Phase III clinical drug trials are used to assess the treatment effectiveness of a new drug on a group of patients. A failed trial leads to the loss of a substantial amount of resources and experimental data, even though there might still be a subset of patients who did benefit from the drug. In addition, a trial might conclude that a drug has a statistically significant benefit, but this effect might be mostly driven by a subset of patients who considerably benefited from the drug, even though others did not. In both cases, it is important to identify these subsets of patients to better understand factors that may influence treatment benefit; this, in turn, can be used to predict the benefit for a given patient at the moment of diagnosis.
In cancer-related treatments, it has been demonstrated that the effectiveness of a drug is influenced by germline variants, which are mutations in a reproductive cell that get copied into every cell in the body. Using this knowledge, Jeroen de Ridder and colleagues propose the use of data corresponding to these variants to predict the treatment benefit of anticancer drugs for specific patients. Nevertheless, this comes with a challenge: germline variation datasets are extremely high-dimensional, which increases the risk of machine learning prediction models to be overtrained.
The authors developed RAINFOREST, which employs a random forest approach to overcome this issue, as random forests have been designed to avoid overtraining. Given that many different cancer treatment options are available, RAINFOREST focuses on predicting whether a patient would have a better survival or benefit outcome on the treatment of interest than on an alternative treatment. This is done by implementing the notion of survival difference to optimize the prediction in the desired direction. In their experiments with a dataset related to a metastatic colorectal cancer clinical drug trial, the authors showed that, while the trial had concluded that the addition of cetuximab did not show any benefits overall, a group of patients that made up 27.7% of the trial population did benefit from cetuximab. Using their approach, the authors also uncovered variants that could potentially explain the positive response to the treatment, and showed that men benefited from cetuximab more than women and that age was not a parameter that affected the impact of the drug. The advent of RAINFOREST is an important contribution to clinical research and allows for a more nuanced identification of drug benefits.
About this article
Cite this article
Rastogi, A. Inspecting treatment benefit in clinical drug trials. Nat Comput Sci 1, 97 (2021). https://doi.org/10.1038/s43588-021-00036-9