Prediction of quantitative function of artificially-designed protein from structural information

Artificially designed proteins are widely used in applications such as optogenetics and biosensing. While experimental optimization of these proteins is effective, it is also costly and labor-intensive. To address this challenge, computational approaches have been developed, primarily relying on sequence-based features. However, protein function is inherently tied to its three-dimensional (3D) structure, and incorporating structural information could enable more accurate predictions and provide deeper biological interpretability. Here, we proposed a structure-based analysis framework called Foldinsight for predicting protein functionalities. In our framework, we first predict protein structures from sequences using AlphaFold2 and then utilize these structures to predict protein properties. Since proteins vary in the number of atoms and lack direct atomic correspondence, we applied molecular field mapping, which captures the energy states surrounding a protein and converts them into fixed-length numerical vectors. This transformation enables the application of machine learning, allowing protein properties to be predicted from structure-derived features. Applying this framework to channelrhodopsin mutants, we achieved predictive performance comparable to sequence-based models. Additionally, our structure-based analysis successfully identified key structural regions contributing to functional differences, highlighting the advantage of incorporating structural data into predictive modeling.