# Signal Peptides in AlphaFold: Prediction and Removal If your AlphaFold model sprouts a loose helix from residue 1, the most likely cause is that the sequence you submitted still carries its N-terminal [signal peptide](/knowledge/molecular-biology/signal-peptide). The mature protein never contains that segment, because the cell clips it off during secretion, so the model displays a piece of the precursor that no crystal structure would ever show. The first thing to check is a signal peptide prediction on your exact input sequence, cross-checked against the curated mature chain annotation for your protein. If a predictor calls a cleavage site within roughly the first 40 residues and the annotation agrees, trim the leader and refold. This article is a practical guide to signal peptides and AlphaFold. It explains what a signal peptide is, how AlphaFold renders one in its output, how to predict the cleavage site before you fold, when removal helps and when it damages the model, and a complete troubleshooting workflow for N-terminal artifacts. A worked tutorial covers cleaning a sequence, submitting it to the AlphaFold server, and running the same job locally with ColabFold. Signal peptide removal is one of the highest-yield [preprocessing](/blog/data-analysis/what-is-data-preprocessing) steps in structural bioinformatics, and nearly every mistake associated with it is avoidable with a five-minute check. ## What Signal Peptides Are and Why AlphaFold Struggles With Them A signal peptide, also called a [signal sequence](/knowledge/molecular-biology/signal-sequence), is a short N-terminal extension that marks a newly made protein for transport into or across a membrane. In eukaryotic secretion, the signal peptide emerges from the ribosome, recruits the signal recognition particle, and delivers the growing chain to the endoplasmic reticulum (ER), where a membrane-embedded enzyme removes it, usually while the protein is still being made. That enzyme is the signal peptidase complex (SPC), a heterotetramer, meaning a complex of four different subunits, whose catalytic subunit is called Sec11 in yeast [1]. Despite their shared job, signal peptides look surprisingly little like one another. They share a conserved tripartite layout, an N-terminal region that is often positively charged, a central hydrophobic core, and a C-terminal region carrying the cleavage site, yet the actual sequences vary widely [1]. The cleavage site follows a loose rule, with small uncharged residues at the minus three and minus one positions relative to the cut, which is the pattern computational predictors search for. Signal peptides are commonly between about 15 and 35 residues long. Several related leaders get confused with signal peptides, and the differences matter for modeling: - A transit peptide targets a protein to mitochondria or chloroplasts and is removed inside the organelle by a different protease. SignalP does not detect these reliably, TargetP does. - A propeptide is removed after folding or activation, often far from the N-terminus, by a protease that recognizes a specific motif. - A bacterial lipoprotein signal peptide is cleaved by a dedicated enzyme called signal peptidase II, and the cysteine immediately after the cut stays with the mature protein and receives a lipid anchor. The key point for modeling is that the mature chain, the protein that remains after the leader is removed, is what exists in the cell's compartments and what structural biologists purify. AlphaFold is normally asked to predict the structure of that mature chain, so anything left over from the precursor is noise. ## How AlphaFold Represents Signal Peptides in Its Output AlphaFold reports two confidence measures that matter here. The pLDDT, or predicted local distance difference test, is a per-residue score from 0 to 100. By the convention in the AlphaFold documentation, scores above 90 indicate very high accuracy, 70 to 90 means the backbone is reliable while side chain orientations vary, 50 to 70 should be treated with caution, and below 50 frequently marks unstructured sequence. The PAE, or predicted aligned error, estimates in angstroms the expected positional error between pairs of residues, which makes it the right tool for asking whether two parts of the model have a defined orientation relative to each other. A full-length prediction of a secreted protein shows a consistent pattern. The signal peptide appears as an N-terminal helix, often with respectable pLDDT because a helix is a real and reproducible feature of these sequences, but the PAE between the helix and the globular domain is high. High PAE means the network is telling you it does not know where the helix sits relative to the rest of the protein, which is exactly right, because in the cell the helix sits in a membrane plane and then disappears. Neither AlphaFold 2 nor AlphaFold 3 embeds the modeled chain in a bilayer, so the orientation a membrane would impose is absent from the calculation. The same pattern appears in the precomputed models at the AlphaFold [Protein Structure](/blog/guides/protein-structure) Database, which are calculated from full precursor sequences and therefore include the signal peptide. There is a constructive side to this story. In recent work on the yeast signal peptidase complex, the catalytic subunit Sec11 carries