NUG starts results in many canonical_extended ORFs

[bmmalone]

Comparing results using NUG starts compared to AUG-only starts, we see many, many more canonical_extended ORFs when using NUG than expected. For example, there are usually more canonical_extended than canonical predictions.

(There is no strand bias, though.)

[bmmalone]

This seems to occur when we have a "close" upstream, in-frame NUG start. The attached image shows an example, but this appears to be much more common than upstream, in-frame AUGs.

[bmmalone]

This is a result of the "select the longest ORF for each stop codon" postprocessing step. Thus, there is not really a simple fix for the behavior. A few ideas are:

• Incorporate the ORF type in the model, where "canonical" is more likely to be translated than others.

• Run both AUG and NUG and subtract out the AUG canonical results from the NUG predictions.

There is not an immediate plan to address this issue.

[m-swirski]

I see the same phenomena and in fact looking at the orf profiles it seems to be correct because of leaky scanning - some fraction of translation starts on each potential start codon and it is what actually should be expected. The only problem is lack of annotation for these leaky-scanning derived isoforms - sometimes as little as 0.1% of translation initiates on particular alternative start codon and the only isoform found in final "filtered.prediction.orfs.bed" is the longest one. Could one work around bayes_factors to delineate between possible starts?