Modifications made in section 4

[chloroplast13.git] / annotated.tex

These  last years  the cost  of  sequencing genomes  has been  greatly
reduced,  and thus  more and  more genomes  are  sequenced.  Therefore
automatic annotation tools are required to deal with this continuously
increasing amount of genomical data. Moreover, a reliable and accurate
genome  annotation  process  is  needed  in order  to  provide  strong
indicators for the study of life\cite{Eisen2007}.

Various  annotation   tools  (\emph{i.e.},  cost-effective  sequencing
methods\cite{Bakke2009}) producing genomic  annotations at many levels
of detail  have been designed  by different annotation  centers. Among
the major annotation  centers we can notice NCBI\cite{Sayers01012011},
Dogma       \cite{RDogma},       cpBase      \cite{de2002comparative},
CpGAVAS                   \cite{liu2012cpgavas},                   and
CEGMA\cite{parra2007cegma}. Usually, previous  studies used one out of
three methods  for finding  genes in annoted  genomes using  data from
these  centers: \textit{alignment-based},  \textit{composition based},
or a  combination of both~\cite{parra2007cegma}.   The alignment-based
method  is used  when trying  to predict  a coding  gene (\emph{i.e.}.
genes that produce proteins) by aligning a genomic DNA sequence with a
cDNA  sequence  coding  an homologous  protein  \cite{parra2007cegma}.
This approach is  also used in GeneWise\cite{birney2004genewise}.  The
alternative   method,   the    composition-based   one   (also   known
as  \textit{ab initio})  is based  on  a probabilistic  model of  gene
structure  to  find genes  according  to  the  gene value  probability
(GeneID \cite{parra2000geneid}).  Such  annotated genomic data will be
used to overcome  the limitation of the first  method described in the
previous section.   In fact, the  second method we propose  finds core
genes  from  large  amount  of  chloroplast  genomes  through  genomic
features extraction.

Figure~\ref{Fig1} presents an overview  of the entire method pipeline.
More    precisely,    the   second    method    consists   of    three
stages:   \textit{Genome    annotation},   \textit{Core   extraction},
and    \textit{Features    Visualization}    which   highlights    the
relationships.  To  understand the  whole core extraction  process, we
describe briefly each  stage below. More details will  be given in the
coming subsections.   The method uses as starting  point some sequence
database  chosen  among   the  many  international  databases  storing
nucleotide sequences, like  the GenBank at NBCI \cite{Sayers01012011},
the    \textit{EMBL-Bank}     \cite{apweiler1985swiss}    in    Europe
or   \textit{DDBJ}   \cite{sugawara2008ddbj}   in  Japan.    Different
biological tools can analyze  and annotate genomes by interacting with
these databases to  align and extract sequences to  predict genes. The
database in  our method must be  taken from any  confident data source
that stores annotated and/or unannotated chloroplast genomes.  We have
considered the GenBank-NCBI \cite{Sayers01012011} database as sequence
database:  99~genomes of chloroplasts  were retrieved.   These genomes
lie in  the eleven type  of chloroplast families and  Table \ref{Tab2}
summarizes their distribution in our dataset.

\begin{figure}[h]  
  \centering
    \includegraphics[width=0.75\textwidth]{generalView}
\caption{A general overview of the annotation-based approach}\label{Fig1}
\end{figure}

Annotation,  which  is the  first  stage,  is  an important  task  for
extracting gene features. Indeed, to extract good gene feature, a good
annotation tool  is obviously  required. To obtain  relevant annotated
genomes, two annotation  techniques from NCBI and Dogma  are used. The
extraction of gene feature, the  next stage, can be anything like gene
names,  gene  sequences, protein  sequences,  and  so  on. Our  method
considers gene  names, gene counts,  and gene sequence  for extracting
core  genes and  producing  chloroplast evolutionary  tree. The  final
stage   allows  to   visualize  genomes   and/or  gene   evolution  in
chloroplast.    Therefore   we   use  representations   like   tables,
phylogenetic  trees,  graphs,  etc.   to  organize  and  show  genomes
relationships,  and  thus  achieve   the  goal  of  representing  gene
evolution.   In addition,  comparing these  representations  with ones
issued from  another annotation tool dedicated to  large population of
chloroplast genomes  give us biological perspectives to  the nature of
chloroplasts evolution. Notice that  a local database linked with each
pipe stage is  used to store all the  informations produced during the
process.

