Extracting gene-gene interactions from gene expression data

A method for generating a gene regulatory sub-network using a custom computing apparatus comprising at least one processor and a memory, the method comprising:

• receiving, in the memory, gene expression data of a plurality of genes measured at two or more time points or conditions;
• accessing, by the at least one processor, the gene expression data;
• and
• by the at least one processor,
• selecting a bicluster of genes and time points or conditions by performing a biclustering on the gene expression data, wherein the bicluster of genes and time points or conditions constitutes a subset of the gene expression data;
• for each gene in the bicluster, calculating a sum of squares of error for fitting the gene to each of the remaining genes in the bicluster, thereby generating a plurality of sums of squares of error comprising a sum of squares of error for fitting every gene to every other gene in the bicluster, wherein a sum of squares of error is determined by Equation (I):
• S
• 12
• 2
• ≡
• 1
• n
• ⁡
• [
• ∑
• i
• =
• 1
• n
• ⁢
• {
• e
• 2
• ⁢
• i
• -
• E
• 2
• ⁢
• i
• ⁡
• (
• e
• 1
• ⁢
• i
• )
• }
• 2
• ]
• (
• I
• )
• wherein:
• S122 denotes the sum of squares of error for fitting a gene (1) to a gene (2);
• n denotes the number of time points or conditions in the bicluster;
• e1i denotes the expression of gene (1) at the ith time point or condition;
• e2i denotes the expression of gene (2) at the ith time point or condition;
• and
• E2i(e1i) denotes the estimate of e2i based on e1i at the ith time point or condition and is determined by Equation (II):
• E2i(e1i)≡α0+α1e1i+α2e1i2+ . . . +αke1ik (II)
• wherein:
• k is a positive integer;
• and
• α0-αk each denotes the coefficient of the kth order fitted polynomial and is determined by solving:
• ∂
• S
• 12
• 2
• ∂
• a
• j
• =
• 0
• ;
• j
• =
• 0
• ,
• 1
• ,
• …
• ⁢
• ,
• k
• ;
• (
• III
• )
• determining a cutoff value from the plurality of sums of squares of error, wherein a sum of squares of error for fitting a first gene to a second gene less than the cutoff value identifies a likely regulatory interaction from the first gene to the second gene;
• and
• building a gene regulatory sub-network with two or more genes in the bicluster, wherein each of the two or more genes has an identified likely regulatory interaction with at least one other gene of the two or more genes.