# AlphaFold vs. ESMFold: Which AI Tool Is Right for You? In one sentence: AlphaFold predicts protein structures from evolutionary information compiled across thousands of related sequences and delivers the highest accuracy, while ESMFold predicts structures from a single sequence using a protein language model and trades some of that accuracy for raw speed. The rule of thumb is simple. Choose AlphaFold when you need the best possible model of one protein or a protein complex and can afford the compute. Choose ESMFold when you need structures for hundreds or thousands of sequences quickly, when your protein has no detectable homologs, or when you are triaging a large batch of designs. Many productive workflows use both, ESMFold for breadth and AlphaFold for depth. Any serious AlphaFold vs. ESMFold comparison has to start with architecture, because that single difference drives accuracy, speed, input requirements, and failure modes. This guide covers all of it, then walks through worked scenarios that show where each tool belongs. ## The Core Difference Between AlphaFold and ESMFold Both tools take an [amino acid sequence](/blog/guides/amino-acid-sequence) as input and return 3D coordinates plus per-residue confidence scores. They get there by completely different routes. AlphaFold builds a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) (MSA), a column-wise alignment of sequences descended from a common ancestor. Residue positions that mutate in tandem across evolution tend to touch each other in 3D space, and this co-evolutionary signal is the strongest available clue about which residue pairs are neighbors. AlphaFold 2 processes the MSA with a transformer architecture (the Evoformer), optionally consults structural templates from the [Protein Data Bank](/blog/guides/protein-data-bank) (PDB), and converts the result into atomic coordinates with a dedicated structure module. AlphaFold 3 keeps the same core idea, adds a diffusion-based generation step, and extends coverage to complexes containing [nucleic acids](/blog/guides/nucleic-acids), small molecules, ions, and chemical modifications. ESMFold removes the MSA entirely. It runs one sequence through ESM-2, a protein language model with 15 billion parameters trained by masked-token prediction over billions of unlabeled sequences. During that training the model absorbs pairwise co-variation statistics implicitly, so its internal attention patterns already encode which residues belong near each other. A folding head reads those representations and emits coordinates directly, with no database search at inference time. Everything practical follows from this split. AlphaFold needs an MSA, which means sequence database searches that consume time and disk space, and it performs best when homologs exist. ESMFold needs nothing but the sequence, which makes it fast and makes it usable on fast-evolving viral proteins, designed proteins, and metagenomic open reading frames with no recognizable relatives. The price is accuracy, which controlled benchmarks confirm is lower on average, and dramatically lower for protein complexes. The distinction also shapes failure modes. AlphaFold tends to produce a plausible-looking model even when evolutionary evidence is thin, because templates and learned priors fill the gap. ESMFold reflects the statistics of the sequences it saw during pretraining, so highly unusual sequences can yield unusual models. Neither behavior is a bug, but both demand that you inspect confidence output before building anything on the coordinates. ## What AlphaFold Is ### The CASP14 Breakthrough At the 14th Critical Assessment of Structure Prediction (CASP14), the competition organizer John Moult announced that the [protein structure](/blog/guides/protein-structure) prediction problem, which had resisted the field for over five decades, had been effectively solved for single protein chains [1]. The evidence was a plot of median Global Distance Test (GDT) scores across 87 of 92 evaluation domains. GDT measures the fraction of residues placed within agreed distance tolerances of the reference structure, and higher is better. AlphaFold 2, developed by Google DeepMind, reached a median GDT of 92.4, a level previously associated with experimental structures, and stood far above every competing method [1]. ### How AlphaFold 2 Works AlphaFold 2 takes three inputs: the target sequence, an MSA of its homologs, and optionally structural templates from the PDB. The network processes the MSA and a paired residue-residue representation jointly, exchanges information between them across repeated Evoformer blocks, and hands the result to a structure module that outputs all-atom coordinates. The system is trained end to end on experimental structures and refines each