\input{population_Table}
        
\subsection{Genome annotation techniques}

For  the first  stage, genome  annotation, many  techniques  have been
developed  to annotate chloroplast  genomes.  These  techniques differ
from  each others  in  the number  and  type of  predicted genes  (for
example:  \textit{Transfer  RNA   (tRNA)}  and  \textit{Ribosomal  RNA
(rRNA)}  genes). Two  annotation techniques  from NCBI  and  Dogma are
considered to analyze chloroplast genomes.

\subsubsection{Genome annotation from NCBI} 

The objective  is to generate sets  of genes from each  genome so that
genes are organized  without any duplication.  The input  is a list of
chloroplast genomes  annotated from NCBI. More  precisely, all genomes
are stored as \textit{.fasta} files  which consists in a collection of
protein  coding genes\cite{parra2007cegma,RDogma}  (gene  that produce
proteins) organized in coding sequences.   To be able build the set of
core    genes,     we    need    to     preprocess    these    genomes
using  \textit{BioPython}  package \cite{chapman2000biopython}.   This
step  starts by  converting  each  genome from  FASTA  file format  to
GenVision \cite{geneVision}  format from DNASTAR. Each  genome is thus
converted in  a list of genes,  with gene names and  gene counts. Gene
name duplications can be accumulated during the treatment of a genome.
These  duplications  come   from  gene  fragments  (\emph{e.g.}   gene
fragments treated  with NCBI) and  from chloroplast DNA  sequences. To
ensure that  all the  duplications are removed,  each list of  gene is
translated  into a  set of  genes.  Note that  NCBI genome  annotation
produces genes except \textit{Ribosomal (rRNA)} genes.

\subsubsection{Genome annotation from Dogma}

Dogma stands for \textit{Dual  Organellar GenoMe Annotator}.  It is an
annotation tool  developed at  University of Texas  in 2004  for plant
chloroplast and  animal mitochondrial genomes.  This tool  has its own
database  for translating  a  genome  in all  six  reading frames  and
queries     the     amino     acid     sequence     database     using
BLAST  \cite{altschul1990basic}  (\emph{i.e.}   Blastx)  with  various
parameters.  Protein  coding genes are  identified in an  input genome
using sequence similarity of genes  in Dogma database.  In addition in
comparison   with   NCBI    annotation   tool,   Dogma   can   produce
both \textit{Transfer RNAs (tRNA)} and \textit{Ribosomal RNAs (rRNA)},
verify their start and end  positions. Another difference is also that
there  is   no  gene  duplication   with  Dogma  after   solving  gene
fragmentation. In  fact, genome annotation  with Dogma can be  the key
difference when extracting core genes.

The Dogma  annotation process  is divided into  two tasks.   First, we
manually annotate chloroplast genomes using Dogma web tool. The output
of this step is supposed to  be a collection of coding genes files for
each genome, organized in GeneVision file. The second task is to solve
the  gene   duplication  problem  and   therefore  we  have   use  two
methods. The first method, based  on gene name, translates each genome
into a set  of genes without duplicates. The  second method avoid gene
duplication  through a  defragment  process. In  each iteration,  this
process  starts by taking  a gene  from gene  list, searches  for gene
duplication, if a duplication is found, it looks on the orientation of
the  fragment sequence.   If it  is positive  it appends  directly the
sequence to  gene files.  Otherwise reverse  complement operations are
applied  on the sequence,  which is  then also  append to  gene files.
Finally, a  check for missing start  and stop codons  is performed. At
the  end  of  the  annotation  process,  all  the  genomes  are  fully
annotated,  their   genes  are  defragmented,  and   gene  counts  are
available.

\subsection{Core genes extraction}

The goal of  this stage is to extract maximum core  genes from sets of
genes.  To find core genes, the following methodology is applied.

\subsubsection{Preprocessing}

In order  to extract  core genomes in  a suitable manner,  the genomic
data are preprocessed with two methods: on the one hand a method based
on gene  name and count,  and on  the other hand  a method based  on a
sequence quality control test.

In the first method, we extract  a list of genes from each chloroplast
genome.  Then we store this list of genes in the database under genome
nam and  genes counts can be  extracted by a  specific length command.
The \textit{Intersection  Core Matrix}, described  in next subsection,
is then  computed to  extract the core  genes.  The problem  with this
method can be stated as follows: how can we ensure that the gene which
is  predicted in  core genes  is the  same gene  in leaf  genomes? The
answer  to this problem  is that  if the  sequences of  any gene  in a
genome annotated  from Dogma  and NCBI are  similar with respect  to a
given  threshold,  then   we  do  not  have  any   problem  with  this
method. When the sequences are  not similar we have a problem, because
we cannot decide which sequence belongs to a gene in core genes.

