Protein structure prediction via deep learningan in-depth review

Two hundred and fourteen thousand. That is how many experimentally determined protein structures were in the Protein Data Bank in 2023. Meanwhile, the TrEMBL sequence repository held over two hundred million entries. That ratio — roughly a thousand sequences for every known structure — is the central problem that a decade of deep learning has been trying to solve. The gap is not from lack of effort. X-ray crystallography, nuclear magnetic resonance spectroscopy, and cryo-electron microscopy are the gold-standard tools for structure determination, and each carries real, practical bottlenecks. Jumper and colleagues, writing about AlphaFold, put it plainly: determining a single protein structure requires months to years of painstaking effort. The review by Meng and colleagues adds the specifics — long cycle times, high costs, and demanding sample preparation. Cryo-electron microscopy has transformed some areas of structural biology, but building atomic models from electron density maps is still challenging without good homologs, and local resolution can be ambiguous even in the best maps. These are not abstract limitations. They translate to calendar time, expensive instruments, and deep technical expertise per target. There is also a fundamental theoretical obstacle. Levinthal's paradox tells us that if a protein sampled all possible conformations randomly, it would take astronomically long to find its native fold — yet real proteins fold reliably in microseconds to seconds. The paradox clarifies two things simultaneously: brute-force search is impossible, and folding must follow constrained pathways through a funneled energy landscape. That same combinatorial complexity is what makes computational prediction hard. From a linear sequence, secondary motifs like alpha helices and beta sheets form locally, but tertiary structure arises from long-range side-chain contacts and three-dimensional packing. Quaternary structure adds another layer entirely. Computational approaches face two linked tasks — learn from sparse experimental ground truth, and do it without anything resembling brute-force conformational search. Which brings us to the data foundation. The Protein Data Bank is indispensable. Meng and colleagues report it held two hundred fourteen thousand one hundred eight structures as of 2023, and deep learning methods use it both as raw training examples and as a source of template structures. Jumper and colleagues trained AlphaFold2 on a Protein Data Bank snapshot from August 2019 and used separate Protein Data Bank releases for template searches, clustered at forty percent sequence identity. Sequence databases are the second pillar. UniProt provides the definitive resource, organized into UniProt Knowledge Base for annotated entries, UniProt Archive as a sequence archive, and UniRef for clustered representatives. AlphaFold2 specifically used UniRef90 as one of its multiple sequence alignment inputs during both training and inference. A third resource class comprises large clustered databases — BFD, Uniclust30, and MGnify clusters — that supply the evolutionary depth modern predictors require. More recently, the AlphaFold Protein Structure Database has made over two hundred million predicted structures openly available, creating a computationally generated catalog that now supplements experimental repositories for downstream applications. The critical caveat is bias in the Protein Data Bank. The review is explicit: the Protein Data Bank is dominated by static, thermodynamically stable conformations and underrepresents dynamic or multistate proteins. That restricted diversity can produce overfitting and limits generalization to transient or alternative conformations. The failure mode is concrete — Abramson and colleagues report that AlphaFold3 predicted a domain-swapped structure for human XCL1 rather than the experimentally observed dimeric conformation, a direct consequence of the model learning from data that underrepresents fold-switching behavior. CASP, the community's assessment harness, benchmarks progress against this backdrop: in CASP14, eighty-eight of the submitted prediction methods explicitly used deep learning; by CASP15, that fraction had grown further. The headline number from CASP14 is striking — AlphaFold2 could accurately predict the three-dimensional structures of ninety-eight point five percent of human proteins. Now let's get into how it actually works — the architecture landscape, because understanding what AlphaFold2 does requires knowing what came before it. Deep neural networks, the simplest class, operate on one-hot or embedded amino-acid inputs and have been applied to primary, secondary, and tertiary structure prediction. The review cites MULTICOM as a deep neural network-based three-dimensional modeling system that combines contact-distance prediction, template-free modeling, and model ranking. Convolutional neural networks process sequence profiles and two-dimensional contact or distance maps, extracting local spatial patterns. A large convolutional residual network from Xu and colleagues predicted correctly folded structures for twenty-six of the thirty-two free modeling targets at CASP13, with