prediction through recycling passes. In practice this yields backbone accuracy that rivals experimental structures for well-covered single domains, with reliability dropping in flexible loops, disordered regions, and proteins with shallow MSAs. AlphaFold 2 has known blind spots worth stating plainly. It models one dominant conformation, so proteins that switch folds or sample multiple states receive a single answer. Intrinsically disordered regions come out with low pLDDT, which is honest reporting rather than failure. And predictions of multi-domain orientation can be context dependent, a point that returns in the limitations section. ### AlphaFold 3 and AlphaFold-Multimer Two extensions matter for daily work. AlphaFold-Multimer adapts the AlphaFold 2 machinery to protein complexes by treating multiple chains jointly. AlphaFold 3 replaced the structure module with a diffusion model and broadened the accepted inputs to include DNA, RNA, ligands, ions, and modified residues, moving the tool toward modeling complete functional assemblies rather than isolated chains. ### The AlphaFold Database DeepMind precomputed AlphaFold models for nearly every catalogued protein and released them as the AlphaFold Database, which hosts approximately 214 million structures covering near-complete proteomes across many organisms [1]. Every entry ships with per-residue pLDDT confidence coloring and predicted aligned error (PAE) data, so a model of your protein of interest often already exists and costs nothing but a search. ## What ESMFold Is ### A Language Model That Learned to Fold ESMFold came out of Meta's ESM (Evolutionary Scale Modeling) team. Its backbone is ESM-2, a transformer trained on unlabeled protein sequences with a simple objective, predict masked-out residues from their context. Training at scale across billions of sequences forces the model to learn which positions co-vary, which residue pairs form contacts, and what plausible backbones look like, all without seeing a single structure during pretraining. ESMFold attaches a folding head to ESM-2 and regresses coordinates directly from the language model's internal representations. No MSA search happens at inference. You paste a sequence, the language model runs once, and coordinates come out. On a single modern GPU this takes seconds for a typical-sized protein, and the approach scales to libraries of millions of sequences that would be impractical to align, let alone fold, with MSA-based pipelines [2]. ### Where ESMFold Wins and Where It Loses The single-sequence design buys speed, simplicity, and coverage of proteins with no homologs. It costs accuracy. Field reviews summarize the tradeoff directly, MSA-based systems such as AlphaFold 3 and RoseTTAFold reach near-experimental accuracy by integrating evolutionary information and geometric constraints, while single-sequence systems like ESMFold gain speed and scalability at the price of some predictive power [2]. How big that price is depends heavily on what you are predicting, as the benchmark numbers below show. Because inference is cheap, language-model folding made structure prediction feasible at metagenomic scale, and together with AlphaFold-derived databases it has put hundreds of millions of predicted models into public circulation, including broad coverage of previously uncharacterized plant proteomes [3]. ## AlphaFold vs. ESMFold: Side-by-Side Comparison | Feature | AlphaFold (2 and 3) | ESMFold | |--|--|--| | Developer | Google DeepMind | Meta (ESM team) | | Core idea | Explicit co-evolution from MSAs, plus templates | Implicit evolution learned by a protein language model | | Input needed | Sequence plus MSA (AF2), templates optional | Single sequence only | | MSA dependence | High for AF2, reduced for AF3 | None | | Monomer accuracy, hard benchmark | 88% (AF2 and AF3) | 76% | | Dimer accuracy, same benchmark | 77% (AF2 and AF3) | 41% | | Complexes | Native support (AF2-Multimer, AF3) | Not designed for complexes | | Relative speed | Minutes to hours per target, MSA search included | Seconds to minutes, no MSA search | | Hardware (local run) | Multi-terabyte sequence databases, high-memory GPU | Single consumer GPU, weights only | | Confidence outputs | pLDDT, PAE, pTM (ipTM for complexes) | pLDDT, pTM | | Response to mutations | Structures nearly invariant to large sequence changes | More sensitive to sequence changes | | Typical failure modes | Disordered regions, shallow MSAs, single-conformation bias | Complexes, marginal sequences, some monomers | | Best suited for | Highest-accuracy single models, complexes, docking inputs | High-throughput scanning, orphan sequences, design triage | | Access routes | AlphaFold Database, AlphaFold Server, ColabFold, local install | ESM repositories and notebooks, portals