High-Content Small Molecule Screen Identifies a Novel Compound 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 novel therapeutic targets for rare diseases. In this study, we developed a high-throughput screening assay to identify molecules that correct aberrant protein trafficking in adaptor 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, C-01, 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 putative molecular targets of C-01 and potential mechanisms of action. 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 Investigational New Drug (IND)-enabling studies.


INTRODUCTION
Despite remarkable advances in our ability to delineate the genetic causes of rare neurological diseases, it is estimated that speci c therapies exist for less than 5% 1 . Thus, there is a signi cant unmet need for developing and implementing novel platforms for drug discovery. Informed by disease-relevant cellular phenotypes, automated and unbiased cell-based high-throughput small molecule screens have the potential to uncover novel therapeutic targets 2, 3, 4, 5, 6 .
Throughout the screen, assay performance was monitored using established quality control metrics for cell-based screens (Z' robust ≥ 0.3, strictly standardized median difference ≥ 3, and an inter-assay coe cient of variation ≤ 10%) 30,31,32 . All assay metrics were calculated for positive and negative controls of the same assay plate to avoid bias by inter-plate variability. Prede ned thresholds were met by all assay plates (Supplementary Fig. 1a and Supplementary File 2). The results of the primary screen are summarized in Fig. 1l and 1m, and the complete dataset is provided in Supplementary File 3.
Of the 28,838 compounds, 26 were excluded due to non-quanti able ATG9A signal, exceptionally low cell counts or imaging artifacts. The remaining 28,812 compounds were evaluated for changes in cell count and the ATG9A ratio. The vast majority (n = 26,961, 93.5%) did not show any signi cant reduction in the ATG9A ratio (de ned as a reduction by at least 3 SD). 1,435 (5.0%) compounds were excluded due to toxicity, de ned as a reduction in the mean cell count by at least 2 SD compared to the negative controls. Only a small subset of 503 compounds (1.7%) reduced the ATG9A ratio by 3 or more SD compared to negative controls (Fig. 1m). Of these, 61 (0.2%) also reduced cell counts, while the remaining 442 (1.5%) showed no toxicity.
In summary, from this high-throughput primary screen, 503 active compounds were identi ed and selected for further testing.
Counter-screen in broblasts from AP-4-HSP patients con rms 16 compounds that lead to a dosedependent redistribution of ATG9A To validate the 503 active compounds identi ed in the primary screen, compounds were retested for dose-dependency using an 11-point dose range (range: 40nM to 40µM) (Fig. 2a). Source data for the secondary screen are provided in Supplementary File 4. All concentrations were screened in biological duplicates and subjected to the same quality control metrics as in the primary screen (Supplementary Fig. 1b and Supplementary File 5). Similar to the results from the primary screen, ATG9A ratios for negative and positive controls showed a robust separation (LoF/LoF mean: 1.4 ± 0.07 (SD), n = 269 wells, vs. WT/LoF mean: 1.12 ± 0.02 (SD), n = 269 wells, Mann-Whitney U test, p < 0.0001, Fig. 2b). Activity in the secondary screen was de ned as the ability to reduce the ATG9A ratio by at least 3 SD in both replicates and at least 2 different concentrations, without exerting toxicity. 51 compounds (10.1%) met these a priori de ned criteria ( Supplementary Fig. 2a,b). After manually verifying image quality and validating doseresponse relationships, compounds were triaged ( Fig. 2a and Supplementary Fig. 2a,b). Seventeen compounds demonstrated a clear and reproducible dose-response relationship, without evidence of image artifacts or auto uorescence. The EC50 for most compounds were in the low micromolar range (median: 4.66µM, IQR: 8.63, Fig. 2). 34 compounds were found to carry auto uorescence or imaging artifacts and were thus excluded from further testing ( Supplementary Fig. 2b). One active compound was unavailable from the manufacturer and was removed.
In summary, a counter-screen in AP-4-de cient patient broblasts con rmed and established dosedependent effects on intracellular ATG9A distribution for 16 compounds (Fig. 2c).
Orthogonal assays in neuronal models of AP-4-de ciency con rm 5 active compounds To validate active compounds from the secondary screen in a human cell line with neuron-like properties, the ATG9A assay was optimized for neuroblastoma-derived SH-SY5Y cells following a 5-day neuronal differentiation protocol with retinoic acid 33 (Fig. 3a). SH-SY5Y cells with stable expression of a AP4B1targeting CRISPR/Cas9 machinery (AP4B1 KO ) 12 served as negative controls while AP4B1-wildtype (AP4B1 WT ) cells were used as positive controls. All 16 active compounds were tested in an 8-point dose range (50nM to 30µM) with a treatment duration of 24h. Quanti cation of the ATG9A ratio in differentiated SH-SY5Y cells showed a robust separation between control conditions (AP4B1 KO : 1.80 ± 0.06 (SD), n = 158 wells vs. AP4B1 WT : 1.17 ± 0.03 (SD), n = 160 wells, Mann-Whitney U test, p < 0.0001, Fig. 3b, Supplementary File 6). Compounds were evaluated based on their dose-dependent reduction of the ATG9A ratio and absence of cell toxicity. Eleven of 16 compounds were excluded due to lacking activity (n = 7), suspicion for artefacts or auto uorescence (n = 3), or obvious changes in cellular morphology (n = 1) ( Supplementary Fig. 3). Of the ve remaining compounds, three restored the ATG9A ratio to levels of wildtype controls (F-01, G-01 and H-01) while two compounds (B-01 and C-01) led to a reduction by at least 3 SD at higher concentrations ( Fig. 3c-h).
