Machine-learning robot could streamline drug development
Testing out newly developed drugs is anextremely time-consuming process, and it can be difficult to getright. Now, a team of scientists at Carnegie Mellon University (CMU)is working to streamline the task, creating a robotically-drivenexperimentation system that's able to reduce the number of teststhat have to be carried out by as much as 70 percent.
When working on a new drug, scientistshave to determine its effects to ensure that it's both an effectivetreatment and not harmful to patients. This is hugely time-consuming,and it's simply not practical to performexperiments for every possible set of biological conditions.
That's where CMU's new robotic systemsteps in. It uses a machine learning approach to choose whichexperiments to conduct, using patterns in the data to accuratelypredict results of experiments without actually carrying them out.
The system is able to conduct selected experiments on its own, using liquid-handling robots and anautomated microscope. Its abilities were put to the test in a study todetermine the effects of 96 drugs on 96 cultured mammalian cellclones, containing different, fluorescently-tagged proteins. A total of 9,216 experiments were possible, each ofwhich involved testing the effects of a drug by taking a picture ofit mixing with the target cell.
The machine began by imaging all 96cells, pinpointing the location of the protein within it. The effects ofeach drug were then recorded in the same way, with the machinelearning algorithm slowly identifying patterns in the location of theproteins, known as phenotypes.
By grouping together similar images,the machine learner was able to identify potential new phenotypeswithout help from the researchers. As more data was gathered, it wasused to form a predictive model, guessing the outcomes of unmeasuredexperiments.
A total of 30 rounds of testing wereundertaken by the automated system, with 2,697 experiments completedout of the possible 9,216. The rest of the outcomes were predicted bythe machine, to an impressive accuracy rate of 92 percent.
The researchers believe that their workproves that machine learning techniques are viable for use in medicaltesting, and could have a big impact on both the practical andfinancial issues faced by the field.
"The immediate challenge will be touse these methods to reduce the cost of achieving the goals of major,multi-site projects, such as The Cancer Genome Atlas, which aims toaccelerate understanding of molecular basis of cancer with genomeanalysis technologies," said senior paper author Robert F. Murphy.
The findings of the research werepublished online in the journal eLife.