such as Foldify | Accuracy percentages come from the controlled benchmark described in the next section [4]. Three rows decide most real choices. The complex accuracy row decides academic projects, because the gap between the two families is widest exactly when two or more chains touch, and ESMFold was never designed for that task. The speed row decides industrial projects, because folding a thousand sequences with AlphaFold 2 means a thousand MSA searches and their GPU-hours, while ESMFold treats the same job as a routine overnight batch. The quiet row is mutation response, where the two tools behave so differently that it deserves its own section below. Both ecosystems are open. AlphaFold 2 code and weights are downloadable and ColabFold exposes them through notebooks, while ESMFold ships through Meta's ESM repositories. AlphaFold 3 adds usage terms that restrict commercial application, which can decide tool choice for industry users on its own. ## Accuracy: What Controlled Benchmarks Show Benchmarks that recycle old training targets flatter every predictor, so the most informative comparison to date used only experimentally determined structures deposited in the PDB between 2022 and 2024, excluding entries with close homologs among earlier releases [4]. After filtering, the evaluation covered 1,666 monomeric and 994 dimeric proteins that were genuinely challenging, meaning the predictors could not simply recall a near-identical structure seen during training. Three tools went head to head, AlphaFold 2, AlphaFold 3, and ESMFold, and the headline percentages appear in the table above. Two patterns matter for tool selection. First, AlphaFold 2 and AlphaFold 3 posted the same headline success rates on both monomers and dimers in this benchmark, so upgrading versions does not automatically buy monomer accuracy. Second, the monomer gap between the AlphaFold family and ESMFold was real but moderate, while the gap on dimeric proteins was severe. A second pattern explains part of the apparent inaccuracy. Most incorrect predictions involved reference structures solved by nuclear magnetic resonance (NMR), a method that reports ensembles of conformations in solution rather than a single crystal lattice. When the analysis was restricted to X-ray and cryo-electron microscopy (cryo-EM) reference structures, monomer accuracy rose to 95 percent for the AlphaFold tools and 83 percent for ESMFold [4]. Your choice of validation set changes the numbers, so match your expectations to your reference set. The benchmark authors also released ProModEv, an interactive portal for exploring the underlying data, which is a useful sanity check before committing to a tool for a specific protein family [4]. ## Speed, Scale, and Throughput Speed differences are categorical, not incremental. An AlphaFold 2 prediction requires building an MSA against large sequence databases, then running a deep network with recycling for minutes to hours depending on length and hardware. ESMFold skips the search entirely, which turns structure prediction into something you can run per sequence in seconds and batch at scale [2]. Precomputation is the third option. If your protein comes from an organism covered by the AlphaFold Database, the model already exists, and downloading it beats computing it [1]. Real projects show what scale looks like. The Viro3D project predicted structures for more than 85,000 proteins from over 4,400 viruses using both AlphaFold 2 via ColabFold and ESMFold, expanding viral structural coverage roughly thirtyfold beyond experimentally determined structures alone [5]. Viral proteins evolve fast and often lack deep alignments, which makes them a natural showcase for single-sequence folding. When you want to compare tools on the same sequence without building infrastructure, portals that wrap multiple predictors remove the friction. Foldify integrates AlphaFold 3, AlphaFold 2, ColabFold, OmegaFold, and ESMFold behind one web interface, runs them on high-performance computing infrastructure, requires no login, and can overlay multiple predictions of the same protein in a single viewer window [6]. ## Confidence Scores: pLDDT, PAE, and pTM ![Three cards on pLDDT as a flexibility indicator, high confidence on short spurious sequences, and confidence failing under mutation.](/article-images/illustrations/alphafold-vs-esmfold-confidence-cautions.webp) Every prediction ships with uncertainty estimates, and reading them correctly matters as much as the coordinates themselves. pLDDT (predicted local distance difference test) scores each residue from 0 to 100 for local structural confidence. The standard interpretation bands used across AlphaFold Database models are, above 90 very high confidence, 70 to 90 well-modeled backbone, 50 to 70 low