High-content screening identifies a small molecule that restores AP-4-dependent protein trafficking in neuronal models of AP-4-associated hereditary spastic paraplegia

Unbiased phenotypic screens in patient-relevant disease models offer the potential to detect therapeutic targets for rare diseases. In this study, we developed a high-throughput screening assay to identify molecules that correct aberrant protein trafficking in adapter protein complex 4 (AP-4) deficiency, a rare but prototypical form of childhood-onset hereditary spastic paraplegia characterized by mislocalization of the autophagy protein ATG9A. Using high-content microscopy and an automated image analysis pipeline, we screened a diversity library of 28,864 small molecules and identified a lead compound, BCH-HSP-C01, that restored ATG9A pathology in multiple disease models, including patient-derived fibroblasts and induced pluripotent stem cell-derived neurons. We used multiparametric orthogonal strategies and integrated transcriptomic and proteomic approaches to delineate potential mechanisms of action of BCH-HSP-C01. Our results define molecular regulators of intracellular ATG9A trafficking and characterize a lead compound for the treatment of AP-4 deficiency, providing important proof-of-concept data for future studies.


Introducfion
The introducfion is well wriften and nicely summarizes the state of scienfific knowledge

Results
General comment: Data visualizafion and interpretafion are valid and robust.Very good and detailed descripfion of the data evaluafion.
Primary screen -How many wells per condifion (compound) were analyzed?-There is a rather large variability in cell counts (Fig. 1g,.Cell count does not match the inifial plafing number (2,000 cells per well) How can this be explained?-ll 110 -111: "LoF/LoF mean: 1.1 ± 0.02 SD, WT/LoF mean: 1.34 ± 0.05 SD" -data swapped Counter-screen -34 compounds were found to have autofluorescence or imaging arfifacts.Was this already analyzed / idenfified in the primary screen?
Secondary screen (SH-SY5Y) General comment: Why not used as "counter-screen assay" as it seems that the separafion between posifive and negafive control is much more robust?(perhaps due to the fact of comparing a full KO to WT instead of a het carrier to a hom pafient) -The inclusion of the DAGLB assay is an important and careful validafion -Why was the "mulfiparametric morphological profiling approach" only performed on 5 compounds?Would it be possible to implement this analysis as a screening tool for future projects, or at least for the counter-screen?-It appears that compounds B-01 and G-01 also alter cellular morphology in a dose-dependent manner (Suppl.Fig. 4, l 184) iPSC-derived neurons -It would be helpful to include a staining (overview) to visualize neuronal morphology especially for those condifions / compounds that show toxic properfies -Would it be possible to quanfify ATG9A levels in the distal part of the axon?Target deconvolufion -RNA sequencing: It is unclear why SHSY-5Y cells were used instead of iPSC-derived neurons as compound C-01 showed only small effects on ATG9A rafios in previous analyses (Fig. 3d) -72h instead of 24h treatment (results of ATG9A rafios for SHSY-5Y cells are missing) -Autophagic flux: Interesfingly, there is no baseline difference between AP4B1-KO and control SH-SY5Y cells (Fig. 8 j-l) Discussion: -Expanding the paragraph describing the limitafions of the primary screen (only one cell line, single readout, 24h treatment, one concentrafion) would be valuable -The key quesfion of whether a redistribufion of ATG9A and DAGLB is disease-relevant with respect to the long distances in pafient neurons (distal axons) needs to be discussed -ll 547-548: "with or without hiPSC-derived astrocytes": Were astrocytes used?Not menfioned elsewhere.

Figure legends:
-l 849: "with hundreds of thousands to millions of cells per experiment" -statement not helpful

Reviewer #2 (Remarks to the Author):
This very large and detailed manuscript by Saffari et al uses a phenotypic screening approach to idenfify small molecule C01 as a promoter of appropriate ATG9A trafficking in cells harboring A4P mutafions established to cause spasfic paraplegia.Figures 1-5 document their high-throughput screen and the various secondary assays they use to priorifize C01 over their other hits.Figures 6-8 use transcriptomic and proteomic approaches to try to provide understanding of how C01 funcfions.

