Researchers from Skoltech (Maxim Fedorov’s group, CDISE) and the German Research Center for Environmental Health, Helmholtz Zentrum München (HMGU) (Igor Tetko’s group) developed an enhanced drug candidate toxicity prediction technology based on the multitask machine learning algorithms and analysis of various types of toxicity data. The new approach allows accurately predicting adverse effects of medicinal compounds. The results of the study were published in the Journal of Chemical Information and Modeling.
Any new drug put on the market must be both effective and safe, so the first series of clinical tests always focuses on safety. According to the Food and Drug Administration (FDA) which oversees the safety of food and drugs in the United States, about 30% of drug candidates are dumped after pharmaceutical companies and scientists have invested tens of millions of dollars and thousands of working hours in their development.
To prevent such a waste, one should develop effective algorithms that will help to pin down the toxic compounds as early as possible.
There is no universal definition of toxicity which can be measured in a broad variety of living organisms, such as mice, rats and monkeys. The toxicity of a specific drug also depends on whether it is taken with meals, administered by injection or applied on the skin.
The authors of the study created a neural network capable of simultaneously predicting several different types of toxicity. To train the model, they used toxicity data on over 70 thousand organic compounds of various kinds that were grouped into 29 types according to the tested animal and the drug administration method.
The researchers compared their model to single-toxicity-type models and showed that the ultimate prediction quality is much higher if many types of toxicity are used when training the model.
One can easily find similarities for the phenomenon at hand in other areas: for example, simultaneous study of related sciences, such as mathematics and physics, helps the student to better understand each of the two disciplines and generally makes the learning process easier. The authors assume that different types of toxicity are also interrelated, which helps the neural network to generate more accurate patterns.
“Multitask training does not always produce a good result, however in our case it makes the prediction much better. Our study does not only testify to the effectiveness of the new approach, but also encourages the revision of outdated computational methods of toxicity prediction,” says the first author of the publication and Skoltech PhD student Sergey Sosnin.
The authors have made their new models available online, so that any chemistry researcher can make a preliminary assessment of a drug candidate in terms of its toxicity for several animal species.
Machine learning and Big Data analysis have already revolutionized many areas of science and are now making their way into toxicology. In the long term, scientists wish to have a way of getting accurate predictions of drug toxicity for humans, which will make the drug development process less costly and more productive.
+7 (495) 280 14 81