# How to Handle Protein Isoforms in AlphaFold An AlphaFold model describes one exact [amino acid sequence](/blog/guides/amino-acid-sequence), while an isoform is one of several sequences that a single gene can produce, so handling isoforms in AlphaFold comes down to picking the right sequence and finding out whether that sequence already has a model. The rule of thumb is to check the AlphaFold Database first, because the 2025 update added isoform-level coverage next to the long-standing canonical models [1]. Download what exists, predict the exact isoform sequence yourself when it does not, and judge every model by its per-residue confidence at splice junctions and termini. This matters because isoforms are not cosmetic. [Alternative splicing](/blog/guides/alternative-splicing) can delete an entire interaction domain, lengthen a coiled coil by a few residues, or swap a C-terminal tail, and each of those edits can change what the protein binds or where it localizes. Isoform-specific rewiring of protein interaction networks is now being modeled at scale with frameworks that combine AlphaFold structures with protein language model embeddings [2], and isoform selectivity is a working goal in drug discovery, where an inhibitor designed against one [histone deacetylase](/knowledge/molecular-biology/histone-deacetylase) isoform can spare its close paralogs [3]. Every one of those analyses starts with the same question, which sequence am I actually folding? ## Isoforms, Canonical Sequences, and the Identifier Problem A gene can yield several mature messenger RNAs through alternative splicing, alternative promoters, and alternative polyadenylation, and the proteins translated from those transcripts are isoforms. Common splicing patterns include exon skipping, mutually exclusive exons, alternative 5 prime or 3 prime splice sites, alternative start sites that extend or shorten the N-terminus, and alternative last exons that swap the C-terminus. Related sequence variation also comes from proteolytic processing, such as [signal peptide](/knowledge/molecular-biology/signal-peptide) cleavage, which produces a mature chain shorter than the gene-encoded precursor. Databases store one representative sequence per protein entry, called the canonical sequence, usually the longest or best-supported isoform. UniProt marks additional isoforms with numeric suffixes, so P12345 is the canonical entry and P12345-2 is a defined isoform with its own sequence page. RefSeq protein accessions at NCBI are isoform-specific, meaning two NP_ accessions from one gene usually correspond to two different isoforms, which makes them convenient handles when you want to talk about isoform panels rather than genes. | Identifier | Example format | What it pins down | Where structures usually come from | |--|--|--|--| | UniProt canonical | P12345 | The representative sequence | AlphaFold Database entry, available for years | | UniProt isoform | P12345-2 | One defined splice variant | AlphaFold Database since the 2025 isoform update [1] | | RefSeq protein | NP_ or XP_ accession | One isoform of one gene product | Map to AlphaFold accessions with batch tools [4] | | Custom or engineered sequence | A FASTA file you write | Whatever you specify | Your own prediction run | One caution before any of this. Not every listed isoform is a protein the cell reliably makes. Some transcripts are flagged as targets of nonsense-mediated decay, a surveillance pathway that degrades transcripts carrying premature stop codons, and some isoforms rest entirely on computational annotation. AlphaFold will confidently fold a sequence that biology never expresses, so read the evidence lines on the UniProt or RefSeq record before you spend compute on a variant. ## Your AlphaFold Isoform Prediction Options at a Glance There are four practical routes to a structure for an isoform, and most real projects combine two or three of them. | Option | What you get | Choose it when | Skill and compute | Main pitfall | |--|--|--|--|--| | 1. Download the AlphaFold Database model | A precomputed single-chain model with per-residue confidence and predicted aligned error, now covering isoforms as well as canonical sequences [1] | The database has your isoform | Minimal, a browser or a few API calls | Grabbing the canonical model and calling it your isoform | | 2. Batch-map accessions and bulk download | Many models plus an ID mapping file that traces inputs back to standard accessions [4] | You work with tens to thousands of RefSeq, UniProt, or locus tag identifiers | Low, web form or API | Unmapped accessions slipping through if you skip the report | | 3. Predict the exact isoform sequence yourself | A model for any sequence you control, including complexes with newer tools | The isoform is absent, novel, engineered, or part of a complex | Medium to high, GPU access helps | Low confidence at novel splice junctions and wobbling termini | | 4. Model trimmed constructs or isoform complexes | Domain-only models for fair comparisons, or co-folded complexes | You need interface-level answers or isoform-shared cores | High, interpretation heavy | Construct choices shifting confidence metrics [5] | The sections