a C-terminal hydrophobic region that was unresolved in human cryo-EM structures of the complex, most plausibly because the region is flexible [1]. AlphaFold predictions of the yeast complex suggested that this region forms a short C-terminal helix positioned within the presumed signal sequence binding site [1]. Topology mapping then confirmed that the helix crosses the ER membrane, and [molecular dynamics simulations](/knowledge/bioinformatics/molecular-dynamics-simulations-of-proteins-and-force-fields) showed that deleting it does not substantially alter the overall architecture of the complex or the nearby membrane environment [1]. The central finding of the study is that this short helix promotes productive engagement of internal signal sequences with the peptidase [1]. The lesson generalizes. AlphaFold can supply concrete, testable structural hypotheses about signal peptide machinery, including regions that experiments have struggled to resolve, but the predictions remain hypotheses until biochemistry confirms them. ## Predicting the Signal Peptide Before You Fold Prediction is cheap, so run it before every fold. Three tool families cover the space: - SignalP predicts cleavable signal peptides and reports the exact cleavage position. The current generation distinguishes Sec pathway signal peptides, Tat pathway signal peptides, bacterial lipoprotein signal peptides, and rare pilin-like types, and it offers separate models for eukaryotes, gram-positive bacteria, and gram-negative bacteria. - Phobius and DeepTMHMM classify each segment as a cleavable signal peptide, a transmembrane helix, or cytoplasmic and extracellular regions, which is the fastest way to tell a removable leader from a permanent membrane span. - TargetP predicts mitochondrial and chloroplast transit peptides, the leaders that SignalP ignores. A sensible routine is to run SignalP plus one topology classifier and compare the answers with curated annotation. Curated protein databases such as UniProt annotate signal peptides and mature chains with explicit evidence tags, so an experimentally verified cleavage site outranks any prediction. Reading the output takes a minute. Suppose you submit a 320 residue protein and SignalP returns a Sec/SPI signal peptide spanning residues 1 through 23 with the cleavage site between residues 23 and 24. The mature chain is therefore residues 24 through 320, a 297 residue polypeptide, and that is the sequence to fold. The result page also reports a probability for the call, and low probabilities deserve skepticism rather than automatic trimming. Two judgment calls come up repeatedly. First, the length of the hydrophobic core. A cleavable signal peptide has a hydrophobic stretch of roughly 7 to 15 residues followed by the cleavage motif, while a transmembrane helix spans about 18 to 25 residues and lacks the cleavage motif. If the hydrophobic segment is long and the predictors disagree, suspect a signal anchor and keep the segment. Second, the organism group. Running the bacterial model on a human protein, or the reverse, produces confident nonsense, so pick the group that matches the source organism. ## Signal Peptide Removal in AlphaFold: When to Trim and When to Keep ![Decision path asking if the question is about the mature protein, whether the segment is transmembrane, and whether it is a bacterial lipoprotein.](/article-images/illustrations/alphafold-signal-peptide-removal-trim-or-keep.webp) Remove the signal peptide when the object of your analysis is the mature protein. That covers most use cases, including docking, mutagenesis planning, comparison with crystal structures of mature extracellular domains, and design of expression constructs that lack the leader. Remove it too when the leader dominates the disordered end of your confidence plot and you keep catching yourself interpreting a phantom helix. Keep the signal peptide when the question involves the precursor itself, such as targeting, membrane insertion, topology, or processing. Keep it when the hydrophobic segment is a signal anchor rather than a cleavable leader, because trimming a transmembrane helix destroys the topology of the entire model. Keep it when you model bacterial lipoproteins, where the mature protein retains the lipidated cysteine and the numbering conventions around that cysteine matter for every downstream annotation. The effect of trimming on model quality is usually modest but positive. Dropping 20 to 30 residues rarely changes the depth of the [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) (MSA), the set of evolutionarily related sequences aligned so that corresponding residues occupy the same columns, because those alignments are built from full database hits either way. What improves is hygiene. The model starts at the biologically meaningful residue, superpositions against mature chain structures align cleanly, and nobody has to remember which end of the model is real. If you model complexes with AlphaFold 3, trim the leaders on every chain consistently so the interface analysis is not polluted by precursor segments. When the decision is genuinely close, fold both versions and compare the pLDDT and PAE values over the residues the two models share. The shared region should agree closely, and if it does not, that discrepancy is itself informative. ## Troubleshooting