Facebook claims that its new artificial intelligence can predict the way drugs interact with each other inside cells quicker than existing methods, enabling speedier discovery of new drug combinations to treat illnesses like cancer, but some researchers say it may not translate into results that will be useful in humans.
The system, developed by Facebook AI Research and the Helmholtz Centre in Munich, Germany, is claimed to be the first easy-to-use AI model able to estimate how different drugs will work in the body. It could speed up our ability to uncover new treatments for diseases like cancer. “Drug research often takes half a decade to develop a compound,” says Fabian Theis at the Helmholtz Centre, one of the authors of the work.
The model works by measuring how individual cells change in response to treatment from a particular set of drugs and recording those responses.
Such an approach could theoretically help tackle cancer tumours, which vary from person to person and react differently to the same treatment, says Eytan Ruppin at the US National Cancer Institute.
The AI factors in variables including the type of drug, what it is used in combination with, the dosage level, the time it is taken and the type of cell it targets. It can then use that information to predict the effect of drug combinations it hasn’t yet seen.
Read more: Facebook AI learned object recognition from 1 billion Instagram pics
The research team behind it says humans can’t make these kinds of predictions: if they were given a pool of 100 different drugs, and asked to choose five to be given in three different doses – not uncommon in cancer treatment – there could be 19 billion possible drug regimes.
The team tested the AI’s predictions against known combinations of drugs and found it was able to accurately forecast cell responses with over 90 per cent accuracy, says Theis. Unsurprisingly, the more drugs put into the model that the AI has seen before, the better its results.
The AI will be released as an open-source tool for the research community to use and develop. “A single model can be trained in a few hours on a single machine,” says David Lopez-Paz at Facebook AI Research.
Andrei Lupas of the Max Planck Institute for Developmental Biology, Germany, calls the results “very promising” but says more work is needed. “The usefulness of the method will now hinge on a rigorous testing under double-blind conditions,” he says.
Ruppin says he is concerned that the claimed results don’t match up to the hype. The AI doesn’t predict whether a cell will live or die, but rather it predicts the changes in the RNA that the cell expresses when treated with a drug. This can show how the interior of a cell responds, but not necessarily whether it will survive or be killed off by the treatment, he says.
He calls it an “important” first step in helping treat cancer, but points out that all the results are in vitro. “We have cured cancer one hundred times in salines and mouse models. They have shown nothing at all that is relevant to patients,” he says.
Journal reference: bioRxiv, DOI: 10.1101/2021.04.14.439903
More on these topics: