SignalP 6.0 is a deep-learning tool that identifies signal peptides in protein sequences and classifies them into all five known biological types, making it the first prediction method to cover the full range of these molecular sorting tags in a single model.1PubMed Central. SignalP 6.0 predicts all five types of signal peptides using protein language models Developed at the Technical University of Denmark, it builds on a lineage of SignalP servers stretching back to 1995 and represents the sixth major overhaul of a tool that has become one of the most widely used in computational biology.2SpringerLink / Methods in Molecular Biology. SignalP: The Evolution of a Web Server Its ability to handle all five signal peptide classes, identify internal regions within those peptides, and work on metagenomic data sets it apart from earlier versions and competing methods.
What Signal Peptides Actually Do
Cells make proteins in the cytoplasm, but many of those proteins need to end up somewhere else: anchored in a membrane, secreted outside the cell, or inserted into a specific compartment. Signal peptides are short stretches of amino acids at the beginning of a protein that act as molecular address labels. They direct the protein to the right export machinery, and once the protein arrives at its destination, the signal peptide is usually snipped off by a specialized enzyme called a signal peptidase. Signal peptidase I, for instance, is the enzyme responsible for releasing most exported proteins from the membrane after they have been threaded through the translocation channel.3PubMed Central. Signal peptidase I: cleaving the way to mature proteins
The basic architecture of a signal peptide is conserved across bacteria, archaea, and eukaryotes, though the details vary. A typical signal peptide has three loosely defined zones: a positively charged stretch at the very beginning, a hydrophobic core in the middle, and a polar region near the cleavage site. These features are what allow prediction tools to spot signal peptides computationally, but they also make signal peptides easy to confuse with transmembrane anchors, which share many of the same hydrophobic characteristics. That confusion was a persistent headache for earlier prediction methods.
The Five Types
Before SignalP 6.0, most prediction tools handled one or two signal peptide types well and ignored the rest. The five types correspond to different combinations of translocation pathways and cleavage enzymes. Understanding what distinguishes them is key to appreciating why a unified model matters.
- Sec/SPI: The most common class. These signal peptides route proteins through the Sec translocon, the general-purpose protein export channel, and are cleaved by signal peptidase I. Most secreted proteins in both bacteria and eukaryotes carry this type.
- Sec/SPII: Found in bacteria, these signal peptides also use the Sec translocon but carry a conserved “lipobox” motif near the cleavage site. Instead of being released into the environment, the mature protein is lipid-modified and stays tethered to the membrane as a lipoprotein. A cysteine residue at the start of the mature protein is essential for this lipid attachment.4PubMed. A lipoprotein signal peptide plus a cysteine residue at the amino-terminal end of the periplasmic protein beta-lactamase is sufficient for its lipid modification, processing and membrane localization in Escherichia coli
- Tat/SPI: These peptides direct proteins to the twin-arginine translocation (Tat) pathway, which exports already-folded proteins across the membrane. The hallmark is a pair of consecutive arginine residues in a conserved motif. After translocation, the signal peptide is still cleaved by signal peptidase I, just as in the Sec/SPI class.5PubMed Central. Specificity of signal peptide recognition in tat-dependent bacterial protein translocation
- Tat/SPII: A rarer combination where the Tat pathway exports a protein that is then lipid-modified and processed by signal peptidase II, yielding a membrane-anchored lipoprotein. These are uncommon but biologically real, and they posed a particular challenge for older tools.
- Sec/SPIII: These signal peptides are processed by a prepilin peptidase rather than signal peptidase I or II. They are found in type IV pilin-like proteins in bacteria and archaea, and their cleavage happens on the cytoplasmic side of the membrane rather than the extracellular side, which is unusual.
Older prediction tools typically handled Sec/SPI well because the training data was abundant. Tat signal peptides were addressed by separate specialist tools like TatP, which could correctly classify about 91% of known Tat signal peptides but was not integrated with Sec-type prediction.6PubMed Central. Prediction of twin-arginine signal peptides Sec/SPIII peptides were largely left out entirely. SignalP 6.0 brought all five under one roof for the first time.
Why a Unified Model Matters
Running separate tools for each signal peptide type creates practical problems. Researchers studying a newly sequenced genome would need to submit their sequences to multiple servers, reconcile conflicting predictions, and decide which tool to trust when one says “Sec/SPI” and another says “Tat/SPI.” A single model that jointly classifies all five types sidesteps that problem. It also learns shared features across the types, which helps with the rarer classes where training examples are scarce. Tat/SPII signal peptides, for instance, are uncommon enough that a model trained only on that class would struggle, but a model that already understands Tat signal peptides and SPII lipoproteins separately can generalize.
The unified approach also matters for metagenomic data, where you have millions of predicted protein sequences from environmental samples with no organism-level annotation. You cannot decide which specialized tool to run if you do not know whether the sequence comes from a bacterium, an archaeon, or a eukaryote. SignalP 6.0 was explicitly designed to handle metagenomic inputs without requiring the user to specify the organism group upfront.1PubMed Central. SignalP 6.0 predicts all five types of signal peptides using protein language models
How It Works Under the Hood
SignalP 6.0 is built on a protein language model, a type of neural network that has been pre-trained on vast numbers of protein sequences to learn statistical patterns in amino acid usage. Think of it as analogous to how large language models learn the rules of English by reading enormous amounts of text. The protein language model learns what “normal” protein sequence looks like, and the signal peptide predictor then fine-tunes that understanding to recognize the specific patterns that distinguish signal peptides from other sequence features.
This is a significant departure from earlier SignalP versions, which relied on handcrafted features like hydrophobicity profiles and position-specific scoring matrices. Those older approaches required experts to decide which sequence properties mattered. The language model approach lets the network figure out the relevant features on its own, which is part of why it performs better on the rarer signal peptide types where human intuition about the distinguishing features is less reliable.
Beyond just classifying a protein as having one of the five signal peptide types or no signal peptide at all, SignalP 6.0 also predicts the positions of the internal regions within the signal peptide. It can identify where the positively charged N-region ends, where the hydrophobic H-region sits, and where the cleavage site falls. This position-level annotation is useful for researchers who want to understand the biochemical logic of a particular signal peptide, not just whether one exists.
How the Training Data Was Built
A prediction tool is only as good as the data it learned from, and signal peptide datasets present a specific challenge: related proteins have similar signal peptides, so if related sequences end up in both the training set and the test set, the model can appear more accurate than it really is. SignalP 6.0 addressed this through a careful partitioning strategy. The developers split sequences into three groups at a maximum of 30% sequence identity, meaning no sequence in one partition shares more than 30% identity with any sequence in another partition.7PubMed Central. SignalP 6.0 predicts all five types of signal peptides using protein language models – Section: Methods Partitioning was done separately for each signal peptide type and for each of the four organism groups (eukarya, gram-positive bacteria, gram-negative bacteria, and archaea), then the sub-partitions were combined. This ensures that the cross-validation results reflect the model’s ability to generalize to genuinely novel sequences.
The constraint might sound like a technical detail, but it has real consequences. Many earlier benchmark results in the signal peptide prediction field were inflated because sequence similarity leaked between training and testing sets. The strict 30% threshold is one reason SignalP 6.0’s reported performance numbers can be taken more seriously than the numbers from some competing tools that used less rigorous evaluation.
Performance Compared to Other Tools
Independent benchmarks show that SignalP 6.0 performs at or near the top of the field. A study that introduced TSignal, a competing transformer-based method, found that TSignal achieved a weighted MCC (Matthews correlation coefficient, a balanced measure of prediction quality) of about 0.852 across all organism groups and signal peptide types, compared to 0.853 for SignalP 6.0.8PubMed Central. TSignal: a transformer model for signal peptide prediction The two methods are essentially neck and neck on overall accuracy, though TSignal showed small improvements on Tat/SPI prediction while SignalP 6.0 had a slight edge on Sec/SPII lipoproteins. The fact that a purpose-built competitor could only match, not clearly surpass, SignalP 6.0 suggests the tool sets a solid baseline for the field.
