Může AI předpovídat struktury proteinů z aminokyselinových sekvencí ?

Pokroky v AI umožnily přesnou predikci proteinových struktur, což byl problém, který vědce pletl po desetiletí. Systémy jako AlphaFold využívají hluboké učení k modelování komplexních biologických interakcí. Tento průlom revolucionalizoval strukturování biologie a procesy objevování léků.

Background

Traditional experimental methods for protein structure determination—such as X-ray crystallography and nuclear magnetic resonance spectroscopy—remain resource-intensive and slow, motivating the development of computational approaches. Classical comparative modeling (e.g., homology modeling) relied on evolutionary conservation and template structures, while fragment assembly methods (e.g., Rosetta) used physical energy functions to guide conformational sampling. Over the past decade, machine learning techniques gradually improved accuracy by learning from solved structures; however, the field lacked end-to-end models capable of inferring folding directly from sequence. A decisive shift occurred with AlphaFold, introduced by DeepMind, which combined deep neural networks with attention mechanisms to predict residue-residue distances and orientations, thereby reconstructing full 3D structures from amino acid sequences in a single forward pass. The system was trained on hundreds of thousands of experimentally determined protein structures from the Protein Data Bank (PDB), alongside genomic data curated by the EBI and UniProt. In the 2020 CASP14 assessment, AlphaFold achieved a median global distance test (GDT) score above 90% on many targets, surpassing previous state-of-the-art by a wide margin, and demonstrated robust performance on orphan proteins lacking homologous templates. Subsequent versions integrated multiple sequence alignments (MSAs), structural templates, and geometric priors to further refine accuracy and generalization. These advances have unlocked new possibilities in structural biology, enabling rapid modeling of entire proteomes and accelerating structure-guided drug design pipelines. By accurately predicting folding landscapes, AI systems now allow researchers to infer protein function, map interaction networks, and anticipate mutational effects at scale.