below explain each option on its own terms, compare them on the criteria that decide real projects, walk through worked scenarios, and finish with the mix-ups that cost people the most time. ## Option 1: Download the Isoform Model from the AlphaFold Database The AlphaFold [Protein Structure](/blog/guides/protein-structure) Database, maintained by EMBL-EBI together with Google DeepMind, holds hundreds of millions of predicted structures and has become the default first stop for protein structure questions [1]. For most of its history it stored one model per UniProt entry, built on the canonical sequence, which is precisely why isoform work used to require custom predictions. The 2025 update changed that picture in three ways. Structural coverage now includes isoforms and not only canonical sequences. The underlying multiple sequence alignments are published together with the structures. And the redesigned entry page integrates annotations with an interactive 3D viewer, adds dedicated domains and summary tabs, and serves everything through the website, FTP, Google Cloud, and updated APIs [1]. A practical checking routine looks like this. 1. Search the database by gene name or UniProt accession. 2. On the entry page, look specifically for isoform entries rather than assuming the first model is the one you need. The new domains and summary tabs make it easier to confirm which domains a given isoform model retains [1]. 3. Download coordinates in mmCIF or PDB format, and take the predicted aligned error (PAE) file as well, because it records how much you can trust the relative positions of residues, expressed in angstroms. 4. Record the database version tied to the UniProt release so the download is reproducible a year from now. Reading the model takes two numbers. The first is pLDDT, a per-residue confidence score from 0 to 100. Under the standard AlphaFold Database coloring, scores above 90 indicate very high confidence, 70 to 90 indicates a confident backbone, 50 to 70 indicates low confidence, and below 50 is treated as likely disordered. The second is PAE, which is about relationships between residues rather than individual residues. Two domains with low PAE between them have a trusted relative placement, while high PAE between domains means the model is unsure how they connect, which is exactly the situation many alternative exons create. Residue numbering deserves its own warning. Model numbering follows the sequence that was actually modeled. If isoform 1 has 720 residues and isoform 2 skips an exon of 31 residues, then every residue after that exon carries a different number in the two models. Align the two sequences first, build an explicit residue map, and only then quote positions in a script or a manuscript. Finally, remember what the model does not contain. AlphaFold Database models are single-chain predictions of an unmodified apo sequence, with no ligands, cofactors, post-translational modifications, or alternative conformational states. Many docking and mechanistic projects therefore start from the database model and then optimize it, the pattern followed in the HDAC11 inhibitor work where the AlphaFold model was refined before docking and the resulting design was validated experimentally [3]. ## Option 2: Batch Mapping and Bulk Downloads for Isoform Panels Isoform studies often begin with a list of accessions pulled from a database query, and those accessions are frequently RefSeq identifiers. The AlphaFoldDB Structure Extractor, a web server and API, was built for exactly this step. It accepts NCBI Taxonomy IDs, RefSeq accessions, old and new locus tags, and UniProt or AlphaFold accessions, then returns structures in PDB, mmCIF, bCIF, or PAE JSON formats together with an ID mapping file that traces every input back to a standard accession and a separate report of unmapped IDs [4]. A mapping-only mode covers cases where you need the correspondence but not the coordinate files, and an API supports integration into automated pipelines. The authors tested the service with individual inputs and with batches of up to 5,000 accessions across the supported identifier types [4]. A compact walkthrough for an isoform panel: 1. Collect the NP_ accessions for the isoforms you care about. RefSeq accessions are the right currency here because they resolve to a single isoform. 2. Submit the list through the web interface or call the API from your pipeline. 3. Read the mapping file carefully. It is your provenance record connecting each downloaded model to the accession you started from, and it is the artifact that makes the analysis auditable. 