Navarrabiomed researchers analyse the limitations of AI in protein structure and reveal evolutionary keys to immune receptors
- The findings of the Protein Crystallography and Structural Immunology Unit have been published in the following international journals: Scientific Reports y Communications Biology.
The Protein Crystallography and Structural Immunology Unit at Navarrabiomed, led by Jacinto López Sagaseta, recently published a study critically analysing the current limitations of artificial intelligence, specifically AlphaFold, in protein structure prediction. In a second study, conducted in collaboration with researchers from the National Cancer Institute in the United States, the discovery of molecular configurations previously unknown in primitive organisms is described, which suggests that key structural elements of the adaptive immune system appeared much earlier than previously thought.
The results, which have been published in scientific articles in Scientific Reports and Communications Biology,, offer new horizons in the design and development of new therapeutic proteins.
Furthermore, the relevance of this research has transcended the local academic sphere. Jacinto López Sagaseta, the researcher who led these studies, has been invited to present these findings at the Protein & Antibody Engineering Summit 2025,conference to be held in Lisbon next November, which will assemble experts from leading centres such as MIT, Scripps, Roche and Genentech, among many others.
1. Severe deviation in protein fold prediction by advanced AI: a case study (Scientific Reports, 2025)
This study critically evaluates the current capabilities of artificial intelligence in protein structure prediction, focusing on AlphaFold.
Using the protein SAML, derived from the marine sponge Geodia cydonium and composed of two domains, the authors compare the structure predicted by AlphaFold with that determined experimentally by X-ray crystallography.
The researchers observed that while AlphaFold accurately predicts the individual structure of each domain (RMSD < 0.9 Å), the AI tool fails to predict the relative orientation of the two domains. This result was confirmed even when expanding the conformational folding space of the protein in AlphaFold predictions. Furthermore, the researchers achieved two forms of crystallisation of RTK, a protein similar to SAML. In both forms, the folding of the protein is identical, indicating that the crystallised conformation is robust and a faithful reflection of the native conformation. The deviation obtained in the predictions is reflected in an RMSD greater than 7.7 Å when the experimental structure and the AlphaFold prediction are overlaid, or in deviations of over 30 Å for equivalent residues when overlaying one of the domains. On a human scale, it is as if, in an anatomical reconstruction, the head were placed at the level of the navel. This analogy highlights how serious an interdomain assembly error can be in structural biology, especially when these models are used to design drugs or antibodies, where precise interaction geometry is critical. Therefore, this study reveals that AlphaFold is an accurate prediction tool for low-complexity structures, but it does not adequately model the spatial arrangement of proteins composed of more than one domain, particularly when clear evolutionary constraints are lacking or when proteins have few homologues in the databases.
These results have direct implications for health and rational drug design. Many therapeutic proteins (antibodies, receptors, enzymes) depend on precise orientation between functional domains, and errors in their prediction can compromise target selection, inhibitor design, or antibody engineering.
In conclusion, this work reiterates that AI generates theoretical models that may deviate from the actual structure of a protein, and highlights the need to prioritise experimental data, especially in sensitive biomedical contexts such as the development of new immunotherapies.
2. Unusual traits shape the architecture of the Ig ancestor molecule (Communications Biology, 2025)
In a second article published in March this year, and in collaboration with experts from the National Institute of Health in the United States, the researchers present a rigorous study of the crystallographic structures of two ancestral Ig-like proteins (SAML and RTK) from the marine sponge Geodia cydonium, representatives of one of the oldest Metazoa lineages. Both proteins have two tandem immunoglobulin domains.
Main findings:
• The N-terminal domain exhibits a unique architecture, distinct from known Ig types, leading to the proposal of a new category: the Early Variable (EV-set) domain.
• The C-terminal domain adopts a C1-set configuration, which until now was considered exclusive to vertebrates. This suggests that the basic architecture of surface immune receptors may have emerged long before the appearance of the adaptive immune system.
This discovery provides a new perspective on the structural evolution of immune systems and could inspire the development of new protein scaffolds for antibody engineering or TCR receptors. Furthermore, understanding how nature has reused these modular platforms for hundreds of millions of years can guide the design of synthetic receptors with evolutionarily optimised properties, useful in immunotherapy, diagnostics and cell therapies.
Conclusions
Both works, which focus on the structure of an ancestral protein, address key aspects of modern structural biology from different angles. The first exposes the limitations of current predictive models, warning of their risks in biomedical contexts. The second explores the evolutionary roots of Ig domains, suggesting that the basic architecture of immune recognition predates the adaptive immune system by a considerable margin.
As a whole, these studies not only enrich our understanding of the origin of immunity, but also offer conceptual and structural tools with great potential for application in modern biomedicine, from the design of therapeutic proteins to the development of diagnostic platforms.
Photo captions:
- Figure 1: Overlaying of the structural model predicted by AlphaFold (yellow) and the structure determined by X-rays (blue). The turquoise spheres indicate equivalent atomic positions in both models. A discrepancy of 32 Å is observed, indicating a significant difference between the computational prediction and the actual conformation of the protein obtained by X-ray diffraction.
- Figure 2: The image shows canonical (lower panel) and non-canonical (upper panel) disulphide bridges, with alternative conformations detected by X-ray diffraction, in the SAML protein from Geodia cydonium. Early Variable and C1 immunoglobulin (Ig) domains are observed, identified for the first time in non-vertebrate organisms, with atoms coloured by type (grey: carbon, blue: nitrogen, red: oxygen, yellow: sulphur). The blue network represents the signal obtained through X-ray diffraction of SAML crystals by researchers at Navarrabiomed.

