Like clothes and shoes, medicines aren’t one-size-fits-all; our genes can affect how our bodies respond to the same drugs. Today, biomedical advances are giving clinicians more insight into these nuances—known as pharmacogenomics—and enabling treatments tailored to a patient’s DNA.
However, current approaches to pharmacogenomics rely on datasets drawn from highly uniform lab-grown cell lines, which often inaccurately represent the varied tumour cells seen in patients, explained Shumei Chia and Ramanuj DasGupta, respectively a Research Fellow and a Senior Group Leader at the A*STAR Genome Institute of Singapore (A*STAR GIS).
“These datasets also involve drug sensitivity tests using dosages often not used in the clinic,” Chia and DasGupta added. “Given their lack of biological and clinical relevance, any findings based on these datasets could be limited in their translatability.”
To address those limitations, Chia, DasGupta and A*STAR GIS colleagues including Iain Tan, Principal Investigator, and Niranjan Nagarajan, Associate Director and Senior Group Leader, worked with international collaborators from institutes such as the National Cancer Centre Singapore; Singapore General Hospital; Siriraj Medical Centre, Thailand; Katholieke Universiteit Leuven, Belgium; Yonsei University College of Medicine, South Korea; and the University of Geneva, Switzerland.
Together, the team developed CAN-Scan: a machine learning (ML) precision oncology platform designed to discover therapeutic targets and response biomarkers for cancer therapies at an individual patient level.
“CAN-Scan is phenotype-driven: it’s built on multi-omic datasets generated from patient-derived cell lines (PDCs), which preserve the source tumour’s biology. These datasets include PDC response data to clinically-relevant drug concentrations,” said Chia.
In developing CAN-Scan, the team started with colorectal cancer (CRC) as a proof of concept. “Despite advances in oncology, patients with CRC today still often receive the standard 5-FU-based chemotherapy, which doesn’t account for the diversity of patient tumours,” Chia and DasGupta explained.
The team built a biobank of 47 CRC PDCs and tested them against 84 FDA-approved drugs to produce a unique pharmacogenomic database. This was then used to train CAN-Scan’s ML models to identify tell-tale molecular signatures of drug resistance and sensitivity in cancer cells. The models’ predictive abilities were validated with the Cancer Genome Atlas and three other independent patient cohorts from Thailand, Korea and Belgium.
Via CAN-Scan, the team identified 11 genes that, with increased expression, were associated with increased 5-FU resistance and poorer outcomes for patients with CRC. They noted that 10 of these genes were clustered on a section of Chromosome 7 also known to harbour cancer-promoting genes, indicating a potential co-amplifying effect.
CAN-Scan’s models also predicted that patients resistant to 5-FU were likely to respond to alternative targeted therapies such as regorafenib, matching the findings of a related clinical trial. “This opens new avenues for personalised therapies based on the genetic makeup of an individual patient’s tumour,” said DasGupta.
Looking ahead, the team aims to uncover novel causes of drug resistance and therapeutic targets, paving the way for alternative strategies against treatment-resistant cancers.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Genome Institute of Singapore (A*STAR GIS).