Key strengths:
The disease context is very interesfing.Addifionally, since many other neurodevelopmental disorders converge on dysfuncfional protein trafficking, the screening approach and C01/other hits may have broader significance than the sefting of spasfic paraplegia.
The screening and probe development efforts in Figs 1-6 are generally rigorous and use a strong mix of biological contexts including differenfiated neuron-like cells and, crifically, neurons derived from iPSCs obtained from pafients.The overall degree of difficulty in execufing this type of high-content screening approach is high, and being able to spotlight a molecule that works consistently across these various contexts is an accomplishment.
C01 itself is a key strength because, in contrast to many academic studies, this molecule has quite strong physicochemical properfies.C01 would be a strong starfing point for CNS applicafions in the eyes of most big pharma med chemists, and these properfies make this discovery more impacfful by opening greater future drug discovery possibilifies.
While some experiments described below may not be ideal imo, it's undeniable that all the data in the manuscript (and there are a lot) are presented clearly and that the utmost effort has been put forth to be transparent and rigorous about the data generated and presented (nofing especially the supporfing figures and data sets).

Perceived weaknesses:
The central weakness is that Figures 6-8 make liftle headway in understanding how C01 funcfions.Target idenfificafion for small molecules is always a challenge, and it's made substanfially more challenging here due to the subtlety of the phenotypes observed and the uM potency of C01; the degree of difficulty arguably couldn't be higher.That said, the transcriptomic analyses labor against the inifial conclusion that there isn't much of a transcriptomic signature for C01; ulfimately, no clear hypothesis emerges and no validafion experiments are performed.The proteomic analyses do idenfify changes after 72h of C01 treatment in RAB proteins that are associated with vesicular trafficking/autophagy.However it's not clear that these changes play a meaningful role in the phenotypes elicited by C01.C01 causes a large reducfion in RAB12 but CRISPR KO of RAB12 doesn't significantly alter ATG9A rafio; conversely C01 lowers ATG9A rafio but causes only a small (significant) reducfion in RAB3B levels.Cells lacking these RAB proteins remain as responsive to C01 as WT cells, strongly hinfing that pathways/targets beyond these RAB proteins are likely dominant for C01's mechanism.This crifique is not a request for addifional data, because the next steps toward target ID would be very challenging long term studies that are clearly beyond the scope of this story.However, I do think it's important to scale back text claims regarding establishing "molecular targets" (including in the abstract and last sentence of introducfion) since no targets of C01 are delineated here.
For validated hits (maybe the top 5 or top 17), I think it's necessary to show structures to enable replicafion or extension by other researchers.Eg others may be interested in disrupters of the TGN for other biological reasons and could be interested in tesfing F01 and H01.Apologies if I overlooked these among the many supporfing files.
I would recommend giving C01 a more formal or at least more descripfive name (eg HMS1234, or just a 6-digit number) because C01 just isn't descripfive enough.Future researchers may report on this cpd and/or future vendors may want to sell this cpd, and it just needs a less ambiguous name.

Reviewer #3 (Remarks to the Author):
Saffari et al. report the finding of a small molecule, C01, able to modulate the trafficking of ATG9A and DAGLB from the trans golgi network to other subcellular regions, as revealed by fluorescent imaging, in AP-4 deficient neurons from hereditary spasfic paraplegia pafients.Using transcriptomics and proteomics analysis, the authors were able to show that the RAB proteins, RAB12 and RAB3C play a role in the mechanism of acfion of C01.Both, ATG9A and DAGLB are cargo proteins of the adaptor protein AP4.AP4 mediates intracellular membrane trafficking and mutafions in the four AP4 subunits have been idenfified as causes of HPG with intellectual disability.Accumulafion of ATG9A, an autophagy protein, in the TGN has been previously suggested as a potenfial cause of the neuropathogenesis of AP4 deficiency.To date, there is a high unmet medical need for HPG.Therefore, finding novel molecules that can progress to the clinic or insights into the disease would be of high significance to the field.The paper by Saffari et al. is well wriften and the figures are rich in informafion but clear.The logic of the study is also clear.Most of the conclusions are overall well supported by the data.The authors use a challenging phenotypic assay to idenfify LMW compounds that are able to correct the phenotype of ATG9A accumulafion in TGN.From the primary hits obtained, most molecules induce toxicity or do not reproduce in other cellular models and only C01 meets the criteria for MoA studies, which the authors do very comprehensively.
However, I believe that in the current version, the work has important limitafions that lower its impact in the field: • The effect of C01 is also present in AP-4 WT cells indicafing that the MoA of C01 is independent of the disease genotype.It would help to know if unrelated targets of AP-4 protein are also redistributed after C01 treatment, as to have an indicafion of whether this pathway would offer a therapeufic window to be explored further in the disease setup.Similarly, the effect of C01 is the mislocalizafion of AMPA receptors, also reported to be involved in the neuropathology of the disease, would be of value • Proteomics and transcriptomics analysis could focus on the differences across WT and KO AP4 models, to invesfigate potenfial disease targets independently of a C01 effect • The molecular target of C01 remains unknown.RAB12 and RAB3C knockouts increase the effect of C01, which devalidate these proteins as the main targets leading to the ATG9A redistribufion phenotype • C01 is a small molecule poorly characterized for in vivo studies and with a weak potency in cells.In the paper there is not an analysis of the SAR in the library tested or explorafion of the chemical space that supports this chemical scaffold.Therefore, the statements made by the authors in the Abstract, Introducfion and Discussion secfion about the potenfial of C01 as IND and therapeufic for AP4-HSP are overstated Minor comments: • How many fibroblast donors were tested and how strong is the variafion of the phenotype • Unclear why the 61 hits showing toxicity in the primary screen were not removed from the primary selecfion • Figure 3b: the assay window looks befter in the neuronal model, why were fibroblast used instead for the primary screen?• The data showed in Figure 3k: there is a great variability and no dose response.I would not conclude the effect of B-01 is also observed for DAGLB

