New AI Method Predicts Cancer Drug Response Using Genomics
AI machine learning algorithm advances precision oncology and pharmacogenomics.
Posted November 12, 2021 | Reviewed by Kaja Perina
A new study published in Nature Machine Intelligence shows an artificial intelligence (AI) machine learning algorithm that can predict cancer drug responses based on gene expression data—demonstrating state-of-the-art pharmacogenomics that may improve the accuracy of precision oncology.
The global market for oncology therapeutics sales is projected to reach USD 250 billion by 2024 with a compound annual growth rate (CAGR) of 12 percent according to McKinsey & Company. An estimated 20 percent of the worldwide pharmaceutical sales, USD 143 billion, was from branded cancer pharmaceutical sales in 2019 based on figures from Evaluate Ltd.
An individual’s DNA can impact whether responses to medication are effective or not. The quantity and type of drug receptor proteins on the surface of cells is determined by a person’s DNA which may impact drug responses. A person’s DNA can impact drug uptake, breakdown, and removal from tissues and cells.
Pharmacogenetics is the study within precision medicine that looks at how variation in a one or more genes can lead to different responses to medications in order to create targeted, more effective drug therapies based on the individual’s genetic information. Pharmacogenomics may also study genetic variation in populations in an effort to understand how race and ethnicity may impact drug responses. In pharmacogenomics, drugs are developed to treat the problem, rather than the symptoms. Both pharmacogenetics and pharmacogenomics aim to provide better, more personalized medicine by customizing therapeutics based on a patient’s individual genetics.
In AI machine learning, massive datasets are needed for training the algorithm. For precision oncology, there is a general lack of publicly-available patient datasets with drug responses. If they are publicly available, these datasets tend to be small in size. To address this dataset issue, preclinical datasets using patient-derived xenografts (PDX), or cancer cell lines may be used as proxies for patients. Patient-derived xenografts are tissue sections or cells that are removed from one species and grafted onto a different one. In personalized oncology, xenografts may be done from samples from the patient to a mouse in order to determine a custom treatment plan. Cancer cell lines are in vitro models systems used to study cancer biology.
However, using preclinical datasets has its drawbacks such as the lack of tumor and immune system microenvironment, and generating massive high-quality datasets is laborious, time-consuming, and costly. Furthermore, these do not address the target domains of future patients that may come to the clinic, thus the drug response for future patients is unknown when algorithms are trained on preclinical datasets.
To address this challenge, the researchers affiliated with Simon Fraser University, Vancouver Prostate Center, Technical University of Munich, the University of Hamburg, and the University of British Columbia created a deep neural network called Velodrome.
Velodrome uses both labeled data of gene expression from cell line and unlabeled data from patient datasets in order to predict drug responses.
“To the best of our knowledge, Velodrome is the first method for semi-supervised out-of-distribution generalization from labeled cell lines and unlabeled patients to different preclinical and clinical datasets,” the researchers wrote.
According to the scientists, their deep neural network achieved state-of-the-art performance for multiple drugs across various preclinical and clinical pharmacogenomics datasets, enabling more accurate precision oncology.
Copyright © 2021 Cami Rosso All rights reserved.