To assess whether these effects were speci c to ATG9A or similar effects were also present for other AP-4 cargo proteins, we turned to a second neuronal AP-4 cargo protein, DAGLB 29 . Similar to ATG9A, the DAGLB ratio (DAGLB uorescence intensity in the TGN vs. in the cytoplasm) showed a robust separation between AP4B1 WT and AP4B1 KO cells (AP4B1 KO : 1.80 ± 0.1 (SD), n = 192 wells vs. AP4B1 WT : 1.36 ± 0.07 (SD), n = 192 wells, Mann-Whitney U test, p < 0.0001, Fig. 3i, Supplementary File 6). All ve active compounds showed activity in the DAGLB assay, suggesting a broader effect on the tra cking of at least 2 AP-4 cargo proteins from the TGN. Again, F-01, G-01 and H-01 ( Fig. 3l-o) resulted in normalization of the intracellular DAGLB distribution, while B-01 and C-01 led to a moderate reduction of DAGLB ratios at higher concentrations (Fig. 3j,k,o).
Since small molecules can have pleotropic effects on cellular functions and organellar morphology, we adapted a multiparametric morphological pro ling approach 34 . Eighty-ve measurements of the nucleus,  cytoskeleton, global cell morphology, the TGN, and ATG9A vesicles were automatically computed for  each image, serving as a rich and unbiased source for interrogating biological perturbations induced by  compound treatment (Supplementary File 6). Principal component analysis was used to reduce dimensionality and cluster images based on their properties ( Fig. 4a and Supplementary Fig. 4). Positive and negative controls clustered closely together and were separated only by the ATG9A signal ( Fig. 4b and Supplementary Fig. 4a). B-01, C-01 and G-01 showed properties comparable to positive and negative controls, suggesting little off-target effects (Fig. 4b Supplementary Fig. 4b,c,e). F-01 and H-01, however, changed cellular morphology in a dose-dependent manner ( Fig. 4b and Supplementary Fig. 4d,f), with changes mainly driven by the rst principal component, accounting for 31.1% of the observed variance ( Fig. 4c). To decipher the phenotypic alterations responsible for these changes, the Pearson correlation coe cients of the rst principal component with each measurement were calculated (Fig. 4d). Features with a correlation coe cient > 0.75 were selected to de ne morphological pro les (Fig. 4e). Interestingly, TGN uorescence intensity and morphology seemed to be the most signi cant drivers for the separation, suggesting that disruption of TGN integrity potentially biased the assessment of ATG9A ratios in cells treated with compounds F-01 and H-01 ( Fig. 4b and Supplementary Fig. 4d,f).
Following these analyses, TGN uorescence intensity and morphological measures such as TGN area and elongation, as well as compactness and roughness, as indicators of the complexity of the TGN, were quanti ed for cells treated with all ve active compounds (Fig. 4f,g). While C-01 showed stable TGN signal and morphology across all assessed measurements, the other compounds depicted some degree of change. Again, F-01 and H-01 seemed to result in TGN changes in a dose-dependent manner while B-01 and G-01 led to only moderate alterations (Fig. 4f,g). Of note, these changes to TGN morphology were not detectable by visual inspection but only delineated through an automated analysis of ~ 600 images containing ~ 30,000 cells per group, showcasing the power of our automated, unbiased, high-throughput platform.
C-01 restores ATG9A and DAGLB tra cking in hiPSC-derived neurons from AP-4-HSP patients Informed by the ndings in differentiated AP4B1 KO SH-SY5Y cells, we next investigated whether these results would translate to human neurons. hiPSCs from patients with AP-4-HSP due to biallelic loss-offunction variants in AP4M1 (NM_004722.4: c.916C > T (p.Arg306Ter) / c.694dupG (p.Glu232GlyfsTer21)) and AP4B1 (NM_001253852.3: c.1160_1161del (p.Thr387ArgfsTer30) / c.1345A > T (p.Arg449Ter)) were generated 35,36 and differentiated into glutamatergic cortical neurons using established protocols 15,37,38 . hiPSC-derived neurons from sex-matched parents (unaffected heterozygous carriers) served as controls ( Fig. 5a and Supplementary File 7). Baseline quanti cation of ATG9A ratios in DIV (day in vitro) 14 neurons treated with vehicle for 24h showed robust separation between patient and control lines, exceeding the differences observed in AP-4-de cient broblasts and differentiated SH-SY5Y cells (SPG50 patient mean: 4.31 ± 0.4 (SD), n = 60 wells vs. heterozygous control: 1.56 ± 0.12 (SD), n = 60 wells, Mann-Whitney U test, p < 0.0001, Fig. 5b). Neurons were treated for 24h in 8-point dose titration experiments. B-01 and G-01 lacked activity on the ATG9A ratio and were thus excluded (Fig. 5d). C-01, F-01 and H-01, by contrast, showed a robust reduction in the ATG9A ratio (Fig. 5e,f). A multiparametric analysis showed that, similar to observations in AP4B1 KO SH-SY5Y cells, only C-01 preserved TGN integrity (Fig. 5f), while F-01 and H-01 impacted TGN morphology, suggesting off-target effects (Fig. 5e). Based on its favorable pro le, C-01 was selected as a lead compound and was re-synthesized for further testing (Fig. 5g). Prolonged treatment of C-01 for 72h to test for ATG9A and DAGLB translocation, demonstrated that C-01 was able to restore ratios of both AP-4 cargo proteins to levels close to controls with an EC50 of ~ 5µM, while maintaining a favorable pro le (Fig. 5h, Supplementary File 7). This greater effect on ATG9A distribution, compared to the ~ 50% reduction of the ATG9A ratio at 24h treatment, suggests a time-and dose-dependent effect. C-01 changed the ATG9A ratio through simultaneously decreasing ATG9A intensities inside the TGN and increasing cytoplasmic ATG9A levels, suggesting ATG9A translocation as the most likely mechanism of action. No changes in TGN morphology or any other cellular measurements were observed, indicating overall preservation of cellular morphology and little off-target effects. A similar pattern was observed with respect to DAGLB translocation (Fig. 5h). These ndings were con rmed in a second set of experiments in hiPSC-derived neurons from a patient with SPG47 ( Fig. 5i, Supplementary File 7), demonstrating that ndings extend to other forms of AP-4-de ciency.