Reviewer #4 (Remarks to the Author):
Thank you for the opportunity to review this elegant manuscript.I congratulate the authors on their approach to idenfifying therapeufic candidates for rare neurological diseases, the ulfimate outcome of which likely will be therapeufics for rare diseases that have unanficipated spillover efficacy in more common neurological diseases.I have a few suggesfions about the biology and the chemistry presented in this paper.
I appreciate that the authors view C-01 as a tool compound to launch a future medicinal chemistry campaign; however, I think the paper is greatly diminished by no in vivo data.C-01 has properfies favorable for brain penetrafion.What peripheral dose achieves 5 uM (the EC50 in neuron derived iPSCs) in mouse brain extracellular fluid?What are the corresponding microscopic, transcriptomic, and proteomic changes?I think results from experiments like these will be necessary for readers to judge the translafional potenfial of the compelling cell culture results in the present version.
It appears that Astellas gave the invesfigators the library but did not disclose structures unfil after screening was completed and only for the hit(s).It would be very helpful to reveal more about how the set was assembled.Astellas should be able to provide more informafion on what was considered, e.g., Tanimoto scores, MW, Lipinski, etc.
General comment: Why not used as "counter-screen assay" as it seems that the separation between positive and negative control is much more robust?(perhaps due to the fact of comparing a full KO to WT instead of a het carrier to a hom patient).
Response: We understand the reviewer's point and agree that the separation of positive and negative controls is even better in SH-SY5Y cells, which is likely due to the homogeneity of this cell line, but also appears to be a feature of other neuronal lines, such as iPSC-derived neurons, perhaps reflecting the increased sensitivity of neurons to loss of AP-4.The screening approach was designed to use predominantly patient-derived cells.In the primary and secondary screen fibroblasts from an AP-4 patient, which are expected to cause a full loss-of-function, were compared to heterozygous controls, which are expected to behave like wildtypes, served as an easily obtainable and patient-relevant cell model (for a detailed assessment of the fibroblast lines compare Behne et al., Hum Mol Genet.2020, doi: 10.1093/hmg/ddz310).AP4B1 knockout SH-SY5Y cells, which are a more neuron-like but arguably less patient-relevant cell type, due to their homogeneity, were used in orthogonal assays and for exploration of potential mechanisms.We believe this approach and combination of multiple in vitro disease models provides key advantages over screens done in a single cell line of any type.
-The inclusion of the DAGLB assay is an important and careful validation.
Response: We thank the reviewer for this comment and agree that evaluating the translocation of a second AP-4 cargo protein adds an important validation step.
-Why was the "multiparametric morphological profiling approach" only performed on 5 compounds?Would it be possible to implement this analysis as a screening tool for future projects, or at least for the counter-screen?
Response: We thank the reviewer for pointing out the value of multiparametric analyses in cell-based assays.Since our goal was to identify compounds that re-distribute ATG9A, while maintaining acceptable cell counts, these two metrics were deemed sufficient as readouts to efficiently identify compounds of interest in the primary and counter-screens.The value of morphologic profiling was highest at the stage of having identified 5 lead compounds that all met predefined criteria.Here, a multiparametric approach helped us profile the impact on multiple cellular phenotypes/structures, beyond our primary readouts, and