Skip to content

Commit

Permalink
added GeMoMa. More changes to come.
Browse files Browse the repository at this point in the history
  • Loading branch information
@lh3
lh3 committed Dec 12, 2022
1 parent 4b04b26 commit 6f2ac0e
Show file tree
Hide file tree
Showing 2 changed files with 48 additions and 34 deletions.
8 changes: 8 additions & 0 deletions tex/miniprot.bib
View file
Original file line number Diff line number Diff line change
Expand Up @@ -382,3 +382,11 @@ @article{Haas:2008tv
title = {Automated eukaryotic gene structure annotation using {EVidenceModeler} and the Program to Assemble Spliced Alignments},
volume = {9},
year = {2008}}

@article{Keilwagen:2019wz,
author = {Keilwagen, Jens and Hartung, Frank and Grau, Jan},
journal = {Methods Mol Biol},
pages = {161-177},
title = {GeMoMa: Homology-Based Gene Prediction Utilizing Intron Position Conservation and RNA-seq Data},
volume = {1962},
year = {2019}}
74 changes: 40 additions & 34 deletions tex/miniprot.tex
View file
Original file line number Diff line number Diff line change
Expand Up @@ -88,8 +88,9 @@ \section{Introduction}
GeneWise~\citep{Birney:1997vr,Birney:2004uy}, Exonerate~\citep{Slater:2005aa},
GeneSeqer~\citep{Usuka:2000vi},
GenomeThreader~\citep{DBLP:journals/infsof/GremmeBSK05},
genBlastG~\citep{She:2011aa}, ProSplign~\citep{Kapustin:2008tq} and
Spaln2~\citep{Gotoh:2008aa,Iwata:2012aa}. Among these, Spaln2 and
genBlastG~\citep{She:2011aa}, ProSplign~\citep{Kapustin:2008tq},
Spaln2~\citep{Gotoh:2008aa,Iwata:2012aa} and GeMoMa~\citep{Keilwagen:2019wz}.
Among these, Spaln2, GeMoMa and
GenomeThreader are the only tools practical for whole-genome alignment. They
can align several hundred proteins per CPU hour and may take a couple of days
to align a few hundred thousand proteins often needed to annotate a genome
Expand All @@ -103,7 +104,7 @@ \section{Introduction}
splice signals, which is not a trivial task, either. On top of these, we need
to fit these complex methods to an efficient implementation with modern
computing techniques. This is partly why we have over a hundred short-read
mappers~\citep{Alser:2021tk} but only two protein-to-genome mappers capable of
mappers~\citep{Alser:2021tk} but only three protein-to-genome mappers capable of
whole-genome alignment.

In this article, we will describe miniprot, a new protein-to-genome aligner
Expand Down Expand Up @@ -437,7 +438,7 @@ \subsection{Evaluated tools}

To evaluate what aligners can map proteins to a whole genome, we randomly
sampled 1\% of zebrafish proteins and mapped with various aligners. Only
miniprot-0.5, Spaln2-2.4.13c~\citep{Iwata:2012aa} and
miniprot-0.5, Spaln2-2.4.13c~\citep{Iwata:2012aa}, GeMoMa-1.9~\citep{Keilwagen:2019wz}
GenomeThreader-1.7.3~\citep{DBLP:journals/infsof/GremmeBSK05} could finish the
alignment in an hour. GenomeThreader found less than 30\% of coding regions in
Spaln2 or miniprot alignment. It is not sensitive enough for the human-fish
Expand All @@ -446,45 +447,48 @@ \subsection{Evaluated tools}
splice sites, it may be still useful for locating coding
regions~\citep{Manni:2021ww}.

In principle, we could localize a protein with a whole-genome mapper above and
then run GeneWise, GeneSeqer and Exonerate in local regions. However, this
would not evaluate mapping accuracy. In addition, \citet{Iwata:2012aa} have
already shown Spaln2 outperformed these older tools. We thus ignored them in
evaluation.

