# AlphaFold for Membrane Proteins: Sequence Preparation Tips By the end of this tutorial you will be able to take any membrane protein of interest, turn it into a clean, correctly scoped sequence file, run a prediction through the AlphaFold server or a local ColabFold installation, and judge how much the resulting model deserves your trust. The workflow applies equally to G protein-coupled receptors, transporters, channels, and viral membrane proteins, because the sequence decisions are the same even when the biology differs. Here is what to have on hand before you start: - A UniProt accession (the examples below use P29972, human aquaporin-1) or a sequence file you can verify against UniProt - A terminal with curl and Python 3.9 or newer - A browser for the AlphaFold server and the AlphaFold [Protein Structure](/blog/guides/protein-structure) Database - Optionally, ColabFold installed locally if you plan to run many predictions or want custom alignments - A structure viewer such as ChimeraX or PyMOL AlphaFold, the DeepMind system introduced in 2021, predicts protein structures with accuracy that is competitive with experimental structures in many cases [1]. For membrane proteins the reward is large, because experimental structures remain harder to obtain than for soluble proteins. The penalty for sloppy sequence preparation is also large, because the alignment machinery amplifies whatever you feed it. The steps below keep that penalty low. ## Why the Sequence You Submit Shapes the Model AlphaFold does not fold a sequence in isolation. It first searches for homologs, builds a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) (MSA), which is a set of evolutionarily related sequences lined up position by position, and it may also pull structural templates from the [Protein Data Bank](/blog/guides/protein-data-bank) (PDB). The network reads coevolutionary patterns out of that alignment to infer which residues contact each other [1]. Skolnick and colleagues describe the result as a highly sophisticated fold-recognition engine that leans on the completeness of the library of known PDB structures [2]. Two consequences follow for membrane proteins. First, the depth and quality of the MSA enabled by your sequence largely determine prediction quality, and analyses of protein-peptide docking show that both paired and unpaired alignments contribute measurable signal while training bias makes novel proteins and novel binding sites harder [3]. Second, proteins with few homologs, or with folds poorly represented in the PDB, are exactly where the method struggles most [3][2]. Membrane proteins add their own twists. Their transmembrane (TM) helices, the hydrophobic segments that span the lipid bilayer, are easy for the network to find but hard to orient correctly, because the training structures come from crystals grown in detergents or lipidic cubic phases that do not fully reproduce a bilayer. One study comparing AlphaFold with the Robetta pipeline on disease-relevant membrane proteins found Robetta generally more accurate at assigning a membrane-compatible topology, and argued that lipid environment, including raft lipids such as gangliosides, shapes membrane protein structure in ways a sequence-only model cannot see [4]. None of this means AlphaFold fails on membrane proteins. It means the sequence you submit and the checks you run afterward carry more weight than they would for a soluble enzyme. ## AlphaFold Membrane Protein Sequence Preparation, Step by Step ![Eight-step flow from fetching the UniProt FASTA through cleaning, topology mapping, database check, prediction and confidence review.](/article-images/illustrations/alphafold-membrane-protein-sequence-sequence-prep-steps.webp) The eight steps below take you from a raw accession to a judged, oligomer-aware model. Every command runs on a standard macOS or Linux terminal. ### Step 1: Fetch the canonical sequence from UniProt Start from the canonical sequence, the representative sequence UniProt displays for an entry, rather than from a [lab notebook](/knowledge/molecular-biology/lab-notebook) construct or an expression plasmid map. UniProtKB is the standard repository for protein sequences, and InterPro, the protein family and domain classification resource, builds its annotations on top of it [5]. Fetch the FASTA with the UniProt REST API: ```bash curl -s "https://rest.uniprot.org/uniprotkb/P29972.fasta" -o raw.fasta grep ">" raw.fasta ``` Expected output (the revision number SV may differ): ``` >sp|P29972|AQP1_HUMAN Aquaporin-1 OS=Homo sapiens