Taken together, C-01 emerged as a robust modulator of ATG9A and DAGLB tra cking in human neurons from patients with AP-4 de ciency.
Target deconvolution using transcriptomic and proteomic analyses delineates putative mechanisms of action for C-01 To explore potential mechanisms of action of C-01 in an unbiased manner, we used a multi-omics approach, combining bulk RNA sequencing and unbiased label-free quantitative proteomics (source data are provided in Supplementary Files 8-10).
First, RNA sequencing was conducted in differentiated AP4B1 WT and AP4B1 KO SH-SY5Y cells treated for 72h with either vehicle or compound C-01 (5µM, Supplementary File 8). Analysis of differential gene expression identi ed few signi cant transcriptional changes in response to C-01 treatment, suggesting that this compound does not elicit major alterations in gene expression or induce many off-target effects ( Supplementary Fig. 5). Since changes in gene expression caused by short-duration small molecule treatments might not reach prede ned cut-offs for standard differential expression analyses, and because compounds might affect groups of genes in shared pathways rather that modifying single target genes, we adapted an unbiased and unsupervised network approach to identify groups of co-expressed genes. Hierarchical clustering of samples showed that treatment with C-01, regardless of cell line, was the main differentiator in our dataset (Fig. 6a). To identify the gene networks responsible for these changes, weighted gene co-expression network analysis (WGCNA) 39,40 was used to group the 18,506 expressed genes into 36 co-expression modules (Fig. 6b). Gene expression pro les within each module were summarized using the "module eigengene" (ME), de ned as the rst principal component (PC) of a module 41 . Within each module, the association of MEs with measured traits were examined by correlation analysis (Fig. 6c). Eight modules that showed an absolute correlation coe cient > 0.5 were selected for further evaluation. For these selected modules, ME based connectivity was determined for every gene by calculating the absolute value of the Pearson correlation between the expression of the gene and the respective ME, producing a quantitative measure of module membership (MM). Similarly, the correlation of individual genes with C-01 treatment was computed, de ning gene signi cance (GS) for C-01. Using the GS and MM, an intramodular analysis was performed, allowing identi cation of genes that have high signi cance with treatment as well as high connectivity to their modules (Fig. 6d). Five modules were signi cantly related to C-01 treatment, de ned as showing an absolute correlation coe cient between MM and GS > 0.5 (Fig. 6e). A list of the genes contained in each module along with their module membership is provided in Supplementary File 9. To summarize the biological information contained in these modules of interest, gene ontology (GO) analysis was performed, which demonstrated enrichment in biological pathways in three out of the ve assessed modules (Fig. 6f). The 'blue module' showed down-regulation of pathways involved in axonogenesis, actin lament organization and proteasome-mediated pathways. The 'light-yellow module' contained genes involved in ER stress response, amino acid metabolism and transcription. Finally, the 'mediumpurple3 module' depicted upregulation of genes involved in vesicular transport, particularly involving TGN and ER-associated transport, as well as membrane and vesicle dynamics. This last module showed the highest gene ratios (de ned as the percentage of total differentially expressed genes in the given GO term) and lowest Pvalues of all differentially regulated pathways across all modules, suggesting the upregulation of alternative vesicle mediated transport mechanisms by compound C-01 (Fig. 6f).
