AI-Powered Discovery of New Molecules Made Easier with Molecular Grammar
Machine Learning algorithms rely on large datasets for discovery. However, predicting molecular properties and generating new molecules requires a substantial amount of training data. Researchers from MIT have developed a solution to this problem by using a small dataset. They created a Machine Learning model called ‘Molecular Grammar’ that learns the language of molecules. This technique analyzes the grammar and information of a small dataset, identifying similarities between molecules with similar structures. The model uses Reinforcement Learning to understand the laws governing molecular similarity. The accuracy and f1 score of the model are impressive, bringing it closer to its goal.
Molecular Grammar consists of two parts: metagrammar and the hierarchical approach. It outperforms several other Machine Learning models, even with a smaller dataset. This technique is not only applicable to molecular datasets but also to graph-based datasets. It can be used for both regression and classification approaches. The research team made a significant breakthrough when they achieved better results by using only half of the training dataset.
The applications of Molecular Grammar are vast and include predicting the physical properties of glass transition temperature. The researchers aim to expand their model to 3D molecules and polymers, leading to the discovery of new molecules and their properties.
For more information, you can check out the Paper and the MIT Article. Stay updated with the latest AI research news, projects, and more by joining our ML SubReddit, Discord Channel, and Email Newsletter. If you have any questions or suggestions, feel free to reach out to us at Asif@marktechpost.com.
Also, don’t forget to explore the 800+ AI Tools in the AI Tools Club!
[Sponsored] Looking for some amazing features? Check out StoryBird.ai, where you can generate illustrated stories from prompts. Check it out here.