Development and validation of a high throughput screening platform to enable target identification in skeletal muscle cells from Duchenne Muscular Dystrophy (DMD) patients

Reviewer #1 (Public Review):

In this report, Hariharan and colleagues describe a protocol based on machine learning to differentiate wild type and DMD myofibers differentiated on a micropattern. They use images generated by immunofluorescence against the proteins of the DAPC complex, (which is disrupted in absence of dystrophin) to train a classifier to recognize healthy and diseased myofibers. They show that combining images generated with utrophin and alpha-sarcoglycan provide the most effective discrimination. They then validate the approach by applying their pipeline to wild type cells treated with DMD siRNA to mimic the DMD mutation or with antisense oligonucleotides to restore the DMD coding frame by exon skipping. They show that their strategy groups the siRNA treated myofibers with the DMD ones efficiently. Grouping of the oligonucleotide treated DMD myofibers with the healthy also works but is less efficient.

Overall the work is well done and I don't have any significant technical critique. I found this highly technical study interesting for a specialized audience.

Reviewer #2 (Public Review):

Summary:
The authors have developed a Myoscreen platform, which is a scalable and physiologically relevant system for generating and characterizing patient-derived myotubes. The platform can be used to accurately predict the DMD disease phenotype in a disease-relevant cell type and has wide applications in the drug development process.

Strengths:
The Myoscreen platform is scalable, meaning that it can be used to generate and characterize a large number of patient-derived myotubes. This is important for drug discovery, as it allows researchers to test a wider range of potential treatments. The Myoscreen platform also uses a physiologically relevant system for generating and characterizing myotubes. This means that the results obtained from the platform are more likely to be relevant to the human disease. This compared for example to using C2C12 myotubes. The Myoscreen platform has been shown to be effective in predicting the DMD disease phenotype. This means that it can be used to identify potential treatments that are likely to be effective in patients with DMD.

Weaknesses:
The study has several limitations. The method and material section could be improved. The authors rely heavily on UMAP to identify differences between non-DMD and DMD donor myotubes. They do not validate their findings using pharmacological small drugs. Additionally, the biological replicates used are extremely low, which raises concerns about the reproducibility of the findings.

Reviewer #3 (Public Review):

Summary: Hariharan et al. establish an analysis pipeline using automated microscopy to detect features identified by morphological profiling from images of common dystrophin complex proteins present in differentiated diseased and unaffected human myoblasts. Ultimately, using a machine learning algorithm to generate high dimensional phenotypes, the authors can distinguish Duchenne patient myotubes from unaffected patient controls based on the morphological features of several Dystrophin complex proteins. Initially analyzed on their own or in pairs the authors identify an optimal combination of Utrophin and a-sarcoglycan and subsequently test their ability to distinguish perturbations of Dystrophin either by knock down (siRNA) in unaffected controls or following treatment with a splicing modifier, vivo-phosphorodiamidate morpholino oligonucleomer (vivo-PMO) to ameliorate the DMD phenotype. It is unclear whether this methodology will see widespread adoption due to the combination of unique methods (micro-patterned plates and machine learning based image analysis) combined with a lack of detail on the specific features responsible for supporting the high dimensional phenotypes generated using their machine learning algorithm.

Strengths: The overall concept of this paper is interesting in that subtle morphological phenotypes, not readily observable by the eye, exhibited by dystrophin complex associated proteins can distinguish DMD samples from unaffected controls. It is interesting In Fig. 3B to see Utrophin and a-Sarcoglycan distinguish DMD and non-DMD lines from each other. This finding is the core of the paper and yet little information on how or why this is detected by image analysis is presented. An argument could be made that Combinations 1-7 all "work" to a certain degree at segregating DMD from non-DMD lines. This finding is exciting and has broad applicability both within and beyond the muscle field.

Weaknesses: Significantly more detail on the 235 features that are identified would greatly benefit the paper. What are the most critical features that give rise to high F-Scores for Utrophin and a-Sarcoglycan? What do the image masks display for the top ~10 features (or 5)? In Fig. 3B what metric(s) is critical in this segregation? What is the effect on the dimensional display if PCA is conducted as opposed to a tSNE?

Biological replicates are lacking to draw conclusions upon. Non-DMD #4 is present in certain figures and absent from others. With 2 replicates (non-DMD) and 2 replicates (DMD) it is difficult to draw statistical conclusions on the data. Non-DMD #4 is identified as a poor line (37% Desmin compared to the other lines being >88%) in Table 1. If this line is a poor line please remove it from the data analysis.

It is not appropriate to calculate Euclidan distance based on tSNE plots. PCA, MDS or UMAP are the appropriate high dimensional visual representations that allow for Euclidian distance calculations. This brings into question the validity of Fig. 4D and Fig. 5D. The link below outlines the limitations of tSNE plots. https://distill.pub/2016/misread-tsne/

It is unclear why treatment (siRNA) results in a statistically significant F-score (>0.9) when comparing non-DMD samples treated with siRNA against Dystrophin with DMD samples. Given that the siRNA knock-down appears to be quite robust this was unexpected and brings into question whether Dystrophin protein is the primary driver for the high dimensional phenotypes observed.