When running Spaln2, we applied option ``-Q7 -T\# -yS -LS -yB -yZ -yX2'' where
``\#'' specifies the species-specific splice model. Option ``-LS'' enables
local alignment and yields sligtly better alignment overall. Option ``-yB -yZ
-yX2'' apparently has no effect for human-zebrafish alignment but it greatly
improves the junction accuracy of the fly-mosquito alignment. We let Spaln2
choose the maximum intron and gene size automatically. Miniprot finds introns
up to 200 kbp in length by default. We changed this value to 50 kbp for
fly-mosquito alignment. We tuned the maximum intron size to 200 kb in the
MetaEuk human-zebrafish alignment, in consistent with the miniprot setting.
fly-mosquito alignment. We tuned the maximum intron size to 200 kb for the
MetaEuk and GeMoMa human-zebrafish alignment, in consistent with the miniprot
setting. We ran GeMoMa with MMseqs2~\citep{Steinegger:2017aa} as the underlying
engine. We evaluated the best unfiltered alignment of each protein as GeMoMa
discarded most alignments in the final output.

In principle, we could localize a protein with a whole-genome mapper above and
then run GeneWise, GeneSeqer and Exonerate in local regions. However, this
would not evaluate mapping accuracy. In addition, \citet{Iwata:2012aa} have
already shown Spaln2 outperformed these older tools. We thus ignored them in
evaluation.

\subsection{Evaluating protein-to-genome alignment}

\begin{table*}[!tb]
\processtable{Evaluation on the human-mouse dataset}
{\label{tab:eval}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}lrrrrrrrrr}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}lrrrrrrrrrr}
\toprule
Genome species & human & human & human & human & human & human & human &fruit fly&fruit fly \\
Protein species &zebrafish&zebrafish&zebrafish&zebrafish&zebrafish& mouse & mouse & mosquito& mosquito \\
Aligner & miniprot& miniprot& Spaln2 & Spaln2 & MetaEuk & miniprot& Spaln2 & miniprot& Spaln2 \\
Splice model & human & general & human & default & N/A & human & human & human &fruit fly \\
\midrule
Elapsed time (sec) & 460 & 471 & 12,716 & 13,024 & 2,518 & 314 & 3,736 & 29 & 2,528 \\
Peak RAM (GB) & 18.0 & 18.6 & 9.2 & 9.6 & 22.0 & 15.3 & 5.6 & 3.2 & 2.7 \\
\# proteins & 30,313 & 30,313 & 30,313 & 30,313 & 30,313 & 21,844 & 21,844 & 13,094 & 13,094 \\
\# mapped & 19,998 & 19,998 & 17,860 & 17,780 & 12,665 & 19,303 & 18,840 & 7,211 & 6,125 \\
\# single-exon & 1,836 & 1,703 & 990 & 606 & 2,230 & 2,810 & 1,975 & 1,308 & 495 \\
\# predicted junc. & 178,096 & 181,169 & 183,519 & 252,893 & 79,656 & 165,458 & 171,241 & 21,178 & 27,582 \\
\# non-ovlp. junc. & 462 & 750 & 1,426 & 18,738 & 216 & 316 & 852 & 459 & 877 \\
\# confirmed junc. & 165,084 & 164,102 & 165,826 & 156,980 & 5,761 & 161,113 & 162,551 & 18,630 & 22,606 \\
\% confirmed junc. & 92.69\% & 90.58\% & 90.36\% & 62.07\% & 7.23\% & 97.37\% & 94.93\% & 87.97\% & 81.96\% \\
\% base SN & 59.92\% & 59.97\% & 57.69\% & 56.28\% & 48.32\% & 89.48\% & 88.62\% & 52.71\% & 50.13\% \\
\% base SP & 95.76\% & 95.28\% & 92.54\% & 84.30\% & 91.58\% & 97.44\% & 95.27\% & 96.78\% & 97.38\% \\
Genome species & human & human & human & human & human & human & human & human &fruit fly&fruit fly \\
Protein species &zebrafish&zebrafish&zebrafish&zebrafish&zebrafish&zebrafish& mouse & mouse & mosquito& mosquito \\
Aligner & miniprot& miniprot& Spaln2 & Spaln2 & GeMoMa & MetaEuk & miniprot& Spaln2 & miniprot& Spaln2 \\
Splice model & human & general & human & default & N/A & N/A & human & human & human &fruit fly \\
\midrule
Elapsed time (sec) & 460 & 471 & 12,716 & 13,024 & 5,787 & 2,518 & 314 & 3,736 & 29 & 2,528 \\
Peak RAM (GB) & 18.0 & 18.6 & 9.2 & 9.6 & 137.3 & 22.0 & 15.3 & 5.6 & 3.2 & 2.7 \\
\# proteins & 30,313 & 30,313 & 30,313 & 30,313 & 30,313 & 30,313 & 21,844 & 21,844 & 13,094 & 13,094 \\
\# mapped & 19,998 & 19,998 & 17,860 & 17,780 & 25,096 & 12,665 & 19,303 & 18,840 & 7,211 & 6,125 \\
\# single-exon & 1,836 & 1,703 & 990 & 606 & 1,969 & 2,230 & 2,810 & 1,975 & 1,308 & 495 \\
\# predicted junc. & 178,096 & 181,169 & 183,519 & 252,893 & 199,252 & 79,656 & 165,458 & 171,241 & 21,178 & 27,582 \\
\# non-ovlp. junc. & 462 & 750 & 1,426 & 18,738 & 5,603 & 216 & 316 & 852 & 459 & 877 \\
\# confirmed junc. & 165,084 & 164,102 & 165,826 & 156,980 & 142,521 & 5,761 & 161,113 & 162,551 & 18,630 & 22,606 \\
\% confirmed junc. & 92.69\% & 90.58\% & 90.36\% & 62.07\% & 71.53\% & 7.23\% & 97.37\% & 94.93\% & 87.97\% & 81.96\% \\
\% base SN & 59.92\% & 59.97\% & 57.69\% & 56.28\% & 61.78\% & 48.32\% & 89.48\% & 88.62\% & 52.71\% & 50.13\% \\
\% base SP & 95.76\% & 95.28\% & 92.54\% & 84.30\% & 86.79\% & 91.58\% & 97.44\% & 95.27\% & 96.78\% & 97.38\% \\
\botrule
\end{tabular*}
}{Protein-to-genome alignments are compared to the annotated genes in ``Genome
Expand All @@ -497,7 +501,7 @@ \subsection{Evaluating protein-to-genome alignment}
coding regions.}
\end{table*}