To assess whether similar themes would emerge on the protein level, we next used unbiased quantitative proteomics in both differentiated SH-SY5Y cells (AP4B1 KO and AP4B1 WT ) and hiPSC-derived neurons (patient with AP4B1-associated SPG47 and control) treated for 72h with either vehicle or compound C-01 (5µM). After quality ltering, 8,141 unique proteins in SH-SY5Y cells and 7,386 unique proteins in hiPSCderived neurons were quanti ed. Differential enrichment analyses for both cell lines are shown in Fig. 7a,b, and source data are provided in Supplementary File 10. As expected, baseline quanti cation of differentially expressed proteins in AP4B1 KO SH-SY5Y cells showed downregulation of AP-4 subunits, AP4B1, AP4E1 and AP4M1, and increased ATG9A levels, as reported in other models of AP-4 de ciency 11,12,13 (Supplementary Fig. 6a). PCA analysis of SH-SY5Y cells demonstrated 4 distinct clusters separated by C-01 treatment (PC1, explaining 12.3% of variance) and genotype (PC2, explaining 8.7% of variance) (Fig. 7a). Testing of vehicle vs. C-01 treated cells showed broadly similar groups of dysregulated proteins in AP4B1 WT and AP4B1 KO SH-SY5Y cells ( Supplementary Fig. 6b-d), suggesting a conserved mechanism of action independent of genotype, which allowed the pooling of cell lines to increase the power of the analysis (Fig. 7a). Similar observations were made for hiPSC-derived neurons ( Fig. 7b and Supplementary  Fig. 6e-h). Here, cell lines were a stronger discriminator, likely due to heterogeneity of the positive and negative controls, as expected in cell lines derived from different individuals. Again, differentially enriched proteins following C-01 treatment in hiPSC-neurons showed a high degree of similarity between patient and control lines ( Supplementary Fig. 6f-h), allowing a combined analysis (Fig. 7b).
Despite the heterogeneity in the neuronal samples, signi cant overlap was observed between the differentially enriched proteins in SH-SY5Y cells and hiPSC-derived neurons. Data sets were thus integrated for a combined analysis, which detected several proteins that were dysregulated across all cell types and genotypes (Supplementary Fig. 6i-l), providing strong evidence that these changes were related to treatment with C-01 (Fig. 7c). Consistent with the overall changes in gene expression, pathway enrichment analysis using the Reactome database 42 highlighted engagement of intracellular tra cking pathways as a potential mechanism of action for C-01 (Fig. 7c). Speci cally, modulation of RAB proteins involved in vesicle transport emerged as a consistent theme across cell types and genotypes, with the strongest evidence for the upregulation of RAB1B and downregulation of RAB3C and RAB12. Notably, while C-01 led to a signi cant change in protein levels of all three RAB protein family members in SH-SY5Y cells, only RAB3C and RAB12 reached signi cance in neurons (Fig. 7d). This overall pattern of RAB protein modulation was further supported by upregulation of the RAB protein geranylgeranyltransferase components A1 (CHM) in SH-SY5Y cells and A2 (CHML) in both SH-SY5Y cells and neurons. CHM and CHML play a vital role for tethering RAB proteins to intracellular membranes 43,44 . Additionally, upregulation of transferrin receptor protein 1 (TFRC) was observed (Fig. 7c), consistent with prior reports showing that reduction of RAB12 associates with increased protein levels of TFRC 45 . Collectively, these ndings suggest a potential role of RAB proteins in regulating vesicle transport in response to C-01 treatment.
RAB3C and RAB12 knockout are involved in C-01 -mediated vesicle tra cking and autophagy RAB3C and RAB12 displayed the strongest and most consistent protein expression changes in both differentiated SH-SY5Y cells and hiPSC-derived neurons following treatment with C-01 ( Fig. 7d) and were therefore selected for further investigation. Correlation analysis revealed a strong correlation (r = 0.93) between the LFQ intensities of these two proteins in both cell types and across different genotypes in response to C-01 (Fig. 7e).
To assess whether a correlation was also present on the transcriptional level, mRNA levels of RAB3C and RAB12 in response to C-01 treatment were analyzed in AP4B1 WT and AP4B1 KO SH-SY5Y cells. While there was a trend toward a reduction of RAB3C and elevation of RAB12 mRNA levels and correlation analysis demonstrated a moderate inverse correlation, none of these changes reached statistical signi cance ( Supplementary Fig. 7). These ndings suggest that RAB3C and RAB12 levels are altered through a posttranscriptional mechanism following treatment with C-01.
To investigate the potential impact of RAB3C and RAB12 on ATG9A translocation in the AP-4-de cient background, we used CRISPR/Cas9-mediated knockouts of RAB3C and RAB12 in AP4B1 KO SH-SY5Y cells (Fig. 8a,b, Supplementary Fig. 8 and Supplementary File 11). We found that knockout of RAB12 did not affect ATG9A translocation, while knockout of RAB3C caused a moderate reduction in the ATG9A ratio ( Fig. 8a). Combined knockout of RAB3C and RAB12 in AP4B1 KO SH-SY5Y cells did not show an additive effect. Interestingly, however, the effects of C-01 on ATG9A translocation were signi cantly enhanced by knockout of RAB3C, but not RAB12 alone. Combined knockout of both genes further augmented the effect of C-01. These ndings suggest that both RAB3C alone, or in combination with RAB12, play a role in C-01-mediated ATG9A redistribution.
