Artificial intelligence has cracked one of biology’s most stubborn puzzles: predicting how a protein folds into its three-dimensional shape from nothing more than its amino acid sequence. For roughly half a century, scientists understood that a protein’s shape dictates its function but had no reliable way to compute that shape from first principles. The breakthrough came from deep-learning systems, most famously AlphaFold, that now predict protein structures with accuracy rivaling laboratory experiments. The ripple effects are already reshaping drug discovery, evolutionary biology, and the design of proteins that have never existed in nature.
Why Protein Folding Was So Hard
Every protein in your body starts as a long chain of amino acids, strung together like beads on a wire. Within milliseconds of being made, that chain crumples into a specific 3D shape. The shape determines what the protein does: whether it carries oxygen, digests food, fights infection, or relays a nerve signal. Get the shape wrong and the protein malfunctions, which is exactly what happens in diseases like cystic fibrosis and sickle cell anemia.
The difficulty of predicting that shape was captured in what became known as Levinthal’s paradox. In principle, a protein chain could twist into an astronomically large number of possible configurations. A random search through all of them would take longer than the age of the universe. Yet real proteins fold in seconds or less. Something guides the chain toward the right answer almost instantly.1PubMed Central. Levinthal’s paradox Nature solves this through physics and chemistry: the chain is biased at every step toward configurations that are energetically favorable, funneling it rapidly toward the final shape. The challenge for scientists was to figure out how to replicate or shortcut that funnel computationally.
The Experimental Bottleneck
Before AI entered the picture, determining a protein’s structure meant painstaking laboratory work. X-ray crystallography, the workhorse technique for decades, requires growing a crystal of the protein and then bombarding it with X-rays. The diffraction pattern reveals atomic positions, but getting a usable crystal can take months or years, and many proteins refuse to crystallize at all. Nuclear magnetic resonance (NMR) spectroscopy works in solution but is limited to smaller proteins. Cryo-electron microscopy, which flash-freezes molecules and images them with an electron beam, has made it possible to determine structures of large complexes that resisted older methods.2PubMed Central. Cryo-electron microscopy and X-ray crystallography: complementary approaches to structural biology and drug discovery
Even with these techniques, progress was slow. By 2020, the Protein Data Bank, the global repository where experimentalists deposit solved structures, held around 170,000 entries. That sounds like a lot, but there are hundreds of millions of distinct proteins across all known organisms. At the rate experimental labs were going, cataloging the full protein universe would have taken centuries.
CASP and the Scoreboard That Drove Progress
Scientists needed an honest way to measure how well computational methods were doing. Starting in 1994, a biennial competition called CASP (Critical Assessment of protein Structure Prediction) provided that scoreboard. Organizers would choose proteins whose structures had been solved in the lab but not yet published. Computational teams would submit their best guesses, and independent assessors would compare those predictions to the hidden experimental answers.3PubMed Central. Critical assessment of methods of protein structure prediction (CASP)-Round XII For years, progress was incremental. Teams used energy functions, statistical potentials drawn from known structures, and evolutionary information mined from related sequences. Predictions improved, but for many targets the results were far from experimental accuracy.
That changed dramatically at CASP13 in 2018, when DeepMind’s first AlphaFold system outperformed every competitor, and then again at CASP14 in 2020, when AlphaFold2 produced predictions so accurate that for many targets, the difference between the AI’s answer and the lab result was smaller than the uncertainty in the experiment itself.4PubMed Central. Critical assessment of methods of protein structure prediction (CASP)-Round XIV The community experiment that had tracked gradual improvement for over two decades suddenly had to reckon with what looked like a solved problem.
How the AI Actually Works
AlphaFold2’s architecture revolves around a component called the Evoformer, which processes two kinds of information simultaneously: the amino acid sequence of the target protein and the evolutionary relationships found in databases of similar sequences from other organisms.5PubMed Central. Toward the appropriate interpretation of Alphafold2 When two positions in a protein tend to change together across species, it signals that those positions are physically close in the folded structure. The network learns to extract these co-evolutionary signals and translate them into spatial coordinates.
