Can a machine do better than human for photocatalysis for organic synthesis?

There is growing importance of nickel/photocatalytic systems in organic synthesis, particularly in the formation of C(sp²)-heteroatom bonds. It highlights the need for the development of efficient methods for the formation of C(sp²)-O bonds, which has lagged behind other types of C(sp²)-heteroatom bond formation despite the significance of phenol derivatives in the pharmaceutical, agrochemical, and materials industries. Recent efforts in the field of organic synthesis have explored the potential of organic photosensitizers. These compounds, particularly organic donor-acceptor photosensitizers (ODAPs), hold the promise to enhance synthetic methods without the need for precious-metal catalysts. ODAPs are characterized by their π-conjugated organic structures, incorporating electron-donor and -acceptor groups. These compounds have demonstrated their prowess in redox or energy-transfer processes, and there’s growing anticipation that they could replace traditional iridium- and ruthenium-based photosensitizers. The beauty of ODAPs lies in their design flexibility. By merely incorporating different electron-donor and -acceptor components into their molecular structure, a wide array of potentially highly active photosensitizers remains unexplored. Researchers have only scratched the surface of identifying these exceptional catalysts amidst the myriad of organic molecules. The transition to using these organic photosensitizers has the potential to open new horizons in catalytic organic synthesis and provide a sustainable alternative to precious-metal-based catalysts.

Machine learning (ML) has gained increasing interest among synthetic chemists for its potential to predict reactivity and selectivity in organic synthesis. By using powerful computational models and algorithms, ML has the capacity to expedite the discovery of catalysts with specific properties, ultimately streamlining the entire research process. ML models for predicting the catalytic activity of organic photosensitizers in organic synthesis remained largely underdeveloped.  To address this, the research conducted by Dr. Naoki Noto, Dr. Akira Yada, Professor Takeshi Yanai, and led by Professor Susumu Saito from Nagoya University, Japan, as published in the Journal Angewandte Chemie International Edition, focused on the development of a novel method for the synthesis of phenol derivatives using a combination of a nickel(II) salt and ODAPs under visible-light irradiation. The research also incorporated ML techniques to predict the catalytic activity of these ODAPs.

The authors conducted synthesis of 60 ODAPs with various electron donor and acceptor groups. They optimized the reaction conditions for the photosensitized nickel-catalyzed synthesis of phenol derivatives, demonstrating that the presence of a photosensitizer is critical for the success of the reaction when irradiated by visible light. They identified highly active ODAPs and less-active ones, providing insight into the structural features that influence their catalytic activity. The study also explores the scalability of the photoreaction, showcasing its potential for industrial-scale production of pharmaceuticals (The smallest amount of ODAP by which the photo-reaction is working: substrate/ODAP = 20000:1). It is noteworthy to mention the researchers incorporated machine learning into the analysis. The authors create ML models to classify the ODAPs based on their catalytic activity and identify promising candidates. They emphasized the need for classification models rather than regression models and evaluated various ML algorithms’ performance in predicting the catalytic activity. The study used a combination of DFT descriptors, RDKit descriptors, and hybrid descriptors to construct ML models, with the hybrid descriptors proving to be the most effective in developing robust models.

The research team afterward conducted a thorough analysis of the results, showing that the hybrid descriptors outperform other descriptors in classifying ODAPs into “very good” and “good” classes. They provided important information into the impact of specific physical properties, such as vertical excitation energies and energy levels, on the ML models’ accuracy. Indeed, the research conducted is an excellent showcase of the potential of machine learning to aid in catalyst design and discovery, streamlining the process and reducing the need for exhaustive experimental investigations. In conclusion, Professor Susumu Saito and colleagues developed a new and novel method for the synthesis of phenol derivatives using ODAPs and nickel(II) salts, supported by machine learning techniques for predicting catalytic activity.