The more meaningful comparison, though, is not between SignalP 6.0 and a single rival but between SignalP 6.0 and the patchwork of older tools it replaced. Before version 6.0, a researcher doing genome annotation would use SignalP 5.0 for Sec/SPI and Sec/SPII, TatP for Tat signal peptides, and some manual heuristic for Sec/SPIII if they bothered at all. Each tool had different input formats, different output conventions, and different error modes. Unifying all of this into a single model with consistent output was arguably a bigger practical advance than the raw accuracy improvement.
The Signal Peptide Versus Transmembrane Helix Problem
One of the trickiest challenges in signal peptide prediction is distinguishing a genuine signal peptide from a transmembrane anchor at the start of a membrane protein. Both features are hydrophobic stretches near the N-terminus of the protein, and to a simple sequence scanner, they look nearly identical. Early computational work showed that roughly 95% of these could be distinguished in gram-negative bacteria using features like amino acid frequency and hydrophobicity patterns in the hydrophobic stretch, along with the starting position of the hydrophobic region.9PubMed. Computational differentiation of N-terminal signal peptides and transmembrane helices Eukaryotic proteins were harder, with correct prediction dropping to about 94%.
SignalP 6.0 benefits from the protein language model’s understanding of broader sequence context. Rather than looking only at the hydrophobic region itself, the model can use information from the entire protein sequence to decide whether a hydrophobic N-terminal stretch is a cleavable signal peptide or a permanent membrane anchor. This contextual information is something that older methods, which focused narrowly on the first 50 to 70 residues, could not easily exploit.
The Twin-Arginine Pathway and Why It Is Hard to Predict
Tat signal peptides deserve special attention because they are biologically fascinating and computationally tricky. The Tat pathway transports proteins that are already folded, sometimes with metal cofactors already bound, across the bacterial membrane.5PubMed Central. Specificity of signal peptide recognition in tat-dependent bacterial protein translocation This is fundamentally different from the Sec pathway, which threads proteins through a narrow channel in an unfolded state. The twin-arginine motif in the signal peptide is the routing label that says “use the Tat door, not the Sec door.”
Predicting Tat signal peptides is difficult for two reasons. First, the twin-arginine motif appears in a broader consensus sequence, but not every pair of arginines in a signal peptide means “Tat.” Sec signal peptides occasionally have arginine-rich N-regions that can fool pattern-matching approaches. Second, a chaperone-based proofreading system exists in the cell to prevent unfolded proteins from accidentally entering the Tat pathway.10PubMed Central. Signal peptide-chaperone interactions on the twin-arginine protein transport pathway This biological complexity means that having a twin-arginine motif in the sequence does not guarantee the protein actually uses the Tat pathway in vivo. Prediction tools can only work with the sequence, not the full cellular context, so some ambiguity is unavoidable.
Engineering Better Signal Peptides for Biotechnology
Beyond basic research, SignalP 6.0 has become a practical tool for biotechnologists trying to produce proteins at industrial scale. When you want a host cell to secrete a therapeutic protein into the culture medium, the signal peptide you attach to it matters enormously. A good signal peptide can mean the difference between milligrams and grams of purified product per liter of culture.