4. Check the unmapped list. Those isoforms need the self-prediction route described next. 5. Download mmCIF plus PAE JSON so the confidence data travels with the coordinates. Before batch-processing a pile of near-identical isoforms, consider clustering the sequences first. Sequence clustering tools collapse redundant or near-identical sequences into representative groups, which keeps you from downloading, predicting, and analyzing the same fold five times [6]. Both sequence-based and graph-based clustering approaches are well described with runnable commands, so this step rarely takes more than an afternoon [6]. ## Option 3: Running Your Own AlphaFold Isoform Prediction Predict the isoform yourself when it is missing from the database, when you study a newly described splice variant, when you engineer constructs, or when you need a complex or ligand-bound state that precomputed models cannot provide. People searching for an alphafold server usually land on one of two very different resources, the AlphaFold Database web interface for downloads and the AlphaFold 3 server for custom jobs, and the two solve different problems. Sequence hygiene comes first, because AlphaFold is a sequence-to-structure function and it has no knowledge of isoform labels. 1. Copy the exact FASTA from the UniProt isoform page or the RefSeq protein record. One wrong residue silently invalidates every downstream claim. 2. Confirm what the splicing event actually is, whether exon skipping, an alternative start site, or an alternative last exon, because the event type predicts where confidence problems will appear. 3. Sanity-check biological reality, including evidence level and any nonsense-mediated decay annotation. 4. Decide the construct. Full isoform, mature chain after signal peptide cleavage, or a single domain, each answers a different question. Tool choices follow from the question. ColabFold packages AlphaFold2 and AlphaFold-Multimer behind a notebook interface and is the usual entry point for custom single-chain and complex predictions. A local AlphaFold2 installation suits large batches and data that cannot leave your institution. The AlphaFold 3 server, run by Google DeepMind, accepts more molecule types than AlphaFold2, including [nucleic acids](/blog/guides/nucleic-acids) and ligands, which matters when an isoform-specific interaction involves DNA or a small molecule, and it operates under non-commercial terms with usage limits and output retention windows, so check the official terms and archive results promptly. For difficult cases, cross-check with a second platform. A systematic study of DNA binding by the competence protein ComP modeled the same complexes with AlphaFold 3, RoseTTAFold2NA, and Chai-1, and found that of six predicted binding modes only two were robust across platforms and proteins [7]. Agreement across independent tools is one of the few reliable markers of a trustworthy model, and that habit transfers directly to isoform work. Multiple sequence alignments deserve special attention for isoforms. The AlphaFold Database now publishes the underlying MSAs together with structures [1], which gives you reference alignments for free. For a novel splice junction, homologous fragments spanning the junction are rare, so the alignment has a gap exactly where you most want information, and low confidence at that point is the honest response. Do not recycle a canonical-sequence MSA for a divergent isoform without inspecting it, because misaligned exon boundaries propagate straight into the model. Template libraries have the same blind spot, since experimental structures rarely cover isoform-specific exons, so templates tend to support the shared domains and stay silent about the junction. Interpretation follows the same two numbers as before, pLDDT for per-residue confidence and PAE for relative placement. Two recurring patterns deserve respect. First, terminal segments misfold more often than cores. Benchmarking of AlphaFold 3 across a broad set of proteins found that overall backbone features were captured well, while minor misfolding of N- and C-terminal segments was common and often did not affect the biological roles of the studied proteins [8]. Because isoforms frequently differ at their termini, treat terminal models as provisional and anchor claims in the folded core. Second, intrinsically disordered regions produce low pLDDT by design, and that is not a failure. Disordered segments do not adopt a single stable structure, so reading a low-confidence tangle as a real conformation is a classic error. Work on alpha-synuclein, a heavily disordered protein central to Parkinson disease, handles this by combining sequence embeddings with AlphaFold coordinates and using per-residue confidence to down-weight unreliable structural regions [9]. Carry the same skepticism into isoform-specific tails. On practical logistics, a single-chain prediction typically finishes in minutes to a few hours depending on sequence length and alignment depth, so a modest isoform panel is an overnight job on one GPU. Pin your software versions, database dates, and Colab notebook versions in your methods section, because isoform predictions are only reproducible if the entire input stack is recorded. ## Option 4: Trimmed Constructs, Domain Boundaries, and Isoform Complexes Trimming serves two purposes. It lets you compare only the region shared by all isoforms, which makes fold comparisons fair, and it focuses compute on the part of the protein your question actually concerns instead of a long disordered tail. Where you cut, however, is a hypothesis with consequences. A cautionary example comes from the acyl carrier protein domains of the mycolactone polyketide synthase. Domain boundaries taken from prior studies