell (rotational symmetry, location from centre to edge of cell). These features were used to train a linear classifier to identify cells as ‘Normal’ or ‘Nocodazole Disrupted’ using manually selected training examples in the presence and absence of nocodazole within Columbus. From this linear classifier we used the regression score between ‘Normal’ and ‘Nocodazole Disrupted’, averaged over a well, as a continuous value quantifying the degree of microtubule disruption, which could then be used to generate XC₅₀ curves.

Data capture using the CellInsight High Content Screening platform

HTS data was acquired using two CellInsight CX5 (Thermo Fisher Scientific) HCS imagers on an automated HighRes CoLAB Flex Cart-to-Cart system. Image acquisition using an Olympus 10x UPlanFL air objective (0.3 NA) was optimised on the CX5 to allow increased throughput by shorter imaging time without compromising data quality. Plates were read in three channels – Blue for cell nuclei staining (Hoescht), Green for BicD2N fluorescence (GFP) and Red for peroxisome fluorescence (RFP) – using six fields of view per well. Image exposure was set at approximately 25% of saturation to prevent over and under exposure. Exposure times for each channel were optimised based on maximum (DMSO) and minimum (10 μM nocodazole in DMSO) control wells across a number of plates on the day of reading. Cells that had lost expression of the PTS-RFP-FKBP transgene were gated out by setting a minimum threshold of RFP fluorescence. GFP expression in the form of average spot area per cell per well in the perinuclear region was calculated for valid objects to generate a feature value (MEAN_CircSpotAvgAreaCh2) for analysis in Genedata Screener (Version 14.0.7-Standard).

HTS data analysis

Compound screening data were analysed through two separate Genedata Screener (Version 14.0.7-Standard) sessions to allow the detection of both activators and inhibitors of dynein-based transport. A robust z-score normalisation was applied using a single control-well group as a reference distribution. Robust z-score was calculated for each well using a formula where the standard deviation and mean are replaced by the robust standard deviation and median, respectively. For a single plate the median and robust standard deviation of the raw values were obtained from the Central Reference well type. As described above, the compound wells on the microtiter plate were used as the central reference with the assumption that only a very small proportion of these would be active and so effectively acted as neutral controls. For each well, the median was subtracted from its raw value and this was divided by the robust standard deviation to obtain the robust z-score:

$$Robustz-score=\frac{x-m}{RSD}$$

x is the raw data value of the well to be standardised

m is the median of the chosen control well group

RSD is the robust standard deviation for the chosen control well group

Robust standard deviation is defined as the median absolute deviation * 1.483

The single-plate robust z-score normalisation was used to remove an offset and a multiplicative distortion per plate.

As described above, robust z-score cut-offs of -5 (inhibitors) and +5 (activators) were used to select compounds for further testing. The generation of a robust Z’ factor was used as a measure of plate quality. Compound wells were again assumed to act as neutral controls and so used in the normalisations for inhibitors (Compounds Minus Inhibitors) or activators (Compounds). Data were transferred to Genedata Screener by a connected link with the AstraZeneca CellInsight Oracle database and mapped to compounds directly via an internal AstraZeneca compound database.

High-resolution confocal imaging

The primary data in Fig. 1, Fig. 5, Supplemental Fig. 6 and Supplemental Fig. 7 were obtained with a Zeiss 780 confocal microscope using a 40 x PlanApo Oil objective (1.3 NA). In Fig. 1c and Supplemental Fig. 6, microtubules in PTS-RFP-FKBP/GFP-BicD2N-FRB U-2 OS cells were visualised with mouse anti-α-tubulin antibody DM1A (Abcam Ab7291, diluted 1:500) and an Alexa Fluor 647-conjugated goat anti-mouse IgG secondary antibody (Invitrogen A21235, diluted 1:500). In Fig. 1, Fig. 5and Supplemental Fig. 7, RFP and GFP signals in PTS-RFP-FKBP/GFP-BicD2N-FRB U-2 OS cells were imaged (see legends for details of rapamycin and compound treatments; note that when compounds were absent, DMSO was used as a vehicle control). In Fig. 1 and Fig. 5, GFP signals were amplified using anti-GFP antibodies (Sigma 11814460001, diluted 1:500) and an Alexa Fluor 488-conjugated donkey anti-mouse IgG secondary antibody (Invitrogen A21202, diluted 1:500). In Supplemental Fig. 7, both RFP and GFP signals were detected directly, i.e. without antibody-based signal amplification.

