Deep Comparative Learning for Estimating Protein Structural Model Accuracy

Recent development of AI-based methods and tools can predict the three-dimensional structures of proteins with unprecedented accuracy, transforming biological sciences and biotechnology research. However, scientists often generate multiple structural models for the same protein and lack reliable tools to determine which predictions are most accurate. Inaccurate model selection can limit the usefulness of protein structure prediction for scientific discovery and practical applications. This project addresses this challenge by developing an advanced AI-based innovative tool that can more accurately assess and rank predicted protein structures. The resulting methods and open-source software will help researchers identify the most reliable structural models for studying protein function and designing new proteins for bioengineering applications. The project advances the national interest by promoting scientific progress in biotechnology and AI and strengthening the computational biology workforce training through education and participation in science, technology, engineering, and mathematics (STEM) activities for undergraduate, K-12, and community audiences. This project is expected to establish a new deep comparative learning framework for the estimation of protein model accuracy (EMA) that predicts the relative and absolute quality of pairs of protein structural models. The research has three objectives: (1) construct a large-scale benchmark consisting of comprehensive datasets of protein structural models for training and evaluating machine learning-based EMA methods; (2) develop deep comparative learning methods for assessing the quality of monomer protein models; and (3) extend these methods to multimer protein complexes. The project will leverage state-of-the-art AI techniques, including protein language models, graph transformers, multi-task learning, and novel pairwise ranking losses, to compare and evaluate protein structures. It will also develop a new protein language model capable of representing multimeric protein complexes, addressing an important limitation of existing models. The resulting benchmark datasets and AI methods are expected to substantially improve the accuracy of protein model quality assessment and provide the scientific community with powerful tools for selecting and ranking protein structural models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.