AN EXAMPLE of the far reach of artificial intelligence (AI) is provided by a recent study in forensic science by Roderic Guigó of the Centre for Genomic Regulation in Barcelona. Although DNA samples are commonly used to help identify bodies, Dr Guigó and his colleagues have trained an AI agent to work out how long a corpse has been dead, based on the expression in the deceased’s cells of messenger RNA, a molecule that is allied to DNA. In life, messenger RNA carries the plans of proteins from genes in a cell’s chromosomes to tiny factories, called ribosomes, where those proteins are then made.

The researchers are collaborators in an international project called GTEx, which since 2010 has collected post-mortem samples from hundreds of donors. By analysing the quality and concentration of RNA sequences produced within cells it is possible to measure how actively particular genes are being read and translated into proteins. This helps to explain how the cells of living organisms work, which is the aim of the GTEx work. The database, however, also notes how long after a donor’s death each sample was collected. This, Dr Guigó and his colleagues realised, meant that they could use the records to study how the activity of genes change after a donor’s death.

The researchers examined the expression of genes in 7,105 tissue samples from 540 donors, all of which were collected within 24 hours of death. In particular, they looked at how the concentrations of RNA corresponding to each gene varied in tissue samples collected at different times. As they report in Nature Communications, they found no significant changes in the expression of the vast majority of genes. The team did find, however, 187 genes that were consistently expressed differently in at least three tissues after death. Of these, the expression of RNASE2, a gene encoding an enzyme that chops up RNA when it is no longer required, was reduced in most tissues. Two further genes, HBA1 and HBA2, which encode oxygen-carrying proteins, tended to be expressed more in several tissues after death, but not in blood.

The team cannot yet explain the complex mechanisms that cause these changes, but wondered that as the patterns of gene expression change with time if it would be possible to determine from a sample when its donor died. To put this to the test, they used data from 399 people to train a machine-learning algorithm, then tested the accuracy of the algorithm’s predictions with expression profiles from the tissues of another 129 individuals. This worked. Samples of lung, thyroid, skin or subcutaneous fat tissue produced particularly accurate results, with samples of two of the four usually being sufficient to pin down to within a couple of hours when their donors had died. To test their algorithm further, they fed the program profiles of samples taken from donors before their death. For these, they found the predictions of the algorithm were random, suggesting that the results they had obtained with post-mortem samples were meaningful.

Establishing the time of death, usually by observing physical changes in the body, can be unreliable. Yet this does not mean RNA is about to join DNA as a stalwart of forensic investigations. For one thing, bodies found at crime scenes are often days or weeks old. The team’s algorithm is only accurate for samples collected in the first 24 hours after death. Nor were the researchers able to assess the effect of environmental factors, such as temperature, on their predictions, and nor were they able to analyse how the cause of death would affect the expression profiles of tissues. None of their donors had, for example, died as a result of violence or in an accident. A promising start then, but more investigations to come.