fDOG presentation

fDOG

feature-aware Directed OrtholoG search

22.03.2021

Vinh Tran

Orthology

Homology

Paralogs <- duplication
Orthologs <- speciation

Applications

Phylogenetic profiles

Gene tree reconstruction, Tree reconciliation

Mallo D. et al. (2016), Syst Biol

Functional annotations transfer

...

RBH-based method

Chen et al. (2007), Plos ONE

The origin of fDOG

HaMStR

fDOG

How fDOG works

How good is fDOG

Benchmark data

78 taxa QfO 2018_04
Human as reference species
20,996 human proteins as seed sequences

Benchmark method

Compared using Quest for Orthologs - Challenge 6

And here is the result from the benchmarking service.
First, the EC conservation test: x-axis is the recall and y-axis is the precision of the evaluation. For the precision they used the average schlicker scores - a semantic similarity score for comparing sets of annotations such as GO terms or EC numbers
This result is in-line with the previous figure, fdog predicted less ortholog pairs than InParanoid, but it has a higher average schlicker score than InParanoid. OMA is the best in this test for the precision, however, when we applied FAS filter on fdog, especially by the mean FAS score, we could gain almost the same precision as OMA pairs.
For the GO conservation test, fdog performed slightly worse than the other tools. But after filtering out the pairwises that have low FAS scores, we reached the highest schlicker score.
These 2 tests shows clearly the trade-off between the recall and the precision of the prediction method.
The last test is the Tree discordance test: here the precision is defined by the Robinson-foulds distance: which is the sum of partitions of data that is implied by one tree but not by the other. The lower the distance, the better. Here fdog without fas filter has the highest recall, but also the lowest precision. However, the RF-Distance of fdog is not much worse than the one of OMA pairs, although OMA-pairs has the lowest number of complete tree sampling. With FAS filter, we could improve the RF-distance for fdog.
So, in overall I would say, fdog performed comparatively good with the other two tools. And depend on our needs, we can further filter the ortholog assessment from fdog to improve the specificity without losing much the sensitivity in comparison to OMA-pairs approach.

How meaningful are the core ortholog groups

... in reflecting the species relationship

Topology test using IQ-TREE

(Topology rejected (-) when p-AU < 0.05, otherwise accepted (+))
18,067 core groups with more than 3 sequences

Approach	Accepted	Rejected
No correction	11,707 (64.80%)	6,360 (35.20%)
Benjamini + Hochberg (BH) correction	13,625 (75.41%)	4,442 (24.59%)
Bonferroni correction	16,204 (88.69%)	2,043 (11.31%)

This is the number of the accepted (not significantly different) and rejected (significantly different) gene trees.
Before correction for multiple testing problem, 65% of the gene trees was accepted. After corrected using BH correction and Bonferroni correction, the number of accepted trees increased up to 89%.
Here we need to applied the multiple testing correction to reduce the false positive rate caused by the running of multiple statistical inference at the same time.

Problem: if one test is performed at the 5% level and the corresponding null hypothesis is true, there is only a 5% chance of incorrectly rejecting the null hypothesis. However, if 100 tests are conducted and all corresponding null hypotheses are true, the expected number of incorrect rejections (also known as false positives or Type I errors) is 5. If the tests are statistically independent from each other, the probability of at least one incorrect rejection is 99.4%.
=> Multiple testing corrections adjust p-values derived from multiple statistical tests to correct for occurrence of false positives
Bonferroni "punishes" all input p-values equally, whereas Benjamini-Hochberg (as a way to control the False Discovery Rate) "punishes" p-values accordingly to their ranking. If you have few tests, it should not make too much of a difference, but if you have many tests (as in this case, 18,000 human proteins), it will make a difference. In this case Bonferroni will produce false negatives (more accepted trees, or in other words it will discard significant observations). Therefore Benjamini-Hochberg is better.

How fast is fDOG

Data set	Core compilation	Ortholog search	FAS calculation
20,995 proteins	19 hrs (*)	67.5 hrs	21 hrs
Q8WZ42	54 hrs (**)	0.2 hrs	20.5 hrs

(*) fdogs.run using 64 CPUs on compute17 (2x AMD EPIC 7601 2.2Ghz, 32 cores)
(**) fdog.run using 1 CPU on compute12 (intel i7-6700K 4.0Ghz)

On average, each human protein needs:

212s for the core compilation
10s for the ortholog search

Note: number of core taxa and search taxa are the same (78 QfO taxa)

Correlation coefficient 0.61 for core compilation and 0.24 for ortholog search
Pearson p-value < 2.2e-16

fdog.run VS fdogs.run

Runtime of 6 randomly selected human proteins using 4 CPUs

An use-case: Where does the archaea locate

Eme et al. (2017), Nat. Reviews Microbiology

The well-known three-domains tree of life
Two-domains tree of life
Eukaryota branches as sister to TACK superphylum
Eukaryota originated from within (or as sister to) the Asgard archaea

48 core genes (FAS >= 0.75, present in at least 50% taxa of each superkingdom)
Super-alignment of 139 sequences, 13,060 aa in length
Tree reconstruction using IQ-TREE

-m MFP: auto identify best fitting model with ModelFinder => LG+R9
-bb 1000: ultrafast bootstrap approximation with 1000 replicates
-bnni: reduce the risk of overestimating branch supports with UFBoot