AlphaFold2 is not the only game in town. ESMFold, developed by Meta AI, takes a different approach: instead of aligning the target sequence against thousands of related sequences, it feeds the sequence directly into a protein language model trained on hundreds of millions of sequences. The model, scaled to 15 billion parameters, infers atomic-level structure from the sequence alone.6PubMed. Evolutionary-scale prediction of atomic-level protein structure with a language model This makes ESMFold substantially faster since it skips the time-consuming step of building a multiple sequence alignment. Benchmarks show that for many proteins the accuracy gap between alignment-free methods and AlphaFold2 is negligible, and the alignment-free predictors can be ten to thirty times faster.7PubMed Central. Balancing speed and precision in protein folding: a comparison of AlphaFold2, ESMFold, and OmegaFold In practice, researchers increasingly pick the tool that fits their need: AlphaFold2 when maximum precision matters, a language-model predictor when they need quick results across thousands of sequences.
From Single Proteins to Molecular Complexes
Proteins rarely act alone. They bind other proteins, wrap around DNA, grab small molecules, and coordinate metal ions. Predicting those interactions is arguably more important for drug design than predicting any single protein’s shape. AlphaFold3, released in 2024, was built specifically for this problem. It uses a diffusion-based architecture that can predict the joint structure of complexes containing proteins, nucleic acids, small molecules, ions, and chemically modified residues, all within a single unified framework.8PubMed Central. Accurate structure prediction of biomolecular interactions with AlphaFold 3
The accuracy gains over previous specialized tools are substantial. For protein-ligand interactions (the bread and butter of drug design), AlphaFold3 outperforms state-of-the-art docking software. For protein-nucleic acid interactions, it beats dedicated nucleic-acid predictors. And for antibody-antigen binding, it surpasses earlier versions of AlphaFold designed specifically for protein-protein complexes.9Nature. Accurate structure prediction of biomolecular interactions with AlphaFold 3 Being able to model all of these interaction types in one system is a qualitative shift: researchers no longer need to stitch together outputs from half a dozen specialized tools.
Drug Discovery Gets a New Toolbox
The pharmaceutical industry has been one of the fastest adopters of protein-folding AI. The traditional drug-discovery pipeline starts by finding a target protein involved in a disease, then screening millions of chemical compounds to find ones that bind to it. Experimental screening is expensive and slow. Virtual screening, where a computer docks candidate molecules into a protein’s binding site, has been used for years but depends on having an accurate structure of the target. AlphaFold has expanded the menu of druggable targets enormously.
Benchmarking studies show that AlphaFold2-predicted structures can be used for virtual screening with performance approaching that of experimentally solved structures, especially after refinement protocols are applied.10PubMed. Benchmarking Refined and Unrefined AlphaFold2 Structures for Hit Discovery For kinases, a large family of proteins frequently targeted by cancer drugs, multi-state modeling using AlphaFold-generated conformations can even outperform crystal structures in discovering diverse molecular scaffolds.11Scientific Reports. Improving docking and virtual screening performance using AlphaFold2 multi-state modeling for kinases
A newer approach called DrugCLIP pushes the speed envelope further. By pairing structure predictions with a contrastive learning framework, it achieves virtual screening speeds up to ten million times faster than conventional docking while maintaining competitive accuracy. In laboratory validation, DrugCLIP identified new inhibitors for previously “undruggable” targets using only AlphaFold2-predicted structures, including one target that lacked any experimental structure or known small-molecule binders.12PubMed. Deep contrastive learning enables genome-wide virtual screening
Understanding Disease at the Molecular Level
Beyond finding drugs, AI-predicted structures are helping researchers understand why diseases happen in the first place. When a genetic mutation swaps one amino acid for another, the effect on the protein’s structure and function can range from negligible to catastrophic. AI tools let scientists quickly model what a mutation does to a protein’s shape and how that might disrupt its interactions. For rare diseases, where experimental structures are often unavailable and funding for structural biology is limited, this is particularly valuable. Researchers have used the AlphaFold database to model the flexibility profiles of proteins linked to rare genetic conditions like infantile-onset ascending hereditary spastic paralysis.13PubMed. AI-based protein structure databases have the potential to accelerate rare diseases research: AlphaFoldDB and the case of IAHSP/Alsin
In bleeding disorders, AlphaFold2-multimer and AlphaFold3 have been used to predict how known mutations in clotting factor VIII and von Willebrand factor disrupt their complex. The AI predictions are consistent with existing cryo-electron microscopy structures, and when combined with molecular dynamics simulations, they can assess the impact of both known and novel mutations with potential applications in precision medicine.14PubMed. Predicting the effects of single pathological mutations in hemophilia A and type 2N von Willebrand diseases using AlphaFold2-multimer and AlphaFold3 The broader vision is to use structure predictions as a first-pass filter: when a patient’s genome reveals a novel variant of uncertain significance, AI can quickly suggest whether that variant is likely to distort the protein’s function.