The second method is based on  the underlying idea: we can predict the
the best annotated  genome by merging the annotated  genomes from NCBI
and Dogma according to a quality test on genes names and sequences. To
obtain all  quality genes  of each genome,  we consider  the following
hypothesis: any gene  will appear in the predicted  genome if and only
if the  annotated genes  in NCBI and  Dogma pass a  specific threshold
of  \textit{quality  control test}.   In  fact,  the Needle-man  Wunch
algorithm  is applied  to compare  both  sequences with  respect to  a
threshold. If  the alignment  score is above  the threshold,  then the
gene will be  retained in the predicted genome,  otherwise the gene is
ignored.   Once    the   prediction   of   all    genomes   is   done,
the \textit{Intersection Core Matrix} is computed on these new genomes
to extract core genes, as explained in Algorithm \ref{Alg3:thirdM}.

\begin{algorithm}[H]
\caption{Extract new genome based on gene quality test}
\label{Alg3:thirdM}
\begin{algorithmic} 
\REQUIRE $Gname \leftarrow \text{Genome Name}, Threshold \leftarrow 65$
\ENSURE $geneList \leftarrow \text{Quality genes}$ 
\STATE $dir(NCBI\_Genes) \leftarrow \text{NCBI genes of Gname}$
\STATE $dir(Dogma\_Genes) \leftarrow \text{Dogma genes of Gname}$
\STATE $geneList=\text{empty list}$
\STATE $common=set(dir(NCBI\_Genes)) \cap set(dir(Dogma\_Genes))$
\FOR{$\text{gene in common}$}
        \STATE $g1 \leftarrow open(NCBI\_Genes(gene)).read()$   
        \STATE $g2 \leftarrow open(Dogma\_Genes(gene)).read()$
        \STATE $score \leftarrow geneChk(g1,g2)$
        \IF {$score > Threshold$}
                \STATE $geneList \leftarrow gene$
        \ENDIF 
\ENDFOR
\RETURN $geneList$
\end{algorithmic}
\end{algorithm}

\textbf{geneChk} is a subroutine used to find the best similarity score between 
two gene sequences after applying operations like \textit{reverse}, {\it complement}, 
and {\it reverse complement}. Algorithm~\ref{Alg3:genechk} gives the outline of 
geneChk subroutine.

\begin{algorithm}[H]
\caption{Find the Maximum Similarity Score between two sequences}
\label{Alg3:genechk}
\begin{algorithmic} 
\REQUIRE $g1,g2 \leftarrow \text{NCBI gene sequence, Dogma gene sequence}$
\ENSURE $\text{Maximum similarity score}$
\STATE $score1 \leftarrow needle(g1,g2)$
\STATE $score2 \leftarrow needle(g1,Reverse(g2))$
\STATE $score3 \leftarrow needle(g1,Complement(g2))$
\STATE $score4 \leftarrow needle(g1,Reverse(Complement(g2)))$
\RETURN $max(score1,score2,score3,score4)$
\end{algorithmic}
\end{algorithm}  

\subsubsection{Intersection Core Matrix (\textit{ICM})}

To extract  core genes, we  iteratively collect the maximum  number of
common  genes   between  genomes  and  therefore   during  this  stage
an \textit{Intersection  Core Matrix}  (ICM) is built.   ICM is  a two
dimensional symmetric matrix where each row and each column correspond
to   one   genome.   Hence,   an   element   of   the  matrix   stores
the  \textit{Intersection Score}  (IS):  the cardinality  of the  core
genes   set  obtained   by  intersecting   one  genome   with  another
one. Maximum  cardinality results in selecting the  two genomes having
the maximum score. Mathematically speaking, if we have $n$ genomes in
local database, the ICM is an $n \times n$ matrix whose elements
satisfy: 
\begin{equation}
score_{ij}=\vert g_i \cap g_j\vert
\label{Eq1}
\end{equation}
\noindent where $1 \leq i \leq n$, $1 \leq j \leq n$, and $g_i, g_j$ are 
genomes. The  generation of a new  core gene depends  obviously on the
value  of the  intersection scores  $score_{ij}$. More  precisely, the
idea is  to consider a  pair of genomes  such that their score  is the
largest element in ICM. These two genomes are then removed from matrix
and the  resulting new  core genome is  added for the  next iteration.
The ICM is then updated to take into account the new core gene: new IS
values are computed for it. This process is repeated until no new core
gene can be obtained.