A converging theme of ATG9A translocation and alteration of RAB protein expression is autophagy. RAB proteins are known modulators of autophagy with key functions in various steps of the pathway 46, 47 . ATG9A, a core autophagy protein, acts as a lipid scramblase and promotes autophagosome formation and elongation 48, 49, 50, 51 . To investigate whether C-01 leads to changes in autophagic ux, AP4B1 WT and AP4B1 KO SH-SY5Y cells were treated with C-01 for 72h and LC3-I to LC3-II conversion was measured by western blotting (Fig. 8c-f and Supplementary Fig. 8a). Levels of LC3-II were signi cantly elevated in all cell lines treated with C-01, suggesting modulation of the autophagy pathway. Co-treatment with ba lomycin A1, which blocks autophagosome-lysosome fusion, led to further LC3-II accumulation, indicating that C-01 increases autophagic ux ( Fig. 8c-f). Blocking the late stages of the autophagy pathway, with either ba lomycin A1 or chloroquine, reversed the effect of C-01 on ATG9A translocation in a dose-dependent manner, suggesting that this process requires intact autophagic ux ( Fig. 8g-i).
Next, since our data suggested a contribution of RAB3C and RAB12 to the effect of C-01, we investigated the impact of RAB3C and RAB12 knockout in AP4B1 KO SH-SY5Y cells with and without C-01 treatment (Fig. 8j-l and Supplementary Fig. 8b-d). Neither RAB3C nor RAB12 knockout alone led to major changes in baseline or C-01-enhanced autophagic ux (Fig. 8j,k). However, combined knockout of RAB3C and RAB12 signi cantly increased the ratio of LC3-II to LC3-I by approximately 36% (Fig. 8l). Upon treatment with ba lomycin A1, both RAB3C knockout alone and combined knockout of RAB3C and RAB12 further increased C-01-mediated LC3-I to LC3-II conversion (Fig. 8j-l). These ndings suggest the possibility that RAB3C and RAB12 modulate C-01-mediated ATG9A tra cking and subsequent autophagy induction.

DISCUSSION
Identi cation of novel therapeutic targets for rare neurological diseases represents a major scienti c and public health challenge 1,4 . The increasing number of rare genetic diseases 52 , the rising rate of diagnoses 53 , and the signi cant burden for patients 54,55 , caregivers 56 and health care systems 57 highlight the urgent need for translational research that moves beyond gene discovery to the identi cation of disease mechanisms and therapies. Unbiased high-content small molecule screens are a platform for drug-repurposing approaches and a starting point for the rationale development of new compounds 1, 2, 3, 4, 5, 6 . Disease-relevant 'screenable' phenotypes across cellular models, including patientderived cells, provide an entry point into developing automated, high-content screening and analysis platforms.
In this study, we develop the rst high-throughput cell-based phenotypic screening platform for a prototypical form of childhood-onset HSP caused by defective protein tra cking. Our platform allows us to determine the subcellular localization of the AP-4 cargo protein ATG9A in several cellular models of AP-4-de ciency. The hypothesis that ATG9A mislocalization is a key mechanism in the pathogenesis of AP-4-HSP is supported by the independent work of the Robinson 12 , Kittler 14 and Bonifacino 11, 13, 58 groups, in addition to our own work 15,21,24,25 , and by the overlapping phenotypes of AP-4 13, 14, 26 and Atg9a 28 knockout mice.
ATG9A is the only conserved autophagy-related transmembrane protein 50 and in mammalian cells cycles between the TGN and ATG9A vesicles, which associate with endosomes 59  Having established a robust and dynamic assay that reliably measures intracellular ATG9A distribution, we systematically screened a large library of 28,864 novel small molecules for their ability to restore ATG9A tra cking from the TGN to the cytoplasm. Following this primary screen, a counter-screen and a series of orthogonal experiments identi ed a novel small molecule, termed C-01, that can restore the intracellular distribution of ATG9A and a second transmembrane AP-4 cargo protein, DAGLB, in neuronal models of AP-4 de ciency, including iPSC-derived neurons from two patients with AP-4-HSP.
Compound C-01 has physicochemical properties that are within the parameters that are optimal for CNS drugs 71 and therefore represents a strong candidate for an in vivo tool compound. In addition, the low molecular weight and topological polar surface area create opportunities for compound optimization. Since the molecular targets of C-01 are unknown, we employed a target deconvolution strategy using transcriptomics and proteomics to de ne the cellular pathways impacted by this novel small molecule. This approach identi ed two central themes: 1) modulation of Golgi dynamics and vesicular tra cking, and 2) engagement of autophagy. At the core of the putative pathways affected by C-01, we identi ed the Rab proteins RAB1B, RAB3C and RAB12, as well as the interacting Rab geranyl transferase subunits CHM and CHML. RAB3C and RAB12 showed the strongest and most consistent association with C-01 treatment in both SH-SY5Y cells and iPSC-derived neurons, and our analyses suggest that these two proteins are involved in C-01-mediated redistribution of ATG9A from the TGN and increase of autophagic ux.
Rab proteins comprise a large family of small guanosine triphosphate (GTP) binding proteins that act as key regulators of intracellular membrane tra cking in eukaryotic cells at several stages, including cytoplasmic cargo sorting, vesicle budding, docking, fusion and membrane organization 72,73 . Rab GTPases function both as soluble and speci cally localized, integral-membrane proteins, the latter being mediated by prenylation. Among the roughly 70 known Rab proteins, more than 20 are primarily associated with the TGN, where they regulate Golgi organization, coordinate vesicle tra cking and interact with various steps of the autophagy pathway 46, 47 .
