The AlphaFold2 revolution

In late 2020, DeepMind's AlphaFold2 stunned the 14th Critical Assessment of protein Structure Prediction (CASP14) by predicting protein structures with near-experimental accuracy on most targets. Its results were a discontinuous jump over every prior method, not an incremental gain. For the first time a purely computational model produced structures usable as biological evidence rather than mere hypotheses.

The CASP14 breakthrough

GDT_TS is the average fraction of Cα (alpha-carbon) atoms falling within 1, 2, 4, and 8 angstrom cutoffs of the native structure after optimal superposition, reported on a 0 to 100 scale. Across all CASP14 targets AlphaFold2 reached a median GDT_TS near 92, with a median backbone error under 1 angstrom. In the assessors' overall ranking by summed z-scores, it scored 244 against about 91 for the next-best group, a more-than-twofold lead with no precedent in the assessment's history.

End-to-end differentiable prediction

AlphaFold1 and other earlier systems predicted intermediate features such as inter-residue distance distributions, then handed those constraints to a separate folding optimizer. AlphaFold2 predicts atomic coordinates directly and trains the whole network by backpropagating a geometric loss through every layer. Its Frame Aligned Point Error loss compares predicted and true atoms in local residue frames, so gradients flow from the final 3D structure back to the input alignment.

The prediction pipeline

The pipeline runs in three stages. First, sequence-database searches build a multiple sequence alignment (MSA) and gather structural templates; then the Evoformer, a stack of 48 attention blocks, jointly refines an MSA representation and a residue-pair representation using triangle attention to keep pairwise geometry consistent. Finally, the structure module applies invariant point attention to place each residue as a rigid backbone frame with side-chain torsions. The whole trunk then feeds its output back to the input and runs again, three times by default, so each recycling pass sharpens the prediction.

Template-based modeling can only interpolate within folds that already exist in the PDB, and threading collapses when no suitable template is found. By learning evolutionary couplings directly from the MSA and reasoning about geometry end to end, AlphaFold2 folds novel and free-modeling targets that template methods cannot address. You can reproduce a full prediction locally with ColabFold, which couples the AlphaFold2 weights to fast MMseqs2-based MSA generation.

colabfold_batch --num-recycle 3 --num-models 5 --amber --templates query.fasta results/

2024-03-11 09:42:15,003 Running colabfold 1.5.5
2024-03-11 09:42:31,771 Query 1/1: query (length 214)
2024-03-11 09:43:02,118 Setting max_seq=512, max_extra_seq=1024
2024-03-11 09:44:55,300 alphafold2_ptm_model_1_seed_000 recycle=3 pLDDT=91.4 pTM=0.86
2024-03-11 09:44:55,332 alphafold2_ptm_model_1_seed_000 took 33.9s (3 recycles)
2024-03-11 09:45:29,001 alphafold2_ptm_model_2_seed_000 recycle=3 pLDDT=90.2 pTM=0.85
2024-03-11 09:45:29,033 alphafold2_ptm_model_2_seed_000 took 33.1s (3 recycles)
2024-03-11 09:46:03,540 alphafold2_ptm_model_3_seed_000 recycle=3 pLDDT=93.7 pTM=0.90
2024-03-11 09:46:03,572 alphafold2_ptm_model_3_seed_000 took 33.6s (3 recycles)
2024-03-11 09:46:37,882 alphafold2_ptm_model_4_seed_000 recycle=3 pLDDT=92.1 pTM=0.88
2024-03-11 09:46:37,914 alphafold2_ptm_model_4_seed_000 took 33.4s (3 recycles)
2024-03-11 09:47:11,220 alphafold2_ptm_model_5_seed_000 recycle=3 pLDDT=89.6 pTM=0.84
2024-03-11 09:47:11,252 alphafold2_ptm_model_5_seed_000 took 33.5s (3 recycles)
2024-03-11 09:47:40,905 reranking models by 'plddt' metric
2024-03-11 09:47:40,906 rank_001_alphafold2_ptm_model_3_seed_000 pLDDT=93.7 pTM=0.90
2024-03-11 09:47:40,907 rank_002_alphafold2_ptm_model_4_seed_000 pLDDT=92.1 pTM=0.88
2024-03-11 09:48:12,455 Done

AlphaFold2's design has already been generalized and superseded. RoseTTAFold from the Baker lab introduced a three-track architecture that couples sequence, pair, and coordinate information, and its descendants add all-atom modeling and diffusion-based generation such as RFdiffusion. AlphaFold3, released in 2024, replaces the structure module with a diffusion head fed by a Pairformer trunk, predicting complexes of proteins, nucleic acids, ligands, and ions within a single model.