confidence with uncertain sidechain placement, and below 50 very low confidence where the region is often genuinely disordered. Color a model by pLDDT in PyMOL or ChimeraX and unstructured tails and shaky loops become obvious immediately. PAE (predicted aligned error) estimates the expected positional error between pairs of residues. Low PAE within a domain means packed, confident geometry. Low PAE between residues in different chains or domains means the model is confident about their relative placement, which is the number to check when judging an interface. High PAE between domains usually means the model does not know how they orient, not that either domain is internally wrong. pTM (predicted template modeling score) estimates global fold similarity to the true structure on a 0 to 1 scale, and its interface variant ipTM serves the same role for complexes. Three cautions come straight from recent evaluations. First, pLDDT behaves partly as a flexibility indicator. A large-scale comparison against molecular dynamics trajectories from the ATLAS dataset, NMR ensembles, and experimental B-factors found that AlphaFold 2 pLDDT correlates reasonably with independent flexibility measures and outperforms B-factors for that purpose, but fails to capture flexibility that appears only in the presence of binding partners [7]. AlphaFold 3 improves slightly on this, and [molecular dynamics simulations](/knowledge/bioinformatics/molecular-dynamics-simulations-of-proteins-and-force-fields) remain the better tool for flexibility questions [7]. Second, high pLDDT does not certify biological reality for short sequences. All three major predictors assign unexpectedly high pLDDT scores to spurious protein sequences, the artifacts of gene-prediction errors that should not fold at all, and discrimination between spurious and real proteins only becomes dependable beyond about 100 amino acids [8]. Treat confidently structured models of short open reading frames with suspicion. Third, confidence metrics can be confidently wrong under sequence perturbation, which the next section quantifies. ## Mutation Sensitivity: The Most Overlooked Difference An adversarial evaluation across 200 proteins exposed a surprising property of AlphaFold 3. Predicted structures stayed essentially invariant when up to 40 percent of residues were mutated, including deliberately destabilizing substitutions, and when up to 10 percent of residues were deleted [9]. The invariance held even for experimentally validated fold-switching proteins, small monomeric proteins known to change fold in response to exactly such mutations, which is the category where AlphaFold is expected to perform best [9]. Confidence metrics did not rescue the situation. They selected the most accurate available structure at most 35 percent of the time, and instead tracked the structural quality of the best template in the training set, suggesting the model leans on memorized patterns more than generalizable biophysics [9]. ESMFold showed greater mutational sensitivity, meaning its output actually responds to sequence changes, though imperfectly [9]. The study attributes the difference to training objectives rather than overall model quality. The practical rule is blunt. Neither tool predicts the structural or functional consequences of mutations, so do not model a point mutant, compare it against the wild-type prediction, and report a conformational change. For mutation effects, use dedicated predictors, free-energy methods, or experiments. ## When to Use AlphaFold (and When ESMFold Makes More Sense) ![Decision path asking about complexes, sequence count and homolog availability, ending in AlphaFold or ESMFold.](/article-images/illustrations/alphafold-vs-esmfold-tool-choice-triage.webp) The choice of AlphaFold or ESMFold reduces to three questions: how much accuracy do you need, how many sequences do you have, and do homologs exist? Choose AlphaFold when: - You need a publication-grade model of one protein or a complex. - The structure will feed docking, mutagenesis planning, or engineering decisions where coordinate errors propagate. - Your protein has a deep MSA, which is true for most proteins from well-studied organisms. - You are modeling an interface between chains, or a complex involving nucleic acids or ligands with AlphaFold 3. Choose ESMFold when: - You need structures for tens to millions of sequences on a deadline. - Your sequence has no detectable homologs, so an MSA would be shallow or empty. - You are triaging de novo designs before synthesis. - You want a fast, independent second opinion on an AlphaFold prediction. When in doubt, run both. Consensus across independent predictors is itself informative, and several published workflows treat disagreement between tools as a flag for manual inspection rather than as a failure. A common production pattern is ESMFold for the wide sweep, AlphaFold for the finalists, and experimental structures for anything that reaches a paper. ## Six Worked Scenarios **Scenario 1: A human kinase model for inhibitor docking.