An artificial intelligence detects the toxicity of chemical substances
We will no longer count the areas of application of artificial intelligence (AI). After astrophysics, video games, medicine or chemistry, she is now testing toxicology and confirms that its effectiveness is well established. Tested on bisphenol S (BPS), the program called AOP-helpFinder developed by Karine Audouze, computer scientist and researcher at Inserm and the University Paris Diderot, and her colleagues, has shown a link between this molecule and the 'obesity. A publication in Environmental Health Perspectives details these results.
Sensitized to the side effects of drugs after working in the pharmaceutical industry, Karine Audouze decided to apply the "new powerful enough methods" of AI to the evaluation of the toxicity of chemical substances, in a way General. The goal is to try to "fill the information gap" on these products, she tells Futura. Bisphenol S appears as a textbook case. It has been used as a substitute for bisphenol A (BPA) since this recognized endocrine disruptor, the cause of an endless sanitary scandal, is banned in baby bottles (2013) and food containers (2015). But knowledge about him remains very poor and he would have effects similar to his cousin, BPA.
The AOP-helpFinder program evaluates the toxicity of products by peeling the mountains of data published in the scientific literature. For this, this "hybrid tool" is based on two methods, explains Karine Audouze. The first is "a text mining method, that is, a search based on words of interest," in this case terms referring to chemical substances, for example bisphenol S or pesticide, and terms describing pathological biological processes.
Then, the program studies the relationship between these terms and determines the strength of this link by assigning it a "weight", calculated by considering the "position" of the words in the scientific publication, and the "distance" between them. Higher scores are given for terms found at the end of the text, as they are likely to be a conclusion or result, while earlier terms are more likely to be search hypotheses. Similarly, words close to each other are "more likely to be in the same sentence, so to be related".
In order for the tool to know which terms to look for, Karine Audouze's team has created "dictionaries" gathering all the known designations of the substance to be studied, here bisphenol S, recorded in the PubChem database, as well as thousands related to diseases and so-called Adverse Outcome Pathways (AOPs). These are biological processes (pathways), at the level of a molecule or a cell, leading to adverse effects (adverse outcome) on the body, such as obesity or cancer. The databases Comparative Toxicogenomics Database (CTD) for diseases and AOP-wiki for AOPs, with 11,850 and 2,000 terms, respectively, were used.
The tool identified in 31 publications links between bisphenol S and several pathologies, a total of 48 PDOs, particularly in relation to obesity, cancer or steatosis, a liver disease. Obesity has caught the attention of researchers because this link is new and compared with bisphenol A. "It is well known that BPA is related to obesity." Cancer, certainly interesting, is "more general and complex to study", according to Karine Audouze.
The program has therefore discovered a potential toxicity of bisphenol S by association of words, without this being proof. "The co-occurrence of both terms [obesity and bisphenol S] was mentioned in six publications," says the researcher, but the link was not directly written. The researchers then manually checked the relevance of this relationship. It turns out that bisphenol S increases the risk of obesity by promoting the formation of adipocytes, cells storing fat.
To better understand the mechanism by which Bisphenol S promotes obesity, the researchers provided additional information to their tool, taken from the US ToxCast Chemicals Database reviewed by the Environmental Protection Agency (EPA). . The tool can be likened to "full of boxes, equivalent to biological mechanisms (PDO), leading to obesity. Some boxes still empty, so we looked for other databases "to complete them, says Karine Audouze.
The researchers put the AOP-helpFinder tool openly on the GitHub platform. It can improve by being used. He has already worked on other molecules, although only his performance on bisphenol S was described in the study. The researchers are now working on a version 2 of the program. They plan to test it "on a set of compounds" and the 29 million articles in biology and medicine available on PubMed, reveals Karine Audouze.
The tool itself is not intended to prove the toxicity or safety of a substance, but to point out the effects to study. "The idea is to give more and more working hypotheses and to go faster and faster on research". Hypotheses which will then have to be validated by experiments. Without completely eliminating them, it could help "reduce animal testing by targeting more targeted tests".
The tool has shown here an effect of bisphenol S on humans, but it is quite "applicable to different species, aquatic or other," to assess the toxicity of all substances documented in the scientific literature, including the drugs and environmental pollutants. "It is based on words," so its operation remains the same whether we speak of human beings or animals, says the researcher.
