Predicting amount of protein adsorption on polymer surface: RIKEN develops AI model to accelerate DDS and biomaterials development

A research team led by Postdoctoral Researcher Shiwei Su and Senior Technical Scientist Nobuyuki Tanaka of the Laboratory for Biologically Inspired Computing at the RIKEN Center for Biosystems Dynamics Research has developed "BB-EIT," an AI model designed to accurately predict the amount of protein adsorbed on polymer surfaces. Training was performed using a combination of physical and biochemical feature quantities such as polymer thickness and surface electrical properties. The model shows generalizability, deriving correct adsorption trends for a wide range of material-protein combinations. Its expected applications include drug delivery systems optimized for specific diseases and advanced diagnostic devices. The results were published in ACS Applied Materials & Interfaces on April 6.

In the development of biomaterials such as artificial organs, protein adsorption to polymers covering the material surface needs to be precisely controlled to meet the conflicting needs of rejection and immobilization. Conventional studies in this area required countless repetitions of trial and error based on researchers' experiences.

In recent years, AI has been applied to material development, but it requires considerable time and cost for acquisition of experimental data. The use of predictive models based on machine learning has been limited to specific proteins and materials.

In this study, the research group combined a large-scale language model with expert knowledge in materials science to develop a system capable of compensating for data scarcity. Polymers which are widely used in biomaterials have low biological toxicity and can be processed and produced in large scales. They are composed of a large number of specific monomers connected to each other like chains. The chemical and physical properties of polymers can be controlled by the selection of monomers.

The study focused on "polymer brushes," which are extensively studied as application forms of polymers allowing precise control of surface properties. Polymer brushes consist of a large number of polymers anchored to a solid surface and stretching upward to form a brush-like shape. The AI model developed in the study can accurately predict how much protein will be adsorbed on the surface of a polymer brush made of a specific monomer.

As the foundation of the model, they adopted "ChemBERTa," a large-scale language model that could be pretrained on a vast amount of chemical structure data and comprehend them as "contexts." By combining physical and biochemical feature quantities such as polymer thickness and surface electrical properties for model training and adopting a data augmentation process that mathematically recombines the notation of structural formulas to increase the data size, the model could achieve a high prediction accuracy, even with a small experimental data set.

The model is capable of deriving correct adsorption trends for a wide range of material-protein combinations. In the prediction process, the structure of the monomer, the smallest unit of the polymer, is first entered into ChemBERTa as a "SMILES string", from which a 768-dimensional vector of chemical features is generated.

This vector is directly coupled with a 5-dimensional vector containing relevant physical and biochemical parameters. The parameters are film thickness, surface hydrophilicity/hydrophobicity and surface potential (a measure of the electrical state of the material's surface) of the polymer brush material, as well as the surface potential and molecular weight of the protein.

For model training, a high-quality experimental data set collected in a previous study was used. In addition, "data augmentation" was adopted to compensate for data scarcity, and Gaussian noise was added to simulate measurement errors. The model showed an extremely high prediction accuracy, achieving an indicator value of 0.88 on the test data set.

Su said, "The latest AI is not just a writing tool. It is now penetrating into our society as a 'promising scientific partner,' supporting the development of antifouling materials enhancing medical safety and biosensors detecting the slightest sign of disease. By combining the latest AI technology and materials science knowledge, we have succeeded for the first time in the world in accurately predicting the amount of protein adsorption on materials within a single framework. Our model has also achieved a high level of generalizability, enabling its application to unknown materials. We expect that the model serves as a powerful foundation that can dramatically accelerate the development of next-generation biomaterials. While promoting its use, we will continue refining the model to develop a performance-enhanced version of BB-EIT."

Journal Information
Publication: ACS Applied Materials & Interfaces
Title: BB-EIT: A Generalized Prediction Model for Protein Adsorption on Polymer Brushes Using Augmented Chemical Embeddings
DOI: 10.1021/acsami.5c25223