We  can observe  that  the ICM  is very  large  due to  the amount  of
data. As a consequence, the  computation of the intersection scores is
both  time and  memory consuming.  However,  since ICM  is a  symetric
matrix we can reduce the  computation overhead by considering only its
triangular  upper part.  The  time complexity  for this  process after
enhancement is thus $O(\frac{n.(n-1)}{2})$.  Algorithm ~\ref{Alg1:ICM}
illustrates the construction  of the ICM matrix and  the extraction of
the  core  genes, where  \textit{GenomeList}  represents the  database
storing all genomes  data. At each iteration, it  computes the maximum
core genes with its two genomes parents.

% ALGORITHM HAS BEEN REWRITTEN

\begin{algorithm}[H]
\caption{Extract Maximum Intersection Score}
\label{Alg1:ICM}
\begin{algorithmic} 
\REQUIRE $L \leftarrow \text{genomes sets}$
\ENSURE $B1 \leftarrow \text{Max Core set}$ 
\FOR{$i \leftarrow 1:len(L)-1$}
        \STATE $score \leftarrow 0$
        \STATE $core1 \leftarrow set(GenomeList[L[i]])$
        \STATE $g1 \leftarrow L[i]$
        \FOR{$j \leftarrow i+1:len(L)$}
                \STATE $core2 \leftarrow set(GenomeList[L[j]])$
                \STATE $core \leftarrow core1 \cap core2$
                \IF{$len(core) > score$}
                  \STATE $score \leftarrow len(core)$
                  \STATE $g2 \leftarrow L[j]$
                \ENDIF
        \ENDFOR
        \STATE $B1[score] \leftarrow (g1,g2)$
\ENDFOR
\RETURN $max(B1)$
\end{algorithmic}
\end{algorithm}

\subsection{Features visualization}

The goal is to visualize results  by building a tree of evolution. All
core  genes generated  represent  an important information  in the  tree,
because they  provide ancestor information of two  or more
genomes. Each  node in the  tree represents one chloroplast  genome or
one predicted core and labelled as \textit{(Genes count:Family name\_Scientific
names\_Accession number)}. While an edge is labelled with the number of
lost  genes from  a leaf  genome or  an intermediate  core  gene. Such
numbers are  very interesting because  they give an  information about
the evolution:  how many genes  were lost between two  species whether
they  belong  to  the  same  family  or not.   By  the  principle  of
classification, a  small number of genes lost  among species indicates
that those species are close to  each other and belong to same family,
while a  large lost  means that we  have an  evolutionary relationship
between species  from different families. To depict  the links between
species   clearly,  we   built   a  phylogenetic   tree  showing   the
relationships based on the distances among genes sequences. Many tools
are    available   to    obtain    a   such    tree,   for    example:
PHYML\cite{guindon2005phyml},
RAxML{\cite{stamatakis2008raxml,stamatakis2005raxml},    BioNJ,    and
TNT\cite{goloboff2008tnt}}.    In   this  work,   we   chose  to   use
RAxML\cite{stamatakis2008raxml,stamatakis2005raxml}   because   it  is
fast, accurate,  and can build large  trees when dealing  with a large
number of genomic sequences.

The procedure used to built a phylogenetic tree is as follows:
\begin{enumerate}
\item For each gene in a core gene, extract its sequence and store it in the database.
\item Use multiple alignment tools such as (****to be write after see christophe****) 
to align these sequences with each others.
\item Use an outer-group genome from cyanobacteria to calculate distances.
\item Submit the resulting aligned sequences to RAxML program to compute 
the distances and finally draw the phylogenetic tree.
\end{enumerate} 

\begin{figure}[H]
  \centering \includegraphics[width=0.75\textwidth]{Whole_system} \caption{Overview
  of the pipeline}\label{wholesystem}
\end{figure}

\section{Implementation}

The different  algorithms have  been implemented using  Python version
2.7,  on  a  laptop  running Ubuntu~12.04~LTS.   More  precisely,  the
computer is a Dell Latitude laptop - model E6430 with 6~GiB memory and
a  quad-core Intel  core~i5~processor with  an operating  frequency of
2.5~GHz. Many python packages  such as os, Biopython, memory\_profile,
re,  numpy, time,  shutil, and  xlsxwriter were  used to  extract core
genes  from large  amount of  chloroplast  genomes.