missed an N-terminal helix later named H0. One isolated ACP fragment folded as expected, while the other, prepared with the old boundaries, failed to fold, and a shorter construct lacking H0 was substantially less thermostable than the N-terminally extended one [10]. The lesson transfers directly to isoform constructs and to any exon-derived domain assignment. Before you trim, run secondary structure prediction and check MSA context, because an apparently extra helix at a domain edge may be a rigid, functional part of the fold rather than linker. Complexes are where isoform handling pays off most. Isoform-specific exons often encode binding surfaces, and modeling the wrong isoform with a partner produces a confident-looking answer about a complex that never exists. The biological motivation is well documented, since isoforms from the same gene can substantially rewire protein interaction networks, which is the premise of prediction frameworks built specifically for isoform-level interactions [2]. Model the isoform and its partner with AlphaFold-Multimer or AlphaFold 3, then interpret the confidence output carefully. That last step has a well-known trap. ipTM is AlphaFold's summary score for interface quality, ranging from 0 to 1, and in common practice values above roughly 0.8 support a predicted interface while values below about 0.6 argue against one. The trap is that ipTM scores whole chains. If one chain carries large disordered or irrelevant regions, those residues shift the score even when the interface prediction itself is unchanged, so trimming a construct can raise or lower ipTM without any change in the modeled interaction [5]. Three practical rules follow. Compare ipTM only between models built with the same construct design. Lean on interface PAE values for local judgments. And know that revised metrics filtering residue pairs by PAE have been proposed specifically to correct this distortion [5]. Small sequence edits can also flip large conformational answers. In histidine kinases, AlphaFold models correctly distinguished cis from trans autophosphorylation arrangements within the dimer, both arrangements occur in nature, the switch has happened repeatedly in evolution, and the determinant turned out to be the length of the helices surrounding the DHp loop, where very small changes flip the mechanism [11]. An isoform that inserts a handful of residues into a helix can therefore change a predicted oligomeric arrangement. That is a genuine isoform effect, and AlphaFold will surface it only if you model the correct sequence rather than the canonical shortcut. ## Pros and Cons of Each Option **Option 1, database downloads.** Pros, free and immediate, consistent pipeline, MSAs now bundled with structures [1], and an ideal baseline for any project. Cons, apo single-chain models only, no custom constructs, isoform coverage is new and still expanding, and everything is fixed to a specific UniProt release. **Option 2, batch mapping and downloads.** Pros, scales to thousands of accessions, preserves provenance through the mapping file, and supports multiple output formats including PAE JSON [4]. Cons, quality depends on accession mapping, unmapped IDs need a fallback plan, and any web service can change its interface over time. **Option 3, custom prediction.** Pros, exact control over the sequence, works for novel variants and engineered constructs, complex and ligand inputs possible with newer tools, and fully reproducible when versions are pinned. Cons, compute and MSA management, recurring junction and terminal errors [8], and confidence interpretation that rewards experience. **Option 4, trimmed constructs and complexes.** Pros, fair isoform-to-isoform comparisons, direct tests of interaction hypotheses, and no wasted compute on disordered tails. Cons, boundary choices change outcomes [10], trimming distorts ipTM [5], and every construct is a hypothesis about biology that needs justification in your methods. ## Choosing Between Options: The Criteria That Decide Real Cases Five criteria settle almost every decision. Coverage asks whether the isoform exists in the AlphaFold Database under the post-update isoform coverage [1]. Sequence fidelity asks whether the model sequence matches your isoform residue for residue, verified by alignment rather than by accession alone. Context asks whether you need a single apo chain or a complex with a partner, nucleic acid, or ligand, because only custom prediction handles the latter. Scale asks whether you need one structure or a thousand, which pushes you toward downloads and batch APIs. Reproducibility asks whether you can rerun the analysis in a year, which favors recorded database versions and pinned prediction software. There is no single best alphafold isoform prediction recipe, because the best choice is whichever one reproduces the exact sequence you care about with confidence you can defend. In practice most projects converge on the same hybrid, database downloads as the baseline, custom predictions for the gaps, and trimmed constructs for controlled comparisons. ## Worked Scenarios: A Compact AlphaFold Tutorial for Isoform Cases These three scenarios work as a compact alphafold tutorial for isoform situations, with steps you can adapt directly. **Scenario 1, comparing three isoforms of a signaling protein.