AI tool can predict which ingredient in foods, cosmetics causes cancer

Indian researchers have developed an artificial intelligence-based model to predict whether substances have cancer-causing ability. Called Metabolkiller, they say that the model can be path-breaking for the food and drug industry to predict carcinogenicity of new molecules instead of finding out after long-term use.
“Whenever companies formulate something new – be it chips or cosmetics – they have to ensure that the ingredients do not cause cancer. Similarly, medicine manufacturers also conduct tests to ensure they are not going to cause cancer. In the last few decades, more than 750 FDA-approved drugs have been withdrawn from the market because it was found out that they were inducing cancer,” said Dr Gaurav Ahuja, one of the researchers at IIIT Delhi involved in the study.
The researchers claim that the model is a game-changer when it comes to carcinogenicity prediction because it looks at six dinct biochemical responses instead of just the chemical properties of the substance.
“Traditionally, any artificial intelligence model for cancer prediction depends on learning from the already known carcinogens. Think of these algorithms like a Google search – when you are looking for something, it searches its database for that thing and if it isn’t there, shows you the closest result. However, the 2,000 or 3,000 carcinogenic molecules that we know about aren’t enough to train an algorithm, resulting in lower accuracy,” said Dr Ahuja. “While we were developing our model, we thought whether we could pass this shortcoming? What if we didn’t depend on just these 2,000 to 3,000 known carcinogens? So, instead we trained our model to recognise six biochemical processes that usually take place in cancer cells,” he added.
The researchers scoured through studies on thousands of molecules – not just the carcinogens – and the biochemical processes that these caused in cells. They trained their algorithm to look for these processes instead.
Metabokiller is trained to figure out whether a molecule can cause six such biochemical processes associated with cancer. First is electrophilicity, which checks whether the molecule carries a positive charge. Dr Ahuja explains that nature DNA, be it human, primate or any other animal, is negatively charged. Hence only a positively charged molecule can be attracted to it and attack it, there leading to cancer. Second, Metabokiller tests the ability to induce proliferation or rapid growth in the cells that is seen in cancerous growth. Third, it looks for the oxidative stress the molecule creates.
“Cancer can be caused an external molecule such as the ones we test or something that happens internally. A molecule might not directly attack the DNA but it can result in increase of reactive oxygen species or ROS that are known to harm the DNA (any anti-oxidants you consume actually work lowering the ROS levels),” said Dr Ahuja.
The fourth process that Metabokiller looks for is genomic instability or the ability of a molecule to damage the DNA. The fifth is epigenetic alteration – “Some molecules may not change the genome sequence of the cells but can lead to epigenetic changes. To explain, think of a gene that, say, gives me long hair. Now, a molecule can prevent the long hair either damaging the gene or through epigenetic changes, turning it off, which means the gene remains unharmed but it doesn’t work.”
The last process that Metabokiller looks for is anti-apoptotic response – the ability to prevent cell death even under stress (such uncontrolled growth of cells is what leads to cancer).After checking for the six processes, the algorithm assigns a prediction score to any molecule. “Anything over 50 per cent is worth validating through the standard experiments for identifying any carcinogen.”
Dr Ahuja explains, “For identifying any substance as cancer-causing, it has to undergo animal testing, then given to human beings and studied for at least five to ten years to see whether it has resulted in cancer. What Metabokiller does is quickly identify the possibly cancer-causing molecules. Instead of one molecule being tested for five to ten years, the AI algorithm can check a thousand molecules in two seconds.”
He adds, “This is essential for say the drug industry to identify the possibly cancer-causing agent at the outset and not putting in money on that. They would, in the end, check only the molecules finalised with the standard experiments. Or while checking industrial waste – the potential cancer-causing molecules can be identified quickly and then those with high prediction scores can be put through the experiment.”
The study, published in Nature Chemical Biology, “observed high synergy between Metabokiller predictions and experimental validations.”
Dr Debarka Sengupta, one of the authors from the department of computational biology at IIIT-D, said that the team has made the AI tool freely available to researchers and on licence to companies. “We are in touch with some pharma companies for the same,” he added. The research was supported the Department of Biotechnology.
According to Dr Juhi Tayal, one of the authors and clinical patholog at Rajeev Gandhi Cancer Institute and Research Centre, “Only 8-10 per cent of the cancers are inheritable (transfer from parents to kids), and the majority of them are caused exposure to carcinogens (cancer-causing compounds). Accurate identification of these carcinogens and their proper detoxification from the body provide an alternative strategy for cancer prevention. There is now growing evidence that these carcinogens are multifaceted and can have a role both in initiation and progression of cancer.”Scients from institutes such as Rajiv Gandhi Cancer Hospital and Research Centre, Rohini, IIT-Ropar and CSIR-IGIB were involved.