Machine learning‐based phenotypic screening for postmitotic growth inducers uncover vitamin D3 metabolites as small molecule ribosome agonists

Abstract Objectives To restore tissue growth without increasing the risk for cancer during aging, there is a need to identify small molecule drugs that can increase cell growth without increasing cell proliferation. While there have been numerous high‐throughput drug screens for cell proliferation, there have been few screens for post‐mitotic anabolic growth. Materials and Methods A machine learning (ML)‐based phenotypic screening strategy was used to discover metabolites that boost muscle growth. Western blot, qRT‐PCR and immunofluorescence staining were used to evaluate myotube hypertrophy/maturation or protein synthesis. Mass spectrometry (MS)‐based thermal proteome profiling‐temperature range (TPP‐TR) technology was used to identify the protein targets that bind the metabolites. Ribo‐MEGA size exclusion chromatography (SEC) analysis was used to verify whether the ribosome proteins bound to calcitriol. Results We discovered both the inactive cholecalciferol and the bioactive calcitriol are amongst the top hits that boost post‐mitotic growth. A large number of ribosomal proteins' melting curves were affected by calcitriol treatment, suggesting that calcitriol binds to the ribosome complex directly. Purified ribosomes directly bound to pure calcitriol. Moreover, we found that calcitriol could increase myosin heavy chain (MHC) protein translation and overall nascent protein synthesis in a cycloheximide‐sensitive manner, indicating that calcitriol can directly bind and enhance ribosomal activity to boost muscle growth. Conclusion Through the combined strategy of ML‐based phenotypic screening and MS‐based omics, we have fortuitously discovered a new class of metabolite small molecules that can directly activate ribosomes to promote post‐mitotic growth.


| INTRODUCTION
Tissue ageing may cause tissue damage and degeneration, in some cases leading to ageing-related diseases that involve tissue/cell atrophy, such as sarcopenia, cachexia, osteoporosis, pancreatic atrophy and Alzheimer's disease. [1][2][3][4][5] In order to restore tissue growth without increasing the risk for cancer, there is a need to identify small molecule drugs that can increase cell growth, without increasing cell proliferation. While there have been numerous high-throughput drug screens for cell proliferation, there have been few screens for postmitotic anabolic growth, for example, in skeletal muscles. Skeletal muscle is composed of postmitotic myofibres. 6 Skeletal muscle mass and strength decrease with advancing age in sarcopenia, as net muscle protein synthesis becomes dramatically decreased. 4,[7][8][9] In twodimensional (2D) culture in vitro, human skeletal muscle myoblasts are able to exit the cell cycle, fuse and differentiate into postmitotic multinucleated myotubes, where the length and diameter distributions of myotubes could be indicators of postmitotic growth. However, aside from a few screens for drugs that promote muscle contraction in response to chemical stimulation 10 and the rescue of dystrophin deficiency, 11 there have been few if any drug screens that exclusively focus on postmitotic growth or cell volume. Due to the stochasticity of myoblast fusion and myotube directionality, automated and accurate extraction of relevant morphological parameters for high-throughput drug screening had remained elusive.
In the field of drug screening, the significant drop in newly approved drugs in recent decades has been partially attributed to failures in the target-based screening model. Evaluation of new drugs approved by FDA has revealed that the number of successful drugs from phenotypic screens exceeded those discovered through molecular target-based screens, despite the overwhelmingly larger amounts of investments into targetbased projects over the same period. 11 This surprising result has contributed to growing interest in phenotypic screens for drug development. Nevertheless, phenotypic screening has its own problems. Lacking a welldefined target(s) makes it difficult to further optimize the drugs, identify molecular biomarkers and select patient subpopulations for clinical trials.
An emerging technology that combines the strengths of phenotypic screens and target-based screens is mass spectrometry (MS)-based thermal proteome profiling (TPP). TPP enables an unbiased and comprehensive search of the targets of a small molecule drug across the entire proteome, by assuming that drug-binding alters its target protein's structure and thus thermal stability in solution, which can be revealed by comparing proteins' abundance in solution, before and shortly after drug treatment, over a temperature range. 12,13 Another advantage of TPP is that it can discover targets in living cells and tissues without requiring compound labelling for pulldowns. 13,14 Thus, TPP affords a high-throughput strategy to fuse the advantages of phenotypic screens with target-based drug discovery.
Here our platform of ML-based phenotypic screens for postmitotic growth combined with MS-based TPP proteomic target discovery, revealed that vitamin D3 metabolites promote postmitotic muscle growth by directly binding and activating ribosomes. The inactive vitamin D3 (cholecalciferol) needs to be metabolized into the inactive 25-hydroxyvitamin D3 (25(OH)D or calcifediol) in the liver, then transported to the kidney where it is finally metabolized to the bioactive form 1,25-dihydroxyvitamin D3 (1,25(OH)2D or calcitriol). 15 Vitamin D3 deficiency is common among older adults and has been linked to muscle weakness and subsequent sarcopenia, potentially leading to increased risk of chronic pain, frailty, falls and death. [16][17][18] Calcitriol exerts genomic effects through the vitamin D nuclear hormone receptor (VDR) to control gene transcription, but it is also widely-speculated to possess rapid and nongenomic effects, although this noncanonical mechanism(s) had remained unclear. [19][20][21] Moreover, VDR expression in muscle tissue is progressively reduced during development and ageing, 22,23 with very low levels remaining in adult skeletal muscle. 24 Our discovery that postmitotic growth can be enhanced by both the "inactive" cholecalciferol and "active" calcitriol, each with VDR K d 's that are orders of magnitude apart, corroborates the importance of noncanonical, VDR-independent mechanisms.  whole cell images were analysed for Hoechst 33342 staining and Digital Phase Contrast signals to determine nuclei numbers and 3D cell volumes respectively, at Days 3, 5, 7 and 9. Images were analysed by using Harmony 4.6 software, and programmed for automated segmentation of nuclei and myotubes. First, nuclei were defined and selected by an optimal area and roundness threshold. Whole myotubes were defined using cell area, length and intensity thresholds, after recognition with linear classifier-based machine learning (ML) algorithms trained and optimized in house.
Specifically, multinucleated myotubes were extracted and calculated by optimizing for specific parameters, including the min/max myotube area, min/max myotube length and minimum intensity thresholds. Finally, the multinucleated myotube area (%) was used as a parameter to identify positive drug hits.