\begin{center}
\begin{table}[b]
\caption{Type of annotation, execution time, and core genes 
for each method}\label{Etime}
{\scriptsize
\begin{tabular}{p{2cm}p{0.5cm}p{0.25cm}p{0.5cm}p{0.25cm}p{0.5cm}p{0.25cm}p{0.5cm}p{0.25cm}p{0.5cm}p{0.2cm}}
\hline\hline
 Method & \multicolumn{2}{c}{Annotation} & \multicolumn{2}{c}{Features} & \multicolumn{2}{c}{Exec. time (min.)} & \multicolumn{2}{c}{Core genes} & \multicolumn{2}{c}{Bad genomes} \\
~ & N & D & Name & Seq & N & D & N & D & N & D \\
\hline
Gene prediction & $\surd$ & - & - & $\surd$ & ? & - & ? & - & 0 & -\\[0.5ex]
Gene features & $\surd$ & $\surd$ & $\surd$ & - & 4.98 & 1.52 & 28 & 10 & 1 & 0\\[0.5ex]
Gene quality & $\surd$ & $\surd$ & $\surd$ & $\surd$ & \multicolumn{2}{c}{$\simeq$3 days + 1.29} & \multicolumn{2}{c}{4} & \multicolumn{2}{c}{1}\\[1ex]
\hline
\end{tabular}
}
\end{table}
\end{center} 

\vspace{-1cm}

Table~\ref{Etime}  presents  for  each  method  the  annotation  type,
execution time,  and the  number of core  genes. We use  the following
notations:  \textbf{N}  denotes NCBI,  while  \textbf{D} means  DOGMA,
and \textbf{Seq}  is for sequence. The first  {\it Annotation} columns
represent the algorithm used to annotate chloroplast genomes, the {\it
Features} columns mean  the kind  of gene feature used to extract core
genes: gene name, gene sequence, or  both of them. It can be seen that
almost all methods need low {\it Execution time} to extract core genes
from large chloroplast genome.   Only the gene quality method requires
several days of computation (about 3-4 days) for sequence comparisons,
once the quality genomes are  construced it takes just 1.29~minutes to
extract core gene. Thanks to this low execution times we can use these
methods to extract core genes  on a personal computer rather than main
frames or parallel computers. The lowest execution time: 1.52~minutes,
is obtained with the second method using Dogma annotations. The number
of {\it  Core genes} represents the  amount of genes in  the last core
genome. The main goal is to  find the maximum core genes that simulate
biological background of chloroplasts. With  NCBI we have 28 genes for
96   genomes,   instead   of    10   genes   for   97   genomes   with
Dogma. Unfortunately, the biological distribution of genomes with NCBI
in core tree do not  reflect good biological perspective, whereas with
DOGMA the  distribution of genomes is biologically  relevant. {\it Bad
genomes} gives  the number of genomes  that destroy core  genes due to
low  number  of  gene  intersection.  \textit{NC\_012568.1  Micromonas
pusilla} is the only genome which destroyed the core genome with NCBI
annotations for both gene features and gene quality methods.

The second important factor is the amount of memory being used by each
methodology.   Table   \ref{mem}  shows  the  memory   usage  of  each
method.  We  used  a  package from  PyPI~(\textit{the  Python  Package
Index})     named     \textit{Memory\_profile}    (located     at~{\tt
https://pypi.python.org/pypi})   to   extract   all  the   values   in
table~\ref{mem}. In  this table, the values are  presented in megabyte
unit and \textit{gV} means  genevision~file~format. We can notice that
the level  of memory which is  used is relatively low  for all methods
and is available  on any personal computer. The  different values also
show that the gene features  method based on Dogma annotations has the
more   reasonable   memory   usage,   except  when   extracting   core
sequences. The third method gives the lowest values if we already have
the   quality   genomes,   otherwise   it  will   consume   far   more
memory. Moreover, the  amount of memory used by  the third method also
depends on the size of each genome.

\begin{center}
\begin{table}[H]
\caption{Memory usages in (MB) for each methodology}\label{mem}
{\scriptsize
\begin{tabular}{p{2.5cm}p{1.5cm}p{1cm}p{1cm}p{1cm}p{1cm}p{1cm}p{1cm}}
\hline\hline
Method& & Load Gen. & Conv. gV & Read gV & ICM & Core tree & Core Seq. \\
\hline
Gene prediction & ~ & ~ & ~ & ~ & ~ & ~ & ~\\
\multirow{2}{*}{Gene Features} & NCBI & 15.4 & 18.9 & 17.5 & 18 & 18 & 28.1\\
              & DOGMA& 15.3 & 15.3 & 16.8 & 17.8 & 17.9 & 31.2\\
Gene Quality  & ~ & 15.3 & $\le$3G & 16.1 & 17 & 17.1 & 24.4\\ 
\hline
\end{tabular}
}
\end{table}
\end{center}