Deep learning methods for protein structure prediction

For fifty years, the protein folding problem sat at the center of structural biology like an unanswered question that everyone agreed was important, yet almost no one could solve. The best computational methods, built on decades of physical intuition and careful statistical heuristics, were wrong more often than right when no evolutionary template was available. Then a single architecture changed that. What follows is the story of how that happened — what the methods were, why they failed, and what the new approaches actually do differently at the level of equations and architecture. Let’s start with the problem itself, because it is genuinely hard. Proteins are linear chains of amino acids whose biological functions are entirely determined by their three-dimensional shapes. That mapping from one-dimensional sequence to three-dimensional structure is what the field calls the folding problem. Structure is described in four levels: primary sequence, local secondary structure, which includes alpha-helices, beta-strands, and turns, tertiary three-dimensional fold, and quaternary assembly of multiple chains. Each level above the first emerges from physical forces: backbone hydrogen bonds for secondary structure, hydrophobic packing and van der Waals contacts for tertiary structure, and subunit interfaces for quaternary complexes. The functional regions — enzyme active sites, ligand binding pockets, and allosteric interfaces — are located at the tertiary and quaternary levels. The gap that drove computational prediction is stark. UniProt contains more than 200 million sequences. The Protein Data Bank holds roughly 200,000 experimentally determined structures. The experimental methods — X-ray crystallography, cryo-electron microscopy, and nuclear magnetic resonance — are slow and expensive, and they strongly favor soluble, well-behaved proteins. The conformational search problem makes things worse: the number of possible backbone geometries grows exponentially with chain length, making exhaustive sampling completely intractable beyond roughly 100 residues. So, the field needed another way in. Traditional methods divided cleanly into two classes, and Qin and colleagues survey both in their review. Template-based modeling, which includes homology modeling, fold recognition, and comparative modeling, works by finding a known structure with sequence similarity to the target, aligning the target onto it, and building a model from that template. Tools like SWISS-MODEL, HHpred, and BLAST operationalized this. The practical rule of thumb that emerged was you need at least around 30 percent sequence identity to a known template to get near-experimental resolution. Below that threshold, accuracy degrades sharply. Template-free or ab initio methods tried to bypass that dependency. Fragment-based assembly, exemplified by Rosetta, works by searching a library of known short structural fragments, assembling them along the sequence, and scoring the resulting models with energy functions that include van der Waals, hydrogen bonding, and electrostatic terms. Monte Carlo sampling then navigates conformational space toward low-energy assemblies. Pure ab initio folding seeks low-energy conformations by molecular dynamics or Monte Carlo on atomistic force fields. Both strategies hit the same wall. The Critical Assessment of Protein Structure Prediction, or CASP, benchmarking competition made these limits visible in hard numbers. Qin and colleagues report that in CASP7, I-TASSER on free-modeling targets, which is the hardest category where no usable template exists, produced correct topologies, defined as a TM-score greater than 0.5, on only 7 of 19 targets. Those targets were capped at 155 residues in length. Even with that constraint, the average root mean square deviation in aligned regions was 13.5 angstroms for free modeling versus 5.0 angstroms for template-based approaches. For context, a carbon atom is about 1.4 angstroms wide. A 13.5 angstrom average error means the coordinates are essentially wrong. When evolution did not provide a guide, accuracy fell off a cliff. Three bottlenecks drove that failure. First, extracting coevolutionary signal from multiple sequence alignments was done with shallow statistical tools, such as direct coupling analysis and its predecessors, that provided imperfect pairwise residue contact estimates. Second, the scoring machinery was a limiting factor. Traditional approaches relied on hand-crafted physical and statistical energy functions that couldn't faithfully discriminate between native and non-native folds. Third, conformational sampling was intractable at scale. Fragment assembly and Monte Carlo searches face exponentially growing search spaces, and long-range interactions between residues far apart in sequence are hard to capture from short local fragments. These three bottlenecks pointed exactly to where data-driven approaches would eventually intervene. The first wave of deep learning in this field used convolutional neural networks applied to contact and distance prediction. Models like DeepCov used fully convolutional architectures on direct coupling analysis-derived coevolution signals, treating residue pairs as two-dimensional images and learning spatial patterns correlated with tertiary structure. Recurrent neural networks and gated variants like long short-term memory networks followed, handling long-range dependencies along the one-dimensional sequence. Bidirectional gated recurrent units, combined with convolutional modules, extracted both global sequence context and local structural motifs. These were real improvements, but they were operating on the same fundamental input — the multiple sequence alignment — just with richer function approximators. The decisive shift was the Transformer and attention mechanism. Self-attention computes relationships between every pair of positions in an