Mobile Menu

Using ML to identify responders vs non-responders

We summarise a recent study, published as preprint in medrxiv, that explored the hypothesis that a data-driven analysis of a randomised control trial population can identify sub-types of patients.

Colorectal cancer

Colorectal cancer (CRC) is responsible for over half a million deaths globally every year. While current therapies can prolong median survival for metastatic patients, these results are only found in a third of patients. Evidence from randomised control trials (RCTs) has supported the idea that advanced CRC patients with mutated RAS/BRAF gene status are non-responders. This means that they do not benefit from treatment based on EGFR targeted antibodies, e.g. panitumumab.

Successful treatment tailored to specific patients requires patient sub-grouping. A common foundation of CRC sub-typing exists. However, it remains to be refined with further omics data and drug intervention clinical outcomes.

ML techniques

In RCTs, analysing differential response to experimental therapies is typically done by dividing patients into responders and non-responders. This is usually based on a single endpoint. In this study, it was hypothesised that an algorithmic approach can overcome the limitations of traditional responders analysis. Additionally, it was proposed that it would provide a framework for patient characterisation based on differential response to a treatment. Specifically, they engineered the raw response data from a RCT and mined the data using unsupervised clustering algorithms. From this, they were able to identify patterns of treatment response that could correspond to distinct patient subgroups. The study was based on data collected for a Phase III RCT, which investigated the effects of panitumumab in CRC patients.

They found that the Partition Around Medoids (PAM) clustering method resulted in the identification of seven sub-types of patients. Each were statistically distinct from each other in survival outcomes, prognostic biomarkers and genetic characteristics. Moreover, conventional responders analysis was proven inferior in uncovering relationships between physical, clinical history, genetic attributes and differential treatment resistance mechanisms.

These findings show that alongside improved characterisation of CRC molecular subtypes, applying ML techniques onto the wealth of data already collected by previous RCTs can support the design of further targeted, more efficient RCTs and better identification of patient groups who will respond to a given intervention.

Image credit: By starline –

More on these topics

Drug Response / Machine Learning / RCTs

Share this article