One of the basic shared goals of most chemistry researchers is the need to predict molecular properties, such as boiling and melting points. Once researchers are able to identify their predictions, they can move forward with their work that brings discoveries that lead to drugs, materials, and more. Historically, however, traditional ways of publishing these forecasts have been associated with significant costs of spending equipment time and wear in addition to funding.
Enter the branch of artificial intelligence known as machine learning (ML). Although ML has reduced the burden of predicting molecular properties to some extent, advanced tools that most effectively drive processes by learning from existing data to make rapid predictions of new molecules require users to have a considerable level of programming expertise. This creates an accessibility barrier for many chemists who may not have the critical computing power needed to navigate their prediction pipeline.
To mitigate this challenge, researchers at MIT's McGuire Research Group created ChemXPloreml, a user-friendly desktop app that helps chemists make these important predictions without the need for advanced programming skills. Freely available, easy to download and work on mainstream platforms, the app is built to work completely offline and helps to hold research data. Exciting new technologies are outlined in a recently published article Journal of Chemical Information and Modeling.
One particular hurdle in chemical machine learning is the conversion of molecular structures into numerical languages that computers can understand. ChemXPloreml automates this complex process with a powerful, built-in “molecular embedding agent” that converts chemical structures into beneficial numerical vectors. The software then implements cutting-edge algorithms to identify patterns and accurately predict molecular properties such as boiling and melting points through an intuitive, interactive graphical interface.
“The goal of Chemxploreml is to democratize the use of machine learning in chemical science,” said Aravindh Nivas Marimuthu, postdoc and lead author of the article. “By creating intuitive, powerful, offline desktop applications, we place cutting-edge predictive modeling directly in the hands of chemists, regardless of the background in programming. This task not only makes the screening process faster and cheaper, but also promotes flexible design for future innovations, but also accelerates search for new drugs and materials.
ChemXPloreML is designed to evolve over time, so when future methods and algorithms are developed it will be seamlessly integrated into your app, allowing researchers to access and implement modern methods at all times. This application was tested with five important molecular properties of organic compounds such as melting point, boiling point, vapor pressure, critical temperature and critical pressure, achieving a high accuracy score of up to 93% at critical temperature. Researchers also demonstrated that the newer, more compact method of representing molecules (vicgae) is roughly as accurate as standard methods such as Mol2Vec, but up to 10 times faster.
“We envision a future in which researchers can easily customize and apply machine learning to solve their own challenges, from developing sustainable materials to exploring complex chemistry in interstellar space,” says Marimuthu. He will be participating in the papers as a senior author of career development in 1943 and an assistant professor at Brett McGuire in chemistry.
