Artificial intelligence is poised to revolutionize made fields, particularly those that directly affect human beings such as the healthcare sector and drug development industry.
For a long time, the process of coming up with new drugs for improving existing ones has involved a long process, which is one of the impediments that the industry faces. During this undertaking, chemists create and tweak molecules in a bid to boost the performance of an existing drug or to create a new one for treating a specific symptom or disease.
Since the drug development process requires a lot of time and the attempts usually result in a drug that that does not function as expected, MIT researchers have come up with a way to automate the exercise using machine learning.
Wengong Jin, an MIT PhD student based at the computer science and artificial intelligence lab, said that the project aims to substitute the unproductive human modification process of designing molecules with an automated one while assuring the validity of the molecules produced.
The research team at Massachusetts Institute of Technology (MIT) used 250,000 molecular graphs to train their machine learning model.
The graphs entail comprehensive images of a molecule’s structure. After training the model, the researchers then allowed it to not only generate molecules but also locate the ideal molecules to create off of and come up with new molecules with upgraded properties.
With that, they realized that their model had the power to complete such tasks more effectively compared to other systems that are designed for automating the process of drug design.
First, when required to produce new, effective molecules, each of the models made proved to be valid. That was an impressive finding, particularly for the MIT researchers since creating invalid molecules is a significant disadvantage that is associated with other drug design automation systems, particularly those that the team made a comparison to during the undertaking.
In fact, the best automation system only had a validity rate of 43.5 percent.
Secondly, when the model was tasked to locate the ideal base molecule, which is otherwise known as a lead molecule, it still outdid other automation systems. The base molecule is known for being easily synthesized and highly soluble.
In this case, the best molecule produced by the researcher ’s model managed to score about 30 percent higher on both desired qualities in comparison the best choice generated by more conventional systems.
Finally, when the model was requested to alter 800 molecules in an attempt to upgrade them for those qualities while keeping them similar to the lead molecule in terms of structure, it generated new, alike structured molecules nearly 80 percent of the time.
What’s more, the molecules were able to score higher, particularly for those both properties, than the original molecules.
Away from the impressive performance of the model, the team intends to evaluate the model on other pharmaceutical qualities. Besides, the MIT team aims at creating a model that can operate using a limited volume of training data.