Following treatment with C-01, the RAB protein family members RAB3C and RAB12 were consistently downregulated in both SH-SY5Y cells and iPSC-derived neurons. Knockout experiments of these two proteins revealed that their loss potentiates C-01-mediated ATG9A translocation and autophagic ux. RAB3C, which is part of the RAB3 superfamily, is primarily expressed in brain and endocrine tissues, where it localizes to the Golgi and synaptic vesicles and is involved in exocytosis and modulation of neurotransmitter release 74 . RAB12 is mainly localized to recycling endosomes where it regulates endosomal tra cking and lysosomal degradation and has been identi ed as a modulator of autophagy 75 . A well-known downstream target of RAB12 is the transferrin receptor (TfR). Knockdown of RAB12 in mouse embryonic broblasts increases TfR protein levels, while overexpression leads to its reduction 45 .
In line with this, we nd that treatment with C-01 reduced RAB12 protein levels while, at the same time, robustly elevating transferrin receptor protein 1 (TFRC). To the best of our knowledge, no interaction between RAB3C and RAB12 has been described so far, however, our data suggest that both proteins are involved in C-01-mediated modulation of vesicle tra cking and autophagic ux.
Our study has identi ed the rst candidate small molecule drug capable of restoring protein mislocalization in AP-4-de cient cells, including human neurons from patients. We acknowledge several limitations of our approach, some of which are inherent to high-throughput screens and some that are speci c to our assay. First, as ATG9A mislocalization is a cellular phenotype of AP-4 de ciency conserved in non-neuronal and neuronal cells both in vitro 11,12,13,14,15,25,76 and in vivo 13,14,27 , we decided to conduct the initial screen in patient-derived broblasts, as a simple cellular model of AP-4 de ciency.
While the use of patient broblasts in the primary screen increases translational relevance, compounds that would have the capacity to correct ATG9A tra cking exclusively in neuronal cells could be missed at this stage. We determined that this risk was outweighed by the bene ts of a robust assay performance and the fact that mechanisms of AP-4-mediated protein tra cking are conserved across tissues and cell types 11,12,13,14,15,35,76 . Second, even though cell-based disease models can, to some extent, mimic the complexity of therapeutic responses in biological systems, the translation to in vivo models is often challenging, particularly for neurodevelopmental and neurodegenerative disease. Considerations such as a lead compound's ability to cross the blood-brain-barrier, target engagement in the central nervous system, therapeutic responses in complex neuronal networks relying on interactions with glia cells, developmental windows amenable to therapy, as well as in vivo off-target effects and toxicity must be considered and explored in future studies. To mitigate some of these risks, we employed unbiased multiparametric pro ling of C-01 which suggested little off-target effects. Future studies are required to exclude pleiotropic effects or off-target toxicity in different cell types or tissues in vivo. Lastly, while C-01 leads to a redistribution of two well-established AP-4 cargo proteins, ATG9A and DAGLB, we are unable to exclude the possibility that other neuron-speci c cargos of AP-4 exist and are important for the pathogenesis of AP-4-HSP. Nonetheless, mislocalization of both proteins is proposed as the major contributor to neuronal pathology caused by AP-4 de ciency, through dysregulated autophagy and endocannabinoid signaling, respectively 11,12,13,14,29 . Our automated high-throughput platform would allow for the rapid interrogation of additional AP-4 cargo proteins in the future.
In conclusion, our ndings provide a solid foundation for lead optimization of C-01 and development in Investigational New Drug (IND)-enabling studies. More broadly, our approach illustrates the development of a small molecule screening platform for a rare neurogenetic disease, leveraging robust cellular phenotypes. We hope this approach will create a paradigm for other rare and more common disorders of protein tra cking. The increase of autophagic ux through C-01 offers the intriguing possibility that this compound could be considered for the treatment of other autophagy-associated diseases.

METHODS
Clinical data from patients with AP-4-HSP

Antibodies and reagents
The
For the generation of glutamatergic neurons, NGN2 transduced hiPSCs were dissociated into single cells using accutase and seeded onto geltrex-coated plates. The following day, NGN2 expression was induced using doxycycline and selected with puromycin. Growth factors BDNF (10 ng/ml, Peprotech, Cat#450- TGN46-positive area). ATG9A uorescence intensity was measured in both compartments in each cell and the ATG9A ratio was calculated by dividing the ATG9A uorescence intensity the TGN by the ATG9A uorescence intensity in the remaining cell body (Fig. 1b): Additional masks for the TGN used for morphologic pro ling included TGN Roughness (shape factor in the MetaXpress software) and the following calculated metrics: Z'-factor robust values and strictly standardized median difference (SSMD) 30 were calculated for each plate and only plates that met the prede ned quality metrics of a Z'-factor robust ≥ 0.3 and SSMD ≥ 3 were included in subsequent analyses.

Western blotting
Western blotting was done as previously described 70  McCarthy and Smyth 78 , implemented in the edgeR package in R. Raw counts were obtained using STAR and low expressed genes were excluded using the method described by Chen et al. 79 . Expression data were normalized using the Trimmed Mean of M-values method implemented in the edgeR package. Genes were considered as differentially expressed according to default options with a false discovery rate (Benjamini-Hochberg procedure) < 0.05 and a log 2 fold change of > 0.3. Gene ontology (GO) enrichment analysis was done using clusterPro ler 80 . Pathways were considered differentially expressed with an FDR < 0.05.