input simultaneously, allowing it to model global dependencies and be parallelized at scale. When applied to protein structure, attention provides a natural way to reason about pairwise residue relationships across an entire sequence in a single forward pass. The architecture that crystallized this insight most forcefully was AlphaFold2's Evoformer. The Evoformer processes two coupled representations simultaneously: an MSA representation with shape N-sequences by N-residues, where the column dimension encodes residue variation across homologs, and a pair representation with shape N-residues by N-residues, encoding directed relationships between every pair of residues. Within each Evoformer block, information flows bidirectionally between these two representations. An elementwise outer product of MSA rows is summed over the MSA dimension and used to update the pair representation — that's the mechanism by which coevolutionary signal gets converted into pairwise geometric constraints. The pair representation then undergoes triangle multiplicative updates and triangle self-attention, operations that enforce geometric consistency by using two edge representations to jointly update a third. AlphaFold2 stacks 48 of these blocks in its trunk, and with recycling — feeding outputs back through the network — generates trajectories of up to 192 intermediate structural predictions per input. Downstream of the Evoformer, the Structure Module translates pair and single residue representations into explicit three-dimensional coordinates. Jumper and colleagues describe each residue as a "residue gas": an independent rotation and translation defining the backbone frame, plus side-chain chi-angle parameters. The module uses invariant point attention, which augments standard attention queries, keys, and values with three-dimensional points expressed in local residue frames, so the attention is invariant to global rotations and translations and has strong geometric locality. Updates to backbone frames are equivariant. Small per-residue heads predict side-chain chi angles and the per-residue confidence metric known as predicted local distance difference test score. The primary training loss is frame-aligned point error, which compares predicted and true atom positions under many different residue-frame alignments, enforcing local orientational correctness and providing a chirality signal. Two training strategies were critical. First, a noisy-student self-distillation: a trained network predicted structures for roughly 350,000 diverse sequences from Uniclust30, a high-confidence subset was assembled, and the architecture was retrained from scratch on a mixture of Protein Data Bank experimental data and those distilled predictions. Second, recycling, which gave the model iterative refinement across cycles. The result at CASP14 was not an incremental improvement. Jumper and colleagues report that AlphaFold2 achieved a median backbone accuracy on the 87-domain CASP14 set of 0.96 angstroms root mean square deviation at the 95th percentile cutoff, compared to 2.8 angstroms for the next-best method. All-atom accuracy was 1.5 angstroms versus 3.5 angstroms for the best alternative. At sub-angstrom backbone error, predictions are at or near atomic resolution for well-predicted domains. On a filtered set of recent Protein Data Bank chains, the median backbone C-alpha root mean square deviation was 1.46 angstroms, with a 95 percent confidence interval from 1.40 to 1.56 angstroms. The confidence metric predicted local distance difference test score was calibrated against empirical accuracy with a near-linear relationship: lDDT C-alpha equals 0.997 times predicted local distance difference test score minus 1.17, meaning the reported confidence score is almost proportional to observed local accuracy. AlphaFold3 represents the next architectural evolution, and it is worth being precise about what changed. Abramson and colleagues describe AlphaFold3 as replacing the residue-frame-centric structure module with a Pairformer trunk, which includes 48 blocks with 128 pair channels and 384 single channels, and a diffusion-based Diffusion Module that generates atomic coordinates directly rather than placing residue frames and then computing torsion parameters. The diffusion module operates on fine-grained per-atom representations and generates coordinates by learning to reverse a noise process. That shift matters because it removes the residue-only abstraction: the model can now jointly generate proteins, nucleic acids, small molecules, ions, and modified residues in a single unified framework, without requiring separate handling of each molecular class. The performance gains on non-protein entities are concrete. On nucleic acid benchmarks, AlphaFold3 showed statistically significant improvements over RoseTTAFoldNA, which had been a leading system for protein-nucleic acid complexes. For RNA, the comparison used 29 structures with a p-value of 1.6 times ten to the negative seventh; for DNA, 63 structures with a p-value of 5.2 times ten to the negative twelfth. On full-complex examples, a bacterial transcriptional regulator bound to DNA and cyclic GMP achieved a full-complex lDDT of 82.8 and GDT of 90.1; a heavily glycosylated human coronavirus OC43 spike protein reached a full-complex lDDT of 83.0 and GDT of 83.1. These are complex, multicomponent systems with covalent modifications — exactly the targets where template-based methods and even AlphaFold2 struggled most. Confidence estimation in AlphaFold3 extends the AlphaFold2 framework. Residue-level predicted local distance difference test score remains. The predicted aligned error matrix gives a per-token-pair expected error, effectively a domain-level confidence map that reveals uncertain interfaces. Interface pTM, which stands for predicted TM score, is used for ranking multi-chain predictions. These metrics are not decorative: AlphaFold3's authors use them to