APARENT - number of predictions seems off

[robertzeibich]

Hi, I am currently running Kipoi APARENT.

The provided example VCF file contains 15,077 variants, but the output only has 416 predictions for delta_logit_distal_prop and delta_logit_proximal_prop. I was expecting 15,077 predictions. Do you know why I only get 15,077 predictions?
Kipoi APARENT

[Hoeze]

Hi @robertzeibich, we currently only evaluate variants that are close to a polyA-site:

models/APARENT/veff/dataloader.py

Lines 84 to 108 in 9dea565

	def get_roi_from_transcript(transcript_start: int, transcript_end: int, is_on_negative_strand: bool) -> (int, int):
	"""
	Get region-of-interest for APARENT in relation to the 3'UTR of a transcript
	:param transcript_start: 0-based start position of the transcript
	:param transcript_end: 1-based end position of the transcript
	:param is_on_negative_strand: is the gene on the negative strand?
	:return: Tuple of (start, end) position for the region of interest
	"""
	# CSE should be roughly around position 70 of the 205bp sequence.
	# Since CSE is likely 30bp upstream of the cut site, we shift the cut site
	# by 100bp upstream and 105bp downstream
	if is_on_negative_strand:
	end = transcript_start + 100
	# convert 0-based to 1-based
	end += 1
	start = end - 205
	else:
	start = transcript_end - 100
	# convert 1-based to 0-based
	start -= 1
	end = start + 205
	return start, end

Some notes:

• I'd strongly recommend to run kipoi predict with the --keep_metadata flag, then you will see which transcript it affects
• Currently there is a bug which will likely lead to wrong results on negative-strand transcripts:
  fix missing reverse-complement in reference sequence fetching for APARENT #336

What would you like to use APARENT for?

[robertzeibich]

Thank you for getting back to me that quickly. I used the keep_metadata parameter and then concatenated the output with a pandas dataframe.

Can you inform me once the bug was fixed?

I want to integrate the poly(A) scores in my whole genome sequence analysis. Perhaps compare the scores against healthy controls (1000 Genomes project) and see if patients with epilepsy and individuals from the 1000 Genomes project cluster somehow. If you have another idea what I could do with the scores, I am all ears. The data I am currently analyzing is whole genome sequenced data from people with epilepsy.

[Hoeze]

Hi @robertzeibich, I did merge now the PR to fix the reverse-complement issue.
Still, I'm not 100% sure if the implementation is perfectly fine, so please look at the predictions with a bit of caution for now :)