MIT’s AI Revolutionizes Material Development and Drug Discovery

MIT-Watson AI Lab has introduced a groundbreaking AI system that revolutionizes the process of drug discovery and material development by accurately predicting molecular properties with minimal data. By employing reinforcement learning to acquire a comprehensive understanding of “molecular grammar,” this system efficiently generates new molecules. Notably, even datasets containing fewer than 100 samples have yielded impressive results.

This AI system has the remarkable capability to predict molecular properties using only a small amount of data, offering a potential acceleration in drug discovery and material development.

Traditionally, the discovery of new materials and drugs involved a time-consuming and costly trial-and-error approach that spanned decades. To streamline this process, scientists have turned to machine learning to predict molecular properties, helping to narrow down the molecules that require synthesis and laboratory testing.

In a collaborative effort between MIT and the MIT-Watson AI Lab, researchers have developed a unified framework that outperforms popular deep-learning approaches by simultaneously predicting molecular properties and generating new molecules in a highly efficient manner.

To teach a machine-learning model to predict the biological or mechanical properties of a molecule, researchers typically expose it to millions of labeled molecular structures, a process known as training. However, due to the expense and challenge of hand-labeling such a vast number of structures, large training datasets are often difficult to obtain, limiting the effectiveness of machine-learning approaches.

In contrast, the system developed by MIT researchers can accurately predict molecular properties using only a small amount of data. It possesses an inherent understanding of the rules governing the combination of building blocks to produce valid molecules. These rules capture the similarities between molecular structures, enabling the system to generate new molecules and predict their properties with remarkable efficiency.

This method has proven superior to other machine-learning approaches, delivering accurate predictions and generating viable molecules, even when presented with datasets containing fewer than 100 samples.

Lead author Minghao Guo, a graduate student in computer science and electrical engineering (EECS), explains, “Our goal with this project is to utilize data-driven methods to expedite the discovery of new molecules, enabling us to train a model to make predictions without relying solely on expensive experiments.”

Guo’s co-authors include Veronika Thost, Payel Das, and Jie Chen from the MIT-IBM Watson AI Lab research staff, recent MIT graduates Samuel Song ’23 and Adithya Balachandran ’23, and senior author Wojciech Matusik, a professor of electrical engineering and computer science and a member of the MIT-IBM Watson AI Lab, who leads the Computational Design and Fabrication Group within the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). The researchers plan to present their findings at the International Conference for Machine Learning.

Mastering the language of molecules

To achieve optimal results with machine-learning models, scientists require training datasets consisting of millions of molecules exhibiting properties similar to those they aim to discover. However, in reality, these domain-specific datasets are often limited in size. To overcome this challenge, researchers employ pretrained models trained on large datasets of general molecules, which are then applied to smaller, targeted datasets. However, these models lack significant domain-specific knowledge, leading to poor performance.

The MIT team took a unique approach by developing a machine-learning system that learns the “language” of molecules, referred to as molecular grammar, using only a small domain-specific dataset. This grammar is utilized to construct viable molecules and predict their properties.

In language theory, grammar rules are employed to generate words, sentences, or paragraphs. Similarly, a molecular grammar consists of production rules that govern the combination of atoms and substructures to generate molecules or polymers.

Just as language grammar can generate a vast array of sentences using the same rules, a molecular grammar can represent an extensive range of molecules. Molecules with similar structures share the same grammar production rules, and the system learns to recognize these similarities.

Given that structurally similar molecules often possess similar properties, the system leverages its inherent knowledge of molecular similarity to predict the properties of new molecules more efficiently.

Guo explains, “Once we have this grammar as a representation for all the different molecules, we can use it to expedite the property prediction process.”

The system learns the production rules of molecular grammar through reinforcement learning, a trial-and-error process in which the model receives rewards for behavior that brings it closer to achieving a goal.

However, considering the countless ways atoms and substructures can combine, learning grammar production rules would be computationally expensive when applied to anything but the smallest datasets.

To address this, the researchers divided the molecular grammar into two parts. The first part, called a metagrammar, is a general and broadly applicable grammar designed manually and provided to the system initially. The system only needs to learn a much smaller, molecule-specific grammar from the domain dataset. This hierarchical approach significantly accelerates the learning process.

Small datasets, significant outcomes

In experiments, the researchers’ new system successfully generated viable molecules and polymers while predicting their properties more accurately than several popular machine-learning approaches, even when utilizing domain-specific datasets containing only a few hundred samples. Unlike other methods, the new system does not require a costly pretraining step.

The technique excelled particularly in predicting physical properties of polymers, such as the glass transition temperature, which typically necessitates expensive experiments involving extremely high temperatures and pressures.

To further advance their approach, the researchers reduced the size of one training set by more than half, containing only 94 samples. Remarkably, their model achieved results comparable to those trained using the complete dataset.

Guo emphasizes the power of grammar-based representation, stating, “This grammar-based representation is very powerful. Moreover, since the grammar itself is a highly versatile representation, it can be applied to different types of graph-form data. We are exploring other potential applications beyond chemistry or material science.”

In the future, the researchers aim to expand their current molecular grammar to incorporate the 3D geometry of molecules and polymers, which plays a crucial role in understanding the interactions between polymer chains. They are also developing an interface that would display the learned grammar production rules to users, soliciting feedback to rectify potential errors and enhance the system’s accuracy.

Reference: Grammar-Induced Geometry for Data-Efficient Molecular Property Prediction

This work is partially funded by the MIT-IBM Watson AI Lab and its member company, Evonik.

Frequently Asked Questions (FAQs) about AI-driven drug discovery

More about AI-driven drug discovery

• MIT News: AI Learns Molecular Language for Rapid Material Development and Drug Discovery
• MIT-IBM Watson AI Lab
• Computational Design and Fabrication Group at MIT
• International Conference for Machine Learning (ICML)