AI will invent drugs much faster than humans

To create a new medicine, today scientists have to test tens of thousands of components to understand how they interact with each other. And this is not the most difficult part. Once a certain substance is found effective against the disease, it will have to go through three more different phases of clinical trials and get regulatory approval.

It is estimated that on average, one new drug entered the market, one needs 1,000 people, 12-15 years, and about 1.6 billion dollars. It seems that there should be a better way – and he is believed to have appeared. Last week, scientists published a paper detailing the artificial intelligence system designed to help find new drugs. It should significantly reduce the amount of time and money spent in this process.

The system is called AtomNet and created its start-up from San Francisco under the name AtomWise. The technology is aimed at rationalizing the initial phase of drug discovery, which involves the interaction of various molecules among themselves, in particular, scientists need to determine which molecules will bind and how much. They use the method of trial and error, scanning tens of thousands of components, both natural and synthetic.

AtomNet reduces this process using deep training methods to predict how molecules behave and how likely they form bonds. The software learns molecular interaction by recognizing patterns, just as AI learns to recognize images.

Remember the three-dimensional models of atoms that many make in high school from foam and tubes to represent the connections between protons, neutrons and electrons? AtomNet uses similar three-dimensional molecular models, including data on their structure, to predict their biological activity.

As AtomWise’s chief operating officer Alexander Levy says, “one can take the interaction between a drug and a large biological system and decompose it into smaller interactive groups. If you study enough historical examples of molecules, you can quickly make accurate predictions. ”

“Fast” can even be an understatement. It is reported that AtomNet can span a million connections per day. Applying modern methods, this would take months.

AtomNet can not invent a new drug or even say for sure whether a combination of two molecules will be an effective medicine. But he can predict how likely a certain compound will work against a particular disease. Scientists use these forecasts to narrow down thousands of options to dozens, to focus their testing where positive results are most likely.

This software has already proven itself, helping to create new drugs for the treatment of Ebola and multiple sclerosis. The last drug was licensed by a British pharmaceutical company, and the drug against Ebola was submitted to a peer-reviewed journal for further analysis.

Although AtomNet is a promising technology that will accelerate the discovery of new drugs, it is worth noting that the future of medicine is also moving towards a proactive rather than reactive approach; Instead of inventing drugs simply for the treatment of sick people, attention shifts to carefully monitoring the state of health and taking the necessary steps that will not let us get sick in the first place.

Last year, the Zuckerberg Foundation donated $ 3 billion to find “a cure for all diseases.” This is an ambitious and somewhat quixotic goal, which nevertheless deserves respect. In another example of a move toward proactive health, the XPRIZE Foundation recently awarded $ 2.5 million to a device designed to facilitate home-based diagnostics and personal health monitoring. Proactive health technology is likely to evolve and grow in popularity.

But this does not mean that reactive medical care will remain in place. In fifty or a hundred years, people will still be sick and need a medicine that will cure them. AtomNet is the first of its kind software. But very soon there will be other methods of using artificial intelligence on this path.

Notify of
Inline Feedbacks
View all comments
Would love your thoughts, please comment.x