Interesting! We were considering to try it on ChromBPNet but realized that in this case the shift problem might be less prominent. You are using global average pooling, right? I assume the effect there is minor. Would be curious to see your current pipeline for indels/SVs.

Plus, we propose in silico deletion (ISD) as an alternative sequence interpretation technique alongside established in silico mutagenesis (ISM). 🔬 🧬

Our "stitching" approach improves indel eQTL classification accuracy and extends to even larger structural variants and tandem repeats.

ML models currently struggle with predicting the effects of indels because of sequence misalignment problems.