AlphaFold and ML Prediction

Core Idea

AlphaFold (DeepMind, 2020) and related deep learning methods (RoseTTAFold, ESMFold, OpenFold) predict protein structures from amino acid sequence with accuracy approaching experimental methods, achieving median backbone RMSD of ~1 Angstrom on CASP14 targets. AlphaFold2 uses a neural network architecture that processes multiple sequence alignments (MSAs) and pairwise residue features through an iterative "Evoformer" module and a structure module that outputs 3D coordinates with per-residue confidence scores (pLDDT). AlphaFold has been applied to predict structures for virtually every known protein sequence (>200 million in the AlphaFold Protein Structure Database), transforming structural biology from an experimental bottleneck to a computationally accessible resource. Limitations include poor performance on intrinsically disordered regions, difficulty with multiple conformational states, and challenges with protein-protein and protein-ligand interactions.

Explainer

The protein structure prediction problem — determining a protein's 3D structure from its amino acid sequence — was one of the grand challenges of computational biology for 50 years. At the biennial CASP competition (Critical Assessment of protein Structure Prediction), the best methods gradually improved from producing vague shapes to reasonable backbone traces. Then AlphaFold2 arrived at CASP14 in 2020 and essentially solved the problem for single-domain proteins, producing predictions indistinguishable from experimental structures for many targets.

AlphaFold2's architecture has two key innovations. The Evoformer module processes a multiple sequence alignment (MSA) of the target protein and its evolutionary relatives, extracting co-evolutionary signals — patterns of correlated mutations that indicate spatial proximity. If positions i and j consistently mutate together across species (when i mutates to a larger residue, j compensates by mutating to a smaller one), they are likely in contact in the 3D structure. The Evoformer uses attention mechanisms to process these signals across the MSA and a pairwise representation, iteratively refining the predicted residue-residue relationships. The structure module then converts these refined representations into 3D coordinates, outputting both the structure and per-residue confidence scores (pLDDT).

The impact has been transformative. The AlphaFold Protein Structure Database provides predicted structures for over 200 million protein sequences — essentially every known protein in UniProt. Structural biology has shifted from "we need to determine this structure" to "we already have a prediction — does it need experimental validation for this particular question?" For many applications (identifying homologs, understanding domain architecture, guiding mutagenesis), AlphaFold predictions are sufficient. For drug design, enzyme mechanism analysis, and studying conformational dynamics, experimental structures remain necessary because the details matter at a level where AlphaFold's predictions may not be reliable.

The limitations are instructive. AlphaFold predicts a single static structure, but many proteins function through conformational changes — an enzyme may need to open and close, a receptor may switch between active and inactive states. AlphaFold typically predicts the most common or most stable conformation, potentially missing functionally critical alternative states. Intrinsically disordered regions are correctly identified (low pLDDT) but their coordinates are meaningless. Protein-protein interactions, protein-ligand binding, and post-translational modification effects are not reliably predicted by AlphaFold2 (AlphaFold3 makes progress here but accuracy varies). And for proteins without evolutionary homologs (de novo designed proteins, orphan sequences), the MSA provides no useful information, and prediction accuracy degrades. AlphaFold has not replaced structural biology — it has redefined the questions structural biologists need to answer, shifting focus from routine structure determination to functional dynamics, molecular interactions, and the biology that predictions alone cannot reveal.