Quantification of fluorescent signals in high magnification confocal images

For the quantification in Fig. 1d, 5b and Supplementary Fig. 7b, 10 to 16 images (212 μm x 212 μm each) were analysed per condition using a custom ImageJ script. For each image, the channel corresponding to the DAPI label was pre-processed using a 3x3 median filter and a rolling ball background subtraction. Segmentation of nuclei was achieved by setting a threshold level with a probability of false alarm of 0.01 using the median and median of absolute deviation as a robust estimate of the mean and variance of the image. A set of signed distance maps was computed for each segmented object and used to define cellular regions associated with each nucleus. The green channel was pre-processed using a Laplacian of Gaussian and the threshold set using the same strategy as described above, thus provided a segmentation for the GFP-BicDN-FRB signal. For each region associated with a nucleus contained entirely in the field of view, the total intensity of each segmented green object was computed and a mean value calculated per cell. To account for different methods of GFP signal detection between experimental series, values within one experimental series were normalised to the median value for the category with the largest median value in that series. Statistical analysis of the results was performed in R.

Results

Cell line generation

A series of cell lines that use the rapamycin-inducible FKBP/FRB heterodimerisation system to control the recruitment of activated dynein complexes to peroxisomes were created. This work was performed by SAL Scientific under an agreement with Aurelia Bioscience Ltd. The composition of expression plasmids was based on the work carried out previously by Kapitein et al.¹, with sequence/plasmid manufacture for transfection further outsourced to Oxford Genetics. Multiple U-2 OS human osteosarcoma cell lines stably expressing FKBP linked to the human PEX3 peroxisome targeting sequence (PTS) and RFP (PTS-RFP-FKBP), as well as FRB linked to mouse BicD2N (residues 1-400) and GFP (GFP-BicD2N-FRB) were established. The U-2 OS cell line was chosen for its low risk (class 1), good transfection efficiency, large cytoplasm and its previous use in dynein studies³³.

Cell lines containing both constructs were assessed for expression levels and localisation of fluorescent proteins in the presence and absence of rapamycin (Supplemental Table 2). A cell line (U-2 OS SP2 CL1) was selected based on strong RFP signal and a clear redistribution of RFP-labelled peroxisomes following the introduction of rapamycin, from dispersed in many small spots in the cytoplasm (often enriched in the general perinuclear region) to concentrated in a single, large spot at the MTOC adjacent to the nucleus (Fig. 1b; Supplemental Fig. 2). GFP-BicD2N-FRB also showed a striking relocalisation in this cell line upon addition of rapamycin, from a diffuse pattern throughout the cytoplasm to concentrated with the relocalised peroxisomes at the MTOC (Fig. 1b; Supplemental Fig. 2). As expected, concentration of peroxisomes and GFP-BicD2N-FRB signal at the MTOC in the presence of rapamycin was abolished by the microtubule depolymerising compound nocodazole³⁴ (Fig. 1c). The robust activities of rapamycin and nocodazole were confirmed by automatically segmenting GFP objects in each cell and calculating the mean of their total intensities (Fig. 1d). Additional controls with U-2 OS lines expressing only PTS-RFP-FKBP confirmed that rapamycin-induced relocalisation to the MTOC was dependent on the presence of GFP-BicD2N-FRB (Supplemental Table 2). The U-2 OS SP2 CL1 line was subsequently bulked up for use by Aurelia Bioscience Ltd and provided to AstraZeneca as a stock of cryopreserved vials.

HTS assay development

An outline imaging assay was created by Aurelia Bioscience Ltd who used the selected GFP-BicD2N-FRB/PTS-RFP-FKBP U-2 OS line to establish an appropriate cell revival protocol, selective growth conditions and seeding density for use. Further preliminary experimentation by Aurelia Bioscience Ltd that was aimed at producing an optimal assay window for the rapamycin response demonstrated the benefits of starving cells of serum for 24 h after an initial 24 h of settling and growth. Removal of protein and growth factors in this manner facilitated cell synchronisation and robust peroxisome relocalisation when rapamycin was added compared to serum-containing conditions. No serum starvation, or starvation periods longer than 24 h, generated variable data with an unworkable assay window.