Table: Symptoms, Causes, Confirmation, Fixes This table covers the failure modes readers encounter most often, ordered roughly from most common to rarest. | Symptom in the model or workflow | Likely cause | How to confirm | Fix | |--|--|--|--| | Loose N-terminal helix, low pLDDT, high PAE to the rest of the model | Signal peptide left in the input sequence | SignalP calls a Sec/SPI leader with a cleavage site in the first few dozen residues | Trim to the mature chain and refold | | Model N-terminus clashes with a crystal structure although the fold matches | Full-length model compared with a mature chain structure | Check the construct statement in the structure entry | Superpose only residues after the cleavage site, or refold the mature chain | | First segment lies across the soluble domain like a membrane span | Signal anchor or transmembrane helix mistaken for a cleavable leader | DeepTMHMM or Phobius labels the segment transmembrane | Do not trim, treat membrane orientation as a separate question | | Confidence collapses after trimming | The cut removed part of the folded mature domain | Compare the trimmed length with curated mature chain annotation and with homologs | Restore the residues and correct the cleavage position | | Mutations and domain boundaries land at wrong positions | Numbering offset after trimming was never applied | Check whether model residue 1 equals gene residue 1 | Add the removed residue count back to every coordinate | | Precomputed database model disagrees with your own run at the N-terminus | Database models are full length, your run used the mature chain | Compare sequence lengths and the first residues of each | Pick one convention and apply it consistently across the project | | Server rejects the job or truncates the input | Sequence exceeds the AlphaFold server input limits | Count residues and check the official server documentation | Submit the mature chain, or run the job locally with ColabFold | | Odd chemistry at residue 1 of a bacterial envelope protein model | Lipoprotein signal peptide, cleavage leaves a lipidated cysteine | SignalP Sec/SPII call | Keep the cysteine in the sequence, handle lipidation outside the structural model | ## A Short Diagnostic Sequence Work these six checks in order before changing anything: 1. Color the model by pLDDT and note whether the problem clusters in the first few dozen residues. 2. Run SignalP on the original full-length sequence with the correct organism group. 3. Run a topology classifier such as DeepTMHMM or Phobius to separate cleavable leaders from transmembrane helices. 4. Read the curated annotation for the mature chain and its evidence level. 5. Decide keep or trim based on the biological question, not on the confidence plot alone. 6. If you trim, record the offset, refold, and remap every residue-level annotation through the offset. ## Why Each Failure Happens, From Most Common to Rarest ### 1. The leader is still in the sequence This is the dominant cause, and the mechanism is worth understanding so the fix feels obvious rather than magical. The multiple sequence alignment behind your prediction contains full-length precursors from many organisms, so the signal peptide columns look like a conserved hydrophobic helix and the network dutifully models one. Nothing tells the model that the segment is removed during secretion, and nothing provides the membrane plane that would orient it. The result is a helix with arbitrary position and orientation, flagged by high PAE to the body of the protein. The mature domain underneath is usually modeled correctly, which is why the artifact is annoying rather than fatal, but any automated analysis that consumes the N-terminus, from contact maps to epitope calculations, inherits the noise. ### 2. You trimmed the wrong kind of leader Transit peptides, propeptides, and signal peptides all sit at or near the N-terminus and all get removed, but by different enzymes at different times and places. Trim a mitochondrial transit peptide because SignalP stayed silent, and you may cut residues that fold back against the mature domain. Trim a propeptide that is structured in the mature protein, and the model loses a real secondary structure element, which shows up as collapsed pLDDT around the cut. The fix is taxonomic, not mechanical. Identify what kind of leader you have before deciding its fate. ### 3. The hydrophobic segment is a signal anchor or an internal signal sequence Some hydrophobic N-terminal segments target the chain to the membrane and then stay there as transmembrane helices. These signal anchors are not cleaved, and trimming them produces a model with a missing helix and nonsensical topology. Complications also exist inside the chain, because some internal signal sequences are engaged and processed by the signal peptidase complex rather than left in place, which is one reason blanket trimming rules fail. When a topology classifier calls the segment transmembrane, believe it, and switch the question from where to cut to how to model a membrane protein honestly. ### 4. The membrane context is missing Even a correctly retained signal peptide or signal anchor is modeled without a bilayer. AlphaFold has no membrane potential, no lipid composition, and no tilt constraints, so the helix orientation you see is one plausible pose among