Decoding the Molecular Language: Predicting Properties through Molecule-Driven Insights

Embarking on the Quest for Discovery can be a winding journey, filled with countless trials, errors, and a few million bucks. But fear not, for there’s nothing that AI and ML (Machine Learning) can’t mend! Until recently, scientists relied on the power of machine learning to forecast molecular properties and narrow down the molecules to synthesize and test in the lab.

But behold, a new solution has emerged! Thanks to the ingenious researchers at MIT and the MIT-Watson AI Lab, we now have a unified framework that not only predicts molecular properties but also churns out brand-new molecules with unrivaled efficiency.

The Strenuous Path of Training & Labeling in Molecular Language

Teaching a machine learning model to predict the properties of a molecule is not an easy task. Researchers have to expose the model to millions of labeled molecular structures in a process called training. However, this can be quite expensive and time-consuming. Plus, manually labeling millions of structures is a daunting challenge.

The real trouble arises when it comes to gathering large training datasets. These datasets are crucial for training the machine learning model effectively. But due to the high cost of discovering new molecules and the difficulties in hand-labeling a massive number of structures, obtaining large training datasets can be quite a challenge. And without ample training data, the performance of machine learning approaches is limited.

In a nutshell, the scarcity of labeled molecular structures hampers the effectiveness of machine learning techniques in predicting a molecule’s biological or mechanical properties. It’s a tricky situation that researchers are actively working to overcome, so we can unlock the full potential of machine learning in molecular research.

Why The MIT System Will Work?

The system possesses a deep understanding of the rules that govern the combination of building blocks to create valid molecules. By capturing the similarities between molecular structures, it becomes a master at generating new molecules and making data-efficient predictions about their properties.

What’s truly fascinating is that this method surpasses other machine learning approaches, whether the datasets are small or large. It works like magic, accurately predicting molecular properties and producing viable molecules even when given less than 100 samples to learn from. It’s a game-changer in the world of molecular research, paving the way for more efficient and effective exploration of new compounds.

Minghao Guo, a computer science and electrical engineering (EECS) graduate student and the lead author explains,

“Our goal with this project is to use some data-driven methods to speed up the discovery of new molecules, so you can train a model to do the prediction without all of these cost-heavy experiments.”

The MIT Team

Guo collaborated with a talented team of researchers for this study. His co-authors include Veronika Thost, Payel Das, and Jie Chen, who are research staff members at the MIT-IBM Watson AI Lab. The team also comprised of recent MIT graduates Samuel Song (’23) and Adithya Balachandran (’23). Leading the group is senior author Wojciech Matusik, a professor of electrical engineering and computer science, and a valued member of the MIT-IBM Watson AI Lab.

Decoding the Molecular Language

• Machine learning models require large training datasets of molecules with similar properties to achieve optimal results.
• In practice, domain-specific datasets are often very small, posing a challenge.
• Researchers resort to using pretrained models trained on large general datasets and apply them to smaller, targeted datasets.
• However, since these models lack domain-specific knowledge, their performance tends to be poor in such cases.

The Molecular Grammar – The Language of Molecules

The MIT team developed a machine learning system that learns the “language” of molecules, known as molecular grammar, using a small, specialized dataset. This system leverages the grammar to generate viable molecules and make predictions about their properties.

In the language theory, just as we generate words, sentences, or paragraphs using grammar rules, a molecular grammar functions similarly. It consists of rules that guide the combination of atoms and substructures to create molecules or polymers.

Like a language grammar that can generate countless sentences using the same rules, a molecular grammar can represent a vast array of molecules. Molecules with similar structures share common production rules within the grammar, and the system learns to recognize these similarities.

Since molecules with similar structures tend to exhibit similar properties, the system utilizes its understanding of molecular similarity to predict the properties of new molecules more efficiently. By leveraging this underlying knowledge, the system improves its predictive capabilities, streamlining the process of exploring and understanding novel compounds.

Guo explains,

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

• The system learns molecular grammar production rules through reinforcement learning, where it receives rewards for behavior that brings it closer to a goal.
• Learning the grammar production rules directly from a dataset with billions of possible combinations is computationally expensive, especially for small datasets.
• To overcome this challenge, the researchers divided the molecular grammar into two parts: a general metagrammar designed manually and a smaller molecule-specific grammar.
• The system is provided with the metagrammar initially, which is widely applicable and serves as a foundation.
• It then focuses on learning the specific grammar from the domain dataset, which speeds up the learning process significantly.
• This hierarchical approach streamlines the learning of grammar production rules and makes it more feasible for larger datasets.

The Power Of Small Datasets

The researchers developed a new system that outperformed several popular machine learning methods in generating viable molecules and predicting their properties. Remarkably, this system achieved high accuracy even when working with domain-specific datasets that contained just a few hundred samples. Unlike other methods, it didn’t require a costly pre-training step.

To showcase the system’s capabilities, the researchers conducted experiments with a significantly reduced training set, consisting of only 94 samples. Astonishingly, the model still achieved comparable results to methods trained using the entire dataset. This demonstrates the robustness and efficiency of the system even with limited training data.

The Science Behind the Grammar Method & the Road Ahead

Guo emphasized the immense power of the grammar-based representation. Its versatility lies in the fact that this grammar can be applied to various types of graph-based data beyond chemistry or material science. The researchers are actively exploring potential applications in other domains, aiming to unlock new possibilities and uncover the full extent of its capabilities.

Looking ahead, the researchers have exciting plans for the future. One of their goals is to expand their existing molecular grammar to incorporate the three-dimensional (3D) geometry of molecules and polymers. This 3D information plays a crucial role in understanding how polymer chains interact with each other, leading to a deeper comprehension of their behavior.

Additionally, they are working on developing an interface that will allow users to explore the learned grammar production rules.

This interactive platform will enable users to provide feedback and correct any potential errors in the rules. By harnessing the collective knowledge and insights of users, they aim to further refine and enhance the accuracy of the system, ensuring its continuous improvement.

The researchers are pushing the boundaries of their molecular grammar approach, seeking to incorporate 3D geometry and harnessing user feedback to create a more advanced and accurate system.

This breakthrough holds tremendous potential for various domains, including chemistry, material science, and beyond. The ability to predict molecular properties with limited data has far-reaching implications, enabling researchers to expedite their explorations and focus on the most promising avenues of discovery.