A major issue in cancer genomics has been distinguishing between harmless mutations and ‘cancer driver’ mutations that give rise to or exacerbate cancer. Many computational methods have been developed to tackle this challenge, spurring Niranjan Nagarajan of A*STAR’s Genome Institute of Singapore to compare the performance of different methods. “We’ve been excited about the concept that you can integrate diverse molecular profiling data for tumors and pinpoint the key alterations among thousands,” says Nagarajan.
Using data from 3,400 tumors spanning 15 cancer types, the team assessed 18 prediction methods which represented five different ways of approaching the problem. The performance of the five approaches varied, with differing results based on the implicit tradeoffs in each. The methods were generally robust to messy data and weren’t misled by fake mutations introduced by the researchers. However, they also failed to predict any drivers in 10 per cent of the patients.
Many of the methods predicted different drivers, which led the team to develop a consensus approach that integrated results from the different methods, known as ConsensusDriver, which consistently outperformed the individual methods. In addition, ConsensusDriver can determine whether unfamiliar mutations are cancer drivers.
“Most groups approach the question from a basic cancer biology perspective of finding key cancer genes. Our interest has always been to go beyond this and see how far we can move towards personalized medicine,” says Nagarajan. “If methods like ConsensusDriver were used for precision oncology, it would double the number of patients for whom we would be able to recommend treatments,” though he adds that precision oncology remains a distant goal requiring extensive clinical validation.
The team has made their research tools freely available online, providing the community with a convenient toolbox to run the 18 predictions methods as well as ConsensusDriver. With improved predictions from consensus-based methods, researchers may be able to find new drug targets by identifying genes which are frequently drivers across different cancer types.
Meanwhile, Nagarajan’s team is developing sophisticated machine-learning techniques to predict drug response in individual patients, as well as validating the predictions using cell cultures and exploring the value of single-cell approaches in predicting cancer drivers and treatments.
Source : A*STAR Research