Massively Multitask Networks for Drug Discovery

A recent paper published February 6, 2015 entitled "Massively Multitask Networks for Drug Discovery" provides us with a new learning framework to improve the speed and success rate of drug discovery synthesizing information from many sources. They found that more data covering more biological processes using multitask neural networks and large scale compute is a better and faster strategy for computational drug discovery.

Machine learning at scale accelerates drug discovery to treat diseases and improve human health. More evidence that data from multiple diseases can be leveraged with multitask neural networks to improve virtual screening effectiveness. Note the algorithms used are the same as Geoffrey Hinton's team used to win the Merck challenge, but with a data set 18x larger.

See paper here. 

Abstract

Massively multitask neural architectures provide a learning framework for drug discovery that synthesizes information from many distinct biological sources. To train these architectures at scale, we gather large amounts of data from public sources to create a dataset ofnearly 40 million measurements across more than 200 biological targets. We investigate several aspects of the multitask framework by performing a series of empirical studies and obtain some interesting results: (1) massively multitask networks obtain predictive accuracies significantly better than single-task methods, (2) the predictive power of multitask networks improves as additional tasks and data are added, (3) the total amount of data and the total number of tasks both contribute significantly to multitask improvement, and (4) multitask networks afford limited transferability to tasks not in the training set. Our results underscore the need for greater data sharing and further algorithmic innovation to accelerate the drug discovery process.