** Suppose isoform 1 has 720 residues, isoform 2 skips an exon that removes 31 residues inside a regulatory domain, and isoform 3 carries an alternative C-terminal tail. Start by querying the AlphaFold Database for the entry and checking which isoform models exist under the updated coverage [1]. For any missing isoform, run ColabFold on the exact sequence. Align all three sequences and build an explicit residue map. Compare pLDDT along the alignment, expecting a dip at the novel junction of isoform 2. Compare PAE patterns between the catalytic and regulatory domains to test whether the skipped exon changes predicted inter-domain geometry, which is often the real mechanistic question. If a partner binds the skipped region, model isoform 1 plus partner and isoform 2 plus partner with the same construct philosophy, and judge the interfaces by PAE rather than by raw ipTM, given how sensitive that score is to construct design [5]. **Scenario 2, an isoform-selective inhibitor.** Suppose you are docking against one member of a paralogous enzyme family and must spare the close relatives. Pull the AlphaFold Database model for the target isoform, then optimize it for docking, refining side chains and restoring the catalytic metal and cofactor chemistry that the apo model lacks, following the HDAC11 precedent [3]. Dock your series, then counter-screen the poses against models of the other isoforms to rationalize selectivity. The proof that this route works is concrete, an optimized HDAC11 AlphaFold model guided the design of an inhibitor with an IC50 of 365 nM that showed selectivity for HDAC11 over other HDAC isoforms, supported by [molecular dynamics simulations](/knowledge/bioinformatics/molecular-dynamics-simulations-of-proteins-and-force-fields) and activity in neuroblastoma cells [3]. Models propose, assays decide, so treat the computational selectivity profile as a hypothesis for in vitro testing. **Scenario 3, forty RefSeq isoforms from a non-model organism.** Start by clustering the protein sequences so near-identical isoforms collapse into representatives and you do not model the same fold repeatedly [6]. Submit the RefSeq accessions to the Structure Extractor, collect the models and the mapping file, and note which IDs came back unmapped [4]. Predict the unmapped ones with ColabFold using their exact sequences. Store everything with the mapping file as the provenance key, so any structure in your analysis folder can be traced back to the accession and isoform it represents. ## Comparing Isoform Structures Once You Have Them Once the models exist, comparison happens at three levels. At the sequence level, a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) anchors every exon difference to specific residues. At the structure level, superposition and per-region deviation measurements come first, and structural similarity searches extend the reach. ISS_ProtSci searches the AlphaFold Database for structural neighbors, validates fast Foldseek precomputations with DaliLite, annotates hits with UniProtKB and Pfam classifications, and reached over 99 percent recall of relevant proteins in its test cases, which is valuable when distantly related domains hide isoform relationships that sequence search misses [12]. At the function level, compare active site environments rather than whole folds. ActSeek searches the AlphaFold Database for proteins whose active sites resemble a seed structure, using a computer-vision-inspired approach, and can reveal whether an isoform preserves or disrupts a catalytic site [13]. Whatever comparison you run, keep it anchored to the residue map from your sequence alignment, because identical residue indices in two isoform models rarely refer to identical residues. ## Common Mistakes and Limitations 1. **Assuming the canonical model is your isoform.** This was the default failure mode for years and remains the most common one. Verify that the modeled sequence matches your isoform before interpreting anything. 2. **Quoting residue numbers across isoforms without a map.** A skipped exon shifts every downstream number. Build the alignment-based mapping once and reuse it. 3. **Confusing pLDDT with PAE.** pLDDT describes confidence in each residue, PAE describes confidence in relative positions between residues. Claims about domain arrangement need PAE. 4. **Comparing ipTM across differently trimmed constructs.** The score responds to disordered and accessory regions outside the interface, so trimming changes it even when the interface prediction does not move [5]. 5. **Recycling a canonical MSA for a divergent isoform.** Misaligned exon boundaries propagate into the model. Inspect the alignment at every junction, and use the MSAs now published with database structures as a reference [1]. 6. **Expecting ligands, modifications, or induced conformations from default models.** Database models are apo and unmodified, so docking and mechanistic projects need an optimization step [3]. 7. **Over-trusting termini.