The paper, called “Massively Multitask Networks for Drug Discovery,” reasoned that because drug discovery takes so long and has high rates of failure, using algorithms to perform some of the process is a much more efficient method. Machine learning has been implemented before in drug discovery — the authors cite a test performed by the Merck Kaggle competition in 2012, in which the teams used machine learning algorithms to test molecular compound models. The authors drew on a number of papers proposing similar methods.
Google conducted an experiment at Stanford’s Pande Lab involving 37.8 million data points across 259 disease data sets and 1.6 million chemical compounds. The team developed algorithms to identify which chemical compound performs best to each disease data set. Although the experiment did not lead to any direct medical discoveries, it showed much more accuracy than the current expectations of machine learning in drug discovery, according to the report.
“Although deep learning offers interesting possibilities for virtual screening, the full drug discovery process remains immensely complicated,” the authors concluded. “Can deep learning — coupled with large amounts of experimental data — trigger a revolution in this field? Considering the transformational effect that these methods have had on other fields, we are optimistic about the future.”