thus helped us triage compounds with potential off-target effects.For future projects of similar scope, multiparametric morphologic profiling can be implemented at any stage, including primary and counter screens, depending on the research question.To facilitate using this approach for future studies, we added a detailed description on how to implement morphologic profiles to the methods section: "The multiparametric morphological profiling strategy employed in this study was adapted from previously published protocols 34 .Single cell measurements of ninety distinct descriptors of shape and intensity for the nucleus (DAPI), the cytoskeleton and global cell morphology (anti-beta-Tubulin III), the TGN (anti-TGN46), and ATG9A vesicles (anti-ATG9A) were acquired and automatically extracted.Single cell data were summarized by computing per-image medians for each variable (Supplementary File 6).Next, a correlation matrix was generated using the Pearson correlation coefficient with complete pairwise observations.Variables with zero variance and observations with missing values were removed.Variables were then transformed to have a mean of zero and a standard deviation of one.Principal component analysis (PCA) was conducted to reduce dimensionality and cluster data based on their properties.To identify the contribution of individual features to the variance within the dataset, correlation analysis was performed between the first principal component, accounting for the majority of the variance within the dataset, and all extracted features.Features displaying a correlation coefficient > 0.75 were selected to define morphological profiles.Profiles were summarized using heatmap visualization.." (lines 615-628) Target deconvolution -RNA sequencing: It is unclear why SHSY-5Y cells were used instead of iPSC-derived neurons as compound C-01 showed only small effects on ATG9A ratios in previous analyses (Fig. 3d).
Response: We understand the reviewer's point.SH-SY5Y cells were used for RNA sequencing due to their genetic homogeneity, eliminating potential biases that arise from different genetic backgrounds.Since iPSC neurons were derived from a patient vs. the sex-matched parents, we anticipated that the transcriptional differences between these two individuals would bias our results.With respect to the transcriptional changes with BCH-HSP-C01 in SH-SY5Y cells, please refer to the next comment.
Response: We thank the reviewer for raising this point.We performed the transcriptomics experiments after 72h of compound treatment since the effect of BCH-HSP-C01 was found to be dose-and time-dependent (possibly due to half-life of proteins, resulting in greater turnover with prolonged treatment).To clarify this point, we added two new datasets and figures: 1 Discussion: -Expanding the paragraph describing the limitations of the primary screen (only one cell line, single readout, 24h treatment, one concentration) would be valuable.
Response: We thank the reviewer for this suggestion and refer to the data we added in response to reviewer 1, comment 2. We now show that BCH-HSP-C01 not only reduces the ATG9A ratio at the TGN, but also increases the amount of ATG9A puncta in neurites, supporting a disease-relevant redistribution of AP-4 cargo proteins.
-The key question of whether a redistribution of ATG9A and DAGLB is disease-relevant with respect to the long distances in patient neurons (distal axons) needs to be discussed Response: We thank the reviewer for this suggestion and refer to our response to reviewer 1, comment 2. At the suggestion of reviewer 1, we have completed additional analyses in iPSC-derived neurons from two individuals with AP-4-related hereditary spastic paraplegia and controls.We quantified the number of ATG9A puncta per neurite length using an automated image analysis pipeline in a high-throughput format.We find that AP-