OX=9606 GN=AQP1 PE=1 SV=2 ``` The sp prefix means the entry is reviewed by UniProt curators, and the header gives the organism, gene name, and annotation status. If curl returns an empty file or a 404 page, the accession is wrong, so re-check it on the UniProt website. If you deliberately want a specific isoform, fetch that isoform by its identifier, because the canonical sequence is not always the isoform your experiment expresses. Confirm the length before anything else: ```bash tail -n +2 raw.fasta | tr -d '\n' | wc -c ``` For aquaporin-1 this prints 301, which matches the Length field on the UniProt entry page. Always compare your count with that field, because a mismatch means extra header bytes or a truncated download. ### Step 2: Read the topology and domain annotations before editing anything Pull the feature table from the same entry: ```bash curl -s "https://rest.uniprot.org/uniprotkb/P29972.txt" | grep -E "^FT (TRANSMEM|TOPO_DOM|SIGNAL|CHAIN|DISULFID)" ``` Expected output for aquaporin-1 is a CHAIN line spanning the full length followed by six TRANSMEM lines, each with its residue range: ``` FT CHAIN 1..301 FT TRANSMEM ... FT TRANSMEM ... ``` Each feature type answers a preparation question: - TRANSMEM marks helical spans that must stay intact in any construct you build - TOPO_DOM marks cytoplasmic and extracellular loops, which tells you which ends face which compartment - SIGNAL marks a cleavable [signal peptide](/knowledge/molecular-biology/signal-peptide), which changes what the mature chain looks like - DISULFID marks cysteine pairs in extracellular domains that the final model should reproduce InterPro adds a second layer here. It classifies the sequence into families and predicted domains, and it now integrates AlphaFold models directly into its entries, which makes it a practical place to decide construct boundaries [5]. Write down the helix ranges and domain boundaries before you touch the file, because step 4 depends on them. ### Step 3: Clean and validate the FASTA Run this short script, saved as clean_seq.py: ```python from pathlib import Path lines = Path("raw.fasta").read_text().splitlines() header = lines seq = "".join(lines[1:]).replace(" ", "").upper() standard = set("ACDEFGHIKLMNPQRSTVWY") odd = sorted(set(seq) - standard) print(f"length: {len(seq)}") print(f"non-standard characters: {', '.join(odd) if odd else 'none'}") seq = seq.replace("*", "") Path("clean.fasta").write_text(f"{header}\n{seq}\n") ``` Expected output: ``` length: 301 non-standard characters: none ``` What the script is protecting you from: - X is the IUPAC code for an unknown residue. AlphaFold treats X as a generic placeholder, a handful of Xs are harmless, but long runs degrade confidence and some interfaces reject them outright. - The asterisk is a translation stop symbol and should be removed. - Lowercase letters can be interpreted as insertion characters in alignment formats, so submit uppercase only. - Spaces and digits from copied sequences corrupt the residue count silently. If your sequence came from a plasmid map rather than UniProt, diff it against the UniProt entry, because affinity tags, linker residues, and cloning scars are invisible in a structure viewer but fully visible to the MSA builder. Small sequence edits can shift predicted structures more than intuition suggests, even when the overall domain architecture survives the edit [6]. Version-control the cleaned file so every prediction traces back to an exact sequence. ### Step 4: Map the topology and choose the construct Run a dedicated topology predictor before modeling. DeepTMHMM classifies every residue as inside, membrane, or outside and reports the spans: ```bash biolib run DTU/DeepTMHMM clean.fasta ``` The job returns a GFF3 file listing each transmembrane region with its residue range, plus a three-line summary of the full topology. Compare those spans with the UniProt TRANSMEM lines from step 2. Agreement between independent predictors is a good sign. Disagreement marks the residues you should scrutinize hardest in the final model. Then decide what to submit: - **Full-length first.** Your first prediction should use the complete canonical sequence, TM helices included. Never trim through a hydrophobic span, because a broken helix forces the network to invent a different register and the entire topology can shift. - **Trim long disordered tails selectively.** Residues that predict as low-complexity and low-confidence waste compute and can dilute the alignment. If a terminus runs past roughly 50 residues with no annotated feature, prepare a second construct without it and compare. - **Handle signal peptides both ways.