long-range contact accuracy exceeding eighty percent — a landmark result for contact prediction. Recurrent networks, which include recurrent neural networks, long short-term memory networks, and gated recurrent units, address sequential dependencies and dominated secondary structure prediction for years. Bidirectional long short-term memory models and gated recurrent unit variants with position-specific scoring matrices handled long sequences and captured both local and some long-range patterns, though they remain limited by their sequential inductive bias. Graph neural networks reframed the problem. Instead of treating the sequence as a one-dimensional string or the contacts as a two-dimensional image, graph neural networks treat residues as nodes and interactions as edges — a natural fit for the relational structure of protein geometry. The review highlights geometric graph neural network work that models distance and dihedral geometric representations via geometric graph convolutions. The Recursive Geometric Network from Chowdhury and colleagues goes further: an end-to-end differentiable model that predicts structures from single sequences without any multiple sequence alignments, reported as superior to AlphaFold2 in certain long-range structural settings. Transformers changed the game most dramatically, and the key insight is how they handle multiple sequence alignments. Earlier co-evolution-based methods derived covariance statistics from alignments and fed those processed features into predictors. AlphaFold2 departed from that pipeline entirely. Jumper and colleagues explicitly note it starts from raw multiple sequence alignments rather than processed features like inverse covariance matrices, and replaces two-dimensional convolutions with attention to capture long-range residue interactions directly. The network processes a two-track representation that iteratively transforms one-dimensional sequence information and two-dimensional pairwise distance information through the Evoformer — a stack of attention modules that communicate between sequence and pair representations. Multiple sequence alignment quality matters enormously for this pipeline. The review highlights DeepMSA2, used in a system called DMFold, which employs iterative search and model-based preselection to produce alignments with better coverage and homologous diversity, improving downstream modeling. A separate but complementary line uses protein language models trained on vast sequence databases, offering an alternative when deep alignments are unavailable. ESMFold from Lin and colleagues is an end-to-end atomic-level predictor operating on individual sequences with up to fifteen billion parameters — no alignment required. These single-sequence models are specifically highlighted as critical for proteins that have no close homologs, where alignment-dependent pipelines become unreliable. Diffusion models represent the newest architectural addition. Standard regression-based predictors produce a single deterministic coordinate set; diffusion models instead learn to sample from a distribution over conformations, modeling multimodal data and generating structurally diverse outputs. The review cites EigenFold from Jing and colleagues, which models the molecule as a system of harmonic oscillators through a harmonic diffusion process to explore protein structural ensembles. These generative approaches are positioned as a route to conformational diversity rather than single-point predictions. AlphaFold3 merges these threads. Abramson and colleagues report that AlphaFold3 replaces the Evoformer with a Pairformer that reduces reliance on multiple sequence alignments while adding a geometric diffusion module that enables probabilistic sampling of complex conformations. AlphaFold3 also extends beyond single protein chains to handle protein-ligand complexes, nucleic acids, and other molecular types — a direct response to the limitation that earlier predictors handled only proteins. The diffusion-based structure module enables better handling of flexible interfaces and allosteric effects compared to AlphaFold2's deterministic coordinate refinement, which used an SE(3)-equivariant transformer to refine atomic coordinates while respecting three-dimensional rigid-body transformations. RoseTTAFold, from Baek and colleagues, achieves comparable performance through a distinct architectural choice: a two-track network that also processes one-dimensional and two-dimensional information but with different attention routing. The review notes its successful use for fitting predicted models into cryo-electron microscopy density maps when local resolution was insufficient for de novo model building — a clear demonstration of how these predictors now support rather than replace experimental methods. The disease-protein applications show what accurate structure prediction actually unlocks. The COVID-19 pandemic provided a high-stakes test: AlphaFold and RoseTTAFold were deployed rapidly to predict structures of SARS-CoV-2 proteins, enabling structure-guided drug discovery under urgent timelines that experimental methods alone could not have met. Broader deployment for emerging disease proteins has followed, where computational structure