Researchers have started using SignalP 6.0 as part of computational pipelines to screen large libraries of candidate signal peptides. One recent study used the tool to screen millions of signal peptide variants derived from mouse and human sequences and from C-region mutants, looking for optimal signal peptides to drive secretion of human serum albumin in Chinese hamster ovary (CHO) cells, one of the most common host systems for producing biopharmaceuticals.11PubMed. Engineering enhanced signal peptides: A high-throughput computational pipeline for optimizing therapeutic protein production in CHO cells
Another group used SignalP 6.0 to guide the mutagenesis of signal peptides in E. coli. They generated hundreds of signal peptide sequences and systematically mutated different regions to find which parts tolerated changes and which did not. The hydrophobic core turned out to be surprisingly tolerant of mutations, as long as the overall hydrophobic character was maintained. One engineered variant successfully mediated secretion of human carbonic anhydrase II, a protein that is otherwise difficult to export from bacterial cells.12Process Biochemistry. Deep learning guided mutagenesis of signal peptide in Escherichia coli
A separate effort built a platform for globally profiling signal peptides and tested hits from high-throughput screening in actual expression systems. Signal peptides that ranked highly in computational and enrichment-based screening yielded 2.4 to 4.8 times more purified protein than the original signal peptide for the same target protein.13ACS Synthetic Biology. Platform for Global Profiling of Signal Peptides for Efficient Protein Production These are the kinds of yield improvements that change the economics of manufacturing a biologic drug or an industrial enzyme.
What SignalP 6.0 Cannot Do
No prediction tool is perfect, and SignalP 6.0 has known blind spots. The most important is that it only handles classical signal peptides, the kind that sit at the very beginning of a protein and are cleaved off after translocation. It does not predict non-classical secretion, where proteins leave the cell through unconventional routes without a recognizable signal peptide. Many biologically important secreted proteins, including some cytokines and growth factors, use these alternative pathways and will be missed entirely by SignalP 6.0.
The tool also cannot predict internal signal sequences, such as those that direct proteins to the endoplasmic reticulum after the ribosome has already started translating the rest of the protein. These signal-anchor sequences remain in the membrane rather than being cleaved, and while they share some features with cleavable signal peptides, they require different prediction methods.
For organisms that are poorly represented in training databases, performance may drop. Archaea, for example, have fewer experimentally verified signal peptides than bacteria or eukaryotes, so predictions for archaeal proteins carry more uncertainty. The same applies to highly divergent organisms from undersampled environments, even though the metagenomic mode is designed for exactly this situation.
Finally, a prediction is a statistical estimate, not a biological guarantee. Whether a signal peptide actually functions in a living cell depends on the host organism’s translocation machinery, the folding kinetics of the mature protein, interactions with chaperones, and other factors that no sequence-based tool can fully capture. SignalP 6.0 tells you what the sequence looks like it should do. Confirming what it actually does still requires an experiment.
Signal Peptidases as Drug Targets
An interesting downstream application of understanding signal peptide biology is the development of antibiotics that target signal peptidases. If you block signal peptidase I in a bacterium, secreted proteins pile up on the membrane, the cell cannot build its outer envelope properly, and it dies. This makes signal peptidase I an attractive target for new antibacterial drugs, and several research groups are pursuing inhibitors. Accurate identification of which bacterial proteins depend on signal peptidase I for their maturation, something SignalP 6.0 excels at, is directly relevant to understanding which cellular processes would be disrupted by such an inhibitor. The secretome of a pathogen, meaning the full set of proteins it exports, can be computationally predicted using tools like SignalP 6.0 and then studied for potential virulence factors or drug-target dependencies.
The lipoprotein pathway is similarly interesting from a therapeutic standpoint. Bacterial lipoproteins processed by signal peptidase II play roles in immune evasion and nutrient acquisition, and the enzyme that attaches the lipid (lipoprotein signal peptidase) is another potential antibiotic target. Knowing which proteins carry Sec/SPII signal peptides in a pathogen’s genome helps researchers prioritize which surface-exposed lipoproteins to study as vaccine candidates or drug targets. The ability of SignalP 6.0 to distinguish Sec/SPII from Sec/SPI with high accuracy is directly useful for this kind of analysis, since misclassifying a lipoprotein as a simple secreted protein would cause a researcher to overlook its membrane association entirely.