| Quantitative real-time polymerase chain reaction
Quantitative real-time polymerase chain reaction (RT-PCR) was performed using a LightCycler480 II sequence detection system (Roche Applied Science). RNA was extracted by TRIzol (Takara) and reverse transcribed with PrimeScript 286 RT (Takara), and quantitative PCR was conducted by using KAPA SYBR FAST qPCR Kits (KAPA Biosystem) following the manufacturer's instructions. All primer sequences were as reported previously. 25

| Thermal proteome profiling
Protein samples were prepared from control (0.01% DMSO) and calcitriol-exposed myotubes. For each group, 2 Â 10 7 myoblasts were seeded in two 15 cm cultured dishes. After 24 h of culturing, growth medium was replaced with 1% KSR medium after one wash with PBS, Waters). The flow rate was 250 μl/min, and column temperature was kept at 50 C. Each fraction was collected and dried in vacuo.

| LC-MS/MS
For liquid chromatography-mass spectrometry (LC-MS) proteomic data collection, the fractions resolved in 0.1% formic acid were separated by nanoLC, and analysed using EASY-nLCTM 1200 system-coupled to a Q-Exactive-HF-X (Thermo Fisher Scientific) mass spectrometer, according to previously reported methods. 13 In brief, 4 μl peptide sample was loaded with a flow rate of 300 nL/minute in a Phenomenex analytical column (75 μm Â 20 cm, Aqua 3 μm C18 125 Å), followed by gradient separation from 1% to 37% 0.1% formic acid in CAN for 63 min.
At 65-min separation, the gradient was increased to 95% 10 mM ammonium formate in 90% CAN and held for 5 min. Methionine oxidation was set as a variable modification.

