Framepool can't handle deletions such as "AT"->""

[KalinNonchev]

Reproduce:

import kipoi

model = kipoi.get_model("../Framepool", source="dir")
inputs = {
    "ref_seq": "AT",
    "alt_seq": ""
}

model.predict_on_batch(inputs)

results in:

ValueError: Error when checking input: expected input_seq to have 3 dimensions, but got array with shape (1, 0)

Suggestion:

• Better error message would be helpful if vcf pre-filtering is required to remove this kind of variants
• Overall, I would just return nan prediction values for incorrect input with a warning

[Hoeze]

@KalinNonchev I agree, there should be a better error message.
But how did you get to an empty alternative sequence?

[KalinNonchev]

deletion such as ref->""

http://grch37.ensembl.org/Homo_sapiens/Transcript/Exons?db=core;g=ENSG00000142621;r=1:15632276-15656094;t=ENST00000375995

short 5'UTR region, which is deleted

[KalinNonchev]

http://grch37.ensembl.org/Homo_sapiens/Transcript/Exons?db=core;g=ENSG00000114115;r=3:139243249-139258561;t=ENST00000483943

Here for example we have only one letter ref 5'UTR seq also, which would result in

inputs = {
    "ref_seq": "A",
    "alt_seq": "G"
}

model.predict_on_batch(inputs)

InvalidArgumentError: Requires start <= limit when delta > 0: 2/1 [[{{node frame_masking_1/arange_2}}]]

Just warning with nan values would be better, instead of exception

[Hoeze]

OK, got it.
Still, I don't think a 1-letter UTR makes sense.

@Karollus Maybe we could add three additional flags:

• Could not predict (for whatever reason)
• Reference sequence too short
• Alternative sequence too short

What do you think?