** Kinases are heavily studied, so the AlphaFold Database almost certainly contains your target already. Search by UniProt identifier, download the model, and check the pLDDT of the active site and activation loop before docking anything. Computing the structure yourself would add days and gain little. This is the classic case for when to use AlphaFold, accuracy matters and evolutionary information is abundant. **Scenario 2: A metagenomic open reading frame with no database hits.** An MSA would be nearly empty, which cripples MSA-based predictors and wastes compute on the search. Run ESMFold first. Then apply the spurious-protein caution from the confidence section, because short predicted sequences can look confidently structured regardless of biological reality [8]. **Scenario 3: Triage for a weekly batch of de novo designs.** A lab synthesizes dozens of candidate binders each week and needs a cheap filter before ordering genes. In a benchmark of 614 experimentally characterized de novo monomers drawn from 11 design studies published between 2012 and 2021, ESMFold's average pLDDT was the best single discriminator between designs that succeeded experimentally and designs that failed, outperforming AlphaFold-based and ProteinMPNN-based metrics as individual predictors [10]. The same study warns that many failed designs carry better confidence scores than successful ones and that metrics are topology dependent, so treat the filter as probabilistic, not deterministic [10]. **Scenario 4: A heterodimer structure for a mechanism paper.** Complex accuracy decides this choice, and the benchmark gap is severe (see the table above). Use AlphaFold-Multimer or AlphaFold 3, report ipTM and inter-chain PAE alongside the model, and validate the interface against any available crosslinking, mutagenesis, or co-immunoprecipitation data. **Scenario 5: A peptide binder screen.** For protein-peptide docking, an evaluation found that ESMFold-based strategies produce acceptable models at rates comparable to traditional docking approaches, generally below AlphaFold-Multimer and AlphaFold 3 but ahead of them in some cases, at a fraction of the compute cost [11]. That combination of quality and efficiency makes ESMFold a sensible first-pass engine in a high-throughput peptide design pipeline, with AlphaFold-based refinement reserved for the shortlist [11]. **Scenario 6: An outer membrane beta-barrel from a bacterial pathogen.** A study of Brucella melitensis outer membrane proteins modeled its targets with five tools, AlphaFold 3, ESMFold, SWISS-MODEL, RoseTTAFold, and TrRosetta, then confirmed membrane insertion with dedicated hydrophobicity and insertion predictors [12]. The multi-tool pattern is deliberate. Membrane proteins stress every predictor, and agreement across methods plus independent insertion checks is the defensible way to publish a model. ## Access Routes: Databases, Servers, and Local Installs Which route you pick depends on batch size and data sensitivity. Servers suit one-off questions. Local installs suit proprietary sequences that cannot leave your infrastructure, and recurring batches where per-job server limits hurt. ### AlphaFold Database Start here before computing anything. Search by protein name or UniProt accession, download the mmCIF or PDB file, and load the accompanying PAE data. Coverage spans near-complete proteomes across many organisms [1], so most well-annotated proteins are already modeled. ### AlphaFold Server Google DeepMind operates AlphaFold Server, a free web interface to AlphaFold 3 that accepts proteins, nucleic acids, ions, and ligands and returns modeled complexes. It exists precisely for researchers without GPU clusters, with daily usage limits and terms that restrict outputs to non-commercial use. For a quick complex hypothesis, it is the lowest-friction route to a state-of-the-art prediction. ### ColabFold and Multi-Model Portals ColabFold wraps AlphaFold 2 and AlphaFold-Multimer in browser notebooks with fast MSA search through MMseqs2, which cuts typical runtime dramatically and requires no local installation. Most AlphaFold tutorial material online walks through ColabFold for exactly this reason, it is the shortest path from sequence to structure. For comparing predictors side by side, Foldify runs AlphaFold 3, AlphaFold 2, ColabFold, OmegaFold, and ESMFold from one web interface with no login, and its comparison mode overlays multiple models of the same sequence in a single viewer [6]. ### Running Locally Local AlphaFold 2 demands real infrastructure, multi-terabyte genetic sequence databases, a CUDA GPU with substantial memory, and patient setup. AlphaFold 3 weights carry additional licensing terms from DeepMind. ESMFold is far lighter, the model weights download once, no MSA databases are needed, and a single consumer GPU handles typical proteins in seconds to minutes. If your workload is many sequences, the lighter footprint compounds quickly. ### Your First Run, Step by Step A minimal first pass that works for either tool: 1. Paste your sequence in single-letter amino acid code. 2. Submit to a server, ColabFold for AlphaFold or the ESM notebook or a portal like Foldify for ESMFold. 