Periodontitis is one of the most common chronic inflammatory diseases worldwide and the leading cause of tooth loss in adults. This disease, caused by pathogenic bacteria, continuously damages the gums and alveolar bone that supports the teeth, so ideal treatment must accomplish three tasks simultaneously: clearing the bacterial infection, controlling the inflammatory response, and promoting the repair of damaged tissue. 

Hydrogels have long been considered an ideal carrier for periodontal treatment due to their excellent local drug delivery properties. However, the challenge lies in developing a hydrogel that possesses both strong antibacterial capabilities and is safe and biocompatible for the human body. In the past, this mainly relied on repeated trial and error, which was time-consuming, costly, and inefficient.

A team led by Professors Xu Hao and Zhao Hang at the West China School of Stomatology, Sichuan University, decided to take a different approach. They attempted to use machine learning to replace blind screening, allowing the computer to first identify the most worthy candidates for testing from a massive number of molecules. Their research findings were published in the International Journal of Oral Science on May 11, 2026.

The research team's overall approach was clear: first, teach the computer to determine whether a molecule possesses the required properties, and then let it perform large-scale screening. They collected nine bioactivity datasets covering antibacterial, toxic, antiviral, and anti-inflammatory information from public databases, used this data to train a machine learning model, and enabled the model to predict the possible bioactivity of a molecule based on thousands of molecular structural features.

Prediction alone is not enough. To more accurately identify target molecules, the team designed two new evaluation metrics. One is called the Molecular Bioactivity Specificity Index, which is used to determine the main biological function tendency of a molecule; the other is called the Composite Molecular Property Score, which integrates multiple ideal properties such as gelation ability, antibacterial activity, and biocompatibility into a comprehensive score, which is directly used to rank candidate molecules.

After computationally screening thousands of molecules, two candidates emerged from the rankings: guanosine monophosphate and deoxyguanosine monophosphate. The computer deemed them the most promising, and the next step was experimental verification.

The laboratory results did not disappoint. Both nucleoside molecules can self-assemble into stable supramolecular hydrogels and possess self-healing and shear-thinning mechanical properties, meaning they have good adaptability and manipulability in practical applications. In antibacterial tests, they effectively inhibited Porphyromonas gingivalis, the main pathogen of periodontitis, while exhibiting good biocompatibility and low toxicity to normal cells.

Animal studies further corroborated this evidence. In a mouse model of periodontitis, both hydrogels reduced the bacterial load in the oral cavity, alleviated the inflammatory response, protected the alveolar bone from further resorption, and promoted tissue repair. Their therapeutic effects were comparable to those of minocycline, a clinically commonly used antibiotic. The study also found that early application of these hydrogels helped slow the progression of periodontitis.

Corresponding author Professor Hao Xu explained that the core of this work lies in using an artificial intelligence prediction model combined with a molecular scoring system to first complete a round of virtual screening, concentrating laboratory resources on a few candidates with the highest scores, thus bypassing traditional large-scale empirical testing. Professor Hang Zhao added that the two molecules ultimately screened not only met the performance standards but also possessed good safety characteristics.

The value of this research extends beyond periodontitis itself. The computational screening framework it established can be replicated in other scenarios. Whether in drug delivery, wound healing, tissue engineering, or regenerative medicine, similar approaches can be applied wherever there is a need to find hydrogel materials with multiple functions and good biocompatibility. As more bioactivity data becomes available, the predictive power of machine learning models will continue to improve, potentially even enabling the design of personalized biomaterials tailored to the needs of individual patients in the future. 

From a broader perspective, this study provides a clear proof of concept: data-driven approaches can indeed accelerate the rational design of biomaterials. The successful validation of GMP and dGMP hydrogels provides a credible starting point for promoting this strategy in the wider biomedical field.