AI model may predict adverse events of combination therapies for various cancer types
Click Here to Manage Email Alerts
An artificial intelligence model appeared to predict adverse events associated with novel combination therapies for various cancer types, according to study results presented at American Association for Cancer Research Annual Meeting.
Rationale and methods
“In our lab, we aim to identify effective multitargeted therapies against cancer and translate the findings to the clinic,” Bart Westerman, PhD, associate professor in the department of neurosurgery at Brain Tumor Center Amsterdam, told Healio. “We found that many obstacles prohibit this translation, one of them being adverse events, which are caused by desired but also undesired targets of drugs. To predict the feasibility of approach in a real-life setting, we created a model to predict the adverse events of drug combinations.”
Westerman and colleagues pooled data from the FDA Adverse Event Reporting System (FAERS) database, which includes 15 million records of adverse events, and developed a method to predict the adverse events of drug combinations for the purpose of selecting those with mild adverse event profiles. They then fed the adverse event profiles into a convolutional neural network algorithm — machine learning that mimics the way human brains make associations between data.
Researchers provided unseen adverse event profiles of drug combinations to the model, which they dubbed “adverse events atlas,” to assess whether the model could recognize and decode them using the latent space descriptors.
Key findings
Results showed the model recognized the new patterns, indicating that measured combined profiles may be converted back into those of each agent included in the drug combination.
Moreover, the model accurately reviewed adverse event profiles for some of the most used combination therapies based on data from FAERS and the U.S. clinical trials database.
Researchers reported limitations of the study, including the potential difficulties in comparing the data with more sparse data and the limited application of the models to clinical practice until further validation.
Looking ahead
“We are currently extracting adverse event data from clinical records in our hospital. This adverse event data is hidden in unstructured text — yet another obstacle that prohibits our understanding of drug combinations in a real-life setting,” Westerman said. “We are using natural language processing and parsing to extract relevant information and, in the future, we aim to link these data to our adverse event prediction model that could have value in the clinic, when proven to do so.”
Westerman added that the researchers are validating the value of their findings by using other unseen data.
“We additionally want to link our model to data from the clinic, which is currently ongoing,” he said. “In parallel, we are building a prediction model to predict the effects of multitarget combinations based on clinical data and will use the adverse event data to better understand on and off targets of drugs in the context of therapy efficacy.”
Perspective
Back to Top
Zahra ‘Nasim’ Eftekhari, MS
Combination therapies are increasingly gaining attention in the era of precision medicine. Various costs are associated with an unsuccessful clinical trial — from potential cost to subjects’ health and care, to the monetary cost and lost time. Therefore, a data-driven approach to optimize selection of therapies with acceptable adverse events that have a high chance of being effectively combined is of great value. With further evaluation, the proposed approach shows promise in predicting adverse events to narrow down the countless list of possible combinations.
The fact that conversion of learned pattern information to the original adverse event profiles showed a high similarity to unseen combination therapy effects is especially interesting. It would be nice to see prediction performance metrics, especially on the ability of the model to predict rare or unexpected adverse events resulting from combination therapies.
The combination of a convolutional neural network and autoencoder is intriguing. It would be interesting to understand the clinical applicability of using a 2D convolutional neural network. The adverse events latent space likely does not have local structure and spatial correlations the way visual space does. One would assume these characteristics are needed for convolutions to work well. However, if the feature reduction step has indeed created an adverse events space with meaningful spatial correlations, this would be a very useful contribution.
The central step is to combine the learned representations of adverse events for multiple drugs. A simple sum, or AND gating, of binary vectors is an interesting approach that seems to capture the additive effect well. Further details are needed to understand how nonlinear synergistic effects can be captured. One avenue would be to characterize the interaction between different drugs by including a few hyperparameters and train the model with adverse event outcomes resulting from combination therapies.
Although further validation is needed for clinical application of this study, it provides a good framework for a data-driven approach to selecting candidate combination therapies for further testing. If this approach is validated, it addresses a difficult problem with a critical need. Machine-learning techniques can spot multidrug interactions in large data sets in a way that humans cannot. The overview figure also gives quite a tantalizing roadmap for the future, in which drug effects can be characterized by their effects on various levels in the hierarchy of the body, and synergistic effects can be identified by this hybrid approach of adverse events-data driven machine learning and physiological/anatomical prior knowledge.
Collapse
Kucukosmanoglu A, et al. Abstract 1908. Presented at: American Association for Cancer Research Annual Meeting (hybrid meeting); April 8-13, 2022.
Read more about
Click Here to Manage Email Alerts