3. Download the structure file and open it in PyMOL, ChimeraX, or the Mol* viewer. 4. Color by pLDDT and inspect low-confidence regions before interpreting anything. 5. Check PAE (AlphaFold outputs) for domain and chain arrangement. 6. Compare against the AlphaFold Database entry if one exists, and against any experimental structure in the PDB. 7. Record which tool and version produced the model, because reproducibility depends on it. ## Downstream Applications: Design, Docking, and Annotation Predicted structures increasingly serve as inputs to second-stage models rather than as endpoints. Antimicrobial resistance pipelines chain AlphaFold or ESMFold predictions into [molecular docking](/knowledge/bioinformatics/docking-algorithms-autodock-glide-and-beyond) engines to screen inhibitors of resistance enzymes [13]. Interaction predictors such as SpatialPPIv2 were built to work robustly with structures from AlphaFold 3, AlphaFold 2, or ESMFold, so the choice of upstream predictor does not lock you out of downstream tooling [14]. Binding-site predictors have likewise been trained and compared using structural features extracted from both AlphaFold and ESMFold models [15]. Large-scale annotation is another growth area. Mining the PDB and the AlphaFold Database by fold similarity connected a single ancient scaffold, the beta-trefoil fold, to thirty-two distinct functional classes, showing how predicted structures extend functional annotation beyond what sequence similarity reveals [16]. Reviews of plant stress biology make the same point at organism scale, where predicted structures from AlphaFold, RoseTTAFold, and ESMFold now support docking and simulation studies of stress-response proteins that never had experimental structures [3]. ## Common Mistakes and Limitations 1. **Treating pLDDT as a flexibility or dynamics measurement.** It tracks confidence, correlates loosely with flexibility, and misses partner-induced changes [7]. Use molecular dynamics for flexibility questions. 2. **Using either tool to predict mutation effects.** AlphaFold's outputs barely respond to massive sequence changes, and its confidence metrics are unreliable in that regime [9]. ESMFold responds more but still cannot be trusted for variant interpretation. 3. **Trusting confident models of short sequences.** High pLDDT on a 40-residue open reading frame tells you very little, because all major predictors assign high confidence to spurious short sequences [8]. 4. **Accepting a single-tool membrane protein model.** Different deep-learning methods returned three distinct conformations for the same Arabidopsis membrane protein, and only molecular dynamics in a lipid bilayer sorted out plausible states [17]. Build consensus and simulate. 5. **Assuming the AlphaFold Database model matches your construct.** Database models represent canonical UniProt sequences, not your tag, isoform, species ortholog, or point mutation. Check the sequence mapping before docking. 6. **Predicting artificial fusion proteins casually.** Domain orientation in chimeras depends on linker length and structural context, and even AlphaFold 2 predictions require careful validation for engineered multidomain proteins [18]. 7. **Comparing your own results against published accuracy numbers without matching benchmark conditions.** Published figures come from specific reference sets and difficulty filters, so a lower success rate on your own NMR-derived targets is not evidence that your pipeline is broken. 8. **Docking into low-confidence pockets.** A binding site inside a 60-pLDDT region is a hypothesis, not a pocket. Prioritize ligands and sites supported by conservation, experimental data, or high local confidence. 9. **Forgetting that a prediction is a hypothesis.