** N- and C-terminal segments are the most common site of minor misfolding even in strong models [8], and isoforms often differ exactly there. 8. **Reading disordered regions as structures.** Low pLDDT in an intrinsically disordered region is correct behavior, and down-weighting those regions rather than interpreting them is the established practice. 9. **Modeling transcripts the cell does not make.** Check evidence levels and nonsense-mediated decay annotations before committing compute to an isoform. 10. **Trimming blind.** Construct boundaries are biochemical hypotheses, and history shows that a missed edge helix can decide whether a construct folds at all [10]. 11. **Skipping version tracking.** Record the database release, software commit, MSA databases, and dates. An isoform analysis that cannot be rerun is not reproducible. 12. **Feeding a transcript sequence where a [protein sequence](/blog/guides/protein-sequence) is expected.** AlphaFold2 folds amino acids, and nucleic acid inputs belong to the newer multi-entity tools, not to a standard single-chain job. One overarching limitation applies to everything above. These models are hypotheses with quantified uncertainty, so any claim that matters for publication, patenting, or therapy needs experimental validation, and consequential projects benefit from a structural biologist's review before the interpretation hardens. ## Frequently Asked Questions ### Does the AlphaFold Database include protein isoforms? Yes. The 2025 update added isoform coverage together with the underlying multiple sequence alignments, alongside the long-standing canonical models, so check the entry page for your specific isoform before predicting anything yourself [1]. ### What is the best AlphaFold isoform prediction workflow? Download first and predict second. Use the AlphaFold Database model when your isoform is covered, otherwise predict the exact isoform sequence with ColabFold or the AlphaFold 3 server, and inspect confidence at splice junctions and termini before drawing conclusions. ### Should I submit the canonical sequence or the isoform sequence to AlphaFold? Submit the isoform sequence you actually study. The canonical sequence is a database convention, usually the longest or best-supported variant, and it is not automatically the biologically relevant one for your experiment. ### Why is my prediction low confidence exactly at the splice junction? A novel exon junction has few or no homologous fragments spanning it, so the multiple sequence alignment gives the network little evolutionary evidence at that point, and low local confidence is the honest response rather than a bug. ### Can AlphaFold predict isoform-specific protein interactions? Yes, with care. Model the isoform and its partner with AlphaFold-Multimer or AlphaFold 3, but judge interfaces with PAE rather than raw ipTM, because trimming changes ipTM even when the interface prediction does not move [5], and dedicated frameworks now target isoform interaction rewiring directly [2]. ### Do I need to rerun a prediction if the isoform is already in the database? No, not for a single-chain apo question. Rerun when you need a complex, a ligand or nucleic acid input, a custom construct, or a sequence version different from the database release. ## Sources 1. [AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage.](https://pubmed.ncbi.nlm.nih.gov/41273079/) 2. [DeepISO: deep learning-powered prediction of protein-protein interaction rewiring generated by alternative splicing.](https://pubmed.ncbi.nlm.nih.gov/41896911/) 3. [Utilization of an optimized AlphaFold protein model for structure-based design of a selective HDAC11 inhibitor with anti-neuroblastoma activity.](https://pubmed.ncbi.nlm.nih.gov/38996352/) 4. [AlphaFold Database Structure Extractor: a web server and API to download AlphaFold structures using common protein accessions](https://pubmed.ncbi.nlm.nih.gov/41291409/) 5. [Rēs ipSAE loquunt: What’s wrong with AlphaFold’s ipTM score and how to fix it](https://pubmed.ncbi.nlm.nih.gov/39990437/) 6. [How to cluster protein sequences: tools, tips and commands](https://doi.org/10.15406/MOJPB.2017.05.00174) 7. [Two different and robustly modeled DNA binding modes of Competence Protein ComP - systematic modeling with AlphaFold 3, RoseTTAFold2NA, Chai-1 and re-docking in HADDOCK](https://pubmed.ncbi.nlm.nih.gov/40338866/) 8. [Benchmarking Google DeepMind’s AlphaFold 3 Performance for Protein 3D-Structure Prediction](https://doi.org/10.26565/2220-637x-2024-43-01) 9. [A Fusion-based Sequence-Structure Protein Encoder of α-synuclein under Structural Uncertainty](https://doi.org/10.1109/ICIRCA69024.2026.11570573) 10. [Modular type I polyketide synthase acyl carrier protein domains share a common N-terminally extended fold](https://pubmed.ncbi.nlm.nih.gov/30787330/) 11. [Phosphorylation Mechanism Switching in Histidine Kinases Is a Tool for Fast Protein Evolution: Insights From AlphaFold Models](https://pubmed.ncbi.nlm.nih.gov/38884545/) 12. [3-D substructure search by transitive closure in AlphaFold database](https://pubmed.ncbi.nlm.nih.gov/40400345/) 13. [ActSeek: fast and accurate search algorithm of active sites in alphafold database](https://pubmed.ncbi.nlm.nih.gov/40713866/)