** For a protein with an annotated SIGNAL feature, prepare one FASTA with the full precursor and one with the mature chain after the cleavage site. Comparing the two tells you whether the peptide stabilizes or distorts the rest of the model. - **Split giant polyproteins thoughtfully.** For viral polyproteins, predictions from the full chain and from the individually cleaved products can differ. Researchers working on Flaviviridae non-structural membrane proteins found that interactions inferred from full-polyprotein predictions largely held when the mature proteins were modeled separately, which supports batching both representations [7]. - **Batch your variants.** ColabFold processes every record in a multi-entry FASTA, so full-length, mature-chain, and trimmed constructs can run in one pass: ``` >AQP1_full <301 residues from clean.fasta> >AQP1_mature ``` ### Step 5: Check the AlphaFold Database before you compute anything Someone may have already predicted your protein. The AlphaFold Protein Structure Database (AFDB), operated by EMBL-EBI with Google DeepMind, launched with over 360,000 structures across 21 model-organism proteomes and expanded to cover most of the UniRef90 sequence space, on the order of 100 million representative sequences [8]. Search your accession at alphafold.ebi.ac.uk. An entry gives you the model, per-residue confidence values, and predicted aligned errors, and the 2025 update aligned the database to the UniProt 2025_03 release while adding isoform coverage, the underlying MSAs, and a redesigned entry page with programmatic access through APIs, FTP, and Google Cloud [9]. Downloading the precomputed MSA matters more than it sounds. You can reuse it to seed local runs, which saves time and keeps your predictions consistent with the database model, a practice emphasized in large-scale interaction screens where alignment reuse keeps results comparable across many submissions [10]. Precomputed models do not eliminate your own runs, because AFDB models single chains and your construct may differ, but they are the fastest sanity check available. ### Step 6: Run the prediction #### Option A: the AlphaFold server The AlphaFold server at alphafoldserver.com is the fastest route for a handful of sequences. Add a Protein chain entity, paste your cleaned sequence, set the number of copies if you want a homo-oligomer, and submit. The service is free and intended for non-commercial use, and it enforces input size limits, so very long chains with large disordered regions may not fit. The server returns a structure with per-residue confidence coloring, a predicted aligned error view, and downloadable coordinate files. For a quick membrane protein question, this route usually answers it in minutes. #### Option B: ColabFold on your own hardware ColabFold packages AlphaFold with a fast MMseqs2-based MSA search and runs from the command line: ```bash python -m pip install "colabfold[alphafold-minus-jax]" colabfold_batch --num-recycle 3 --model-type alphafold2_ptm clean.fasta out_aqp1/ ``` Each option matters: - **num-recycle 3** runs three refinement passes. Recycling lets the network revise its internal representation, and three passes is the common working default. - **model-type alphafold2_ptm** selects single-chain weights. For complexes you must switch to the multimer weights, covered in step 8. - The MSA comes from the default MMseqs2 search against UniRef and environmental sequences. To test how the model behaves without evolutionary information, set msa-mode to single_sequence. When the run finishes, the output directory holds five ranked models, a scores file for each, and the MSA that was used: ``` ranked_0.pdb ranked_0_scores.json ranked_1.pdb ranked_1_scores.json ... ranked_4.pdb ranked_4_scores.json ``` ranked_0.pdb carries the best internal ranking. To reuse your own alignment instead of the automatic search, hand ColabFold an A3M file as the input. A3M is a FASTA-like alignment format in which lowercase letters mark insertions relative to your query sequence, and the AFDB downloads from step 5 give you a ready-made starting point [9]. ### Step 7: Judge the model with pLDDT, PAE, and topology checks AlphaFold reports two confidence measures you should read on every membrane protein. The predicted local distance difference test (pLDDT) is a per-residue score from 0 to 100, calibrated so that residues above 90 are highly accurate while residues below 50 often correspond