Designing Proteins That Never Existed
Predicting the structures of natural proteins was the original challenge, but the technology has enabled something even more ambitious: designing entirely new proteins from scratch. RFdiffusion, built by fine-tuning a structure prediction network on denoising tasks, generates protein backbones that fold into stable structures and perform specified functions. Researchers have used it to design symmetric protein assemblies, metal-binding proteins, and proteins that bind tightly to chosen targets, with hundreds of designs confirmed experimentally.15PubMed Central. De novo design of protein structure and function with RFdiffusion
The latest iteration, RFdiffusion3, extends this capability to the full complexity of biological chemistry. It generates protein structures in the context of ligands, nucleic acids, and other non-protein atoms, modeling every atom explicitly. This makes it possible to design, for example, enzymes with precisely positioned active sites or proteins that grip a specific stretch of DNA. Experimentally validated designs already include novel DNA-binding proteins and cysteine hydrolases.16PubMed Central. De novo Design of All-atom Biomolecular Interactions with RFdiffusion3 In principle, this opens a route to custom-built therapeutic proteins, biosensors, and catalysts that evolution never explored.
What the AI Still Gets Wrong
The success stories are real, but the limitations matter just as much for anyone relying on these tools. The most fundamental one is that AlphaFold predicts a static snapshot, but proteins are not static objects. They breathe, flex, and sometimes rearrange dramatically to carry out their functions. Intrinsically disordered regions, stretches of protein that do not settle into any single structure, are especially poorly handled. AlphaFold2’s static prediction framework simply cannot capture the conformational diversity that defines these regions.17PubMed Central. Modeling intrinsically disordered regions from AlphaFold2 to AlphaFold3 Given that an estimated third of the human proteome contains significant disordered regions, this is not a minor gap.
The distinction between structure prediction and the protein folding problem itself is also sharper than most headlines suggest. Structure prediction asks “what does the final folded shape look like?” The folding problem, more broadly, includes “how does the chain get there?” AlphaFold2 answers the first question brilliantly but was not designed to address the second.18Nature Communications. Rapid estimation of protein folding pathways from sequence alone using AlphaFold2 Newer tools like PathFold aim to predict entire folding pathways directly from sequence, which could shed light on how misfolding diseases like Alzheimer’s and Parkinson’s progress.19PubMed Central. PathFold: Predicting the Entire Protein Folding Pathway from Protein Sequence Alone
Another practical shortcoming: AlphaFold is surprisingly poor at predicting the effect of point mutations on protein stability. When researchers tested whether AlphaFold’s confidence metrics could predict how a single amino acid swap changes a protein’s thermodynamic stability or function, they found very weak or no correlation with experimentally measured values.20PubMed Central. Using AlphaFold to predict the impact of single mutations on protein stability and function This means that while AlphaFold can give you the wild-type structure beautifully, you cannot simply mutate the input sequence and expect the output to tell you whether the mutant protein will be stable. Specialized tools like PROST have been developed to fill this gap, using AlphaFold2 features as inputs to a separate predictor trained specifically for stability changes.21PubMed. PROST: AlphaFold2-aware Sequence-Based Predictor to Estimate Protein Stability Changes upon Missense Mutations
How AI and Lab Work Reinforce Each Other
A common misconception is that AI predictions will replace experimental structural biology. In practice, the two are converging. A tool called ROCKET integrates X-ray crystallography and cryo-EM data directly into OpenFold’s inference process, using the AI’s learned representations as a prior. This hybrid approach allows barrier-crossing structural rearrangements, like drug-induced conformational changes, that conventional computational refinement struggles with. The resulting models match the accuracy of manually curated experimental structures.22PubMed Central. AlphaFold as a Prior: Experimental Structure Determination Conditioned on a Pretrained Neural Network For experimentalists, this means spending less time on tedious refinement and more on the biology. For computational biologists, it means their predictions are constantly being calibrated against fresh data.