Key strengths:
The disease context is very interesting.Additionally, since many other neurodevelopmental disorders converge on dysfunctional protein trafficking, the screening approach and C01/other hits may have broader significance than the setting of spastic paraplegia.
The screening and probe development efforts in Figs 1-6 are generally rigorous and use a strong mix of biological contexts including differentiated neuron-like cells and, critically, neurons derived from iPSCs obtained from patients.The overall degree of difficulty in executing this type of highcontent screening approach is high, and being able to spotlight a molecule that works consistently across these various contexts is an accomplishment.
C01 itself is a key strength because, in contrast to many academic studies, this molecule has quite strong physicochemical properties.C01 would be a strong starting point for CNS applications in the eyes of most big pharma med chemists, and these properties make this discovery more impactful by opening greater future drug discovery possibilities.
While some experiments described below may not be ideal imo, it's undeniable that all the data in the manuscript (and there are a lot) are presented clearly and that the utmost effort has been put forth to be transparent and rigorous about the data generated and presented (noting especially the supporting figures and data sets).
Response: We thank the reviewer for their positive assessment and for pointing out the strengths of our manuscript.
Perceived weaknesses: The central weakness is that Figures 6-8 make little headway in understanding how C01 functions.Target identification for small molecules is always a challenge, and it's made substantially more challenging here due to the subtlety of the phenotypes observed and the uM potency of C01; the degree of difficulty arguably couldn't be higher.That said, the transcriptomic analyses labor against the initial conclusion that there isn't much of a transcriptomic signature for C01; ultimately, no clear hypothesis emerges and no validation experiments are performed.The proteomic analyses do identify changes after 72h of C01 treatment in RAB proteins that are associated with vesicular trafficking/autophagy.However it's not clear that these changes play a meaningful role in the phenotypes elicited by C01.C01 causes a large reduction in RAB12 but CRISPR KO of RAB12 doesn't significantly alter ATG9A ratio; conversely C01 lowers ATG9A ratio but causes only a small (significant) reduction in RAB3B levels.Cells lacking these RAB proteins remain as responsive to C01 as WT cells, strongly hinting that pathways/targets beyond these RAB proteins are likely dominant for C01's mechanism.This critique is not a request for additional data, because the next steps toward target ID would be very challenging long term studies that are clearly beyond the scope of this story.However, I do think it's important to scale back text claims regarding establishing "molecular targets" (including in the abstract and last sentence of introduction) since no targets of C01 are delineated here.
Response: We acknowledge the reviewer's point and have scaled back our wording in the respective paragraphs.
For validated hits (maybe the top 5 or top 17), I think it's necessary to show structures to enable replication or extension by other researchers.Eg others may be interested in disrupters of the 5g: "Molecular Weight: 256.3 kDA" -cannot be correct, 256.3 DA (g/mol) -Fig.8l:Blots for RAB3C and RAB12 are missing Methods: ) A times series experiment with different concentrations of BCH-HSP-C01 showing that the maximum effect on ATG9A translocation is reached at 72h. 2) A dose response curve of BCH-HSP-C01 after 72h treatment showing a maximum reduction of the ATG9A ratio of around 12 SD compared to the negative control.Please refer to revised Figure 5g&h and the following new text passage: "To investigate the time-and dose-dependent effect of BCH-HSP-C01, we used AP4B1 KO SH-SY5Y cells to conduct time series experiments with different concentrations of BCH-HSP-C01 (Fig. 5 g,h, Supplementary File 6).All concentrations tested show a maximal effect on ATG9A translocation after 72-96h of treatment (Fig. 5g,h, Supplementary File 6), exceeding the effects seen after 24h (Fig. 3d)." (lines 201-205) -Autophagic flux: Interestingly, there is no baseline difference between AP4B1-KO and control SH-SY5Y cells (Fig. 8 j-l) Response: We agree with the reviewer that the lack of visible difference in autophagic flux at baseline seems unexpected at first.However, these findings by western blotting are consistent with our proteomics data from DMSO treated AP4B1 WT and AP4B1 KO SH-SY5Y cells that show no significant change in protein levels of LC3B.We acknowledge that this contrasts with previously published western blot and proteomics data from AP-4-KO HeLa cells (Davies, et al.Nat Commun.2018 Sep 27;9(1):3958.PMID: 30262884; Mattera et al.Proc Natl Acad Sci U S A. 2017 PMID: 29180427).However, our findings are in line with other published data on LC3 levels in AP-4-KO neurons, where no changes in basal levels of LC3B were detected (Fig. 7D-F, de Pace et al.PLoS Genet.2018 PMID: 29698489).Thus, the impact of AP-4 deficiency on LC3 levels appears to be cell-type specific.
4 deficient neurons have a decreased number of ATG9A puncta compared to controls.Treatment with BCH-HSP-C01 significantly increased the number of ATG9A puncta in neurites to levels similar to controls.This important finding and dataset have been added.Please refer to the revised version of Figure 5 (Panels k-m), the revised version of Supplementary file 7 as well as the following paragraph that has been added to the main text: "Prior work in neurons isolated from AP-4-deficient mice 13, 14 highlighted depletion of axonal ATG9A pools.In hiPSC-derived neurons from individuals with SPG50 and SPG47, we observed a reduction of ATG9A puncta density in neurites.BCH-HSP-C01 treatment for both 24h and 72h restored neurite ATG9A puncta density to levels similar to control (Fig. 5 k-m, Supplementary File 7)." (lines 219-222) Minor comments: Figures: -Fig.5g: "Molecular Weight: 256.3 kDA"cannot be correct, 256.3 DA (g/mol) Response: We thank the reviewer for catching this mistake.The information on compound structures and molecular weights of all five compounds tested in iPSC-derived neurons has been added to new Supplementary Fig. 4. Response: We thank the reviewer for this comment and have added panels for Rab12 and Rab3c to this figure.Please refer to the revised version of Fig. 8l and Supplementary Fig. 10d.

:
Figure legends:l 849: "with hundreds of thousands to millions of cells per experiment"statement not helpful