to disordered regions rather than confident structure [1]. The predicted aligned error (PAE) is a residue-by-residue matrix expressed in Angstroms that describes how confidently the relative positions of two regions are known, and AFDB exposes both measures for every stored model [8]. Compute the mean pLDDT from the PDB B-factor column, where AlphaFold stores the per-residue values: ```python vals = [ float(line[60:66]) for line in open("out_aqp1/ranked_0.pdb") if line.startswith("ATOM") and line[12:16].strip() == "CA" ] print(f"residues: {len(vals)} mean pLDDT: {sum(vals)/len(vals):.1f}") ``` Example output: ``` residues: 301 mean pLDDT: 92.4 ``` Then run the membrane-specific checks: - **Walk the pLDDT profile against your TM list.** Every annotated helix should sit in a high-confidence stretch. A helix with sagging confidence is a candidate for a wrong register. - **Read the PAE blocks between helices.** Low PAE values between TM segments mean their relative packing is settled. High values mean the network is guessing, which is common for helices connected only by long loops. - **Confirm the topology direction.** Check that the inside and outside assignments from DeepTMHMM and UniProt match the model. Disagreement here is precisely the failure mode reported in the Robetta comparison study of membrane proteins [4]. - **Color the model by B-factor in your viewer.** Both ChimeraX and PyMOL can color by the B-factor column, which turns the pLDDT profile into an immediate visual audit. ### Step 8: Model the functional oligomer Many membrane proteins only make sense as assemblies. Aquaporin-1 is a tetramer, many channels are dimers, and a monomer model of an obligate oligomer can look superb while missing the biological unit. Benchmarking of AlphaFold-Multimer shows strong performance on dimers and more variable results on larger complexes, with the proposed pDockQ2 score estimating the quality of each interface in a multimer [11]. Before you pick a stoichiometry, look for experimental evidence. Native mass spectrometry approaches such as top-down analysis of intact membrane protein complexes can reveal oligomeric state and sequence information in a single experiment [12], and stoichiometries are often already annotated in UniProt or in the PDB entry of a homolog. To model an oligomer with ColabFold, put each chain in the same FASTA file and switch the weights: ``` >AQP1_chain1 <301 residues> >AQP1_chain2 <301 residues> >AQP1_chain3 <301 residues> >AQP1_chain4 <301 residues> ``` ```bash colabfold_batch --model-type alphafold2_multimer_v3 --num-recycle 3 tetramer.fasta out_tetramer/ ``` The scores files now include an interface predicted template modeling score (ipTM), which summarizes confidence in the arrangement between chains. Values around 0.8 or higher generally indicate a trustworthy interface, while low values mean the relative chain placement is speculative. On the AlphaFold server, setting the copy number for the protein entity to 4 runs the equivalent job. ## Parameters Worth Knowing | Parameter or file | Typical choice | What it controls | Practical note | |, - |, - |, - |, - | | Canonical sequence | UniProt REST FASTA download | Which residues enter the MSA | Length must match the UniProt Length field exactly | | msa-mode | mmseqs2_uniref_env (default) | Source of the alignment | single_sequence probes orphan behavior | | templates option | on for conserved folds | Whether PDB structures guide the model | Turn off if known templates bias the fold | | num-recycle | 3 | Number of refinement passes | Higher values cost time, rarely hurt | | model-type | alphafold2_ptm or alphafold2_multimer_v3 | Network weights for monomer versus complex | Set explicitly whenever chains exceed one | | max-msa | 64:128 when memory-bound | Alignment depth per pass | Cuts memory on long chains | | Custom MSA | A3M file as input | Reuse a precomputed alignment | AFDB now supplies MSAs for its models [9] | | Confidence readouts | pLDDT, PAE, ipTM | Where the model trusts itself | Judge per residue and per interface, never on averages alone | ## Errors You Will Hit and How to Fix Them ![Decision path asking whether a database model exists, how many chains, and whether memory or size limits force trimming.](/article-images/illustrations/alphafold-membrane-protein-sequence-model-type-triage.webp) ### Out of memory on long membrane proteins The run dies with a JAX message like `RESOURCE_EXHAUSTED: Out of memory while trying to allocate ...