AlphaFold3 further bridges this divide by accepting experimental inputs alongside sequence data: post-translational modifications, partial structural information from experiments, and annotations about the cellular environment such as lipid bilayer context for membrane proteins.23PubMed Central. Analysing protein complexes in plant science: insights and limitation with AlphaFold 3 Feeding these additional constraints into the model improves prediction accuracy in exactly the cases where a pure sequence-to-structure approach would stumble.
Illuminating the Dark Proteome and Deep Evolution
A huge fraction of known protein sequences have no experimentally determined function, a realm sometimes called the “dark proteome.” Traditional methods for assigning function rely on finding a well-studied protein with a similar sequence, but when sequence similarity is low these methods fail. AI-generated structure predictions offer a way around this, because two proteins can share a recognizable 3D fold even when their sequences have diverged beyond recognition. Using structural comparisons powered by AlphaFold, researchers have bridged the gap between sequence-based and structure-based protein classification systems, providing functional insights into previously uncharacterized proteins.24Journal of Molecular Biology. Bridging the Gap between Sequence and Structure Classifications of Proteins with AlphaFold Models A recent tool called FANTASIA, which pairs language models with predicted structures across roughly a thousand animal proteomes, assigns predicted functions to virtually all proteins, including up to half that had no annotation through traditional methods.25PubMed Central. FANTASIA leverages language models to decode the functional dark proteome across the animal tree of life
This has also opened a window into evolutionary history. Protein language models trained on hundreds of millions of sequences learn a representation space that mirrors biology at multiple scales, from the chemical properties of individual amino acids up to remote evolutionary relationships between protein families.26PubMed Central. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences Newer remote homolog detection methods using these embeddings have revealed evolutionary connections that sequence searches alone never could, such as the discovery that certain transmembrane protein families exist across bacteria, archaea, and eukaryotes rather than being limited to one branch of life.27Current Biology. How Protein Folding AI Is Solving a Decades-Old Problem These findings are rewriting our understanding of which protein architectures are ancient and which arose more recently.
Post-Translational Modifications and Drug Interactions
Cells constantly decorate their proteins with chemical tags after they are made: phosphate groups, sugar chains, methyl groups, and dozens of others. These post-translational modifications can reshape a binding pocket, create new interaction surfaces, or toggle a protein between active and inactive states. Until recently, predicting how modifications alter structure and drug binding was almost entirely out of reach for computational tools. AI-based structure prediction has changed that. In one large-scale study, researchers used AlphaFold3 and related tools to generate over 14,000 models of modified human proteins with docked drug molecules, mapping how specific modifications can influence drug binding across all known human drug targets.28PubMed Central. Leveraging AI to explore structural contexts of post-translational modifications in drug binding This kind of analysis would have been unimaginable five years ago.
Who Owns a Protein Designed by an Algorithm
As AI-designed proteins move from academic papers toward commercial products, uncomfortable legal questions are surfacing. Patent law in most jurisdictions requires a human inventor. When an AI system autonomously generates a novel protein that could become a therapeutic, the question of who gets listed on the patent application is not academic. Current intellectual property frameworks were built for human-driven invention and are poorly equipped for a world where the creative act is performed, in part or in full, by a neural network.29Journal of Ethics and Legal Technologies. The Legal Status of AI-Generated Biotech Inventions: Who Owns Life Engineered by Algorithms? The situation is further complicated by biotechnology’s “product of nature” doctrine: courts have long held that naturally occurring genes and proteins cannot be patented. When an AI designs a protein that closely resembles something evolution might have produced, the boundary between discovery and invention blurs in ways that existing case law does not clearly resolve. These questions are already being tested in patent offices around the world, and the answers will shape how freely the benefits of protein-folding AI spread through medicine, agriculture, and industry.