several. Two practical consequences follow. Do not read oligomerization interfaces or binding grooves off the face of a floating helix, and do not assume the modeled helix tilt matches the membrane. For serious membrane protein work, treat the AlphaFold output as the fold of the chain and layer topology predictions and experimental constraints on top. ### 5. The cleavage site is off by a few residues, or the N-terminus is processed again Predictors occasionally miss the exact cut by one or a few residues, especially when the minus three and minus one positions deviate from the canonical motif. Mature N-termini can also be trimmed further after the signal peptidase acts, by aminopeptidases, by pyroglutamate formation, or by removal of the initiator methionine upstream of the leader. Alternative initiation and [alternative splicing](/blog/guides/alternative-splicing) can produce longer or shorter leaders than the canonical transcript suggests. When the identity of the first mature residue matters, for example in an enzymology or immunology claim, establish it experimentally by mass spectrometric sequencing of the purified N-terminus or by Edman degradation rather than by prediction. ## A Step-by-Step AlphaFold Tutorial: Clean the Sequence, Then Fold ### Step 1. Fetch the canonical sequence Download the reviewed [protein sequence](/blog/guides/protein-sequence) in FASTA format, the plain text format with a header line beginning with a greater-than symbol followed by the one letter [amino acid sequence](/blog/guides/amino-acid-sequence). Save it as query.fasta. Work from a reviewed entry rather than a prediction-only record whenever one exists. ### Step 2. Predict the signal peptide Paste the sequence into the SignalP web server, select the organism group that matches your protein, and submit. The result lists each predicted type with a probability and reports the cleavage site as a position pair, for example a cut between residues 23 and 24. For batch work, install the command-line version and follow the README that ships with it, because the flags for organism group, output format, and output directory have changed between releases, and copying flags from old forum threads is a reliable way to generate errors. ### Step 3. Cross-check curated annotation Open the curated entry for your protein and read the feature table. Look for a signal peptide feature and a chain feature, note the coordinates, and note the evidence tags. Agreement between prediction and annotation settles the question. Disagreement means you go with the experimental evidence, and if none exists, you treat the boundary as provisional. ### Step 4. Trim with a script, not by hand Manual editing of FASTA files is where off-by-one errors breed. A short Biopython script is deterministic: ```python from Bio import SeqIO record = SeqIO.read("query.fasta", "fasta") offset = 23 # cleavage occurs after residue 23, mature chain starts at residue 24 record.seq = record.seq[offset:] record.id = record.id + "_mature" SeqIO.write(record, "mature.fasta", "fasta") ``` Store the offset in a small text file next to the sequence. Every downstream mapping depends on it. ### Step 5. Sanity check the trimmed sequence Three checks take seconds. The length should equal the annotated mature chain length. The first residues should match the annotated mature N-terminus. A hydrophobic stretch at the new N-terminus is a red flag that you cut in the wrong place. ### Step 6. Submit to the AlphaFold server The AlphaFold server, maintained by Google DeepMind and Isomorphic Labs for AlphaFold 3, runs predictions from a browser and is free for non-commercial use. Sign in, accept the terms of use, paste the mature FASTA into the protein field, add any ligands, ions, or partners the biology requires, name the job, and submit. The server enforces daily job limits and input size limits, so a trimmed sequence also buys headroom. Finished jobs return coordinate files with per-residue pLDDT and the full PAE matrix. One caution has nothing to do with chemistry. Anything you upload to a public server leaves your control, so check your institution or company data policy before submitting confidential or unpublished sequences. ### Step 7. Or run locally with ColabFold For batches, ColabFold packages AlphaFold with fast sequence search and a queue-friendly batch mode. After installing it and preparing its databases as described in its documentation, a single command folds the mature chain: ``` colabfold_batch mature.fasta results_mature ``` The same invocation accepts a directory of mature chain files for overnight runs. ### Step 8. Remap everything through the offset ![Worked calculation showing that a cleavage after residue 23 makes mature residue 1 equal precursor residue 24, and residue 297 equal 320.](/article-images/illustrations/alphafold-signal-peptide-removal-offset-remap.webp) Convert mature chain numbering back to precursor numbering with one rule, original position equals mature position plus the offset. With a cleavage site after residue 23, mature residue 1 is precursor residue 24. Apply this to variant positions, domain boundaries, active site residues, and primer design coordinates, and apply it in code so it happens the same way every time. **Safety note.