Assay development within the UKC4LD determined a requirement for cells to be washed in situ to remove sufficient media for effective serum starvation. Assay variability was greatly affected by the number and stringency of washes and the equipment used, so multiple washing systems and protocols were explored before selecting the Biotek EL406 plate washer. The final conditions involved gentle removal of approximately 90% of cell media and replacement with serum free media, which minimised cell loss and greatly reduced well-to-well variability in peroxisome relocalisation.

We next assessed the ability of the CellInsight imaging system to evaluate peroxisome relocalisation. First, images were generated by confocal microscopy with an Olympus 40x air objective (UPLS apo, 0.95 NA) using the Yokogawa CV7000 (Fig. 2a, 2b). This system allowed detection of the concentration of GFP-BicD2N-FRB and PTS-RFP-FKBP at the MTOC in rapamycin-treated cells. Comparison of these images with the widefield images taken using the CellInsight system with a 10x air objective (UPlanFL, 0.3 NA) showed concordance (Fig. 2c). Inhibition of relocalisation by nocodazole was also clearly detectable with the low magnification imaging (Fig. 2d). Collectively, these data validate the use of the higher throughput CellInsight imaging system over the high-magnification confocal approach.

Figure 2. Confirmation of Utility of CellInsight Imaging and Establishment of Screening Parameters.

Images generated by high-magnification confocal microscopy were compared with low-magnification widefield images from the CellInsight. (a, b) 40x (0.95 NA) Yokogawa CV7000 confocal images of RFP and GFP signals in PTS-RFP-FKBP/GFP-BicD2N-FRB U-2 OS cells that were (a) unstimulated (DMSO only) or (b) stimulated with 2 nM rapamycin (150 min treatments before fixation). Note concentration of RFP and GFP fluorescence at a discrete spot in (b). (c, d) 10x (0.3 NA) widefield CellInsight images of GFP signal in PTS-RFP-FKBP/ GFP-BicD2N-FR U-2 OS cells treated with (c) 2 nM rapamycin or (d) 2 nM rapamycin and 10 μM nocodazole. Introduction of the microtubule disruptor gave a clear reduction in spot area (d), which was detectable using the CellInsight. Cells were treated with nocodazole or vehicle (DMSO) for 180 min before fixation, with rapamycin present for the last 150 min. In (a) and (d), the relatively dispersed fluorescent signal in the absence of rapamycin or in the presence of nocodazole is typically not detected with these imaging parameters. (e) Determination of a suitable rapamycin concentration and timecourse for Screening. Activating the cell system using a sub-maximal rapamycin concentration and assay endpoint generated signal windows suitable for identifying activating and inhibiting compounds of dynein-based transport in a single screen. Final conditions of 2 nM rapamycin and a 150-min timepoint (*) were selected for HTS.

To generate a suitable assay endpoint measure, test data comprising unstimulated wells, rapamycin-stimulated neutral control wells (DMSO treated) and rapamycin-stimulated inhibitor (10 μM nocodazole in DMSO) control wells were generated. Analysis of these data was carried out using Columbus image analysis software (Perkin Elmer) to create an output that could be assessed using a principal component analysis (PCA) method within Genedata Screener. All output features generated by Columbus were assessed for control separation and found to be most significant for components related to compactness of the fluorescent signal of GFP-BicD2N-FRB (Supplemental Fig. 3a-3c). The compactness is a measure of the spatial concentration or dispersion of fluorescence intensity within cells and is therefore intuitively similar to the quantification used by Kapitein et al. based on RFP signal on peroxisomes¹. In our experimental system, the RFP signal gave an inferior signal window to the GFP signal as peroxisomes tended to be enriched in the general vicinity of the nucleus in the absence of rapamycin, whereas this was not the case for GFP-BicD2N-FRB. The features isolated were compared to those available within the CellInsight Screener application (Thermo Fisher) and found to be most similar to the mean area of GFP spot size per cell per well (MEAN_CircSpotAvgAreaCh2), which was subsequently chosen as the assay feature measurement for high throughput experiments.

We reasoned that by configuring the cell system using a sub-maximal rapamycin concentration and/or an early timepoint it should be possible to create an assay window permissive for detecting activators and inhibitors of dynein-based transport in a single assay. Optimisation of assay conditions resulted in the establishment of parameters that gave a similar activity window size for both activators and inhibitors for use as an HTS (150 min incubation with 2 nM rapamycin; Fig. 2e).

