Emerging technologies in the scientific community are helping researchers achieve more goals and make discoveries. Revolutionary tech such as artificial intelligence (AI) and machine learning (ML) have already disrupted various industries, from manufacturing to retail and beyond.
ML has expedited the discovery process, especially for grad students at the Massachusetts Institute of Technology (MIT) and staff members at IBM Research. MIT News reported that a new generative graph grammar model allows researchers to create novel molecules.
Here’s more about this discovery at MIT and its implications for the future.
New Discovery of Novel Molecular Materials at MIT with Generative Graph Grammar Models
According to an article published by MIT News, MIT students and IBM researchers used a generative graph grammar model to build synthesizable molecules that fall into the same chemical class as the training data.
Traditional deep learning techniques must use extensive datasets for their training models. Additionally, many of these datasets contain only a handful of example components, meaning the technology cannot generalize or generate molecular models that could be made in the real world.
Researchers overcame this problem by using a generative graph model and treating atom formations and chemical bonds as a graph. The next step is developing a graph grammar, which is defined as a linguistic representation of systems and structures for word ordering. It contains a sequence of rules for the model to build molecules like monomers and polymers.
Once the grammar and production rules have been established, the model uses reverse engineering on its example and even creates new compounds. The method is cited as being systematic and data-efficient and, in other words, “a language for creating molecules.”
The authors of this paper include MIT professor of electrical engineering and computer science Wojciech Matusik, MIT graduate student and lead author Minghao Guo, MIT graduate student Beichen Li and IBM research staff members Veronika Thost, Jie Chen, and Payal Das.
Matusik, Chen, and Thost are three members of the team affiliated with the MIT-IBM Watson AI Lab. The new research method, coined data-efficient graph grammar (DEG), will be presented at the International Conference on Learning Representations.
Guo adds that the DEG method was tested against existing state-of-the-art models and techniques, and it outperforms them by a large margin.
However, it is important to note that some of the molecules created using the DEG technique are valid, but some are not. The goal would be to identify the minimum amount of production rules to maximize the synthesizable percentage of molecules.
Implications of the DEG Discovery
It’s expected that this new model for molecular generation can be applied in many ways because it can adapt to more than just chemical structures. The graph in the model is considered a flexible representation, so other entities can use this form.
For example, this model can be applied to:
- Buildings and construction projects
- Electronic circuits
DEG may be a revolutionary approach for future projects, and it could further automate innovation. Creating and synthesizing molecules like monomers and polymers could also have many medical applications.
For example, a process called osseointegration must occur during dental implants, where the bone must heal around it. Surface coatings play an important role here. It may be possible to create new ones with growth factors, peptides or messenger molecules with this DEG approach, ultimately making the implant and healing process more efficient.
The researchers at MIT and IBM focused on using less than 33 samples each, including acrylates, chain extenders, and isocyanates. However, they’re hopeful that the DEG approach may be extended to any chemical class.
In addition, the researchers suggest that switching up the order of production rules in the learning language can create other molecules. The DEG method could carve out a place in medicinal chemistry by creating the molecules needed for organic compounds. It may be possible to develop new pharmaceutical agents or drugs for widespread use.
MIT and IBM’s DEG Technique and Future Opportunities
The authors of the paper plan to scale up the generative graph grammar learning process and produce new chemicals with specific, desired properties. The DEG technique is still in its early stages, but it seems as though it will have plenty of applications in various industries. It will be interesting to see how these researchers plan on further developing DEG so more use cases can emerge.