AI methods for reconstructing protein complex structures from cryo-EM data and predicting mutated protein structures

The National Institute of General Medical Sciences awarded the University of Missouri System $2.285 million on September 1, 2026, under the Biomedical Research and Research Training program (CFDA 93.859) to develop artificial intelligence methods for reconstructing protein complex structures from cryo-electron microscopy data and predicting mutated protein structures.

The award funds research to address two key gaps in structural bioinformatics: the lack of automated, accurate methods for reconstructing protein complex structures from cryo-EM data, and the absence of reliable techniques for predicting three-dimensional structures of protein mutants induced by amino acid mutations. The investigator will develop generative AI methods—including multi-modal transformers, generalized protein language models, diffusion models, and flow matching models—to integrate multiple complementary data modalities: one-dimensional protein sequences, protein evolutionary information, two-dimensional protein shapes, and three-dimensional experimental or predicted structural data. The approach will leverage self-supervised learning to harness large amounts of unlabeled cryo-EM data, addressing limitations of current methods that rely on limited datasets and deliver suboptimal performance.

Work is performed at Columbia, Missouri, with a period of performance from September 1, 2026, through August 31, 2030. This is a MIRA (Maximizing Investigators' Research Award) research program, which the National Institute of General Medical Sciences prioritizes over traditional R01 awards as part of its strategic funding approach.

Abstract

With recent advancements in artificial intelligence (AI) such as alphafold, which have enabled high-accuracy prediction of protein tertiary structures, accurately determining the quaternary structures of large protein complexes from cryo-electron microscopy (cryo-EM) data and predicting the structures of protein mutants are emerging as key challenges in structural bioinformatics. Building on the PI's extensive expertise and experience of developing machine learning and AI methods for protein structure prediction and cryo-EM-based structure determination, this mira research program seeks to develop advanced AI methods to address the two key gaps in the field: (1) the lack of automated and accurate methods for reconstructing protein complex structures from cryo-EM data, and (2) the absence of reliable techniques for predicting the three-dimensional (3D) structures of protein mutants (equivalently structural changes induced by amino acid mutations). In the two areas, we propose to develop innovative generative AI methods - such as multi-modal transformers, generalized protein language models, diffusion models, and flow matching models - to integrate multiple modalities of complementary data. These modalities include 1D protein sequences, protein evolutionary information, 2D protein shapes, and 3D experimental or predicted structural data. Our integrative approach will enable the automatic and accurate reconstruction of cryo-EM protein structures (area #1) and the prediction of protein mutant structures (area #2). This will address the limitations of current methods, which rely on limited data and thus deliver suboptimal performance. Moreover, by leveraging self-supervised learning - a key driver behind the success of large language models (llms) such as chatgpt - we will, for the first time, harness large amounts of unlabeled cryo-EM, protein sequence, and structural data to significantly enhance cryo-EM structure modeling and protein mutant structure prediction. The generative AI methods will be trained using both self-supervised and supervised learning techniques to perform multiple essential tasks via multi-task learning. These tasks include cryo-EM structure reconstruction, mutant structure prediction, confidence estimation, and stability prediction for mutated structures, resulting in more accurate and comprehensive predictions than previously possible. The pretrained generative AI models will serve as foundational AI models in the field, enabling accurate cryo-EM protein structure determination and precise mutant structure predictions. These models will also facilitate downstream applications such as protein function analysis, protein interaction prediction, dynamic protein conformation modeling, and phenotype prediction through transfer learning, fine- tuning, and few-shot learning. Finally, we will make these generative AI methods available as free, user-friendly, open-source tools, promoting widespread adoption and further advancements.