Artificial intelligence applied to precision medicine

The analysis of DNA and genomic data, as well as the extraction of variants or mutations, is quite complex. The automatic detection of which of them may be of potential risk (either because they have a high probability of contributing to the development of diseases such as cancer, or because they present unusual characteristics) and therefore it is convenient to analyze them in depth, is of great importance. .

Taking this into account, researchers from the Visual Telecommunications Application Group (GATV) of the Polytechnic University of Madrid (UPM), in Spain, have developed a system that provides, through artificial intelligence, a probability that the variants under study are potentially malignant (and therefore of interest for analysis) or may be benign variants within the human genome. The work developed, and the proposed tool, can help the study of genomic alterations in human cancers that allow targeted therapies for precision oncology based on next-generation sequencing personal data.

These researchers from the GATV of the UPM have carried out a study using several traditional techniques, and other new ones, of automatic learning (a modality of artificial intelligence) to classify somatic mutations.

As a result, they have developed a classification tool collecting a large number of known variants, both malignant and benign, collected in endorsed clinical studies and made available to the general public in open databases.

Symbolic artistic recreation of the concept of artificial intelligence, here represented by the binary code of a computer program, used in the fight against cancer, here represented by a cancer cell. (Illustration: Jorge Munnshe for NCYT from Amazings)

To learn more about each of those mutations, they have used an annotation software tool called ANNOVAR, which brings together other databases and algorithms that provide additional details about the mutations. 70 annotations (that is, 70 numerical values ​​that describe each variant collected) have been used as input for different artificial intelligence models whose objective is to obtain a probability that each of these mutations is benign or malignant.

The same data set has been tested on already existing classifiers that also use artificial intelligence models, among other proposals, to compare the effectiveness of the detectors. After comparing the solutions using known classification metrics for machine learning, it can be concluded that the implications of these results are considerably relevant, since they show that this proposed classification tool outperforms the rest of those studied.

The results obtained have been very promising, since the best artificial intelligence models have been able to correctly classify around 80% of potentially dangerous mutations. As Anaida Fernández García, a researcher who has been part of the work team, points out, “from a medical perspective, these tools are very useful when carrying out a genomic analysis of a patient, since the percentage of potentially malignant variants is usually very high.” small compared to benign ones, so being able to find them quickly and automatically supposes a huge reduction in working hours that can be dedicated to the direct study of these potential detections”.

All this work has been carried out thanks to the European project Genomed4All led by the UPM.

The new study is titled “DrOGA: An Artificial Intelligence Solution for Driver-Status Prediction of Genomics Mutations.” And it has been published in the academic journal Precision Cancer Medicine. (Source: UPM)