** If the trimmed construct moves from the laptop to the bench for cloning and expression, [recombinant DNA](/blog/guides/recombinant-dna) work falls under your institution's [biosafety program](/knowledge/diagnostics/emerging-tech/biosafety-program-implementation-a-roadmap-for-diagnostic-laboratories). Confirm that the host strain and vector combination are covered by your approved registrations, and verify that the leader you removed is not performing a function your expression system still needs, such as secretion or periplasmic export. ## Common Mistakes and Limitations - Trimming on the strength of a single predictor. One tool, one organism group, one vote. Require agreement between a dedicated signal peptide predictor and a topology classifier, or between prediction and curated annotation, before cutting. - Reading pLDDT as a verdict on the signal peptide. A helix with pLDDT in the 70s is still a helix whose placement the model does not vouch for. The PAE matrix, not the pLDDT track, carries the orientation information. - Assuming the AlphaFold Protein Structure Database models are mature. They are full length. Every interpretation you apply to your own trimmed model has to be applied in reverse to database models. - Breaking numbering. The most expensive mistake in this workflow is silently mixing mature and precursor numbering across a project. Variants get misreported, primers amplify the wrong fragment, and the error survives review because the sequences look almost identical. - Comparing across different boundaries. A full-length model superposed wholesale onto a mature chain crystal structure wastes the N-terminal residues and distorts the global root mean square deviation, the standard superposition distance metric. Trim one side or the other first. - Trimming transmembrane segments. The costliest version of the mistake. A signal anchor removed from the input produces a confidently wrong model of a protein that does not exist. - Expecting AlphaFold to reason about processing. AlphaFold predicts structure. Cleavage sites, processing order, and maturation chemistry belong to dedicated predictors and experiments. The Sec11 study is a good template for the correct division of labor, with AlphaFold supplying a structural hypothesis and biochemistry testing it [1]. A final limitation deserves its own sentence. Every rule in this article is a starting point, and the mature N-terminus of any specific protein is an empirical fact that prediction can suggest but only experiment can confirm. ## Prevention Practices for Future Projects - Make the reviewed mature chain the project reference sequence, and derive everything else from it. - Store the cleavage evidence, the predictor outputs, and the numeric offset in a machine-readable file beside the sequence, not in someone's memory. - Run two orthogonal predictors automatically in your pipeline and flag disagreements for human review. - Fold both the full-length and mature versions when the distinction could matter, and archive both. - Inspect pLDDT and PAE plots before drawing any conclusion that involves an N-terminus. - Route every residue-level annotation through the offset programmatically, so renumbering is a function and not a habit. ## Frequently Asked Questions ### Should I remove the signal peptide before running AlphaFold? Usually yes, if your goal is the structure of the mature protein. Trim the leader at the predicted cleavage site, confirm the position against curated annotation, and submit the mature chain. ### Does leaving the signal peptide in ruin the whole model? No. The mature domain usually folds correctly either way. The leftover peptide appears as a low-confidence N-terminal helix with an arbitrary orientation, which misleads any analysis that touches the N-terminus. ### How do I find the exact cleavage site? Run a dedicated predictor such as SignalP and cross-check curated annotation. The output names the cut position directly, for example between residues 23 and 24, so the mature chain starts at residue 24. ### Is a transmembrane helix the same as a signal peptide? No. A cleavable signal peptide is removed after targeting, while a transmembrane helix stays in the mature protein. Some hydrophobic N-terminal segments are signal anchors that must not be trimmed. ### Can AlphaFold predict where the signal peptidase cuts? No. AlphaFold predicts structure, not processing. Cleavage site prediction is the job of tools such as SignalP. ### Why does the AlphaFold Protein Structure Database model look different from my crystal structure? Database models are computed from the full precursor sequence and include the signal peptide. Superpose only the residues after the cleavage site when you compare against a mature chain structure. ### Does trimming change model quality? Usually slightly for the better. Removing a leader that is absent from the mature protein removes noise, and confidence in the mature domain stays the same or improves. The larger benefit is clean, correctly numbered output. ### How do I map residue numbers after trimming? Add back the number of residues you removed. If the cleavage site falls after residue 23, mature residue 1 corresponds to residue 24 of the original sequence. ## Sources 1. [The Sec11 C-terminal short helix promotes productive engagement of internal signal sequences with the signal peptidase complex.](https://pubmed.ncbi.nlm.nih.gov/42586426/)