Facebook today detailed what it claims is the first single AI model capable of predicting the effects of drug combinations, dosages, timing, and other types of interventions like gene deletion. Developed in collaboration with Helmholtz Zentrum München, Facebook says the model could accelerate the process of identifying combinations of medications and other treatments that might lead to better outcomes for diseases.
Discovering ways to repurpose existing drugs has proven to be a powerful tool to treat diseases including cancer. In recent years, doctors have seen success with “drug cocktails” to combat malignant conditions and continue to explore personalized treatments for patients. But finding an effective combination of existing drugs at the right dose is extremely challenging, in part because there are nearly infinite possibilities. Researchers would have to try from 5,000 to 19 billion solutions to find the optimal regimen given a pool of 100 drugs.
Facebook’s open source model — Compositional Perturbation Autoencoder (CPA) — ostensibly addresses this with a self-supervision technique that observes cells treated with drug combinations and predicts the effect of new combinations. Unlike supervised models that learn from labeled datasets, Facebook’s generates labels from data by exposing the relationships between the data’s parts, a step believed to be critical to achieving human-level intelligence.
CPA’s predictions take hours as opposed to the years that might elapse with conventional methods, allowing researchers to select the most promising results for validation and follow-up, according to Facebook.
In biology, RNA sequencing is used as a way to measure the gene expressions of cells at the molecular level and study the effects of perturbations including drug combinations. Academia and industry have released RNA sequencing datasets containing up to millions of cells and 20,000 readouts per cell to facilitate biomedical research.
Facebook leveraged these datasets to train CPA using an approach called auto-encoding, in which data is compressed and decompressed until summarized into patterns useful for prediction. CPA first separates and learns the key attributes about a cell, such as the effects of a certain drug, combination, dosage, time, gene deletion, or cell type. It then independently recombines the attributes to project their effects on the cell’s gene expressions. For example, if one of the datasets had information on how drugs affect different types of cells A, B, C, and A+B, CPA would learn the impact of each drug in a cell-type specific fashion and then recombine each in order to extrapolate interactions between A+C, B+C, and A+B.
To test CPA, Facebook says it applied the model to five publicly available RNA sequence datasets with measurements and outcomes of drugs, doses, and other confounders on cancer cells. Benchmarked in terms of the R2 metric, which represents the accuracy of the gene expression predictions, Facebook claims that CPA “stayed consistent” between training and testing — an indication of robustness. Moreover, CPA’s predictions of the effects of drug combinations and doses on cancer cells matched those found in the testing dataset “reliably.”
Facebook believes that CPA can “dramatically” accelerate the process of identifying optimal combinations of treatments, as well as pave the way for new opportunities in the development of medications. Toward this end, the company is making available APIs and a software package designed to let researchers plug in datasets and run through predictions.
“Our hope is that pharmaceutical and academic researchers as well as biologists will utilize [CPA] to accelerate the process of identifying optimal combinations of drugs for various diseases,” Facebook program manager Anna Klimovskaia and research scientist David Lopez-Paz wrote in a blog post. “In the future, [CPA] could not only speed up drug repurposing research, but also — one day — make treatments much more personalized and tailored to individual cell responses, one of the most active challenges in the future of medicine to date.”
While Facebook claims that CPA is novel in its architecture, it isn’t the first algorithm engineered to predict drug interactions. In July 2018, Stanford researchers detailed an AI system that can anticipate the effects of drug combinations by modeling the more than 19,000 proteins in the body that interact with each other and with medications. Researchers at the MIT-IBM Watson AI Lab, Harvard School of Public Health, Georgia Institute of Technology, and IQVIA more recently created an AI tool called CASTER that estimates potentially harmful and unsafe drug-to-drug interactions. A separate Harvard group has proposed applying AI to identify candidates for drug repurposing in Alzheimer’s disease. And researchers at Aalto University, University of Helsinki, and the University of Turku in Finland created a machine learning model that projects how combinations of drugs might kill various cancer cells.