HTS assay validation and screening

To determine whether the optimised assay met the AstraZeneca HTS validation criteria for single shot screening³⁵, a preliminary set of 1500 compounds followed by a statistically more powerful set of 7000 compounds were screened in duplicate on different days. Cells were preincubated with compounds at a single point concentration of 10 μM for 30 min followed by a further 150 min incubation in the presence of 2 nM rapamycin. Results were compared using a set of statistical tools created in Spotfire (Tibco). The average RZ’ was 0.53, with an average % CV (Coefficient of Variation) of 7 across a plate of neutral controls, which meets our validation criteria (RZ’ >0.5 and % CV <10). The level of agreement between the data from the duplicate runs was visualised using a Bland-Altman plot, where the mean of the data for each compound was plotted on the x-axis against the difference between the two results on the y-axis. The data generated showed good agreement between assay runs and validated the ability of the assay to detect both inhibiting and activating compounds (Fig. 3a). The mean of the difference between the data sets was close to zero, indicating very little intrinsic run-to-run variation. The limits of agreement (the 95% limits of agreement give the expected difference between the two measurements for 95% of future measurements) was quite large (approx. ±35%), but in the main the compounds classed as inhibitors and activators showed good agreement.

Figure 3. Validation and Primary Screen Data Summary.

(a) Bland-Altman plot comparing repeated validation data sets. ~ 7000 representative compounds were screened at a single concentration (10 μM) on two separate occasions and the data compared for agreement. The plot shows the mean results for each compound against the difference between the results. The unbroken line represents the median value for all difference data and the dashed lines are ± 1.96*SD, which shows the limits of agreement. Complete agreement of data sets would generate a median value of 0 with all points lying on the line. (b) Distribution of data generated from the screening of 505,500 compounds from the AstraZeneca compound collection at a single concentration of 10 μM.

505,500 compounds from the AZ collection were then screened using the developed assay. The average hit rate was 0.46% for inhibitors (2105 compounds) and 0.10% for activators (430 compounds) using a robust z-score cut-off of -5 for inhibitors and +5 for activators (Fig. 3b and Supplemental Table 3). Available hit compounds were taken through to confirmation XC₅₀ screening.

HTS assay performance

High throughput screening was carried out over a series of 30 runs, with an average robust Z’ value of 0.44 (± 0.10) and an average assay window of 2.1 (± 0.2) across the campaign. Use of the HighRes CoLAB Flex Cart-to-Cart CellInsight system (Supplemental Fig. 1) gave a low rate of plate failure and loss (0.09%) (Supplemental Fig. 4). Introduction of a workflow incorporating a 4oC storage environment, followed by incubation to elevate plate temperature to room temperature prior to reading, ensured signal stability over time and low data variability.

Hit confirmation by XC₅₀ screening and high-resolution imaging

XC₅₀ validation screening was carried out as a 10-point half log dilution series with a top and bottom concentration of 30 μM and 1 nM, respectively. Data were processed using Genedata Screener. In XC₅₀ testing, 1786 inhibitors (85%) and 308 activators (72%) from the original HTS were confirmed as active (>30% inhibition or activation at top dose) (see Fig. 4a, 4b and Supplemental Fig. 5a, 5b for example data), as were 695 inhibitors and 113 activators in near-neighbour screening of available compounds with similar chemical structures (Supplemental Table 3). XC₅₀ analysis with several compounds that are known to interfere with microtubule integrity (combretastatin, nocodazole, colchicine, and Vermox (mebendazole)), and are therefore expected to inhibit peroxisome relocalisation, provided positive controls for this analysis (Fig. 4c and Supplemental Fig. 5c, 5d).

Figure 4. Example Curve Data, Summary of Positive Controls, and Compound Clustering.