| Protein synthesis assays
Human myoblasts in six-well plates (100% confluency) were incubated in 1% KSR medium. After 24-h differentiation, the cells were incubated in fresh 1% KSR medium containing 50 μg/ml CHX (cycloheximide, to block further protein synthesis), in the presence or absence of 100 nM calcitriol. Importantly, CHX was added to the cells 1 h before calcitriol. After CHX ± calcitriol incubation, the human myoblasts were washed with PBS and replaced with 1% KSR ± calcitriol for further differentiation. After terminal differentiation, the cells were washed with PBS and collected for Western blotting. Western blots for myosin heavy chain (MHC), MYOD1 and MYOG, were used to measure steady state protein levels, as described above.
Nascent protein synthesis was quantified by measuring incorporated O-propargyl-puromycin (OP-puro) into polypeptides. Briefly, 4000 human myoblasts (each well) were seeded in a gelatin-coated 96-well μClear microplate (Greiner), then incubated in 1% KSR medium for myotube differentiation. After 72-h differentiation, the cells were incubated with 10 μM OP-puro (MedChem) and 0.01% DMSO (as the negative control) or 100 nM calcitriol for 0.5 h at 37 C and 5%CO 2 , while 1% KSR without OP-puro served as the unlabeled control. Immediately after incubation, a click reaction was performed between click-iT Plus OPP Alexa Fluro 594 and the OP-Puro, according to the manufacturer's protocol (C10457; Invitrogen). The stained cells were imaged by Nikon fluorescence microscopy. Average fluorescence intensity was qualified in ImageJ, the data were obtained from at least 10 images from three independent experiments.
Western blotting for puromycin was also used to quantify nascent protein synthesis. Human myoblasts (2 Â 10 6 ) in a six-well plate were grown, differentiated and incubated with 10 μM OP-puro ±0.01% DMSO or 100 nM calcitriol as described above. Immediately after incubation, the cells were washed with cold PBS three times, followed by protein lysis and BCA quantification. Blots were incubated with anti-puromycin clone 12D10 (1:3000; Sigma).

| Ribo Mega-SEC and uHPLC
Human myoblasts in two 15-cm dishes were differentiated into mature myotubes for 96 h, then incubated with 1% KSR medium containing 50 μg/ml CHX for 5 min at 37 C and 5% CO 2 . After incuba- were merged and processed by using Chromeleon 7.0 software.

| Statistical analysis
All experiments were performed ≥3 times in triplicate and data are represented as mean value ± SEM for statistical comparisons. Statistical analysis was performed by using GraphPad Prism 9.0 and Microsoft Excel. Significance of differences was assessed by a paired two-tailed Student's t-test to compare two groups. A value of p < 0.05 was considered statistically significant.

| Identification of natural compounds that promote muscle volumetric growth
Our goal was to establish a reliable and high-throughput screening strategy for identifying drugs that boost human skeletal muscle postmitotic growth ( Figure 1A). First, we treated primary human skeletal myoblasts that had exited mitosis with a large metabolite compound and small molecule library. Then, we analysed the cells using Digital Phase imaging and ML algorithms, to estimate 3D cell volumes in a high-content, high-throughput analysis system. Digital phase image construction uses a computational strategy to generate quantitative phase images with multiple red LED-based brightfield images. The computational algorithm calculates the rate of change in light intensity distribution between different brightfield images at different focal planes, to obtain the refraction-induced phase shift intensities of the specimen.
that is, optical thickness is proportional to the phase shift intensity at point x.
where w is a weight vector for j parameters that is learned from a set of labelled training samples, and f is a threshold function, which maps all values of w Â x above a certain threshold to the myotube Class 1 and all other values to the nonmyotube Class 0, that is  Figure 1A). To validate our ML algorithm, we compared the result given by ML with the analysis result by an experienced researcher, and the outcome was highly consistent. Moreover, our linear classifier's segmentation of multinucleated myotubes could robustly assess the progrowth phenotype of the anabolic growth-inducing factor IGF-1, and the anti-growth phenotype of the inflammatory growth-suppressive cytokine IL-1β ( Figures 1B,C). Furthermore, we observed a strong correlation between the ML-determined OPD signal and the area of immunostained MHC+ myotubes (R = 0.88, p < 0.0001; Figure 1D). In addition, the ML analysis correctly identified postmitotic myotubes and excluded all mitotic Ki67+ muscle cells ( Figure S1A). These findings indicate that our platform's 3D volumeand ML-based quantification of myotubes could reliably assess postmitotic growth.
For the high-throughput drug screen, we seeded primary human myoblasts onto 384-well plates at high density, and screened natural small molecule compounds from the TargetMol library ( Figure 1E). Primary hits for postmitotic growth were selected using the 3D volumeand ML-based recognition of myotubes, and calculating the fold change in myotube percentage, relative to that of DMSO vehicle. The top primary hit was calcitriol (1,25(OH)2D3) or the bioactive form of vitamin D3 ( Figure 1E). Pioglitazone, one of the other primary hits, has been shown to be a PPARg agonist that significantly improved skeletal muscle functions in a reserpine-induced fibromyalgia rat model. 27 Another primary hit, cholecalciferol, is an inactive precursor of calcitriol which requires sequential hydroxylation in the liver and then the kidney before it binds avidly to the vitamin D receptor, 28 and has also been shown to improve muscle strength and performance in athletes and old adults. 29,30 The identification of pioglitazone and cholecalciferol as positive hits further supported the validity of our primary screening and data analysis, although the identification of both calcitriol and cholecalciferol despite their vastly different K d 's for VDR (3.8 pM and >20 mM, respectively) also suggested that VDRindependent mechanisms might be operating. 28 Figure S1B). Importantly, the mean myotube 3D volume was significantly higher in the calcitriol group than the DMSO control group (Figure 1F,G). Other parameters such as myotube width, length, total area and myotube percentage area, were also significantly higher in the calcitriol group ( Figure 1G). These observations showed that calcitriol dramatically promoted human myotube hypertrophy. Myotube terminal differentiation was further estimated by the fusion index, based on MHC staining at Day 7 ( Figure 1H). Compared with the DMSO vehicle control group (70.80% fusion), calcitrioltreated myoblasts differentiated into multinucleate myotubes to a slightly higher degree (75.08% fusion, Figure 1I). Similarly, mRNA profiling of myogenesis markers also showed that calcitriol induced the expression of myogenesis markers such as ACTA1 and MYH7 at Day 4 ( Figure S1C). These results strongly suggest that calcitriol mainly promotes myotube growth and maturation, but also promotes myotube terminal differentiation. Based on the findings above, we chose calcitriol as the final lead compound.

