Protein structure prediction powered by artificial intelligencefrom biochemical foundations to practical applications

Fifty years. That's roughly how long the protein folding problem sat as one of biology's great unsolved puzzles. Then, at CASP14 — the 2020 edition of the biennial blind structure prediction competition — AlphaFold2 posted a median backbone root-mean-square deviation of 0.96 angstroms on the test domains. The next best method sat at 2.8 angstroms. That isn't incremental progress. That's a different game entirely. To understand what actually happened — the architecture, the ideas it built on, and what it still cannot do — you have to start not with the algorithm, but with why experiments alone could never get us there. Three experimental methods have dominated structural biology for decades: X-ray crystallography, nuclear magnetic resonance spectroscopy, and cryo-electron microscopy. Each is genuinely powerful. Each hits a hard wall. Crystallography can deliver atomic-level resolution, but it demands high-quality crystals, and a substantial fraction of biologically important proteins — membrane proteins, large flexible assemblies, and intrinsically disordered systems — simply won't crystallize reliably. Because crystallography measures a lattice average, conformational dynamics and alternative states are frequently invisible or smeared out. Nuclear magnetic resonance works beautifully in solution and captures dynamics, but the technique runs into spectral overlap and signal attenuation as molecules grow. The practical ceiling is roughly 30 kilodaltons, or about 270 amino acids, which excludes the multi-domain assemblies that carry out most cellular signaling. Cryo-electron microscopy has been the exciting expansion of the last decade, and it genuinely excels at large complexes where crystallization fails. But it remains expensive, computationally intensive in reconstruction, and it loses resolution for small proteins or for regions that are conformationally heterogeneous. Collectively, these three methods are low-throughput and high-cost. Acquiring a single high-quality structure can take months. Meanwhile, sequence databases have grown at genomic speed. The gap between known sequences and known structures just kept widening. That gap drove two parallel traditions in computational prediction. Template-based and homology methods work by finding structural relatives in the Protein Data Bank and using those solved structures as scaffolding. As Jumper and colleagues summarize, these methods work well when a close homolog has been solved, but fall far short of atomic accuracy in the majority of cases where a close homologue has not been solved experimentally. For novel folds, they fail. Physics-first approaches — trying to simulate folding from thermodynamic principles — face a different problem: computational intractability. Even moderate-sized proteins require enormous sampling of conformational space, and the energy functions are never quite accurate enough. The crucial bridge between these classical approaches and modern deep learning was co-evolutionary contact prediction. Weigt and colleagues in 2009 and Marks and colleagues in 2011 showed that correlated substitutions across deep multiple sequence alignments reveal pairs of residues that are spatially close in the folded structure. If position A and position B co-vary across thousands of species, they are almost certainly in contact — evolution is preserving the interaction. Jones and colleagues' PSICOV method extracted these sparse inverse-covariance signals statistically, giving structure predictors geometric constraints that didn't require a template. Deep learning then turbocharged this: Wang and colleagues demonstrated that neural networks could turn covariation into far more reliable contact maps, and deep-learning-guided predictions at CASP13 showed the field the direction it was heading. That's the ramp. CASP14 is where you see what's at the top of it. Jumper and colleagues reported those CASP14 numbers clearly. AlphaFold2's median backbone root-mean-square deviation at ninety-five percent coverage was 0.96 angstroms, with a confidence interval of 0.85 to 1.16. The next best method produced a median of 2.8 angstroms, with a confidence interval of 2.7 to 4.0. For all-atom accuracy, AlphaFold came in at 1.5 angstroms versus 3.5 for the best alternative. The paper describes this as accuracy competitive with experimental structures in a majority of cases. That phrase is doing a lot of work. It means the computation is, in many instances, as informative as spending months in the lab. So what is AlphaFold2 actually doing? The network has two major components: the Evoformer trunk and the structure module. The Evoformer jointly processes two arrays. The first is the multiple sequence alignment representation — rows are sequences from homologous proteins, and columns are residue positions. The second is a pairwise representation encoding relationships between every