Example XC₅₀ curve data for an activator (a) and inhibitor (b) of dynein-based transport created using cell-based imaging assay data (see Supplemental Fig. 5a, 5b for additional examples). Data was generated as 10-point concentration response curves (bottom dose 1 nM, top dose 30 μM; one well (six fields of view) per concentration) using average green spot area (GFP-BicD2N-FRB) as a measure. % activity for inhibiting compounds was determined by normalising data to uninhibited DMSO only controls (0%) and nocodazole (10 μM) inhibited controls (100%). % activity for activating compounds was determined by normalising data to uninhibited DMSO only controls alone. (c) Effect of known microtubule modulating compounds in the XC₅₀ U-2 OS peroxisome trafficking cell assay. Microtubule destabilising compounds disrupted peroxisome localisation, as expected (see Supplemental Fig. 5c, 5d for examples of IC50 data). Compounds that stabilise microtubules also had this effect, suggesting that microtubule dynamics are important for peroxisome relocalisation in this assay. The actin-disrupting compound cytochalasin D was not active in the assay. (d) Cluster diversity of inhibitors. After elimination of microtubule disruptors identified with the CellInsight images, HTS hits were clustered with a Tanimoto distance cut off <0.3 to create more than 50 diverse clusters (see also Supplemental Table 4). The graphical representation highlights the structural diversity of clusters thought to contain five or more members by grouping compounds where the proximity of points is relative to structural similarity. Commercially available disruptors of the cytoskeleton are included in this analysis for comparison. Points that lie in the grey surrounding area are classed as singletons with no structural similarity to the clustered compounds. Darker spots are a result of overlapping points where multiple compounds have very similar structures.

Inhibitor XC₅₀ compound plates were rescreened in a tubulin immunostaining assay to identify and eliminate disruptors of microtubules, which will inhibit dynein-based transport indirectly. Plates were stained with a primary anti-tubulin antibody and a far-red fluorescent secondary antibody, before imaging using the CellInsight system. The fine, filamentous nature of microtubules meant that it was not straightforward to analyse their integrity using low magnification, widefield microscopy and automatic image processing methods. However, we were able to develop a method that discriminated between control and nocodazole-treated cells using Columbus software (Perkin Elmer) (see Methods). Using this procedure, 1352 of the 2481 inhibitor hits (54%) from the XC₅₀ assay plates and near-neighbour screen were scored as microtubule disruptors (Supplemental Table 3).

The accuracy of the CellInsight-based microtubule analytics was further assessed by manually treating cells with a subset of compounds and assessing microtubule targeting activity by high-resolution confocal imaging (40x oil objective, 1.3 NA) (Supplemental Fig. 6). All nine tested compounds previously scored as microtubule disrupting agents were confirmed to strongly affect microtubule integrity. Nine of 16 compounds previously scored as not disrupting microtubule integrity were validated by high magnification imaging, with the remaining seven compounds causing substantial microtubule disruption. Collectively, these data indicate that, although a fraction of microtubule-disrupting compounds were missed in the low magnification CellInsight data, a sizeable proportion of such compounds were identified.

From the set of activator compounds and those inhibitor compounds scored as not disrupting microtubules in the CellInsight data, structure clustering and chemistry triaging was carried out to remove those with undesirable structures (e.g. quaternary ammonium salts, metal-chelating complexes, surfactants) and physicochemical properties (e.g. MW<600, cLogP <5.5) or low purity (cut-off of < 85%). A filter for frequent-hitter analysis was also applied, which eliminated compounds with abnormally high frequency of activity in previous HTS screens at AstraZeneca (active in >15% of previous screens) and are therefore candidates to have pleiotropic effects. Potential dyes (e.g. coumarin-based dyes) and compounds with structures that closely resembled rapamycin were also removed. A shortlist for follow-up testing was generated that contained 842 inhibitors and 37 activators (Supplemental Table 3). Clustering analysis indicated that none of the activators had structural similarities. Structure-based clustering of inhibitors was, however, possible (Supplemental Table 4; Fig. 4d) and this is described in more detail in the Discussion.

The activity of a small number of shortlisted activator compounds was confirmed by testing them in further XC₅₀ screens containing varied amounts of rapamycin (Table 1). Two of 12 compounds tested were scored as activators of peroxisome relocalisation in the absence of rapamycin, meaning that they potentially act analogously to rapamycin by recruiting GFP-BicD2N-FRB to these organelles. One other compound did not reproducibly activate peroxisome relocalisation in the presence of rapamycin. Treatment of the parental, unmodified U-2 OS cell line allowed the detection and removal of two additional compounds that were fluorescent and therefore gave false positive results. In summary, these analyses indicate that seven of the 12 tested compounds are strong candidates to be bona fide activators of dynein-based transport. Extrapolating from these results, a substantial fraction of the remaining 25 shortlisted activator compounds will also be in this category.

Table 1.