** The strongest workflows pair predictions with crosslinking data, cryo-EM maps, mutagenesis, or activity assays rather than treating coordinates as ground truth. The bottom line for the AlphaFold vs. ESMFold decision is that the tools are complements more than competitors. AlphaFold sets the accuracy ceiling, ESMFold sets the throughput ceiling, and mature pipelines use both with the confidence scores read carefully at every step. ## Frequently Asked Questions ### Is ESMFold as accurate as AlphaFold? No. Controlled head-to-head benchmarks put AlphaFold 2 and AlphaFold 3 ahead on monomeric proteins, with a much wider gap on protein complexes, where ESMFold is not designed to compete. The gap narrows for single chains with few or no homologs, which is where ESMFold's single-sequence design helps most. ### When should I use AlphaFold instead of ESMFold? Use AlphaFold when accuracy justifies the compute, when you are modeling a complex or an interface, when a deep multiple sequence alignment exists, or when the model will drive docking or engineering decisions. Check the AlphaFold Database first, since a model of your protein may already exist. ### Can ESMFold predict protein complexes? Not reliably. ESMFold was built for single chains. For dimers, larger assemblies, or complexes with nucleic acids and ligands, use AlphaFold-Multimer or AlphaFold 3. ### Do I need a GPU to run these tools? For local runs, effectively yes. ESMFold runs comfortably on a single consumer GPU, while local AlphaFold 2 wants more GPU memory plus multi-terabyte sequence databases. Without suitable hardware, use the AlphaFold Server, ColabFold, or a multi-model portal such as Foldify. ### Are predicted structures good enough for docking? As starting points, usually yes, especially in high-confidence structural cores. Treat low-confidence loops and predicted sidechain orientations with suspicion, relax and rescore docked poses, and prefer an experimental structure for the binding site when one exists. ### What does pLDDT mean and how should I read it? pLDDT is a per-residue confidence score from 0 to 100. High scores mark well-modeled residues, mid-range scores mark uncertain regions, and the lowest scores usually flag disorder. Read it as a guide to which regions deserve trust, not as a verdict on the whole model. ## Sources 1. [From CASP13 to the Nobel Prize: DeepMind's AlphaFold Journey in Revolutionizing Protein Structure Prediction and Beyond.](https://pubmed.ncbi.nlm.nih.gov/41047692/) 2. [Protein structure prediction powered by artificial intelligence: from biochemical foundations to practical applications.](https://pubmed.ncbi.nlm.nih.gov/41878008/) 3. [Review of protein structure-based analyses that illuminate plant stress mechanisms.](https://pubmed.ncbi.nlm.nih.gov/40727428/) 4. [Comparative evaluation of the prediction accuracy of AlphaFold and ESMFold for monomeric and dimeric proteins.](https://pubmed.ncbi.nlm.nih.gov/41551928/) 5. [Viro3D: a comprehensive database of virus protein structure predictions.](https://pubmed.ncbi.nlm.nih.gov/40958060/) 6. [Foldify: Web Application for Protein Structure Prediction.](https://pubmed.ncbi.nlm.nih.gov/42359896/) 7. [Flexibility or uncertainty? A critical assessment of AlphaFold 2 pLDDT.](https://pubmed.ncbi.nlm.nih.gov/41005302/) 8. [Folding the unfoldable 2: using AlphaFold and ESMFold to explore spurious proteins.](https://pubmed.ncbi.nlm.nih.gov/42370329/) 9. [Adversarial Sequence Mutations in AlphaFold and ESMFold Reveal Nonphysical Structural Invariance, Confidence Failures, and Concerns for Protein Design.](https://pubmed.ncbi.nlm.nih.gov/42376644/) 10. [Evaluating zero-shot prediction of monomeric protein design success by AlphaFold, ESMFold, and ProteinMPNN.](https://pubmed.ncbi.nlm.nih.gov/41556605/) 11. [Protein-Peptide Docking with ESMFold Language Model.](https://pubmed.ncbi.nlm.nih.gov/40053869/) 12. [Identification and Characterization of Outer Membrane Proteins and Membrane Spanning Protein Complexes in Brucella melitensis.](https://pubmed.ncbi.nlm.nih.gov/41582728/) 13. [Artificial intelligence in protein-based detection and inhibition of AMR pathways.](https://pubmed.ncbi.nlm.nih.gov/41286386/) 14. [SpatialPPIv2: Enhancing protein-protein interaction prediction through graph neural networks with protein language models.](https://pubmed.ncbi.nlm.nih.gov/41255828/) 15. [Predicting protein-carbohydrate binding sites: a deep learning approach integrating protein language model embeddings and structural features.](https://pubmed.ncbi.nlm.nih.gov/41608988/) 16. [Tracing the function expansion for a primordial protein fold in the era of fold-based function prediction: β-trefoil.](https://pubmed.ncbi.nlm.nih.gov/40608768/) 17. [Consensus structure prediction of A. thaliana's MCTP4 structure using prediction tools and coarse grained simulations of transmembrane domain dynamics.](https://pubmed.ncbi.nlm.nih.gov/40663537/) 18. [Fusion of two unrelated protein domains in a chimera protein and its 3D prediction: Justification of the x-ray reference structures as a prediction benchmark.](https://pubmed.ncbi.nlm.nih.gov/35833233/)