Deep Learning Advances May Pave Way to AI-Made Drugs

Researchers in the field of artificial intelligence may pave way to AI that can create new drugs by learning how to mix and match molecules in effective patterns. This project may hold new applications for deep learning models as this may prove just how well they work in identifying patterns in complex natural structures.

Harvard chemistry professor Alan Aspuru-Guzik said humans could be better chemists if they have AI assistants. The system aims to allow a deep neural network to "explore intuitively" using chemical knowledge, like how chemists would. The system has a database with thousands of drug molecules and, using deep learning, will try to work out and see what fits with patterns.

This is called a generative model and is already used in autocomplete features. He said this is similar to an image recognition database, where if used with records from the American Chemical Society, could use about 100 million chemical structures to find potential new drugs.

According to the MIT Technology Review, the software will exhaust giant pools of candidate molecules using rules from chemists. The simulations will try and identify useful structures. This can generate plausible structures by combining properties of existing compounds and can be asked to suggest molecules that strongly displays certain properties. A new drug from an AI may not be too far off.

This project is the latest in a line of AI automating jobs that hopes to not replace chemists. In the field, AI can help professionals spend more time focusing on human-oriented jobs. This means there is a need for human oversight for less of the workload.

A good example of this is Luminance's "robot lawyer AI" which, as Inverse reported, focuses on mergers and acquisitions. It checks through documents and flags potential clashes. In the same regard, IntelligentX helps brewers collate beer feedback and even suggest new recipes.