Summary of Compound Activities and Modes of Action for Potential Activators of Dynein-Based Transport when Screened in Different Cell Lines at Varying Rapamycin Concentrations. XC₅₀ values for MEAN_CircSpotAvgAreaCh2 (GFP) as calculated by Genedata Screener are shown in parentheses. Compounds were screened in PTS-RFP-FKBP/GFP-BicD2N-FRB or wild-type U-2 OS cells using 0, 1 or 2 nM rapamycin. Curve shape (increasing or inactive) was determined by the Genedata Screener data analysis package and used as a measure of compound activity alongside XC₅₀. Compounds giving positive data in the wild-type cell assay were deemed to give an artefactual fluorescent signal. Compounds giving positive data with 0 nM rapamycin in the PTS-RFP-FKBP/GFP-BicD2N-FRB cells have the potential to have rapamycin-like activity (i.e. mediating association of GFP-BicD2N with peroxisomes).

	Genedata Data Mode
	U-2 OS SP2 CL1 (PTS-RFP-FKBP/GFP-BicD2N-FRB)			U-2 OS Wild type
Comp ID	0 nM Rapamycin	1 nM Rapamycin	2 nM Rapamycin	0 nM Rapamycin	1 nM Rapamycin	2 nM Rapamycin	Conclusion
1	Inactive	Inactive	Inactive	Inactive	Inactive	Inactive	False Positive
2	Increasing (>30 μM)	Increasing (22.3 μM)	Increasing (23. 1 μM)	Inactive	Inactive	Inactive	Rapamycin independent
3	Increasing (>30 μM)	Increasing (>30 μM)	Increasing (>30 μM)	Inactive	Inactive	Inactive	Rapamycin independent
4	Increasing (1.3 μM)	Increasing (2.9 μM)	Increasing (3.0 μM)	Increasing (12.8 μM)	Increasing (6.7 μM)	Increasing (6.5 μM)	Fluorescent Compound
5	Inactive	Increasing (0.7 μM)	Increasing (5.2 μM)	Inactive	Inactive	Inactive	Confirmed Activator
6	Increasing (2.1 μM)	Increasing (2.45 μM)	Increasing (2.9 μM)	Increasing (3.0 μM)	Increasing (1.3 μM)	Increasing (1.9 μM)	Fluorescent Compound
7	Inactive	Increasing (0.46 μM)	Increasing (0.3 μM)	Inactive	Inactive	Inactive	Confirmed Activator
8	Inactive	Increasing (>30 μM)	Increasing (20.9 μM)	Inactive	Inactive	Inactive	Confirmed Activator
9	Inactive	Increasing (27.2 μM)	Increasing (>10 μM)	Inactive	Inactive	Inactive	Confirmed Activator
10	Inactive	Inactive	Increasing (8.7 μM)	Inactive	Inactive	Inactive	Confirmed Activator
11	Inactive	Increasing (>30 μM)	Increasing (10.1 μM)	Inactive	Inactive	Inactive	Confirmed Activator
12	Inactive	Increasing (21.0 μM)	Increasing (>10 μM)	Inactive	Inactive	Undefined	Confirmed Activator

Finally, we further validated shortlisted compounds in new peroxisome relocalisation assays that were analysed using high-resolution confocal imaging (40x oil objective, 1.3NA). We first selected four inhibitor compounds that were not related to each other structurally and were amongst those found not to disrupt microtubules in our previous high-resolution analysis. Strong inhibition of dynein-based transport by each of these compounds was evident in the raw images (Fig. 5a) and this was confirmed by automated analysis of mean GFP spot intensity per cell (Fig. 5b). We next selected one shortlisted activator and confirmed that it accelerates dynein-based trafficking to the MTOC in the peroxisome relocalisation assay using the same analysis pipeline (Supplemental Fig. 7). These data further corroborate the ability of the screening platform to identify modulators of dynein-based cargo transport in vivo.

Figure 5. Further Validation of a Subset of Inhibitors with High-Resolution Imaging.

(a) High-magnification confocal images (40x oil objective, 1.3 NA) of GFP-BicD2N-FRB/PTS-RFP-FKBP U-2 OS cells illustrating inhibition of rapamycin-dependent clustering of GFP and RFP signals by four inhibitors validated in the XC₅₀ screen. DMSO control (top row) shows location of fluorescent signals in the absence of both rapamycin (2 nM) and an inhibitor (10 μM). Cyan arrowheads point to examples of intense clustering of RFP and GFP at the MTOC in the presence of rapamycin only. Yellow arrowheads show examples of rapamycin-induced enrichment of GFP on RFP-labelled peroxisomes that fail to relocalise to the MTOC in the presence of an inhibitor, indicating that the compounds do not disrupt association of FKBP with FRB. (b) Quantitative analysis of mean intensity of GFP spots per cell confirms that compounds E-H disrupt rapamycin-dependent relocalisation. Mean GFP spot intensity values are normalised to the median value for rapamycin only. Number of cells analysed for each condition are shown in italics. Boxes show interquartile range (25^th-75^th percentile of values) and horizontal line is the median. Vertical lines illustrate 1.5x the interquartile range with outliers shown with circles. Statistical significance (compared to the rapamycin only sample) was evaluated using a pairwise t-test with p-values adjusted for multiple comparisons using the false discovery rate correction (****, p <0.0001).

