A novel automated discovery platform developed by a team of North American researchers eases the creation of custom reticular frameworks for use in gas storage, separation, catalysis and a host of other important chemical processes.
Such frameworks include metal organic frameworks (MOFs) and covalent organic frameworks (COFs), both of which are effectively tailored sponges that can form into a vast number of different molecular arrangements. MOFs are increasingly gaining traction as absorbing materials for removing carbon dioxide from flue gas and other combustion processes.
The researchers, based at the University of Toronto in Canada and Northwestern University, Evanston, Ill., are using machine learning and artificial intelligence (AI) to craft the best building blocks in the assembly of such frameworks so they target specific applications.
Described in a recent issue of Nature Machine Intelligence, the work started in a conventional way by constructing a large number of MOF structures on the computer and simulating their performance using molecular-level modeling.
From this, the researchers generated a “training pool” of MOF candidates that could be used in a specific carbon dioxide separation application. In the past, each member of the pool would be screened computationally until a list of top candidates could be isolated for further study.
However, for this work, they created an automated nanoporous materials discovery platform powered by a supramolecular variational autoencoder (VAE) for the generative design of the reticular materials.
“What’s new here is that the automated materials discovery platform developed in this collaborative effort is more efficient than such a ‘brute force’ screening of every material in a database. Perhaps more importantly, the approach uses machine learning algorithms to learn from the data as it explores the space of materials and actually suggests new materials that were not originally imagined,” explains study co-author Randall Snurr, professor and chair of the Department of Chemical & Biological Engineering in the McCormick School of Engineering at Northwestern.
“Designing reticular materials is particularly challenging, as they bring the hard aspects of modeling crystals together with those of modeling molecules in a single problem,” says senior co-author Alán Aspuru-Guzik, research chair in theoretical chemistry at Toronto. “This approach to reticular chemistry exemplifies our emerging focus … of accelerating materials development by means of AI. By using an AI model that can ‘dream’ or suggest novel materials, we can go beyond the traditional library-based screening approach.”
The study authors conclude that their model shows high fidelity in capturing MOF structural features and that the autoencoder has a promising optimization capability when jointly trained with multiple top adsorbent candidates identified for superior gas separation.
“MOFs discovered here are strongly competitive against some of the best-performing MOFs/zeolites ever reported,” write the authors.
Another branch of the industry pursuing VAEs is cheminformatics. Here, the concept of molecular similarity implies that molecules with similar structures tend to have similar properties. So, for example, knowing that a certain molecule shows a desirable chemical activity, the question then arises how many — if any — candidates in potentially massive online collections might have similar properties.
However, defining such similarity is a far from simple procedure and different methods have their own inherent problems. They can be highly variable, and prone to both false negatives and false positives, for example.
To try and overcome this, a team of researchers from the universities of Liverpool and Manchester in the U.K. and Lyngby in Denmark are using a VAE to target over six million drug-like molecules and natural products.
Writing in a recent issue of Chemical Biology, the researchers say their approach provides a rapid and novel metric for molecular similarity that is both easily and rapidly calculated.
As well as leveraging a new method of encoding the molecules themselves, the researchers note that methods which involve VAE are generative and allow for the creation of entirely new molecules. “This opens up a considerable area of chemical exploration, even in the absence of any knowledge of bioactivities,” they write.