Patient datasets often contain a plethora of ‘-omics’ data, but the corresponding drug response information is limited and not suitable for drug discovery. Using integrated in vitro high throughput screening data and patient breast tumour molecular information, researchers have developed a virtual drug response modelling technique that enables drug discovery, with simultaneous biomarker identification for a patient population such as triple negative breast cancer.
Triple negative breast cancer (TNBC) drug development
TNBC is a defined subset of breast cancers that are inherently difficult to treat. This is because they are defined by the absence of a distinct molecular target. Research has shown that out of all breast cancer patients, those with TNBC have the worst 5-year survival rate. Most TNBC patients are still being treated with cytotoxic chemotherapies. There is a clear need to identify new therapies for TNBC to help reduce mortality in these patients.
Traditional drug development pipelines are time-consuming, and it takes a number of years for target identification, validation, and design optimization of the lead candidate compounds. These traditional methods are indispensable in the generation of new therapeutic compounds. However, to explore the potential of using existing drugs to treat cancer new approaches are needed.
In vitro drug response modelling
To test existing drugs for phenotypic changes in cancer cell lines in vitro screening can be used. For use in precision medicine, in vitro screening data can be used as inputs for machine learning models to obtain predictions of patient drug response. The researchers behind this study had previously developed an approach to predict drug response in patients. This method was shown to be accurate in clinical studies. They built predictive models between baseline gene expression values in cell lines and their drug efficacy metrics.
Generating their in vitro response model
The researchers used expression data from The Cancer Genome Atlas (TCGA) and Cancer Cell Line Encyclopaedia (CCLE). They filtered the data for common genes between the two datasets and then integrated it using ComBat. Next, they performed feature selection by removing 20% of genes with the lowest variation in gene expression across the samples. They then power transformed AUC values from The Broad Institutes Cell Therapeutics Response Portal (CTRP), and fit a linear regression model between the CCLE gene expression and corresponding cell line AUC values for every drug independently. After the models were fit, they inputted the homogenized TCGA patient gene expression data into the models. This enabled the researchers to obtain a drug sensitivity estimate for each patient to every drug in CTRP.
Using their model to identify TNBC therapies
Moving forwards with this study, they aimed to alter the traditional drug response modelling paradigm and identify drugs targeted towards a particular patient population. More specifically, they began by looking at the patient population they would like to see responses in and tested for compounds predicted to target this patient population. They aimed to use their method to fill in the pharmacological data that is often missing from clinical patient datasets, providing a virtual drug screen of patients to hundreds of drug compounds. This will allow the identification of trends such as, imputed drug response, clinical features, and patient subtypes.
Applying their model to identify a potential treatment for TNBC, they built models based on cell line transcriptome data. They then applied their model to patient tumour data to obtain predictive sensitivity scores for hundreds of drugs in 1000 breast cancer patients. The researchers then examined the relationship between predicted drug response and patient clinical features.
Researchers identified a number of compounds-of-interest, including the Wee1 inhibitor AZD-1775. This drug is predicted to have preferential activity in TNBC. They validated this finding using independent cell line screening data and pathway analysis. The researchers then co-administered a TNBC xenograft mouse model with AZ-1775 and standard-of-care paclitaxel. This resulted in inhibition of tumour growth and increased survival.
In conclusion, this research has provided a framework to turn any transcriptomic dataset into a dataset for drug development. This method can be used to quickly generate meaningful drug discovery hypotheses for a cancer population of interest, such as TNBC.
Image Credit: FreePik