Mask not computed correctly. Possible off-by-1 error?

[bw2]

broadinstitute/SpliceAI-lookup#59
has several examples of apparently inconsistent masking. For example, for
1-237727120-T-C
the mask only affects the Acceptor Gain score, but should arguably be applied to both the Acceptor Gain and Donor Gain:

Am I not interpreting the results correctly?

[kishorejaganathan]

This is unfortunately caused because of the fact that the mask implementation is a super-hacky one and not precise. We only look for the closest exon boundary

SpliceAI/spliceai/utils.py

Line 62 in f364513

dist_ann = (dist_tx_start, dist_tx_end, dist_exon_bdry)

and check whether that position is reported in the output

SpliceAI/spliceai/utils.py

Lines 185 to 188 in f364513

	mask_pa = np.logical_and((idx_pa-cov//2 == dist_ann[2]), mask)
	mask_na = np.logical_and((idx_na-cov//2 != dist_ann[2]), mask)
	mask_pd = np.logical_and((idx_pd-cov//2 == dist_ann[2]), mask)
	mask_nd = np.logical_and((idx_nd-cov//2 != dist_ann[2]), mask)

This worked most of the time when the distance parameter is less (like the default 50) as you are unlikely to see two ends of an exon within such a short distance, but completely falls apart when the distances are large. I would recommend ignoring the mask property for larger distances for the time-being.