kmer_gapped_kernel.m

《模式分析的核方法》一书中的源代码

function K = kmer_gapped_kernel(S,k,maxgaps,lambda)

%function K = kmer_gapped_kernel(S,k,maxgaps,lambda)
%
% Computes the k-mer gapped kernel between the strings stored in S;
% w gaps are allowed. The features are weighted by an
% exponential lambda^-m, where m is the actual number of gaps for
% that feature.
%
%INPUTS
% S = a cell array containing all strings
% k = an integer indicating the size of the k-mer v(not including the gaps)
% maxgaps = an integer <k indicating the number of gaps allowed
% lambda = the forget factor
%
%OUTPUTS
% K = the kmer kernel evaluated between all strings
%
%
%For more info, see www.kernel-methods.net

%
%Author: Tijl De Bie, march 2004, bug fixed october 2004 (thanks to Neil Lawrence).

nS=length(S);

alphabet=[];
for j=1:nS
    for i=1:length(S{j})
        alphabet=union(alphabet,S{j}(i));
    end
end

na=length(alphabet);

for j=1:nS
    ngaps{j}=zeros(length(S{j}),1);
end
K=zeros(nS,nS);
for i=1:na
    i
    present=[];
    clear newS newind newngaps
    ct=0;
    for j=1:nS
        newind_store=find(S{j}(1:end-k+1)==alphabet(i));
        newngaps_store=zeros(size(newind_store));
        if ~isempty(newind_store)
            ct=ct+1;
            present=[present ; j];
            newS{ct}=S{j};
            newind{ct}=newind_store;
            newngaps{ct}=newngaps_store;
        end
    end
    if ~isempty(present)
        K(present,present)=K(present,present)...
            +iterates(newS,newind,newind,1,k,newngaps,maxgaps,lambda,alphabet,alphabet(i));
    end
end


function K=iterates(S,indstart,indend,it,maxit,ngaps,maxgaps,lambda,alphabet,seq)

%function K=iterates(S,indstart,indend,it,maxit,ngaps,maxgaps,lambda,alphabet,seq)
%
% A help function for kmer_wildcard_kernel.m
%
%
%Author: Tijl De Bie, march 2004.

na=length(alphabet);
nS=length(S);


if it==maxit
    for i=1:nS
        number(i,1)=0;
        for j=1:length(indstart{i})
            number(i,1) = number(i,1)...
                +align_seqs(S{i}(indstart{i}(j):...
                min(indstart{i}(j)+maxit+maxgaps-1,length(S{i}))),seq,lambda);
        end
    end
    K=number*number';
    return
end

K=zeros(nS,nS);
for i=1:na
    ct=0;
    present=[];
    clear newindstart newindend newS newgaps
    for j=1:nS
        newindstart_store=[];
        newindend_store=[];
        newngaps_store=[];
        for k=1:length(indstart{j})
            for offset=1:min(maxgaps-ngaps{j}(k)+1 , length(S{j})-indend{j}(k))
                if S{j}(indend{j}(k)+offset)==alphabet(i)
                    newindstart_store=[newindstart_store ; indstart{j}(k)];
                    newindend_store=[newindend_store ; indend{j}(k)+offset];
                    newngaps_store=[newngaps_store ; ngaps{j}(k)+offset-1];
                    break
                end
            end
        end
        if ~isempty(newindend_store)
            ct=ct+1;
            newindstart{ct}=newindstart_store;
            newindend{ct}=newindend_store;
            present=[present ; j];
            newS{ct}=S{j};
            newngaps{ct}=newngaps_store;
        end
    end
    if ct~=0
        K(present,present)=K(present,present)...
            +iterates(newS,newindstart,newindend,it+1,maxit,newngaps,maxgaps,...
            lambda,alphabet,[seq alphabet(i)]);

    end
end



function align_seqs=align_seqs(s,t,lambda)

% Compute alignment score of a short sequence t to a long one s, where gaps
% are allowed only in t, and gaps are exponentially penalized with a factor
% lambda.

%Note: this can be made more efficient by computing only a thick diagonal
%of the DP table!

M=zeros(length(t)+1,length(s)+1);

%for j=1:length(s)-length(t)+1
%    M(1,j)=1;
%end
M(1,1)=1;
for j=2:length(s)+1
    for i=2:length(t)+1
        M(i,j)=(t(i-1)==s(j-1))*M(i-1,j-1)+((i==length(t)+1)*1+(i~=length(t)+1)*lambda)*M(i,j-1);
    end
end
align_seqs=M(end,end);