Srsf10 and the minor spliceosome control tissue-specific and dynamic SR protein expression

Reviewer #2:

This paper describes an unusual regulatory feedback process in the alternative splicing of the Srsf10 gene involving competition between major and minor intron splice sites. Several RNA binding proteins self-regulate their protein levels by inclusion of alternative "poison exons" that include premature stop codons. Srsf10 appears to have a poison exon 3 who's exclusion requires minor spliceosome activity. This part of the paper seems solid. They then show that there is a general correlation of the expression levels of many SR proteins including Srsf10. This has been described before and similar correlations can be seen in ribosomal protein genes indicating that they are under coordinate regulation. The authors then remove the poison exon from the Srsf10 gene and argue that the results support a key role for Srsf10 and, by extension, the minor spliceosome in controlling the mRNA levels of SR genes in general. The key results shown in Figure 4E do not, to me, support this idea. The authors speculate that alterations in mRNA levels are due to NMD. A direct demonstration would be more convincing. With the data they show, I think they should focus on the Srsf10 case alone rather than pushing for a more global mechanism.

We agree that the effects observed in Figure 4F are rather subtle. However, we find it quite remarkable that the expression of all SR proteins is increased in the Crispr cell line and for 8/12 this is statistically significant. We would therefore like to keep this figure, but have modified the text to say “…was sufficient for subtle upregulation…”. We have also weakened the conclusion, which now reads: “…which is sufficient to change expression levels of the other SR proteins in a tissue- and differentiation state-specific manner.” In the previous version it was “of all other SR proteins”. We hope that these adjustments address the reviewer’s concern.

Reviewer #3:

[…]

Overall, we really liked this work. The authors should be congratulated for a thorough line of thoughtful experiments in support of their regulatory model as mentioned above. The manuscript is clearly written and we enjoyed reading it. We have few general comments/suggestions that should be addressed/clarified.

1) Are the changes in Rnpc3 observed in tissues in the same range as done in the titration experiments?

In cell culture siRNA mediated knockdown of Rnpc3 results in reduction of the Rnpc3 mRNA levels to 50%. Under this conditions Srsf10 levels are reduced to ~75%. Using the slope of the linear regression fit for Rnpc3 and Srsf10 across different tissues we can calculate how strong a 2-fold reduction in Rnpc3 effects Srsf10. Based on this calculation Srsf10 expression is reduced to 55%, which is in good agreement with the 75% observed in cell culture.

2) Expression computation and correlations: It's not clear how these were computed and whether these were done properly. The authors state TPM were derived by Whippet (Figure 3—figure supplement 1) but Whippet is designed only for splicing changes. It's not clear how Whippet would give full transcripts, and more importantly weighted gene level TPM values. Furthermore, TPM is not a proper measure to compare across many different experiments/conditions (It's not as bad as RPKM but still not great). Between sample normalization should be applied as implemented in DESeq and TMM. See for example https://haroldpimentel.wordpress.com/2014/12/08/in-rna-seq-2-2-between-sample-normalization/

Thank you for raising these points, which are important for our conclusions and thus merit further clarification. First, the Whippet Quant function also calculates weighted TPM values on gene level (see Sterne-Weiler, 2018 and the GitHub documentation). This is a great advantage of the Whippet tool, as it allows alignment, splicing, and gene expression analysis in one step/one pipeline.

However, to be more confident with our results we repeated the analyses using Salmon to determine GE levels and DESeq2 to obtain normalized expression counts. These analyses were performed by an additional bioinformatician, Alexander Neumann, who is now included as co-author. Comparing expression across different tissues, we observe comparable correlations using Salmon derived gene expression levels, which are now included as Figure 3—figure supplement 1B (Srsf10) and 1C (Gapdh). When removing a single outlier value, we also observe a significant correlation between Rnpc3 and Srsf10 in the DESeq2 normalized values (R² = 0.7509, p < 0.0001) but as we observe large variations (e.g. the three Rnpc3 values in epididymis 1: 222; 2: 17921; 3: 2655 and Srsf10 1: 2522; 2: 28859; 3: 6516) we would rather not include these into our manuscript. This is also mentioned in the Materials and methods part: “DESeq2 based normalized expression counts dramatically increased the variance of replicate tissue samples (but not of ES cell differentiation), which is why we chose to analyze only the un-normalized values for the tissue comparisons.”

In ES cell differentiation, we also observe comparable correlations using Salmon derived gene expression levels (new Figure 3—figure supplement 1I). Here however, normalization via DESeq2 strongly increases the correlation (new Figure 3—figure supplement 1J), making us confident that the DeSeq2 analysis in principle works (but is not suitable for the different tissue samples). Taken together, these additional analyses are fully consistent with the conclusion that Rnpc3 and Srsf10 expression indeed correlate across different tissues and development stages.

3) Figure 4: We understand where these p-values come from, but we are still worried about possible artifacts in the normalization procedures that might affect the results (also see above). Another way to compute a p-value and address the above concern is to compute an empirical p-value compared to sampling a large set of similarly expressed genes and computing the correlation values for them. True, some may be bona fide targets as well, but presumably this population of targets is rather small and the Sr proteins correlations stand out.

The R² and p-values for correlations of SR-protein and RNPC3 expression across different tissues indeed stands out with most R² values above 0.6 (average 0.74) and each p-value < 0.0001. These correlations are slightly reduced with Salmon derived correlations (R² average 0.6) but remain always highly significant (each p-value < 0.0001). Additionally, we would like to point out that in Figure 3D we have calculated correlation coefficients for 587 minor intron containing genes and 629 only major intron containing genes using the same procedure. This resulted in mean R² values of 0.24 and 0.16, respectively. We are therefore confident that the correlation coefficients for SR proteins are indeed true, as they stand out from only major intron containing genes and also from other minor intron containing genes. Also in ES cell differentiation a mean R² of 0.37 for SR proteins is clearly higher than the correlation for all major intron genes (0.16).