Variant Effect Prediction

[KhudayarFarmanli]

Hello,

I am using Evo 2 to calculate delta likelihood scores (by using alternate and wildtype sequence with 8,192 bp window) to see if the variant is likely to be pathogenic or benign. I do analysis with SNPs and Indels. The results for SNP seem to be quite logical and useful but for Indel when the length is higher (>50 bp let's say by average), results seem to be biased since the are getting generally the highest and lowest delta scores. It does not seem reliable at all and I want to ask what is your experience in this issue? Do I have to use it only for <10 bp Indels to rely on the results?

Thanks in advance.

[garykbrixi]

Hi @KhudayarFarmanli, scoring longer indels and complex variants is an open question. In the paper we keep ref and alt sequences the same length (8,192 bp) by adjusting the sequence boundary after the indel. Otherwise the longer indels would have delta log likelihoods with a sequence length bias, which is likely what you are seeing. I haven't tried a length calibration, but one alternative to keeping the length the same using context padding/slicing would be to instead regress out length effects from a neutral dataset. Curious what you find!

[KhudayarFarmanli]

Hi @garykbrixi , thank you very much for your explanation. But I think I was not clear about my issue. I also keep ref and alt sequences the same length to keep the consistency as you recommended in your paper. The problem is directly about inserted or deleted sequence length, if the indel is by average longer than 50 bp, the results do not seem plausible. I have to recheck the paper, but I think you have used maximum 10 bp indel in your experiments. What do you think?

[garykbrixi]

Thanks for the clarification. For clinvar section we consider variants that affect less than 64bp length, most of which are significantly shorter, while the mutagenesis examples were 10bp as you said. We don't have analysis for longer variants or length effects, so I'll be curious to know if you end up characterizing this more.

[KhudayarFarmanli]

I have done few more analyses and my observations about Indels are as before. Mostly, the lowest and the highest delta scores are coming from Indels and they increase lowest and highest end of the distribution. In this case, I do not know how much reliable are the delta likelihood scores for Indels.