| Effects of calcitriol on myoblast proliferation, myotube growth and adhesion
To confirm calcitriol's abilities to regulate proliferation, growth and blasts. 34 To investigate the effect of calcitriol on human skeletal muscle myoblast proliferation, we performed cell counting and Ki67 immunofluorescence staining. Although the Ki67+ proportion was slightly increased in calcitriol-treated myoblasts at Day 5, it was not statistically significant ( Figure 2D). Moreover, cell counting revealed that calcitriol treatment suppressed primary human myoblast proliferation ( Figure 2D). Thus, calcitriol promotes muscle cell growth without increasing cell proliferation.
It is also well-known in the field that within 1-2 weeks of culture, most primary myotubes will undergo delamination 35 and cell death due to imperfect metabolic conditions and inadequate extracellular matrix (ECM) biosynthesis in vitro. As expected, primary human myoblasts underwent delamination after just 5 days of differentiation, and flocculent aggregates of multinucleate myotubes were found floating in the culture medium by Day 7 ( Figure 2E). In contrast, calcitriol-treated myotubes showed significantly stronger ECM adhesion during differentiation ( Figure 2E). Here, we further investigated if the adhesion effect of calcitriol would be abrogated by supplementing the culture surface with type I collagen during myotube differentiation. We found that calcitriol-treated myotubes still showed stronger ECM on type I collagen-coated culture wells at Day 5, whereas DMSO-exposed myotubes were already undergoing delamination at Day 3 ( Figure S2A). Moreover, both mRNA profiling and Western blots of MHC confirmed that calcitriol promotes myotube ECM adhesion, longer-term growth and maturation even in the presence of type I collagen (Figures S2B and S2C).

| Identification of protein targets for calcitriol using TPP
The identification of both cholecalciferol and calcitriol in our primary screen, which have vastly different binding affinities K d 's for VDR, 28 and the fact that VDR is only weakly expressed in adult muscle, 24 suggested that VDR-independent mechanisms might be operating. To identify novel binding targets of calcitriol in an unbiased manner, we used the latest MS-based techniques for TPP. The thermal stability of drug-binding versus nonbinding proteins was measured by quantification of the relative amounts of soluble proteins remaining in the supernatants at different temperatures ( Figure 3A).
Hierarchical clustering of the TPP proteomic results showed that the thermal stability profiles were highly similar across the entire proteome with and without brief calcitriol treatment ( Figure 3B), with the exception of 58 proteins in the middle of the heatmap, which displayed significantly different thermal stability profiles around 37-44 C after brief calcitriol treatment ( Figure 3B).
Intriguingly, most of these were ribosomal proteins, indicating that the ribosome proteins' solubility at 37-44 C were affected by brief treatment with calcitriol, and suggesting that the ribosome complex itself could bind to calcitriol (Figure 3B right). Representative ribosomal proteins' (RPL15, RPS8, RPL6 and RPL37A) solubility at 59 versus 37 C showed a significant difference after calcitriol treatment, compared to DMSO treatment ( Figure 3C). In contrast, as negative comparative controls, the proteins ENY2, ANP32E, PGGT1B and PTK7 showed a variety of thermal stability trends, but little differences after calcitriol treatment ( Figure 3C). When examined in greater detail, the solubility versus temperature (or "melting") curves of the ribosomal proteins RPL15, RPS8, RPL6 and RPL37A were significantly different between calcitriol-and DMSO-treated samples, whereas no differences in solubility were observed among the negative control proteins ENY2, ANP32E, PGGT1B and PTK7 over the whole TR ( Figure 3D). These observations suggested that the ribosome complex binds directly to calcitriol to produce these thermal stability changes.

