A program that teaches itself to recognise the DNA patterns in silenced genes could help us better understand many diseases.
Mammals, including us, have two copies of each gene: one from the mother and one from the father. Normally, both are expressed, but occasionally one is imprinted, or silenced, which gives the other the deciding role. If this active gene is a mutation that would otherwise have been recessive, it can lead to disease. Conversely, if the silenced gene is harmful in some way, imprinting can be beneficial.
Investigating this phenomenon is difficult, however, as often the only distinguishing feature of an imprinted gene is that it has a methyl group attached, and this doesn’t show up in ordinary gene scans. Although 1 per cent of the 20,000 genes found on chromosomes 1 to 23 are estimated to be imprinted, only 40 have actually been found. “If you went through all of our genes, you’d have to search 100 to find one imprint,” says Alexander Hartemink at Duke University in Durham, North Carolina.
Advertisement
To speed up the search, Hartemink and his colleagues turned to an artificial intelligence technique called machine learning, in which a computer program is fed examples of imprinted and non-imprinted genes and teaches itself how to tell the difference between them. Machine learning has previously been used to create software that can identify DNA patterns known to be associated with increased susceptibility to cardiovascular disease or diabetes.
The researchers started by defining 7000 patterns in DNA sequences that they thought might possibly be present around imprinted genes. They then “trained” two different machine-learning programs by showing them 40 known imprinted genes. Once the programs had learned the patterns found in these genes, they were able to narrow the list of sequences down to around 800 likely candidates. This included sequences left over from virus attacks and repeating cytosine-guanine nucleotide strings, which often seem to be associated with the methyl groups.
When they turned the programs loose on the entire human genome, they were able to identify 156 genes that contained some of the likely sequences (Genome Research, ). So far the group has shown that two of these are indeed imprinted. One of them, DLGAP2, appears to play a role in suppressing bladder cancer when active, while the other, KCNK9 may contribute to bipolar disorder and some forms of cancer.
“The programs identified 156 genes that were likely to have been silenced”
The next step will be to test whether the rest of the 150-odd genes identified by the software really are imprinted. Hartemink suspects around 95 per cent are, but even if that falls to 50 per cent, he says it is still a useful tool. The two imprinted genes were found on the eighth chromosome, which was not previously thought to contain imprints, highlighting the software’s potential to speed up discovery of imprinted genes.
Genetics – Keep up with the pace in our continually updated special report.