pair of residue positions. These two tracks exchange information continuously through every Evoformer block. In the multiple sequence alignment track, attention operations are biased by the pairwise array, so spatial hypotheses directly influence which evolutionary signals get attended to. The updated multiple sequence alignment then feeds back into the pairwise array via an outer-product update summed across the sequence dimension. This happens inside every block — the two tracks are in constant conversation. Within the pairwise array, AlphaFold enforces geometric consistency through triangle-centered updates. The idea is that any embedding corresponding to a real three-dimensional structure must satisfy triangle inequalities: if residue A is close to B and B is close to C, there are constraints on how far A can be from C. Two operations — a triangle multiplicative update and triangle self-attention — implement this consistency check throughout the Evoformer. The result is a pairwise representation that is geometrically coherent before a single three-dimensional coordinate has been predicted. To encourage the network to actually learn evolutionary covariation rather than overfit to sequence statistics, the training objective masks random positions in the multiple sequence alignment and asks the network to reconstruct them jointly with predicting the structure — a BERT-style objective applied to the alignment. This is what forces the Evoformer to extract phylogenetic signal. Then comes the structure module. Rather than predicting distances and deriving coordinates indirectly, it works in explicit three dimensions. Each residue gets its own rigid-body frame — a rotation and translation that parameterizes the backbone geometry. These frames are updated iteratively through invariant point attention, or IPA. IPA augments standard attention queries and keys with three-dimensional points produced in each residue's local frame, making the attention operation invariant to global rotations and translations while maintaining a strong locality bias. After each IPA update, the residue frames themselves are modified within local coordinate systems, and the backbone propagates. Peptide bond geometry is then encouraged by a violation loss during fine-tuning and enforced exactly by a post-prediction Amber relaxation step — though Jumper and colleagues note that relaxation removes stereochemical errors but does not improve global distance test or lDDT scores. The quality was already there. The per-residue confidence score, known as pLDDT, comes from small networks on the final activations. Jumper and colleagues show that across chains, the linear fit is: lDDT-Cα equals 0.997 times pLDDT minus 1.17, with a Pearson correlation of 0.76. High pLDDT means high local backbone accuracy. Low pLDDT — especially the extended low-confidence regions characteristic of intrinsically disordered proteins — tells you the model is uncertain, and that signal is real. At scale, this matters enormously. The AlphaFold Protein Structure Database, described by Varadi and colleagues, now contains over 214 million precomputed structures. Tunyasuvunakool and colleagues reported confident structural coverage — pLDDT above 70 — for approximately 98.5 percent of the human proteome. Single-chain prediction was a revolution. But most biology happens in complexes, and most drug targets involve ligands, nucleic acids, or protein-protein interfaces. That's where AlphaFold3, described by Abramson and colleagues, represents the next major shift — not just an extension but a change in generative paradigm. AlphaFold3 replaces the structure module's deterministic coordinate prediction with a diffusion-based head. The trunk is now called the Pairformer, composed of 48 blocks, with 128 channels in the pair representation and 384 in the single representation. The Pairformer feeds into a Diffusion Module that operates on per-atom representations and is trained as a generative process — it learns to predict all-atom coordinates by iteratively denoising from a random distribution. At inference, the module runs 20 iterations of this mini roll-out. The key consequence: by working at the all-atom level from the start, AlphaFold3 can handle protein-DNA, protein-RNA, and protein-small-molecule systems within a single unified framework. The benchmarks Abramson and colleagues report are striking. On RNA prediction, a paired Wilcoxon signed-rank test against RoseTTAFold2NA gives a p-value of 1.6 times ten to the negative seventh in favor of AlphaFold3. For DNA systems, the p-value is 5.2 times ten to the negative twelfth. For ligand accuracy, the primary metric is the percentage of pocket-aligned ligand root-mean-square deviations under 2.0 angstroms. One concrete example is a bacterial CRP or FNR regulator bound to DNA and cyclic GMP, with Protein Data Bank entry 7PZB, predicted with a full complex lDDT of 82.8 and a global distance test score of 90.1. For a 4,665-residue heavily glycosylated human