predictions have guided rapid drug repurposing by identifying binding sites and conformational features before experimental structures were available. Now, how do we know when a prediction is right? The answer depends critically on which metric you use, and each captures something different. RMSD, which stands for root mean square deviation, is the most intuitive: the square root of the mean squared distance between matching atom pairs after optimal superposition, reported on backbone atoms. The problem is that averaging over all atoms makes RMSD sensitive to a few large errors and those errors can make a largely correct model look bad — especially for large proteins where a few misplaced loops dominate the average. TM-score, introduced by Zhang and Skolnick, corrects for this by normalizing distance contributions by a length-dependent factor, making the score roughly independent of protein size. TM-score rewards global topology rather than penalizing for local deviations, which is why it better captures whether a model has the right overall fold. GDT-TS or Global Distance Test total score complements both by reporting the fraction of carbon-alpha atoms within four distance cutoffs: one, two, three, and four angstroms after superposition, then averaging. Because it spans multiple windows, GDT-TS can recognize models with many locally correct fragments even when the overall superposition is imperfect. Local accuracy gets its own set of metrics. LDDT or local distance difference test is superposition-free and computed per residue by comparing pairwise distances within a local shell. pLDDT is the model's predicted LDDT — a per-residue confidence score. AlphaFold3's predicted confidences are reported to be well calibrated against actual LDDT values. For interfaces and complexes, interface LDDT and DockQ are the relevant measures. DockQ is a composite score for protein-protein interfaces; a common success threshold used in benchmarks is a DockQ greater than zero point twenty-three. AlphaFold3's analysis shows that DockQ and interface LDDT are highly correlated, with the DockQ categories "correct" and "very high quality" corresponding to interface LDDT values of approximately twenty-three point six and seventy-seven point six, respectively. And then there is what none of these metrics capture: conformational ensembles, dynamics, or functional states. AlphaFold3's paper is direct about this — the model predicts single static conformations, and multiple random seeds do not recover a true solution ensemble. The review by Meng and colleagues makes the same point more broadly: static structural coverage is only a first step toward understanding protein function, and predictions of a single lowest-energy conformation are not a substitute for knowledge of the conformational landscape. That leads directly to the genuine open problems. Abramson and colleagues report that four point four percent of AlphaFold3's benchmark predictions show chirality violations — the model occasionally places atoms in stereochemically incorrect configurations. Steric clashes remain, concentrated in very large protein-nucleic complexes with more than one hundred nucleotides and more than two thousand residues total. The diffusion module introduces hallucinated order in intrinsically disordered regions, typically flagged by very low pLDDT but still present as an artifact. Jumper and colleagues quantify a different failure mode: accuracy falls substantially when median alignment depth drops below about thirty sequences, with diminishing returns past roughly a hundred sequences. AlphaFold2 also performs notably worse for chains whose shape is defined mainly by interactions with other chains — heterotypic contacts — rather than self-interactions. The review by Meng and colleagues adds structural gaps. Multi-domain protein prediction remains difficult: while single-domain accuracy is largely solved, correctly assigning domain boundaries and modeling inter-domain arrangements is still a bottleneck. FUpred, from Zheng and colleagues, attempts to address domain boundary detection through contact-based methods, but this remains an active research problem. And interpretability is essentially unsolved — these are deep black-box models, and understanding what structural features drive their predictions remains an open challenge. The picture that emerges is this: deep learning, particularly transformer-based architectures using raw multiple sequence alignments as input, has genuinely transformed protein structure prediction. AlphaFold2's ninety-eight point five percent coverage of human protein structures in CASP14 is a real achievement. AlphaFold3's extension to complexes and its geometric diffusion module represent the field's next step. But the remaining failures are not random noise — chirality violations, steric clashes in large assemblies, shallow-alignment degradation, heterotypic-contact difficulty, and the complete absence of dynamics are systematic. The gap between predicting a static fold and understanding how a protein actually functions in a cell remains substantial, and closing it will require methods that go beyond single-conformation coordinate prediction entirely. 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.