| Calcitriol promotes protein translation through direct binding of the ribosome
To confirm our TPP data analysis results with an orthogonal method, we performed a Ribo Mega-SEC analysis of myotube ribosome proteins using uHPLC. Ribo Mega-SEC is a recently developed approach for characterizing changes in ribosome assembly in response to drugs. 26  In particular, brief treatment of pure ribosomes with pure calcitriol reproducibly decreased the fast-migrating 80S peak, and increased the slow-migrating 40S peak ( Figure 4A right) Figure 4B). Cell morphology observations on myotube differentiation corroborated these findings ( Figure S4). In addition, we also measured nascent protein synthesis by using a nonradioactive probe for protein incorporation of fluorescently-labelled O-propargyl -puromycin, in the absence or presence of calcitriol. Compared to the DMSO vehicle control, calcitriol stimulation significantly increased the incorporation of fluorescent puromycin after just 3 h of incubation, indicating that calcitriol can rapidly promote nascent protein synthesis in human myotubes ( Figure 4C). Finally, Western blot analysis for puromycin incorporation further confirmed that calcitriol stimulation significantly increased nascent protein synthesis within 0.5 h ( Figure 4D). These data further support our findings that calcitriol can rapidly promote protein translation in human skeletal muscle, well before the genomic effects of VDR-dependent transcription can begin to take effect.  45 This has been a nontrivial problem in the field of muscle cell culture, and calcitriol's preventive effects on myotube delamination might represent a new strategy to address this problem for in vitro drug screening and tissue engineering.

| DISCUSSION
Given that both cholecalciferol and calcitriol emerged in our highthroughput screen, despite their disparate binding affinities for VDR, we were motivated to search for VDR-independent targets of calcitriol, using TPP proteomics. 46 Perrin et al. have previously used TPP to identify the targets of panobinostat in rat liver, lung, kidney and spleen, and revealed that ZNF512 binds to panobinostat. 46 TPP also enabled the high-throughput discovery of drug targets and downstream effectors, 47 including transmembrane protein targets. 48 In TPP, the temperature gradient gradually denatures and precipitates proteins irreversibly in general. But individual proteins with different structures, especially after binding small molecules, might have different thermal solubility or 'melting' curves. 13,49 While traditional approaches assumed most protein solubility or 'melting' curves follow a sigmoidal shape, deviations from sigmoidal curves have also been found to be very common, 50-52 especially for protein complexes.
Using hierarchical clustering and maximum fold change analyses, we identified many novel targets of calcitriol, of which an overwhelming majority are ribosomal proteins. Importantly, Ribo Mega-SEC, puromycin incorporation and functional assays orthogonally confirmed the ribosome complex as a novel target of calcitriol, and verified that TPP data analysis is a reliable strategy for high-throughput target discovery after ML-based phenotypic screening.
Many previous studies have focused on ribosomal inhibitors such as the natural compounds puromycin and cycloheximide, which bind very tightly to lock ribosomes in certain conformations, but none have uncovered ribosomal agonists hitherto. To our knowledge these are the first archetypal members of small molecule ribosome agonists, which could find applications in the treatment of muscle degenerative diseases of ageing. In comparison, proteasome modulators, which also regulate steady state levels of protein abundance, have seen some success with muscle atrophy due to sepsis or denervation, but largely failed to reverse sarcopenia and cachexia in vivo. 53 Indeed, previous studies have already shown that vitamin D3, through genomic or nongenomic effects, can also enhance the development and growth of brain, bone, immune, ovarian follicle, skin, hair and sperm cells. [54][55][56][57][58][59][60] Our results suggest that ribosome agonists might provide an alternative approach for preventing and treating tissue degeneration in the future.