One of the greatest challenges of current medicine is predicting how a patient will respond to a given drug. In an ideal world, where time, money and – most importantly – the patient’s well-being and survival are not an issue, we would simply either keep trying different treatments until hitting the jackpot, or perhaps harvest the patient’s cells and try a range of treatments in vitro. The problem is, of course, that the world is not ideal and such in vitro testing is usually not practical and, in general, especially in the case of many of the most debilitating diseases, patients often don’t have time to waste.
It is then not really surprising that many researchers have been trying for oh-so-many-decades to come up with efficient methods that can predict chemotherapeutic response. One of the most common ways of doing this at the moment is to use results from the given therapeutic’s clinical trial to construct a predictive gene panel. And these gene panels are effective, to a degree. One limitation, however, is that most panels are drug-specific, which makes them not that helpful in the great high-throughput-y scheme of things.
And some of the most interesting, and sometimes most scandalous, efforts have been directed at solving this problem in the last decade. To start with, a number of studies by Anil Potti published in the mid-2000s claimed to be able to predict clinical outcomes based on gene expression patterns. This enthusiasm, unfortunately, turned out to be a bit premature – a large number of Potti’s articles have since been retracted following allegations of irregularities.
More recently the topic returned in a different setting: one of the ‘DREAM’ challenges for computational biologists announced in 2012 was dedicated to finding a way of using genomic information to estimate sensitivity of cancer cell lines to a number of drugs. In general, projects such as the National Cancer Institute cancer sensitivity screen, which looked at the drug response in 60 cancer cell lines, or the Sanger Institute’s Genomics of Drug Sensitivity in Cancer (GDSC) effort, which looks at over a thousand cancer cell lines and their response to a few hundred or so therapeutics, are great resources for researchers trying to come up with a generic machine learning method capable of predicting clinical outcomes.
Recently, one such method was published by Julio Saez-Rodriguez and colleagues in PLOS ONE: these authors constructed a model which learns not only from GDSC drug sensitivity data, but also uses information on the effectiveness of therapeutics with a similar mechanism of action to predict in vitro drug sensitivity. While this model works extremely well for predicting sensitivity in cell lines, its application in a clinical setting is limited, since there usually is no or very little knowledge of the patient’s response to similar drugs.
Which is where a recent publication in Genome Biology by Stephanie Huang, Nancy Cox and Paul Geeleher comes in: in this article published last week the authors describe a model, also trained on the GDSC in vitro gene expression data, which predicts drug sensitivity in vivo. The model uses the patient’s baseline gene expression data and is tested on three independent clinical trial cohorts (docetaxel for breast cancer, bortezomic for myeloma, and erlotinif for NSCLC). For each of these trials, a prognostic gene panel has been constructed based on in vivo response to the drug. And yet, Huang et al. model is as efficient – and in some cases even more efficient – in predicting the outcomes, as these specific panels.
The Genome Biology article is only the beginning of the road to the clinical application of Huang et al.’s method. But it is a good and promising start, and we shall be looking forward to seeing what our readers make of it.