Network connectivity analysis
To identify transcriptional changes in co-expressed groups of genes following compound treatment, a weighted gene co-expression network analysis (WGCNA) was performed.
Raw counts were generated, and low expressed genes were removed as described above. Data were normalized using variance stabilizing transformation as described by Anders et al. 81 . Batch effects were removed using the limma package in R 82 . Preprocessed data were then analyzed using the WGCNA package in R 83, 84 . In brief, pairwise Pearson correlations were calculated between all genes and genes with a positive correlation were selected to form a "directed" correlation matrix. Next, the correlations were raised to a power to approximate a scale free network. The adequate power was chosen based on soft thresholding aiming for a high scale independence above 0.8 by keeping a mean connectivity between 200 and 500. Genes were then grouped based on topological overlap and clusters were isolated using hierarchical clustering and adaptive branch pruning of the hierarchical cluster dendrogram, giving rise to groups of co-expressed genes, so called modules. Gene expression pro les within each module were summarized using the "module eigengene" (ME), de ned as the rst principal component of a module.
Within each module, association of MEs with measured clinical traits was examined by correlation analysis. For these selected modules, module eigengene based connectivity was determined for every gene by calculating the absolute value of the Pearson correlation between the expression of the gene and the respective ME, producing a quantitative measure of module membership (MM). Similarly, the correlation of individual genes with the trait of interest was computed, de ning gene signi cance (GS).
Using the GS and MM, an intramodular analysis was performed, allowing identi cation of genes that have high signi cance with treatment as well as high connectivity to their modules. The biological information contained in modules of interest was summarized with gene ontology (GO) enrichment analysis using clusterPro ler 80 . Pathways were considered differentially expressed with an FDR < 0.05.
Sample preparation for mass spectrometry Cells were lysed for whole proteome analysis in RIPA lysis buffer (Thermo Fisher Scienti c, Cat# 89900) supplemented with cOmplete protease inhibitor (Roche Cat# 04693124001) and PhosSTOP phosphatase inhibitor (Roche Cat# 4906845001) and sonicated in a Bioruptor® Pico Sonication System (one single 30 seconds on/off cycle at 4°C). Protein concentrations were determined using a Pierce BCA Protein Assay Kit (Thermo Fisher Scienti c Cat# 23225). Lysates were stored at -80° C until further processing. To generate peptide samples for analysis by mass spectrometry, 30-50µg protein were precipitated by overnight incubation in 5 volumes of ice-cold acetone at − 20° C and pelleted by centrifugation at 10,000×g for 5 min at 4° C. All subsequent steps were performed at room temperature. Precipitated protein pellets were air-dried, resuspended for denaturation and reduction in digestion buffer (50 mM Tris pH 8.3, 8M Urea, 1 mM dithiothreitol (DTT)) and incubated for 15 min. Proteins were alkylated by addition of 5 mM iodoacetamide for 20 min in the dark. Following reduction and alkylation, proteins were enzymatically digested by addition of LysC (1 µg per 50 µg of protein; Wako, Cat# 129-02541) for an overnight incubation. Samples were then diluted four-fold with 50 mM Tris pH 8.3 before addition of Trypsin (1µg per 50µg of protein; Sigma-Aldrich, Cat# T6567) for 3 hours. The digestion reaction was stopped by addition of 1% (v/v) tri uoroacetic acid (TFA) and samples were incubated on ice for 5min to precipitate contaminants, which were pelleted by centrifugation at 10,000×g for 5min. Acidi ed peptides were transferred to new tubes, before puri cation by solid-phase extraction using (v/v) acetonitrile), and stored at − 20° C until analysis by mass spectrometry.

Mass spectrometry
Mass spectrometry was performed on an Exploris 480 mass spectrometer coupled online to an EASY-nLC 1200, via a nano-electrospray ion source (all from Thermo Fisher Scienti c). Per sample, 250 ng of peptides were loaded on a 50 cm by 75µm inner diameter column, packed in-house with ReproSil-Pur C18-AQ 1.9 µm silica beads (Dr Maisch GmbH). The column was operated at 50° C using an in-house manufactured oven. Peptides were separated at a constant ow rate of 300nL/min using a linear 110min gradient employing a binary buffer system consisting of Buffer A (0.1% [v/v] formic acid) and Buffer B (80% acetonitrile, 0.1% [v/v] formic acid). The gradient ran from 5 to 30% B in 84min, followed by an increase to 60% B in 8min, a further increase to 95% B in 4min, a constant phase at 95% B for 4min, and then a washout decreasing to 5% B in 5min, before re-equilibration at 5% B for 5min. The www.uniprot.org). Label-free quanti cation was performed using the MaxLFQ algorithm 88 . Matching between runs was enabled to match between equivalent and adjacent peptide fractions, within replicates. LFQ minimum ratio count was set to 1 and default parameters were used for all other settings. All downstream analyses were performed on the 'protein groups' le output from MaxQuant.
