I started to add recently the gap partition in my ML analysis of ribosomal genes because it improves the results. To perform my analysis I use the online version of RAXML at Cipres Gateway. I add the gap partition as binary data [1/0] with ascertainment_correction_lewis.
However in my last analysis, where few short sequences were present together with other longer, I obtained an unexpected result. The short sequences clustered together, even if not related, in the wrong position and they were showing very long branches. An other short sequence was in the correct position but also showed a very long branch. The latter sequence covered the ssu-its part, where most of the gaps where present, while the formers covered part of the lsu.
I deduced that the abnormal topology of the tree depended on the missing data in the gap partition, codified by '0' like the absence of gap. In other words the short sequences [not covering the its] had a big part of gap partition identical because of a long list of '0'.
There is a way to overcome this problem without trigging the sequences at the same lenght?
Dear Franco,
Have a look at https://www.ncbi.nlm.nih.gov/books/NBK20261/
Article Fast Mapping of Short Sequences with Mismatches, Insertions ...
Thank you Muhammad Ali but I don't see how this could help with the settings of RAXML.
Make the shorter sequence long by repetition as need be. If unsatisfactory, state the sequences
This may not really answer your question as I'm not sure what you mean by gap partition. I found your question when I recently encountered what might be a similar issue, in which a partitioned, concatenated RAxML analysis (3 partition, GTRGAMMA) gave very implausible branch lengths when run with the "-M" (estimate individual per-partition branch lengths) option. The best trees inferred for each partition separately (done in same analysis) were fine and broadly congruent, yet the overall best tree was implausible and had an unreasonably high substitution-per-site axis/extremely long branches. I suspect this had something to do with the high proportion of nonrandom missing data, in the form of both missing and shorter sequences... or just user error, not quite sure. Either way, I was able to get a plausible tree by not estimating branch lengths per partition (ie, not using -M), but I ultimately chose to do single-gene analyses separately with more careful taxon sampling because there were few genes and I didn't quite trust the concatenated version when there was so much missing data.
Hi Alison, thank you for your answer. Actually I solved my problem long time ago. My issue was that gaps, when absent, were coded in the same way (=0) of missing data. Now I have changed the program to codify the gaps in binary data and I use 0 for gap absent and - for missing data.
- Badges
- Science topic
- Similar topics
- Force