It should soon be easier to tell a chemical’s toxicity without killing animals


THE number of chemicals that might come into contact with a human being is staggering. The European Chemical Agency (ECHA) recognises over 130,000 molecules. Its American counterpart recognises 85,000. Testing all of these for toxicity is well-nigh impossible. Animal testing, in particular, is controversial, slow, costly and often cruel. Nor is it reliable. Its results are often irreproducible.

Things would be better if there were some way to predict the likely toxicity of a substance before animals get involved. That would permit the riskiest-looking to be prioritised. To this end, toxicologists like Thomas Hartung of Johns Hopkins University, in Baltimore, have been trying for years to find objective links between a chemical’s molecular structure and its biological activity. And now Dr Hartung thinks he has one. It relies, as do so many advances these days, on machine learning.

A way to link molecular structure and biological activity does already exist. It is called “read-across” and attempts to infer the hazards of an untested chemical by comparison with those of a tested one with a similar structure. In 2015 read-across was accepted as an alternative to animal testing for meeting the ECHA’s Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) requirements. But read-across depends on expert analysis and opinion, making it subjective and also difficult to generalise beyond small, well-studied groups of chemicals.

Dr Hartung believes machine learning, with its power to find patterns in large quantities of data, could help close the gap. His right-hand man in this work is Thomas Luechtefeld, a computer scientist who joined him as a PhD student in 2013. To tap into machine learning’s capabilities, the two of them first needed lots of good data. When Mr Luechtefeld started work, these were unavailable. He had adequate data for only about 250 chemicals. In 2014, however, he began to build a database that overcame this limitation, by downloading 816,048 toxicity studies on 9,801 compounds registered with REACH.

He spent a year training an algorithm to read these studies, process the text they contain and extract pertinent information. This algorithm automatically correlates chemical features like the presence of particular groups of atoms with measures of hazard such as the median lethal dose in an animal test, allowing all chemicals in the database to be compared. The result, which the two researchers reported in 2016, did indeed provide some insight into the prevalence of different types of toxicity. But to make more general predictions, they needed a larger data set still.

Mr Luechtefeld has therefore spent the past year scouring public data sets like those from PubChem, which is run by America’s National Institutes of Health. He now has relevant data on 80,908 chemicals and is able to correlate their features with 74 types of hazard. These are not just medical threats. They also include such things as fire hazard and potential to harm the ozone layer.

His latest algorithms focus on nine types of toxicity, including skin irritation, eye irritation and mutation-causing potential, which are conventionally assessed by animal trials. Using data from tested substances these algorithms are able to estimate the toxicity of untested ones. Instead of a single number, such as the median lethal dose in an animal test, they provide a probability that a substance is hazardous enough to worry about. Anything that scores above 0.8 should be regarded as a problem without further ado. Anything below 0.2 can be regarded as safe. Chemicals scoring between those values should be treated with caution until more data come in to push their scores up or down.

Mr Luechtefeld is now Dr Luechtefeld, having obtained his PhD a few weeks ago. He and Dr Hartung claim, in a recent paper, that the algorithm’s assessments are more accurate than animal testing. By this they mean that if a given molecule’s toxicity, as predicted by the algorithm, is compared with its read-across result, the two are more likely to coincide than are two independent animal tests on that molecule.

They are now waiting to hear from the authorities whether their method will be formally adopted, alongside conventional read-across, as a legal alternative to animal tests. Regardless of whether it is, though, what they have come up with should help understanding of the underlying mechanisms of toxicity. And that will be an important step forward.