Proteomic downstream data analysis Differential enrichment analysis of proteomics data was done using the DEP package in R. Preprocessing and quality ltering was performed separately for SH-SY5Y cells and hiPSC-derived neurons. Proteins that were only identi ed by a modi cation site, or matched the reversed part of the decoy database, as well as commonly occurring contaminants were removed. Duplicate proteins were removed based on the corresponding gene names by keeping those with the highest total MS/MS count across all samples. All following steps were done separately for each cell type (SH-SY5Y cells ( Fig. 7a and Supplementary  Fig. 6a-d) and hiPSC-derived neurons ( Fig. 7b and Supplementary Fig. 6e-h) and for the pooled dataset ( Fig. 7c and Supplementary Fig. 6i-l).   set (70% of data) and a test set (30%). A generalized linear model was trained using the training set. The performance of the prediction model using the test set is shown in (k). The AUC is 0.96. (l) Impact of 28,864 compounds applied for 24h at a concentration 10µM. Z-scores for the primary metric, the ATG9A ratio, are shown. All data points represent per well means. The mean of the positive control (WT/LoF) is shown as a green dotted line. The green shaded areas represent ± 1 SD. Active compounds were a priori de ned as those reducing the ATG9A ratio by at least 3 SD compared to negative controls. Toxicity was de ned as a reduction of cell count of at least 2SD compared to the negative control. 501 compounds show activity by reducing the ATG9A ratio by more than 3 SD. (m) Distribution of Z-scores of all non-toxic Counter-screen in broblasts from AP-4-HSP patients con rms 16 compounds that lead to dosedependent redistribution of ATG9A.
(a) Overview of the counter-screen of the 503 active compounds identi ed in the primary screen. To assess for dose-dependent effects, compounds were screened in AP-4-HSP patient-derived broblasts in 384-well microplates using 11-point titrations ranging from 40nM to 40µM. All concentrations were screened in duplicates. Active compounds were a priori de ned as those reducing the ATG9A ratio by at least 3SD compared to negative controls, in more than one concentration. Toxicity was de ned as a reduction of the cell count of at least 2 SD compared to negative controls. 51 compounds demonstrated a clear and reproducible dose-response relationship and raised no suspicion for auto uorescence on automated and manual review. 34 compounds showed auto uorescence or resulted in imaging artifacts. One active compound was unavailable from the manufacturer and was therefore excluded from subsequent testing. (b) Baseline differences in the ATG9A distribution in WT/LoF (n=269) vs. LoF/LoF  Orthogonal assays in AP4B1 KO SH-SY5Y cells con rm 5 active compounds. (a) Overview of the orthogonal screen of 16 active compounds in differentiated AP4B1 KO SH-SY5Y cells, a neuronal model of AP-4 de ciency. Active compounds were a priori de ned as those reducing the ATG9A ratio by at least 3 SD compared to negative controls. Toxicity was de ned as a reduction of cell count of at least 2 SD compared to the negative control. (b) Baseline differences in ATG9A ratios of AP4B1 WT vs. AP4B1 KO SH-SY5Y cells were quanti ed from 160 AB4B1 WT and 158 AB4B1 KO wells from 5 assay plates. Statistical testing was performed using the Mann-Whitney U test. Positive and negative controls showed a robust separation (p < 0.0001). (c-g) Dose-response curves for ATG9A ratios in AB4B1 KO cells treated with different compounds. Data points represent per-well means from 3 different assay plates. Dashed lines show mean Z-scores for positive (green) and negative (salmon) controls.
Shaded areas represent ± 1 SD. (h) Representative images of the intracellular ATG9A distribution for individual compounds. The merged image shows beta-3 tubulin (grey), DAPI (blue), the TGN (red) and ATG9A (green). The TGN and ATG9A channels are further separately depicted in greyscale. Scale bar: 10µm. (i) Baseline differences of DAGLB ratios in AP4B1 WT vs. AP4B1 KO   Compound C-01 restores ATG9A and DAGLB tra cking in iPSC-derived neurons from AP-4-HSP patients.
(a) Overview of the testing of 5 active compounds in iPSC-derived cortical neurons from a patient with AP4M1-associated SPG50 compared to heterozygous controls (same-sex parent). Active compounds were de ned as those reducing the ATG9A ratio by at least 3 SD compared to negative controls (patient- Figure 6 Target deconvolution using bulk RNA sequencing and weighted gene co-expression network analysis in AP4B1 KO SH-SY5Y cells treated with C-01.  Target deconvolution using unbiased quantitative proteomics in AP4B1 KO SH-SY5Y cells and AP-4-HSP patient-derived iPSC-neurons treated with C-01. (a -c) Differential protein enrichment analysis. Statistical testing was done using protein-wise linear models and empirical Bayes statistics. Proteins were considered as differentially enriched with a false discovery rate of < 0.05 and a log 2 fold change > 0.3. (a) SH-SY5Y cells: 8141 unique proteins were analyzed. PCA of the top 500 variable proteins shows robust separation between experimental conditions. The volcano plot summarizes differential protein enrichment for AP4B1 WT and AP4B1 KO cells pooled into two groups, vehicle vs. C-01 treated. Differentially enriched proteins are depicted in black. Proteins with the most consistent enrichment pro les across all experimental conditions (see Supplementary Fig. 6a-d)   (a) AP4B1 KO SH-SY5Y cells were transfected for 72h with RNPs targeting RAB3C, RAB12 or both compared to NLRP5 as a non-essential control. Vehicle vs. C-01 treatment at a concentration of 5µM was administered for 24h. Each experimental condition was tested in 18-28 wells from 3-5 independent plates.