Facebook AI and Helmholtz Zentrum München, the German Research Center for Environmental Health, announced today the introduction of a new machine learning tool aimed at helping accelerate the discovery of effective new drug combinations for fighting disease and cancers.

Malignant tumors and complex diseases often require a combination of drugs, known as “drug cocktails,” that are tailor-made to fight them. These need to be specially formulated to attack multiple types of cells, prevent drug resistance and even deal with harmful side effects.

Today’s announcement is the release of an AI tool that predicts the effects of drug combinations using a model for dosages, timings and other types of interventions, such as gene knockout or deletions.

The model, known as Compositional Perturbation Autoencoder, is being released in open source with an easy-to-use application programming interface and Python package for developers. The work is detailed in a preprint on bioRxiv, the open-access preprint repository for research papers, and the researchers plan to submit it to journals.

The essence of the model is that it uses historical observational data about drug combinations on cell types in order to predict drug behavior at the molecular level, such as different dosages and timings. CPA can then use a novel self-supervision technique to observe cells treated with a finite number of drug combinations and predict the effect of unseen combinations.

Image: Facebook AI

With CPA pharmaceutical researchers have a tool that can guide experimental processes to help narrow down billions of potential choices that originally would have taken months or years into simulations that take mere hours.

Then, out of those simulations, they could render combinations of 100 drugs and doses with the highest hypothetical potential best outcomes and run them against in vitro cell-lines to see how they work in the real world.

“Our field has been successful in putting together cell atlases for different organs,” said Dr. Fabian Theis, director of the Institute of Computational Biology at Helmholtz Munich. “This search space – across cell types, drug combinations as well as patient variation – is incredibly large, and can never be explored in full experimentally, so machine learning is crucially needed here.”

The vision of Facebook AI and Helmholtz Munich is to provide researchers and biologists the tools they need to prototype and simulate drug combinations rapidly using the data they already work with on a daily basis in simulation. The tool is open source and does not require ML expertise.

The tool could be put to use fighting diseases such as COVID-19, complex diseases such as cancer – for example, fine-tuning chemotherapy or immunotherapy cocktails – and other interventions for diseases that require specialized drug treatments.

Even further in the future, the machine learning tool could open the way for personalized medicine tailored to individual cell responses, a challenging issue at the forefront of medicine.

Image: Pixabay