DUGi

Avaluació comparativa d’eines d’intel·ligència artificial per a la predicció de complexos proteïna-lligand

http://dugi.udg.edu/item/http:@@@@hdl.handle.net@@10256@@28659

In recent years, artificial intelligence has become a key tool in the field of structural biology, transforming how proteins and their molecular interactions are studied. This work aims to deeply analyze the effectiveness and applications of two advanced AI tools—AlphaFold3 and Chai-1—in the context of structural and functional protein prediction. The selected model system is the epoxide hydrolase protein, well known for its biochemical and pharmacological relevance. This protein has a large number of experimentally resolved structures in complex with known inhibitors, making it an excellent reference model. To validate these tools for the study of epoxide hydrolase, the first step involved evaluating the ability of AlphaFold3 to predict three-dimensional structures from amino acid sequences in FASTA format, comparing its predictions with experimentally resolved structures available in the Protein Data Bank. The comparison was carried out using functionally centered root mean square deviation (RMSD) calculations, showing a very high level of agreement and highlighting near-atomic accuracy. In parallel, Chai-1 was used to address functional scenarios such as ligand competition and the simulation of saturation conditions within the active site of the protein. This tool allows for the simultaneous introduction of multiple ligands, analysis of their spatial distribution, and detection of positional shifts based on binding affinity, without requiring explicit binding energy calculations. The results show that Chai-1 is capable of predicting realistic behaviors in terms of binding, affinity, and exclusion of different ligands from the active site, particularly under high-concentration conditions. The combined use of AlphaFold3 and Chai-1 provides a complementary perspective that integrates structural precision with functional simulation, establishing a solid foundation for rational drug design and therapeutic target exploration, and demonstrating the complementary value of both tools

9