Google Looks To Deep Learning To Ramp Up Pace Of Drug Discoveries


Google Research and Stanford University have co-published a paper describing how deep learning combined with a broad array of medical datasets could increase the pace of new drug discoveries.

The paper, called “Large-Scale Machine Learning for Drug Discovery,” cites several early examples of pharmaceutical labs that have turned to deep learning neural networks to augment the cost-intensive and time-intensive processes that currently underpin drug discovery research. Currently, pharmaceutical companies start the research process by selecting a biomolecular target that the new drug will be designed to interact with. Next, tens of thousands of chemical compounds are created and tested to see what effect they have on the target. This process, known as high-throughput screening, can take years and cost millions of dollars depending on how difficult it is to isolate and measure the biomolecular target’s reaction. It is here that pharmaceutical companies have been experimenting with machine-learning systems as a potential substitute.

Initially, machine-learning systems were given very small datasets, focused entirely on a single disease and biomolecular target. These early experiments were theoretically successful, but prediction accuracy was low and the real-world use cases for the technology were limited. Now, Google has proposed that by combining all of these disease-specific data sets into a broad neural network and developing a far more complex deep learning algorithm, a more accurate high-throughput screening tool could be created.

Underpinning its research is an experiment Google conducted with the Pande Lab at Stanford, which was designed to measure the accuracy improvements gained as additional disease-specific data sets were added to the neural network. For this, Google built a neural network comprised of 37.8 million data points spanning 259 disease-specific datasets and 1.6 million chemical compounds. Next, algorithms were developed to process the data and identify the best chemical compounds to target specific biomolecular targets. As expected, these neural networks performed far better than the small-scale machine learning algorithms that are being used today.

The findings, though they did not lead to any direct medical discoveries, do have immediate real-world implications. While 259 datasets was enough to establish an improvement, the paper contends that far more data is needed to scale the system up for commercial purposes. The paper explains, “obtaining greater amounts of data is of critical importance for improving the state of the art. Major pharmaceutical companies possess vast private stores of experimental measurements; our work provides a strong argument that increased data sharing could result in benefits for all.”

While the use of deep-learning neural networks may indeed reduce the cost and time associated with drug discovery research, the process is still inherently complex and costly, and even though Google’s findings could potentially expedite one step of the process, the overall pathway for a drug to make it from the lab to the market is still going to be a long and costly one.

Enjoy HIStalk Connect? Sign up for update alerts, or follow us at @HIStalkConnect.

↑ Back to top

Founding Sponsors

Platinum Sponsors