flag failure modes explicitly, including a chirality-violation rate of 4.4 percent on the PoseBusters benchmark and atomic clashes that concentrate in large assemblies — particularly protein-nucleic acid complexes exceeding 100 nucleotides or total sizes above roughly 2,000 residues. The data and evaluation infrastructure that enables all of this is worth examining. The key databases include the Protein Data Bank for experimental ground truth, UniProt and UniRef90 for sequence data, Uniclust30 for clustered sequence search, and BFD and MGnify for metagenomic sequences that dramatically expand the depth of multiple sequence alignments. Some large sequence collections used in training contain over 2.5 billion sequences. Multiple sequence alignment depth matters enormously: Jumper and colleagues found that prediction accuracy falls markedly when median multiple sequence alignment depth drops below roughly 30 effective sequences. Complementary to multiple sequence alignment-dependent approaches, self-supervised protein language models, such as ESM-based models and ESMFold, use masked-language objectives on unlabeled sequence corpora to produce embeddings that support single-sequence predictions without multiple sequence alignments, with predicted local distance difference type confidence applied to their outputs as well. Evaluation uses multiple metrics because each captures different failure modes. Root mean square deviation measures coordinate deviation directly but is sensitive to large localized errors; Jumper and colleagues use a variant that excludes the worst 5 percent of C-alpha positions. TM-score is length-normalized and emphasizes global fold similarity — values above 0.5 generally indicate the same fold. Local distance difference test score is superposition-free and local, which is why it underlies predicted local distance difference test score confidence. DockQ and interface local distance difference test score quantify interface quality for complexes. None of these alone is sufficient; together they triangulate accuracy at different scales. The Protein Data Bank's biases create a ceiling on what can be learned from it. Qin and colleagues note that experimental structures heavily favor soluble proteins, leaving membrane proteins underrepresented. Models trained primarily on soluble protein data struggle on membrane-embedded conformations — a direct limitation for drug discovery, since many high-value targets are membrane proteins. AlphaFold-Multimer has been applied to over 500,000 predicted complexes, with 70 percent of predictions within 4.0 angstroms root mean square deviation of experimental structures, but large multicomponent assemblies and organelle-scale systems still stress current approaches. Intrinsically disordered regions remain a persistent failure mode with a specific character. AlphaFold3's authors document spurious structural order — hallucinations — in disordered regions. Those regions are typically assigned very low predicted local distance difference test score, but the model can still fabricate ordered geometry that does not reflect the dynamic, ensemble reality of the disordered state. Running multiple random seeds does not produce an approximation of the solution ensemble. Abramson and colleagues give the concrete example that some E3 ubiquitin ligases are predicted in a closed, ligand-bound conformation even when the input data are in their unbound state. Static prediction and dynamic reality remain mismatched. For drug discovery, the picture is similarly mixed. Qin and colleagues cite Zhang and colleagues' finding that AlphaFold-based virtual screening increased hit rates by 30 percent over traditional methods, and Liu and colleagues using AlphaFold models to design SARS-CoV-2 main protease inhibitors. AlphaFold3's advances in protein-small molecule modeling open real possibilities for structure-based drug design against targets without experimental structures. But the stereochemical failure modes and the risk of overconfidence in predicted holo conformations are genuine cautions. The directions the papers point toward are concrete. Diffusion and flow-matching generative models are the immediate frontier — AlphaFold3 already adopts a diffusion head, and the review discusses Schrödinger bridges and stochastic interpolation as related approaches for sampling conformational ensembles rather than single static structures. Multimodal language models that jointly represent sequence, structure, and functional annotations are a longer-range goal. On the compute side, the review cites multi-GPU parallelization strategies that reduced training time by approximately 70 percent and cut error fluctuations by approximately 20 percent — necessary infrastructure for scaling to larger assemblies and real-time prediction pipelines. RNA structure prediction remains an open domain where even AlphaFold3's improvements still leave substantial room for progress. The trajectory from CASP7's 13.5 angstrom average free-modeling errors to AlphaFold2's sub-angstrom backbone accuracy to AlphaFold3's unified atomic diffusion across molecular classes is not simply a quantitative improvement. It is a change in the nature of what is being predicted — from residue-level fold topologies to full-complex atomic coordinates, including ligands, nucleic acids, and post-translational modifications. What the review by Qin and colleagues captures is the architectural logic of that progression: how multiple sequence alignment-based coevolution signal, attention-based pairwise reasoning, geometric modules, and now diffusion-based generation each addressed a specific bottleneck in the earlier approach. The problems that remain — disordered regions, membrane complexes, and conformational ensembles — are not residuals of the same old difficulties. They are new questions that only become visible once the old ones are mostly solved. 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.