`. Long chains with big disordered regions drive memory use through the attention layers. Fix it by trimming the disordered termini, lowering the alignment depth with the max-msa option (for example 64:128), reducing the model count with num-models 1, or sending the job to the AlphaFold server instead. ### Invalid or ambiguous residues The run aborts during setup and the log points at a nonstandard character, or the server flags the sequence in the browser. Replace B, Z, X, and other ambiguity codes with the most likely residue or trim the affected region, then rerun the cleaning script until the check prints none. ### Wrong model type for the job A multi-chain FASTA run with monomer weights either errors out or produces chains that ignore each other. Set model-type to the multimer weights whenever the input has more than one chain. ### Chains silently merged Two sequences without separate FASTA headers become one long chain, and the model grows a phantom helix at the junction. Every chain needs its own header line beginning with the greater-than symbol. ### Server rejections The AlphaFold server refuses the submission with a message about sequence validity or size. Strip nonstandard characters, split oversized inputs into domains, and keep the non-commercial terms in mind. ### Empty or truncated downloads The length check from step 1 disagrees with the UniProt Length field. Refetch the file, confirm the accession character by character, and never hand-edit a sequence to force the numbers to agree. ## Common Mistakes and Limitations 1. **Submitting the expression construct instead of the protein.** Affinity tags, fusion partners, and linker residues change the MSA and can produce confident-looking helices that exist only in your vector. Strip to the canonical or mature chain. 2. **Reading only the mean pLDDT.** Averages hide localized failures. A model can average above 85 while one loop or one helix sits at 45. 3. **Trimming through a transmembrane helix.** Hydrophobic spans are load-bearing. Cut elsewhere or not at all. 4. **Ignoring oligomeric state.** Monomer-first habits from soluble protein work transfer badly, because multimer benchmarks show interface quality varies widely and needs its own scoring [11]. 5. **Expecting the bilayer.** AlphaFold sees no membrane, no membrane potential, no cholesterol, and no raft lipids. The Robetta comparison frames these environmental factors as part of what its authors call the epigenetic dimension of protein structure, beyond sequence [4]. Position the finished model into a membrane with tools such as the OPM and PPM servers, and treat any topology disagreement as a red flag. 6. **Over-trusting predictions for orphan sequences.** Viral and parasite membrane proteins often have shallow alignments, and the models partly reproduce structures seen in training, which hurts novel folds and binding sites [3]. The feline calicivirus LC protein study shows the practical response, running AF2 and AF3 alongside other methods such as QUARK and then testing the putative TM domain experimentally [13]. 7. **Treating the single model as the single truth.** Channels and transporters sample multiple conformations, and a static prediction is one state. Flow-based extensions of AlphaFold generate ensembles that capture conformational variability better than MSA subsampling, and reviews of AlphaFold 3 emphasize that large complexes, dynamics, and mutation effects remain hard [14]. 8. **Editing sequences casually between runs.** Even small perturbations can change predicted structures measurably [6]. Freeze the FASTA, record its checksum, and version every change. 9. **Skipping the database check.** AFDB covers most UniRef90 representatives [8] and now ships isoforms and MSAs [9], so hours of local compute are sometimes redundant. 10. **Forgetting extracellular constraints.** Annotated disulfide pairs and glycosylation sites in extracellular domains should be reflected in the model, and mismatches there usually signal a deeper problem with the fold or the register. ## Sensible Next Steps Once a model passes the checks, three directions pay off. Refine and contextualize with molecular dynamics in an explicit bilayer, which restores some of the environmental physics the prediction ignores. Explore conformational states with ensemble methods rather than treating ranked_0 as the only answer. And use the structure as a feature source for downstream questions, for example transporter substrate prediction pipelines that combine AlphaFold-derived structures, protein language model embeddings, and structural encodings such as FoldSeek 3Di sequences [15]. Whatever you do next, record the exact FASTA, the software version, the flags, and the date, because both the sequence databases behind the MSA search and AFDB itself keep updating [9]. ## Frequently Asked Questions ### Should I include the signal peptide in the sequence I submit to AlphaFold? Prepare both the full precursor and the mature chain and compare the two predictions. The mature chain is usually the biologically relevant construct, but keeping the peptide in the first run helps you confirm that the model respects the annotated cleavage site. ### Do I need an MSA, or can AlphaFold work from a single sequence? It can run from a single sequence, but accuracy usually drops, especially for membrane proteins with few homologs. Test the difference by setting msa-mode to single_sequence in ColabFold and comparing the confidence profiles. ### Which confidence score should I trust for transmembrane helices? Trust the per-residue pLDDT together with the PAE values between helices, never the mean score alone. A helix with high pLDDT but high PAE relative to its neighbors is packed speculatively even when it looks beautiful. ### Can the AlphaFold server handle my full transporter? Usually yes within its input size limits, and it is the quickest route for a first look. Very long chains with large disordered regions may not fit, in which case trim the tail or move to a local ColabFold run. ### How do I model a membrane protein dimer or tetramer? Put every chain in one FASTA file, switch to the multimer weights in ColabFold, or set the copy number on the AlphaFold server. Judge the result with the ipTM score and, for larger assemblies, per-interface scores such as pDockQ2. ### My protein has no close homologs. Will AlphaFold still work? Sometimes, but treat the output with caution, because shallow alignments and folds missing from the training data are known weak points. Compare against other predictors, use templates deliberately, and validate key features experimentally. ### Does AlphaFold place the model in the membrane for me? No. The prediction contains no bilayer, so orient the model in the membrane afterward with positioning tools such as OPM or PPM and check that hydrophobic spans sit inside the lipid phase. ### Where can I get ready-made membrane protein models and their MSAs? The AlphaFold Protein Structure Database at alphafold.ebi.ac.uk hosts predictions searchable by UniProt accession, and its 2025 update added isoform coverage and downloadable underlying MSAs alongside the models. ## Sources 1. [Highly accurate protein structure prediction with AlphaFold](https://pubmed.ncbi.nlm.nih.gov/34265844/) 2. [AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function](https://pubmed.ncbi.nlm.nih.gov/34586808/) 3. [Training bias and sequence alignments shape protein-peptide docking by AlphaFold and related methods](https://pubmed.ncbi.nlm.nih.gov/41084281/) 4. [The Epigenetic Dimension of Protein Structure Is an Intrinsic Weakness of the AlphaFold Program](https://pubmed.ncbi.nlm.nih.gov/36291736/) 5. [InterPro: the protein sequence classification resource in 2025](https://pubmed.ncbi.nlm.nih.gov/39565202/) 6. [Exploring the Predictive Capabilities of AlphaFold Using Adversarial Protein Sequences](https://doi.org/10.1109/TAI.2024.3353708) 7. [AlphaFold Prediction of Protein-Protein Interactions in the Flaviviridae Proteomes](https://pubmed.ncbi.nlm.nih.gov/41155451/) 8. [AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models](https://pubmed.ncbi.nlm.nih.gov/34791371/) 9. [AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage](https://pubmed.ncbi.nlm.nih.gov/41273079/) 10. [Using AlphaFold Multimer to discover interkingdom protein-protein interactions](https://pubmed.ncbi.nlm.nih.gov/39152709/) 11. [Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes](https://pubmed.ncbi.nlm.nih.gov/37405868/) 12. [Top-down mass spectrometry of intact membrane protein complexes reveals oligomeric state and sequence information in a single experiment](https://pubmed.ncbi.nlm.nih.gov/25970171/) 13. [The Feline calicivirus Leader of the Capsid (LC) Protein Contains a Putative Transmembrane Domain, Binds to the Cytoplasmic Membrane, and Exogenously Permeates Cells](https://pubmed.ncbi.nlm.nih.gov/39205293/) 14. [Analysing protein complexes in plant science: insights and limitation with AlphaFold 3](https://pubmed.ncbi.nlm.nih.gov/40402396/) 15. [Application of Protein Structure Encodings and Sequence Embeddings for Transporter Substrate Prediction](https://pubmed.ncbi.nlm.nih.gov/40807401/)