• Deep learning: SignalP 5.0 is based on convolutional and recurrent (LSTM) neural networks. The deep recurrent neural network architecture is better suited to recognizing sequence motifs of varying length, such as signal peptides, than traditional feed-forward neural networks (as used in SignalP 1-4).
• Conditional random field: The neural networks of SignalP 5.0 are combined with a conditional random field (CRF). The CRF imposes a defined grammar on the prediction and obviates the need for the post-processing step (Y- and D-scores) used in earlier versions of SignalP.
• Transfer learning: Instead of training separate networks for each organism group, SignalP 5.0 exploits the fact that signal peptides from different domains of life are to some degree similar. Thus, SignalP 5.0 is trained on data from all groups while an extra input unit informs the network about the origin of the sequences.
• Archaeal option: Thanks to transfer learning, SignalP 5.0 is able to make predictions also of signal peptides from Archaea, even though the data set is limited.
• Lipoprotein signal peptides: SignalP 5.0 can now differentiate between "standard" signal peptidase I-cleaved signal peptides (Sec/SPI) and signal peptidase II-cleaved lipoprotein signal peptides (Sec/SPII) in Bacteria and Archaea. Previously, we referred to the LipoP server for this prediction.
• Tat signal peptides: SignalP 5.0 can now differentiate between "standard" signal peptides translocated by the Sec translocon (Sec/SPI) and "Tat" (Twin-Arginine Translocation) signal peptides translocated by the Tat translocon (Tat/SPI) in Bacteria and Archaea. Previously, we referred to the TatP server for this prediction. However, SignalP 5.0 cannot predict lipoprotein signal peptides translocated by the Tat translocon (Tat/SPII) since we did not find any confirmed examples of these while constructing the data sets.

• SignalP 5.0 improves signal peptide predictions using deep neural networks
  José Juan Almagro Armenteros, Konstantinos D. Tsirigos, Casper Kaae Sønderby, Thomas Nordahl Petersen, Ole Winther, Søren Brunak, Gunnar von Heijne and Henrik Nielsen.
  Nature Biotechnology, , 2019.

• For the web page, an option to set the D-score cutoff values so that the sensitivity is the same as that of SignalP 3.0.
• Option included to set the minimum cleavage site position i.e. Ymax position - default value is 10.
• For the signalp package an option has been included to specify a temporary directory (-T dir).
• For the signalp package an option has been included to show signalp version (-V).
• Documentation rewritten.

Main publication:

• SignalP 4.0: discriminating signal peptides from transmembrane regions
  Thomas Nordahl Petersen, Søren Brunak, Gunnar von Heijne and Henrik Nielsen.
  Nature Methods, 8:785-786, 2011.

• Improved discrimination between signal peptides and transmembrane regions.
• No HMM method - only one prediction.

Main publication:

• SignalP 4.0: discriminating signal peptides from transmembrane regions
  Thomas Nordahl Petersen, Søren Brunak, Gunnar von Heijne and Henrik Nielsen.
  Nature Methods, 8:785-786, 2011.

• D-score. Improved quality of prediction.

Main publication:

• Improved prediction of signal peptides: SignalP 3.0.
  Jannick Dyrløv Bendtsen, Henrik Nielsen, Gunnar von Heijne and Søren Brunak.
  J. Mol. Biol., 340:783-795, 2004.

• Incorporation of a hidden Markov model version: SignalP V2.0 comprises two signal peptide prediction methods, SignalP-NN (based on neural networks, corresponding to SignalP V1.1) and SignalP-HMM (based on hidden Markov models). For eukaryotic data, SignalP-HMM has a substantially improved discrimination between signal peptides and uncleaved signal anchors, but it has a slightly lower accuracy in predicting the precise location of the cleavage site. The user can choose whether to run SignalP-NN, SignalP-HMM, or both.
• Retraining of the neural networks: SignalP-NN in SignalP V2.0 is trained on a newer data set derived from SWISS-PROT rel. 35 (instead of rel. 29 as in SignalP V1.1).
• Graphics integrated in the output: SignalP V2.0 shows signal peptide and cleavage site scores for each position as plots in GIF format on the output page. The plots provide more information than the prediction summary, e.g. about possible cleavage sites other than the strongest prediction.
• Signal peptide region assignment: SignalP-HMM provides not only a prediction of the presence of a signal peptide and the position of the cleavage site, but also an approximate assignment of n-, h- and c-regions within the signal peptide. These are shown in the graphical output as probabilities for each position being in one of these three regions.
• Automatic truncation: in SignalP V1.1, we recommended that you should submit only the N-terminal part of each protein, not more than 50-70 amino acids. SignalP V2.0 now offers to truncate your sequences automatically.

Main publication:

• Prediction of signal peptides and signal anchors by a hidden Markov model.
  Henrik Nielsen and Anders Krogh.
  Proceedings of the Sixth International Conference on Intelligent Systems for Molecular Biology (ISMB 6), AAAI Press, Menlo Park, California, pp. 122-130, 1998.

Main publication:

• Identification of prokaryotic and eukaryotic signal peptides and prediction of their cleavage sites.
  Henrik Nielsen, Jacob Engelbrecht, Søren Brunak and Gunnar von Heijne.
  Protein Engineering, 10:1-6, 1997.