MIT researchers developed the EquiBind deep learning model, which is 1,200 times faster than its counterparts to attach molecules to proteins to create drugs.

Before starting drug development, researchers must first find molecules that can “bind” with specific target proteins. However, this process requires significant financial and computational resources. In addition, scientists said it takes decades to develop and test a new drug, and 90% of discoveries fail.

According to the study’s lead author, Hannes Stark, current tying methods ligand with protein it’s like “trying to put a key in a lock with lots of keyholes”.

“Typical models take a lot of time and evaluate each “fit” before choosing the best option. In contrast, EquiBind directly predicts the exact position of a switch in one step, without prior knowledge of the protein’s target pocket, known as “blind matching.”

According to the researchers, the model has built-in geometric reasoning that helps it learn the underlying physics of molecules and make better predictions when faced with new, unknown data.

Pat Walters, Relay Therapeutics data director, suggested that scientists test the system on an existing drug and protein used to treat lung cancer, leukemia, and gastrointestinal tract tumors. According to him, the algorithm successfully bound ligands to proteins, which traditional methods had failed.

“EquiBind offers a unique solution to the placement problem involving both pose estimation and anchor location identification,” Walters says.

The researchers plan to present the algorithm at the International Machine Learning Conference (ICML). Stark stated that the team plans to gather more feedback on the system from industry experts to improve the system.

Recall that in March, an AI algorithm released more than 40,000 types of chemical weapons in six hours.

In November 2021, Alphabet Holdings launched a new AI-powered drug discovery company.

In July, DeepMind’s artificial intelligence modeled 20,000 human protein structures.