We aligned zebrafish proteins to GRCh38 with miniprot, Spaln2 and MetaEuk
We aligned zebrafish proteins to GRCh38 with miniprot, Spaln2, GeMoMa and MetaEuk
(Table~\ref{tab:eval}). When we apply human-specific splice models to both
miniprot and Spaln2, miniprot is doing slightly better than Spaln2 at the base
level and on the junction specificity. Spaln2 finds 0.5\% more confirmed junctions,
Expand All @@ -509,12 +513,14 @@ \subsection{Evaluating protein-to-genome alignment}
create an intron even if the alignment is weak. Second, the Spaln2 developers
observed that heuristics may be doing better than strict DP around short
introns or exons. In one case, Spaln2 correctly created an exon with one amino
acid. Miniprot under the current setting would never produce such an alignment.
acid. Miniprot under the current setting would not produce such an alignment.

For both miniprot and Spaln2, species-specific models improved alignment though
the default Spaln2 model performed worse. MetaEuk did not pinpoint exact splice
junctions, as is expected. It also aligned fewer proteins and had lower
base-level sensitivity. We therefore did not evaluate it on other datasets.
the default Spaln2 model performed worse. GeMoMa did better than the default
Spaln2 but were not as accurate as Spaln2 with a proper splice model. MetaEuk
did not pinpoint exact splice junctions, as is expected. It also aligned fewer
proteins and had lower base-level sensitivity. We therefore only evaluated
miniprot and Spaln2 on other datasets.

For the human-mouse alignment, Spaln2 again has higher junction sensitivity and
miniprot is better on other metrics. On the more challenging fly-mosquito
Expand Down

0 comments on commit 6f2ac0e

Please sign in to comment.