Researchers have developed a machine-learning (ML) framework to identify robust drug biomarkers and thereby, predict anti-cancer drug efficacy.
For successful treatment, it is crucial to identify molecular biomarkers to classify cancer patients according to drug sensitivity. Most biomarkers associated with drug response are found within patient cohort data. However, conducting clinical trials is extremely expensive and time-consuming. As a result, accurate discovery of robust biomarkers from preclinical models is becoming increasingly important.
Large-scale pharmacogenomic screenings of preclinical models have become very useful in the discovery of clinically relevant biomarkers. Additionally, studies have found that ML algorithms trained from preclinical model data are predictive of cancer patient drug responses. However, these tend to fail in predicting drug sensitivity within human tumours. Also, limited training data can impact the performance of ML techniques. Moreover, input heterogeneity can pose key challenges in most biological studies. Therefore, a method to reduce biological heterogeneity and select relevant features, while developing an efficient model for ML, is important in order to make robust predictions. Several studies have demonstrated that network-based approaches can reduce biological complexity and improve performance of ML methods to predict therapeutic outcomes.
Three-dimensional (3D) organoid culture models have been shown to recapitulate human tumours at both molecular and phenotypic levels. This further supports their use in drug biomarker discovery. Nonetheless, researchers have yet to determine a method that systematically identifies biomarkers from organoid models to predict drug responses.
In this study, published in Nature Communications, researchers integrated pharmacogenomic data derived from 3D organoid culture models and network-based methods to develop an ML framework for the prediction of patient-drug responses. To test the prediction performance of the model, the team analysed drug treatment responses of colorectal and bladder cancer patients to 5-fluorouracil and cisplatin, respectively.
Using this method, the team identified biomarkers that could accurately predict the drug responses of 114 colorectal cancer patients and 77 bladder cancer patients. They also confirmed their biomarkers using external transcriptomic datasets of drug-sensitive and drug-resistance cancer cell lines.
This work presents a potential method to predict cancer patient drug responses using pharmacogenomic data derived from organoid models. The team envision that the translation of predictive biomarkers identified in preclinical models to human tumours will continue to be an active area of research. Additionally, they expect that as organoid models become more sophisticated, biomarker discovery for cancer therapies will advance quicker.
Image credit: By vectorjuice – www.freepik.com