Planning to integrate published and new data to explore lots of ideas related to chromatin patterns, co-txnal splicing.
- Basic characterization of stuff we can do without considering variation between individuals
- what sequence features can regulate H3K36me3, RNAPII-phosphorylation. Much literature about how these things are established over gene bodies, H3K36me <--> RNAPIISer2P interactions, Ser2 phosphorylation and H3K36me3 accumulates as transcription proceeds to 3' end... But not much literature on the sequence features that influence these patterns.
- How does splicing influence chromatin and transcription. Quantifying H3K4me3 and TSSs could be of interest E-mats mechanism. Can we see this effect. Characterize distance constraints of mechanism
- H3K9me3 usual model with regards to splicing is that it slows polymerase, promotes exon inclusion. So do colocalized sQTL/H3K9me3QTL/polymeraseQTL generally have concordant effect directions with that. Same with H3K36me3
- Can we anchor some splicing/transcription/chromatin causal graphs for colocolized QTLs based on well-mapped QTL positions: Like if a QTL is in a TF binding site in an enhancer, versus if it is in a splice site. What are directions of effects? I'm envisioning an analysis like McVicker et al), Fig3, but with TFBS, as well as splice sites and other RBPs with good PWMs, which we have strong prior beliefs in their primary effect on splicing, and correlating that to effect sizes for both intron retention, and chromatin
- Measuring variants that effect splicing/transcription speed. Do they associate with PolII phosphorylation?
- Assess contribution of any new molecular phenotypes to complex traits
- Are there any GWAS variants that we can see as sQTL/eQTL in chRNA but not polyARNA. That is interesting in itself I think. Presumably GWAS variants work closer to cytoplasm, so maybe these still have consistent effects in polyA fraction that are just smaller or harder to estimate due to NMD or other rapid decay mechanisms.
- ~50 fractionations done so far, in batches of 20-30 each for cell growth and fractionation, then a batch of 48 for chRNA-seq library prep.
- Currently growing cells for another fractionation batch. Planning one more batch after that (finish in a month-ish).
- Library prep for chRNA-seq + Cut&Tag another couple weeks after that. So long as the preliminary Cut&Tag data look good, I think we should have all the data sequenced sometime in July.
- see code/config/samples.tsv for full list of published and in-house samples that I have already started processing.
Will have data for all the following assays for YRI panel (and geuvadis has europeans too)
- H3K4me3, H3K4me1, H3K27Ac ChIP-seq: Grubert et al
- RNA-seq: geuvadis
- chRNA-seq: in-house
- H3K36me3, RNAPII-Ser5P, RNAPII-Ser2P, possibly H3K79me2 & H3K9me3 Cut&Tag: in house (still need to check data quality from tests, but will probably prep libraries for all of those marks for YRI panel)
- CTCF ChIP-seq: Ding et al
- Could also consider pro-cap/TSS initiation: Kristjánsdóttir et al
- 4sU, RNA methylation, DNA methylation, DNAseI
- adapter clipping (fastp)
- alignment w/ STAR (RNA-seq) or hisat2 (ChromatinProfiling), used WASP filtering
- QC: multiqc, qualimap (check that chRNA maps intron mapping rate)
- chRNA characterization: meta-intron coverage. Enrichment for novel junctions, NMD-specific junctions
- Obtaining genome-wide splicing timing estimate from the intron slope thing (something I showed you a preliminary analysis for) now that I think of it is probably too complicated because of unknown amount of fragmentation. More fragmentation before library prep will lead to a flatter slope for non-biological reasons, biasing estimates in ways that are hard to reasonably correct for. This shouldn't really matter for QTL calling, since we are just looking for relative differences by genotype.
- eRNA representation in chRNA-seq
- co-transcriptional splicing is more efficient for constitutive introns than ones flanking cassette exons (previously I showed this using 3SS ratio method (Normalized to polyA-RNA-seq) for quantifying co-txnal splicing efficiency). Based on the much research on splicing efficiency correlating with highly active transcription, I would expect to see some correlation with gene expression too, which I haven't yet looked for.
- ChromatinProfiling peak calling to identify phenotype features for testing with macs2... Testing peak calls with all individuals jointly, or with each individual and then merging.
- For RNA-seq, top 14000 protein coding genes (as measured by median CPM in geuvadis) counted with featureCounts and standardized and inverse-normalized to association testing.
- For ChIP-seq, same process with the macs2 peak calls.
- For H3K36me, (don't have data yet) I may test out a couple ideas for determining test features. Like just using the same 14000 expressed gene regions, and then breaking those regions into 1000bp windows
- Splicing junctions quantified with leafcutter pipeline.
- For intron retention took constitutive introns (annotated introns that do not overlap any annotated exons) in the 14000 expressed genes and calculated a the coverage (reads per feature length) of the intron over coverage of the host gene exon regions. Also filtered out introns where this ratio is >10 or <1/100. ~77,000 introns.
- Intron slope thing that I showed before. I wonder how much of those slope affects are more related to transcription initiation than elongation rate.
- QTL testing done with QTLtools, grouping splicing phenotypes by cluster to get Cluster-level tests (if I turn out breaking broad H3K36me3 phenotypes into windows, will group these too). All phenotypes standardized and inverse-normalized.
- Will include 0, 1, 2 or 3 genotype PCs as covariates, after manually inspecting PC plots for stratification. Will include expression PCs too: Planning to include the number of PCs for which the variance explained by a PC exceeds the variance explained from PCA on a permuted phenotype matrix (each row randomly shuffled). This is what Rasqual paper did, and from my testing this seems to give a reasonable number of PCs.
- For chromatin phenotypes I could consider doing the rasqual pipeline which I have working, I'll have to see how much more power it gets.
- testing a lot of the questions will involve looking for colocalizing molQTLs. Probably worth using a formal colocalization stats method for this. I'm not yet sure of how best to do this with all the different assays we have. HyPrColoc looks promising
short term TODO:
- choose hyprcoloc parameters and colocalize trait clusters by gene.
- more critically investigate intron retention qtl pipeline. I have performed this on polyA RNA-seq (GEAUVADIS) and chRNA-seq. I haven't spent much time manually inspecting the hits.
- Fix the pygenometracks scripts that I have been using to plot QTLs
- Repeat splice site SNP enrichment analysis for sQTLs for YRI GEAUVADIS subset.
- Something like McVicker's correlation of splice site PWM changes to chromatin phenotype changes
- Redo heatmap of temporal phenotypes, using intronic coverage for chRNA. Also, similarly plot temporal phenotypes in PC space
- more general enrichment analysis for QTLs; chromHMM and other regional annotations.