The jump from 2018 to 2020 defines the field. At CASP13 in December 2018, the AlphaFold system achieved a median GDT_TS of 58.9 for Free Modeling targets. By CASP14 in 2020, the AlphaFold2 model reached a median GDT score of 92.4 across all targets. This 2020 result exceeded previous best results by over 20 points for single-domain targets. The AlphaFold2 model achieved an average GDT-TS of 90.0 across 30 single-domain targets, which exceeded previous best results by over 20 points, and in 25 of the 30 cases, the GDT-TS exceeded 80. I see the 2020 results as the arrival of atomic-level accuracy. One target, T1024, achieved a GDT-TS of 95.2 and a backbone RMSD of 1.1 A. Target T1064 achieved a GDT-TS of 93.8 and a backbone RMSD of 1.4 A, whereas Target T1052 achieved a GDT-TS of 90.7 and accurately modeled long loop insertions.
| Competition | Year | Median GDT Score |
|---|---|---|
| CASP13 | 2018 | 58.9 |
| CASP14 | 2020 | 92.4 |
Mechanics of the Evoformer and energy landscapes
The architecture uses an Evoformer stack and a structure module. The Evoformer stack extracts information using row- and column-wise attention on the MSA and triangle multiplicative updates on the pair features. The structure module produces 3D coordinates through invariant point attention, which treats each residue as a rigid triangle with learnable 3D translations. The model learns through end-to-end differentiability using loss components for distogram error, torsion angles, and frame-aligned point error. The internal recycling scheme is a design element to gradually increase structure prediction accuracy. The prediction is repeated until the convergence of both structure prediction and its confidence scores. The system follows a "local first, global later" mechanism. For six small proteins including ubiquitin and the SH3 domain, iterative predictions showed a specific sequence. In protein G, the process starts with a native C-terminal hairpin formation, followed by a register-shifted N-terminal hairpin, before finding the native fold. For protein L, the N-terminal hairpin forms first, and subsequent iterations fold the C-terminal hairpin. These predictions match experimental psi-values.
The requirement for MSAs causes problems when homologous sequences are sparse. When MSA depth drops below 10 sequences, accuracy falls. The MSA-Augmenter addresses this by generating de novo homologous protein sequences. For 81 artificially downsampled targets with a depth of 5, the MSA-Augmenter increased the LDDT from 53.45 to 66.32. This increase of 12.87 LDDT units shows the model can mitigate performance loss in low-homology regimes. The model uses a two-track network architecture where information at the 1D sequence level and the 2D distance map level is iteratively transformed and passed back and forth. If you are looking for a way to bypass the need for MSAs, the model’s ability to use a learned biophysical energy function suggests it can rank candidate structures accurately without coevolutionary data.
Nobel recognition and biological utility
The 2024 Nobel Prize in Chemistry recognizes David Baker, Demis Hassabis, and John Jumper. Hassabis and Jumper used artificial intelligence to predict the structures of almost all known proteins. Baker designed artificial proteins with specific, targeted functions, such as the Top7 protein designed in 2003. This recognition builds on the work of Christian B. Anfinsen, who concluded in 1961 that amino acid sequences control protein structure. It also addresses the challenge of Levinthal’s paradox, which notes that random folding would take longer than the lifetime of the universe. The AlphaFold Database, which the European Molecular Biology Institute maintains, contains over 200 million predictions. The database has over 3.4 million users from 190 countries. A recent expansion added 1.7 million homodimers to the collection. These complexes consist of two interacting strands of the same molecule.
I observe that while AlphaFold2 matches X-ray crystallography accuracy for many targets, it still fails to predict conformational variability or fold flexibility. The model also lacks co-factors and posttranslational modifications. The AlphaFold2 predictions are not always perfect in terms of chemical geometry. Can these models eventually capture the full dynamics of protein-protein interactions?




