Automatically catch intervals which extend beyond the chr?

[cafletezbrant]

Hi gReLU team,

Here is a reprex of a pair of SNPs that I am trying to predict using Borzoi. The prediction interval of the first is contained within the chr, but not the 2nd.

import grelu 
import grelu.resources
import pandas as pd

df = pd.DataFrame({
    'ref': ['G', 'C'],
    'alt': ['A', 'T'],
    'chrom': ['chr17']*2,
    'start': [80920469, 81005659],
    'end': [80920470, 81005660],
    'pos': [80920469, 81005659]
})
grelu.data.dataset.VariantDataset(variants=df, seq_len=524288, genome='hg19')

This gives the error

AssertionError: All input sequences must have the same length.

It would be great for this to be caught automatically somehow :)

[HelloWorldLTY]

Hi, I have a similar quesetion here. I am using the borzoi and grelu.data.dataset.AnnDataSeqDataset. I have checked that my datasets contain the gene with all seq len as 524288, but stilll receive the assertion error:

File /home/tl688/.conda/envs/evo/lib/python3.11/site-packages/grelu/sequence/format.py:414, in convert_input_type(inputs, output_type, genome, add_batch_axis)
    412         return strings_to_one_hot(inputs, add_batch_axis=add_batch_axis)
    413     elif output_type == "indices":
--> 414         return strings_to_indices(inputs, add_batch_axis=add_batch_axis)
    416 # Convert indices
    417 if input_type == "indices":

File /home/tl688/.conda/envs/evo/lib/python3.11/site-packages/grelu/sequence/format.py:251, in strings_to_indices(strings, add_batch_axis)
    247         return arr
    249 # Convert multiple sequences; they must all have equal length
    250 else:
--> 251     assert check_equal_lengths(
    252         strings
    253     ), "All input sequences must have the same length."
    254     return np.stack(
    255         [[BASE_TO_INDEX_HASH[base] for base in string] for string in strings]
    256     ).astype(np.int8)

AssertionError: All input sequences must have the same length.

[cafletezbrant]

Update: the error I was reporting is actually independent of sequence length / extending beyond chromosome boundary.

I just walked the call stack for grelu.data.dataset.VariantDataset() and actually the source of my error is in VariantDataset()._load_alleles(), which internally calls grelu.sequence.format.strings_to_indices(). Basically, internally it is expected that I am loading alleles of a common length (presumably, this is expected to be a SNP, rather than indel).

Here is an updated reprex:

from grelu.sequence.format import strings_to_indices
alleles = ['A', 'C', 'AC', 'G']
strings_to_indices(alleles)
## AssertionError: All input sequences must have the same length.
## VERSUS
alleles = ['A', 'C', 'G']
strings_to_indices(alleles)
## array([[0],
##       [1],
##       [2]], dtype=int8)

So the simple fix for me as a user is just to focus on SNPs, at least for now.

[avantikalal]

Addressed in #95