Artificial intelligence source data to accelerate drug discovery

0
23

A new cryptosystem allows pharmaceutical companies to collaborate with academic laboratories to develop new drugs faster without revealing any confidential data to competitors.
The core of this computing system is an artificial intelligence program called neural network. Artificial intelligence studies which drugs interact with various proteins in the body to predict new drug-protein interactions.
More training data produces a smarter AI, which has been a challenge in the past because drug developers often do not share data because of intellectual property considerations. Researchers reported in the October 19 issue of the Journal Science that the new system allows artificial intelligence to keep information secret while also pulling data sets together, encouraging collaboration to develop faster drugs.
Identifying new drug protein interactions can identify potential new therapies for various diseases. Or it could reveal whether the drug interacts with unexpected protein targets, which could indicate whether the drug may cause specific side effects, said Ivet Bahar, a computational biologist at the University of Pittsburgh who was not involved in the work.
In the new AI training system, data collected from the research team is assigned to multiple servers, and each server owner appears to see only random numbers. “This is where cryptography happens,” says computer scientist David Wu of the University of Virginia in Charlottesville, who was not involved in the work. Although no participant can see the millions of drug-protein interactions that make up the training set, servers can use this information together to teach neural networks to predict previously invisible drug-protein interactions.
“The work is far-sighted,” said Jian Peng, a computer scientist at the University of Illinois, who was not involved in the study. “I think it will lay the foundation for future biomedical cooperation.”

MIT computational biologist Bonnie Berger and his colleagues Brian Hie and Hyunghoon Cho assessed the accuracy of their systems by training neural networks of about 1.4 million drug protein pairs. Half of them were extracted from a suture database of known drug-protein interactions; the other half were non-drug-protein interactions. When new drug-protein pairs are found to interact, artificial intelligence picks the set of interactions with 95% accuracy.
To test whether the system can identify drug-protein interactions that have so far been unknown, Berger’s team trained neural networks to study nearly two million drug-protein pairs: a full-needle dataset of known interactions, plus the same number of non-interaction pairs. Well-trained AI provides some interactions that have never been reported before or have been reported, but are not in the stitching database.
For example, artificial intelligence has found that estrogen receptor proteins interact with a drug called droloxifene, which treats breast cancer. Neural networks have also found interactions between the leukemia drug imatinib and the protein erb4 thought to be associated with different types of cancer that have never been seen before. The researchers confirmed this interaction with laboratory experiments.
This secure computing network may also encourage more cooperation in areas other than drug development. Hospitals can share confidential health records to train artificial intelligence programs to predict patients’prognosis or formulate treatment strategies, Peng said.
“Whenever you want to study a large number of people about behavior, genomics, medical records, legal records, financial records — anything sensitive to privacy, these technologies are very useful,” Wu said.

LEAVE A REPLY

Please enter your comment!
Please enter your name here