google-deepmind/alphafold

AlphaFold is the open source code released by Google DeepMind for AlphaFold version 2, the system that predicts the three-dimensional shape of a protein from its amino-acid sequence. This repository contains the inference pipeline, meaning the part that takes a sequence and runs the trained model to produce a predicted structure. Model weights are downloaded separately. The package also includes AlphaFold-Multimer, an extension for predicting complexes made of more than one protein chain. The README notes that the multimer variant is a work in progress and is not expected to be as stable as the single-chain version. There is a technical note for an updated AlphaFold v2.3.0 and a CASP15 baseline set of predictions included with the repo. Running AlphaFold has heavy system requirements. You need a Linux machine, a modern NVIDIA GPU, and up to about 3 TB of disk space to store the genetic sequence databases it needs as input. The typical workflow is to install Docker and the NVIDIA Container Toolkit, clone the repo, run a download script that pulls roughly 556 GB of databases, build a Docker image, and then run a Python script pointing at a FASTA file containing the protein sequence you want to predict. The model relies on several public sequence databases such as BFD, MGnify, UniRef90, UniRef30, and the PDB, plus extra ones like UniProt and PDB seqres when running AlphaFold-Multimer. A reduced-database preset is offered for users who cannot store the full set. The repo points users to the AlphaFold paper for the scientific method and asks that work using the code cite that paper. The full README is longer than what was shown.