AI Protein Structure Prediction: How AlphaFold Changed Biology

The Protein Folding Problem

Proteins are molecular machines that perform virtually every function in living cells. They catalyze chemical reactions, transport molecules, provide structural support, transmit signals, and defend against pathogens. A protein's function is determined by its three-dimensional structure, the precise way its chain of amino acids folds into a compact shape. Understanding this structure is essential for understanding how the protein works and for designing drugs that interact with it.

The folding problem is this: given the linear sequence of amino acids (the primary structure), predict the three-dimensional arrangement of atoms (the tertiary structure). A typical protein chain of 300 amino acids has an astronomical number of possible conformations. Levinthal's paradox, stated in 1969, pointed out that if a protein sampled conformations randomly, it would take longer than the age of the universe to find its correct structure. Yet real proteins fold in milliseconds to seconds. Nature has solved this problem; scientists struggled to replicate the solution computationally for half a century.

Experimental structure determination uses X-ray crystallography, cryo-electron microscopy (cryo-EM), or nuclear magnetic resonance (NMR) spectroscopy. These techniques are powerful but slow and expensive. Crystallography requires growing protein crystals, which can take months and sometimes fails entirely. Cryo-EM has become faster but still requires specialized equipment costing millions of dollars. Before AlphaFold, the Protein Data Bank (PDB) contained about 180,000 experimentally determined structures, covering only a small fraction of the millions of known protein sequences.

How AlphaFold Works

AlphaFold 2, released by DeepMind in 2020, uses a deep learning architecture called the Evoformer that processes two types of information: the target protein sequence and multiple sequence alignments (MSAs) of evolutionarily related proteins. The MSA is key because evolution preserves structural contacts: amino acids that are close together in the folded protein tend to co-evolve, changing in coordinated ways across species. AlphaFold learned to read these co-evolutionary signals and translate them into distance and angle predictions for every pair of amino acids.

The architecture processes information through multiple rounds of attention-based updates, iteratively refining its predictions. The output is a set of 3D coordinates for every atom in the protein, along with a confidence score (pLDDT) for each residue that indicates how reliable the prediction is. Regions with pLDDT above 90 are typically very accurate (less than 1 angstrom error), regions between 70 and 90 are generally correct in overall fold but less precise in details, and regions below 50 are often disordered loops where even the protein itself does not adopt a single fixed structure.

AlphaFold 2 achieved a median GDT score of 92.4 out of 100 in the CASP14 (Critical Assessment of protein Structure Prediction) competition, where scores above 90 indicate accuracy comparable to experimental methods. This represented a leap from the previous best of about 70, and many scientists described it as solving the protein folding problem for single-domain proteins.

AlphaFold 3, released in 2024, extended predictions beyond single proteins to complexes: protein-protein interactions, protein-DNA interactions, protein-ligand interactions, and post-translational modifications. This broader scope makes it directly useful for drug discovery (predicting how a drug molecule binds to a protein) and for understanding the molecular machines that carry out cellular functions (which are typically multi-protein complexes rather than individual proteins).

Other AI Structure Prediction Tools

AlphaFold is the most well-known but not the only AI structure prediction system. ESMFold, developed by Meta AI, uses protein language models instead of multiple sequence alignments. This makes it much faster (predictions in seconds rather than minutes) but slightly less accurate for proteins with many homologs. ESMFold excels for orphan proteins that have few evolutionary relatives, where MSA-based methods struggle because there is not enough co-evolutionary information.

RoseTTAFold, developed by David Baker's lab at the University of Washington, uses a three-track architecture that simultaneously processes sequence, distance, and coordinate information. RoseTTAFold has been particularly influential in protein design: the same lab used related AI methods to design entirely new proteins with no natural counterpart, creating custom enzymes, biosensors, and therapeutic proteins.

OmegaFold, ColabFold, and OpenFold provide additional options with different tradeoffs between speed, accuracy, and ease of use. ColabFold makes AlphaFold accessible through Google Colab notebooks, requiring no local installation or GPU hardware. This democratization has been enormously important: researchers at institutions without computational resources can now predict protein structures for free using a web browser.

Practical Applications

Drug Discovery

Structure-based drug design requires knowing the 3D shape of the target protein, particularly the binding site where a drug molecule could attach. Before AlphaFold, researchers could only design drugs against the roughly 180,000 proteins with experimental structures. Now they can design against essentially any protein. This has opened up thousands of previously "undruggable" targets, including many proteins involved in cancer, neurodegeneration, and infectious disease.

Pharmaceutical companies use AlphaFold predictions to identify potential binding pockets, run virtual screening against those pockets, and optimize lead compounds to fit the pocket better. When the AlphaFold prediction has high confidence in the binding site region (pLDDT above 80), the predicted structure is accurate enough for computational docking studies that would normally require an experimental structure.

Understanding Disease Mechanisms

Many genetic diseases are caused by single amino acid changes (missense mutations) that alter protein structure and function. AlphaFold predictions help researchers understand why specific mutations cause disease: does the mutation destabilize the protein fold, disrupt a binding interface, block an active site, or create a new aggregation-prone surface? This structural interpretation transforms a list of disease-associated mutations into mechanistic understanding that guides therapeutic development.

Enzyme Engineering

Industrial biotechnology relies on enzymes to catalyze chemical reactions under mild conditions: breaking down plastic waste, producing biofuels, manufacturing pharmaceuticals, and processing food. AI structure prediction enables rational enzyme engineering, where researchers modify specific amino acids to improve enzyme stability, activity, or substrate specificity based on structural understanding. Combined with generative AI that proposes mutations, this approach designs better enzymes in weeks instead of the months or years required by traditional directed evolution.

Evolutionary Biology

Comparing predicted structures across species reveals evolutionary relationships that sequence comparison alone cannot detect. Proteins with similar structures but different sequences (remote homologs) indicate shared evolutionary ancestry that diverged so long ago that sequence similarity has been lost. AlphaFold predictions for entire proteomes (all proteins in an organism) enable large-scale structural phylogenetics that maps the evolution of protein folds across the tree of life.

Limitations and Caveats

AlphaFold predicts static structures, but proteins are dynamic. Many proteins change shape when they bind ligands, interact with other proteins, or respond to cellular signals. AlphaFold typically predicts one conformation (usually the most stable), and it may miss alternative conformations that are biologically important. For proteins that function as molecular switches, cycling between active and inactive states, a single predicted structure captures only half the story.

Intrinsically disordered regions, segments of protein that do not fold into a stable structure, are flagged by low confidence scores but cannot be meaningfully predicted. About 30% of the human proteome is intrinsically disordered, and these regions often play crucial roles in signaling and regulation. AI prediction tools are transparent about this limitation, which is why checking the confidence scores is essential before using any prediction.

Membrane proteins, which sit within the lipid bilayer of cell membranes, are predicted less accurately than soluble proteins because the training data contains fewer experimental membrane protein structures. This is a significant limitation because many important drug targets (GPCRs, ion channels, transporters) are membrane proteins. Prediction accuracy is improving as more membrane protein structures are determined experimentally and added to training sets.

AlphaFold predictions should be validated experimentally for high-stakes applications. Using an unvalidated prediction to design a drug that enters clinical trials would be scientifically irresponsible. The prediction is a hypothesis about the structure; the experiment confirms or refutes it. For exploratory research, understanding biological mechanisms, or prioritizing targets for further study, the predictions are reliable enough to guide decisions without experimental validation of every structure.