Artificial intelligence prediction of carcinoembryonic antigen structure and interactions relevant for colorectal cancer

Abstract

Carcinoembryonic antigen (CEA) is used as a biomarker for colorectal cancer. It is expressed during fetal development but in healthy adult cells the expression is low. Due to its size and the high degree of glycosylation, there are no structures available for mature CEA. By employing novel structure prediction methods, we aim to investigate CEA tertiary structure and interactions.

Alphafold 3 server has increased the accuracy of structure predictions and allows for modelling of glycans in proteins and complexes. Models were created for a monomeric CEA, dimeric CEA and for CEA in complex with the antibody Tusamitamab. The structure of the monomeric glycosylated CEA exhibit two bends, one in the domain interface B1–A2 and one in the domain interface B2-A3. The dimer structure pairs in a parallel manner, with direct contacts in the N and the A2 domains of the two chains. The complex of CEA with Tusamitamab closely resembles the EM structure of the complex that was released after the training of Alphafold 3 was completed.

Overall, the investigations give new angles to investigate for CEA. The predicted bend, primarily in the B2 and A3 domain interface, would allow for dimer formation of CEA from both the same cell as from adjacent cells and could help to explain the outstanding issue on how it can fulfil both tasks. The prediction of the antibody binding to CEA was accurate, the all-atom RMSD was 1.3 Å. This is encouraging for other antibody – protein complexes predictions as the complex structure was not part of the training set for Alphafold 3.