3D modeling programs based on artificial intelligence (AI) have revolutionized structural biology, often predicting protein structures with an unprecedented level of confidence. However, these structures remain models, and it is always interesting to compare the results of different predictors (AlphaFold3, RoseTTAFold, ESMFold...) and not to limit one’s choice to the model generated with the best confidence score ; It can also be crucial to introduce experimental data, such as known intermolecular regions, in the selection of the best model. We have created a web server, named "FoldScript", to address these issues (https://foldscript.ibcp.fr/​). It can be used to synthesize in 1D, 2D and 3D the multi-model information generated by the AI. We have tested this web software on numerous viral protein complexes studied by our team, for which low or high confidence scores had been obtained. We show that, while AI can provide relevant answers, a critical look remains crucial. Indeed, the "top score" solution proposed by the AI is not always the "best".

​