What would you get if you cross ibuprofen with aspirin? No one is sure, but a team of MIT researchers headed by Alán Aspuru-Guzik has developed an artificial intelligence program that could provide an answer to this by recommending a molecular structure that amalgamates properties of both the medications.
The program could assist in finding new drug compounds. So far, researchers have relied on software that crawls through huge pools of candidate molecules with the help of rules crafted by chemists to predict or identify useful structures. This technique often requires humans, the precision of simulations, and the necessary processing power.
The new artificial intelligence is, however, independent and uses deep learning instead of lengthy simulations. It uses its own experience, created by training machine-learning algorithms with information on countless drug-like molecules.
Humans will turn out to better chemists with this program since it explores things intuitively using the chemical knowledge it acquires, similar to a chemist, said Aspuru-Guzik. The AI was trained on nearly 250,000 drug-like molecules and was able to generate conceivable new structures with the help of the knowledge of properties of current drugs and combining them. The software could even identify certain properties of molecules, such as solubility, for instance. According to Aspuru-Guzik, the chemical knowledge of the AI will be improved if it's fed more amount of data.
Vijay Pande from Stanford University hints that the software could be used widely by playing a role in optimizing or discovering drug candidates or other fields like catalysts or solar cells. One challenge that could crop up, according to Pande, is that the technique involved in teaching the software to learn chemistry. Researchers are yet to find the best data format that can be fed into the software. Text, speech, and images may not be a great way to encode chemical structures, he said.