Discussion

A dual-labelled cell line expressing a rapamycin-inducible system for linking activated dynein complexes to peroxisomes was generated to monitor dynein-based transport within live cells. Corroborating previous results^1,2, chemically controlled tethering of dynein to peroxisomes with this system allowed tightly regulated translocation of these organelles towards the perinuclearly-positioned MTOC.

Developing an imaging assay that can identify both the inhibition and activation of dynein-based transport and is suitable for screening more than 500,000 compounds required several adjustments to previously described protocols. Robust and reproducible peroxisome relocalisation was achieved by serum starvation, whilst the use of multiple imaging platforms allowed validation of data acquisition and selection of features in image analytics that ensure a robust assay (Robust Z’ ~ 0.5). This report is the first demonstration of the HighRes CoLAB capability to generate complex workflows through a flexible combination of different automation and instrumentation. The use of these systems in this project demonstrates that a high level of reliability can be achieved, which is compatible with out-of-hours operation.

Following successful assay validation, a 500,000 compound HTS was completed, resulting in the identification of 2105 candidate inhibitors and 430 candidate activators. XC₅₀ screening (30 μM top dose) confirmed the activity of 1786 inhibitors and 308 activators (>30% activity) from the original screen and 695 inhibitors and 113 activators from a near neighbour screen.

The hits from the XC₅₀ analysis were assessed using a high throughput tubulin immunostaining assay to identify potential microtubule disruptors. Future work will involve using high-resolution confocal imaging to confirm which shortlisted compounds have microtubule disrupting activity. Whilst such compounds would be excluded from efforts to develop direct inhibitors of the dynein transport machinery, they may be candidates to develop as microtubule targeting agents for therapeutic purposes^36,37.

Inhibitors of dynein-based transport that were not scored as microtubule disruptors, and activators confirmed through XC₅₀ screening, were assessed by a process of triaging to remove compounds with undesirable structures and poor physicochemical properties or which were frequently found in other cellular HTS screens. This resulted in 842 inhibitor and 37 activator compounds.

The shortlisted molecules were subjected to clustering analysis based on structural similarity using a Tanimoto distance cut off <0.3. Whereas all activators were singletons, a large number of diverse clusters (>50) were observed for the inhibitors (Supplemental Table 4). The diversity of these clusters is analysed in Figure 4d, where the top clusters are illustrated as a function of similarity and distance to one another. This plot shows that these clusters are diverse from one another, as well as from most of the known scaffolds of commercial microtubule-targeting compounds, except Vermox (mebendazole) and analogues. From this clustering analysis it can be inferred that novel structures were identified in the screen.

We also established methods for the identification of undesirable activation modes in the assay for dynein-based transport, including the ability of compounds to activate transport in the absence of rapamycin or give an artefactual fluorescent signal. These assays will be extended to the full set of activators from the XC₅₀ analysis in the future.

Finally, we performed additional analysis of a subset of shortlisted inhibitors and activators with high-resolution confocal imaging, which corroborated the ability of the screening platform to identify interesting lead compounds.

In summary, innovative assay development, the use of cutting-edge HTS automation, and sophisticated imaging analysis tools resulted in a robust assay capable of detecting two outcomes in a single screen: activation and inhibition of dynein-based cargo transport in cells. Identified compounds were subjected to a series of cascade assays that iteratively provide information about compound mode of action. The primary objective for the future will be to characterise the molecular mechanisms by which the shortlisted activating and inhibiting compounds modulate dynein-based transport. This work is ongoing and includes in vitro assays for the activity of compounds towards the purified dynein motor in the presence and absence of dynactin and activating adaptors. It is anticipated that this project will yield further publications and form the basis of an expanded drug discovery collaboration between AstraZeneca and academic partners.