evec not invertible and Numerical underflow in Lie Markov DNA models

[dbajpp0]

Dear colleagues,
I am assessing which is the best substitution model for full mitochondrial protein-coding sequences as well as some partitions by using both standard and Lie Markov DNA models.
This is one example of the commands used:
iqtree2 -s 13PCGnt.fasta -nt 12 --mrate G,R,I,I+G,I+R --merit AIC --mset JC,F81,K2P,K3P,K81uf,HKY,TrN,TNef,TIM,TIMef,TVM,TVMef,SYM,GTR,1.1,2.2b,3.3a,3.3b,3.3c,3.4,4.4a,4.4b,4.5a,4.5b,5.6a,5.6b,5.7a,5.7b,5.7c,5.11a,5.11b,5.11c,5.16,6.6,6.7a,6.7b,6.8a,6.8b,6.17a,6.17b,8.8,8.10a,8.10b,8.16,8.17,8.18,9.20a,9.20b,10.12,10.34,12.12 --safe --seed 1234567 -pre mito
During the assessment of the Lie Markov DNA models, there are many messages "WARNING: evec not invertible"
and "WARNING: Numerical underflow for non-rev lh-branch Noname", particularly for those Lie Markov models with more parameters such as 9.20b,12.12. The run finishes without errors or core dumps but I am not sure if parameters estimated for those Lie Markov models and hence the AIC values are correct. This fact is important in my case since Lie Markov model 12.12 is generally the best fitting model for my data/partition. Is it save to use the results in despite of both warning?
Thanks in advance.
Best,
Joan

[bqminh]

You can safely ignore "evec not invertible". If you want to know more: IQ-TREE implements several methods to do matrix exponential when computing the probability transition matrix $P(t)=exp(Qt)$ from the rate matrix $Q$. One method is based on eigenvalue decomposition. For non-reversible models, this method can sometime be unstable numerically, and when instability is detected, this warning will be printed. But then in such case, IQ-TREE will fall back to another stable (but slower) technique called scaling-squaring. So everything is fine and I should have removed this warning.

"Numerical underflow for non-rev lh-branch": This happens when site-likelihood is exactly zero. In such case one cannot take $log(0)$, right?. Mathematically this is -infinity. But IQ-TREE will replace -infinity with a very negative number (I can't remember this negative number), but we tried set it to a negative value such that this site-likelihood will be smaller than any other sites, and at the same time printing this warning. So this is another mechanism that I hope to evade the problem. It happens during model parameter optimisation, but it also means it only happened to some combination of parameters that are so unlikely, and we can ignore having exact likelihood values of such combination, as long as they are lower than the MLEs. Hopefully, the final model parameters are sensible (and have a much higher likelihood).

My recommendation in such cases: Look at the model parameter values that IQ-TREE estimated carefully and see if there is any weird parameter. For example, are they at the boundary (too small or too large)? Is the log-likelihood of 12.12 model worse than other (simpler) models? If not, then you can safely ignore the WARNINGs.

[dbajpp0]

Thanks a lot for your reply. The best likelihood value was obtained with the 12.12 model in despite of having many “WARNING: evec not invertible” and few “WARNING: Numerical underflow for non-rev lh-branch” so I guess I can use the results.

I found another problem by mixing standard and Lie models to find the best model for a partitioned analysis:

iqtree2 -s 13PCGnt.fasta -nt AUTO --mset HKY,TN,K3Puf,TIM,TVM,GTR,4.5a,4.5b,5.6a,5.6b,5.7a,5.7b,5.7c,5.11a,5.11b,5.11c,5.16,6.6,6.7a,6.7b,6.8a,6.8b,6.17a,6.17b,8.8,8.10a,8.10b,8.16,8.17,8.18,9.20a,9.20b,10.12,10.34,12.12 --mrate I,G,R -pre part18 --safe -t PARS --seed 13579 --merit BIC -Q 18partitions.nex -n 0

I get the best model for each partition and the output files (model.gz, .log, .best_scheme.nex, bestscheme, and .parstree) but then I got a crash before the tree search (i.e. missing .iqtree and .treefile).

Full partition model BIC score: 338419.166 (LnL: -163214.675 df:1289)
Partition information was printed to part18.best_scheme.nex
Partition information in Raxml format was printed to part18.best_scheme

All model information printed to part18.model.gz
CPU time for ModelFinder: 1845.800 seconds (0h:30m:45s)
Wall-clock time for ModelFinder: 426.804 seconds (0h:7m:6s)

NOTE: 25 MB RAM (0 GB) is required!
Estimate model parameters (epsilon = 0.010)
ERROR: STACK TRACE FOR DEBUGGIERROR: STACK TRACE FOR DEBUGGING:SCK TRACE FSTACK TRACE FOR DEBUGGING:O
ERROR:
ERROR: R DEISTACK TRACEOR DEBUGNG:GING:GING:
ERROR:
ERROR: STACK TRACE FOR DEBUGGING:
ERROR: STACK TRACE FOR DEBUGGING:
ERROR: STACK TRACE FOR DEBUGGING:
ERROR:
ERROR: *** IQTRETREE REEC
ERROR:
C
ERROR:
ERROR: *** IROR:
R*
ERROR:
ERROR:
ERROR:
ERROR:
ERROR:
Q-TREE CRAS** ISHES WH***** IQ SIGNAL * IQ** IQQ** -TRET-THISEGMENTATION FAULTE CRREE CRASHESASHREES WITES WITH SIGNAL
ERROR: WITH SHE ERROR: WITH SHE -Q-TSEGT FCRMEN or RASHIGNALTATION FAULTEE CRASHES WITH SIGNAL REE CRASHES WITH SIGNALbug
ERROR: ES WIT SIGNAL SS WIT SIGNAL H SGrSEGMNL** Eeport please send to develoFor bug report GEGMENTATEplMEpINeaseEGMENTATION FAULTd to developers:
ERROR: *** Log file: part18.log
ERROR: *** TATIOAlignmers:ERROR: *** TATIOAlignmers:NAeON FAULTAeON FAULT
ERROR: Tnt ERROR: SEGMENTAT
ERROR: INfis* FATT
ERROR: ** NI( ** OLoif po**gorN FAULTb repo For *AUL ssible) For bug report pleasT
ERROR:
ERROR: bug report eafeafil
ERROR: e send to devel

The PartitionFinder analyses using both standard and Lie models also crashes even before the ModelFinder is done!
Any suggestions to perform PartitiomFinder and ModelFinder analyses with both standard and Lie models?