One thing that jumps out to me is that you're merging BAMs from different SRA studies. I'd be very cautious about interpreting subtle coverage patterns after doing that.
If the signal becomes stronger as you add more datasets, that doesn't necessarily mean it's biological. It could also mean you're accumulating protocol-specific biases that happen to align at the same genomic location. Different library prep methods, fragmentation strategies, PCR amplification levels, read lengths, and even alignment settings can all create coverage artifacts that become surprisingly consistent when enough samples are pooled.
The fact that many reads appear to start at exactly the same position is what I'd investigate first. In a typical RNA-seq dataset I'd generally expect a bit more heterogeneity unless there's a genuine transcript boundary, a strong processing site, or some technical bias.
A few things I'd check:
- Whether the feature is present within each individual sample rather than only after merging.
- Whether it appears in datasets from different labs and protocols.
- Strand-specific coverage, if available.
- Existing transcript annotations, including alternative TSSs and isoforms.
- Long-read RNA-seq or CAGE/TSS datasets for that locus, if any exist.
Personally, I'd be hesitant to call the low-level upstream coverage noise just because it's lower. RNA-seq often captures real low-abundance transcription, alternative start sites, cryptic transcripts, or incompletely annotated UTRs. The harder question is whether that signal represents the transcript you're interested in or some overlapping transcriptional event.
If you want to dig deeper, tools like IGV are great for manual inspection, and I've also found it useful to quantify coverage across individual samples rather than looking only at merged BAMs. Sometimes what looks like a clean biological pattern in the merged track turns out to be driven by a handful of samples.
