Teaching artificial intelligence Chemistry so that it can find new drugs

What if you mix Aspirin and Ibuprofen together? Both drugs have analgesic and antipyretic effects. Professor Aspuru-Guzik from Harvard University is also not sure . But he could ask a special assistant. It is an artificial intelligence system that can effectively support people in the search for new drugs.

Today, the most advanced pharmaceutical research is based on software platforms and computer simulations.

Software is actually written down from the rules discovered by chemists. It is used to collect data from a huge repository, containing potential molecules that can be used to make drugs. Meanwhile, computer simulations are used to predict useful molecular structures that have never existed in reality.

Teaching artificial intelligence Chemistry, it can create new drugs for humans.

Both of these tools have their own limitations. Software is limited by its original rules, sometimes not updated. Meanwhile, computer simulations are limited by the accuracy with which it is set. Pharmaceutical research still relies heavily on human factors.

On the way to find new tools in this area, Aspuru-Guzik and his team have developed an artificial intelligence system that can take on the role of humans. Even if it is not possible to become a chemist, it will at least become an effective assistant with thinking and intuition of science.

Aspuru-Guzik's artificial intelligence system can "imagine" itself on molecular structures. This process does not depend much on people and does not require long simulations. The reason is that it can cultivate its own experiences from machine-learning algorithms and data of hundreds of thousands of molecules.

"It makes discoveries more intuitive, using the knowledge learned, just like what a chemist would do," Aspuru-Guzik said. "People can become more sophisticated chemists, if there is such a system as an assistant."

Professor Aspuru's system can produce billions and billions of molecules.

Aspuru-Guzik Professor's system is built, using a machine learning technique called "deep learning". This technique is now very popular in the technology industry, especially computer and computers. However, in the field of natural science, it is quite new.

Deep learning uses something like the generative model . In it, it is provided with a huge database and uses what is learned to create new data. In the field of informatics technology, this model is often used to create images, words or text. For example, Google's Smart Reply feature with email.

However, last month, Aspuru-Guzik and his colleagues at Harvard University, University of Toronto and Cambridge University proved that deep learning can go beyond its current limits. They published research findings that deep learning can be applied to both natural science research.

A deep learning model was created, and scientists trained it with 250,000 drug-like molecules. Later, this system can create new molecular structures by combining the properties of existing drugs. Not only that, it can also provide data about the properties of a new type of molecule, such as solubility or it can easily be produced in practice.

Can in future, artificial intelligence replace the role of scientists?

Vijay Pande, a Professor of Chemistry at Stanford University, said the new study has continued to show us: Artificial intelligence will change the field of natural science research. It suggests that a deep learning software can take chemical knowledge as input.

This will definitely help a lot of scientists."I think it can be widely applied , " Professor Pande said. "It can play a role in finding and optimizing potential drugs, or in many other areas such as solar energy or catalyst research."

Indeed, right now, researchers have continued to test this artificial intelligence system on a new database in material science. They train it with organic LED molecules, which make curved screens on electronic devices.

Although the potential is huge, researchers will have to continue to improve their systems. It must be trained to improve chemical skills, because some of the molecular structures that the system creates are evaluated by scientists as meaningless.

Pande said that a challenge to make software that can learn chemical knowledge is to have a good source of input data. This may be something that the research team has not done yet.

Images, speech, and text have been proven to be the perfect data sources for artificial intelligence systems. So the software today can recognize images, voices and translate too well the texts into different languages. However, when working with chemical structures, they have not yet shown their abilities.

Professor Aspuru-Guzik from Harvard University.

Aspuru-Guzik and his colleagues are also thinking about that. They continue to study more features to reduce the error rate of this artificial intelligence system. Aspuru-Guzik also hopes that more data can be added to his system, improving its chemical knowledge. This is similar to the image recognition feature that must be built on a database with millions of photos.

Now, looking at the field of natural sciences, Aspuru-Guzik has seen about 100 million published researches related to chemical structure. All of them are being stored by the American Chemical Society. In the future, he hopes to put all of this data into a version of his artificial intelligence program. If Aspuru-Guzik succeeds, we can hardly imagine what his "assistant" can do.