google-deepmind/alphafold ? reverse-engineered prompt

Build me a working AlphaFold app that can take a protein sequence file and predict the 3D structure from it.

I want the whole thing set up so I can run it on Linux with Docker and a modern NVIDIA GPU, with clear steps for downloading the required genetic databases and model weights, building the container, and then running a prediction from a FASTA file. It should save the resulting structure files and any useful confidence outputs into an output folder I can inspect afterward.

Please also include support for the multimer case if that’s part of the project, and make sure the setup is practical for a real machine with lots of storage, since the databases are huge. If there are any current dependency or install details that are easy to get wrong, check the latest docs online and make the instructions as smooth as possible.