Proceedings of the International scientific and practical conference “Science in the Modern World” (January 19-21, 2026) / Publisher website: www.naukainfo.com. - Cambridge, United Kingdom, 2026. - 203 p.

8 deep neural network architecture - centered on the Evoformer block that combines attention mechanisms with iterative geometric refinement. AlphaFold integrates multiple sequence alignments, evolutionary covariation, and learned physical constraints to predict inter-residue distances and angles, ultimately generating full atomic models. Unlike traditional computational methods, AlphaFold performs end-to-end structure prediction without requiring explicit energy minimization or experimental input, enabling rapid and scalable predictions across entire proteomes [5, p. 708]. AlphaFold’s transformative potential was demonstrated during the CASP14 competition (2020), where it achieved near-experimental accuracy for many protein targets, substantially outperforming all other computational methods [3, p. 586]. This performance marked a turning point, suggesting that computational predictions could rival experimental techniques for a wide class of globular proteins. Figure 1. Performance comparison of median Global Distance Test - Total Score (GDT- TS) from the CASP14 protein structure prediction competition. AlphaFold 2 achieved a median GDT-TS of 92.4, indicating near-experimental quality predictions, substantially higher than other competing methods (60 and under). Higher GD-TS values indicate closer agreement between predicted and experimentally determined structures [4, p. 1886]. Following this success, the release of the AlphaFold Protein Structure Database provided open access to hundreds of thousands of predicted structures. This resource 0 20 40 60 80 100 median GDT-TS Competitor Average AlphaFold