coronavirus OC43 spike protein, AlphaFold3 returned a full complex lDDT of 83.0 and a global distance test score of 83.1. These are large, heterogeneous systems that were essentially inaccessible to single-chain methods. The diffusion approach introduces its own failure modes, and Abramson and colleagues document them honestly. The chirality violation rate on the PoseBusters benchmark is 4.4 percent even with a chirality penalty in ranking. Atomic clashes persist, particularly for protein-nucleic acid assemblies exceeding 100 nucleotides combined with more than 2,000 total residues. And the diffusion model tends to hallucinate spurious structural order in disordered regions — though hallucinated regions are typically flagged by low pLDDT, so the confidence signal remains useful. These technical capabilities are now feeding real applications, though with important caveats the literature is explicit about. In drug discovery, AlphaFold models feed virtual screening and structure-based design pipelines at major pharmaceutical companies, including Pfizer, Novartis, and AstraZeneca, which are integrating AlphaFold and RoseTTAFold models into internal workflows for target identification and candidate screening. But pLDDT assesses local structural consistency, not functional correctness. High confidence does not guarantee correct binding-site geometry, side-chain conformations, or ligand placement. Studies using AlphaFold3 structures directly for molecular docking without accounting for conformational flexibility have reported lower hit rates in virtual screening. Hekkelman and colleagues developed AlphaFill specifically to address the missing-ligand problem, enriching database models with cofactors transferred from related experimental structures. For enzyme engineering, the pipeline is becoming increasingly defined. ESMFold is used to rapidly generate backbone predictions. Geometric graph neural networks and attention-based models identify key catalytic residues. Molecular dynamics simulations probe active-site dynamics and substrate-binding stability. On the design side, ProteinMPNN — a message-passing neural network described by Dauparas and colleagues — generates highly foldable, stable sequences. It outperforms prior approaches in speed, accuracy, and experimental success rate. The paper calls it one of the most powerful tools for de novo design of stable enzyme sequences. But experimental validation remains indispensable. Predicted active-site geometry and conformational dynamics are still not accurate enough to bypass testing. In disease research, AlphaFold2 predictions combined with molecular dynamics have been used to study beta-sheet arrangements in amyloid-beta aggregates relevant to Alzheimer's disease and to investigate mutation-induced stability changes in tumor-associated proteins, including the p53 protein, KRAS, and BRCA1. These aren't proof-of-concept demonstrations anymore. These predictions are informing hypotheses about pathogenic mechanisms and therapeutic targeting. The honest accounting of limitations comes back to a few recurring themes. Current models produce a single static conformation, not an ensemble. Abramson and colleagues are explicit that multiple random seeds do not approximate a thermodynamic ensemble, and they document specific failure cases — E3 ubiquitin ligases where AlphaFold3 predicts the closed state for both apo and holo systems, when the apo form should be open. The models don't capture the conformational heterogeneity that determines function in many regulatory proteins. Multiple sequence alignment depth matters in ways that constrain where these methods work. Jumper and colleagues report that accuracy falls substantially when median multiple sequence alignment depth drops below about 30 effective sequences, with gains flattening above roughly 100. For proteins from poorly sampled lineages or truly novel folds without evolutionary relatives, the predictors are operating with degraded input. Confidence scores don't fully convey this. pLDDT reflects internal model consistency, not functional or interaction correctness, and Tunyasuvunakool and colleagues caution explicitly that it should not be treated as a guarantee of binding-site accuracy. The forward path visible in the literature runs through three directions: integrating physics-based simulation with deep learning predictions to sample energy landscapes and conformational ensembles rather than single structures; developing explicit multi-state prediction frameworks that train on or generate multiple conformations rather than one; and building automated design-to-validation pipelines that treat AI predictions as hypotheses to be tested, not answers to be deployed. The field has traveled from low-throughput structure guessing to routinely useful, high-accuracy single-structure prediction. The gap between static prediction and dynamic understanding — that's where the next decade's work lives. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.