pocketz.doi.bio/docs/signalp6/README

SignalP 6.0 — Research Notes

Research notes on SignalP 6.0, the DTU Health Tech tool that predicts signal peptides and their cleavage sites across all domains of life using a protein language model. Compiled 2026-07-14 from the DTU service page, the Nature Biotechnology paper, the PMC full text, and the fteufel/signalp-6.0 repository.

Source clip in this vault: 2026-07-15182916serviceshealthtechdtudk_SignalP-6.0 Downloaded assets: papers/…SignalP-6.0_*.pdf, raw/…SignalP-6.0/

One-line summary

SignalP 6.0 is the first tool that predicts all five known signal-peptide types in one model. It couples a 30-layer BERT protein language model (ProtBERT, pretrained on UniRef100) to a conditional random field (CRF) decoder that labels every residue, so it also returns the cleavage site and the n-/h-/c-region structure of the signal peptide.

Contents

| File | What's in it | |------|--------------| | 01-overview.md | What signal peptides are, the five types, why v6.0 matters | | 02-architecture.md | Model architecture, training data, cross-validation | | 03-usage.md | Web server + standalone CLI: install, arguments, outputs | | 04-benchmarks-and-context.md | Performance vs SignalP 5.0, limitations, related tools, references | | 05-build-notes.md | Result of cloning + building the GitHub repo (2026-07-14) | | 06-retraining-cost.md | What it would cost to remake the weights ourselves (fine-tune, not pretrain) |

Fast facts