DeepTMHMM

A signal peptide and a transmembrane helix look nearly the same from sequence alone: roughly twenty hydrophobic residues near the N-terminus. Confusing the two does not merely add one wrong segment — it flips the predicted orientation of everything downstream, so a receptor's extracellular loops come back labelled cytoplasmic. The second long-standing gap is structural. Membrane proteins come in two folds, alpha-helical bundles and beta barrels, and topology predictors have historically specialised in one, leaving users to run several tools and reconcile their disagreements.

DeepTMHMM, from the Technical University of Denmark, BioLib, the University of Copenhagen and Stanford University School of Medicine, answers both with a single sequence-to-sequence labeller. Every residue receives one of six labels — signal peptide, inside (cytosol), alpha membrane, beta membrane, periplasm, or outside (extracellular space and the ER, Golgi and lysosomal lumen) — and the resulting label string is the topology. Protein type classification, segment boundaries, N-terminal orientation and signal peptide cleavage site all fall out of one prediction rather than a stack of specialised tools.

The name descends from TMHMM (Krogh et al., 2001), where a hidden Markov model encoded which label sequences are biologically legal. DeepTMHMM keeps that grammar and replaces the hand-specified emissions underneath it with a pretrained protein language model. Its successor, DeepTMHMM2, extends the same design to re-entrant regions, interfacial helices and membrane-type assignment.

#Key Features

• One model for both structural classes: Alpha-helical and beta-barrel transmembrane proteins, with and without signal peptides, plus globular proteins, are handled by a single network rather than a pipeline of specialists.
• Grammar-constrained decoding: A conditional random field scores the whole label sequence instead of each position independently, and expanding its state space beyond the six labels lets it reproduce realistic segment length distributions and legal compartment transitions. Either Viterbi or marginal decoding can be used.
• Cleavage sites, not just detection: Most topology methods detect a signal peptide but place its cleavage site poorly; DeepTMHMM was the best of the compared methods on cleavage site accuracy, slightly ahead of SignalP 6.0.
• Language model transfer in place of alignments: Because the encoder starts from representations learned over 250 million sequences, the method needs far less labelled structural data than an alignment-based predictor and needs no MSA at inference.
• Proteome scale: The hosted server accepts up to 10,000 sequences per submission, and the authors recommend running locally for anything larger.

#Technical Details

The encoder is three components in series: the pretrained ESM-1b protein language model, a bidirectional LSTM, and a dense layer with dropout, feeding a CRF decoder. Training used 3,574 sequences — 2,000 globular, 1,000 signal peptide plus globular, 387 alpha-helical transmembrane, 106 signal peptide plus alpha transmembrane, and 81 prokaryotic beta barrels — with CD-HIT homology reduction within each type at 30% sequence identity. Performance was measured in five-fold cross-validation, three folds training, one validating and one testing, with test performance assessed once after optimisation. DeepTMHMM leads overall protein type classification and beta-barrel detection. For alpha-helical topology it trails TOPCONS2 slightly and edges out CCTOP and DMCTOP, both of which contain more than 30% of the test proteins in their own training sets; it is best on alpha proteins carrying a signal peptide, and first on beta barrels ahead of BOCTOPUS2 and PRED-TMBB2. The authors note that benchmarking on their full set favours the compared methods a priori, since those training sets overlap it.

#Applications

Topology is the cheapest useful annotation of a membrane protein: it says which loops face outward and are therefore reachable by an antibody or a drug, and which face the cytosol. Running proteome-wide, DeepTMHMM supports target triage in pharmacology, functional annotation of metagenome-assembled genomes across all domains of life, and quality control on secretome predictions where signal peptides and transmembrane helices are routinely confused. It is used through the DTU Health Tech and BioLib web servers, and has been packaged as a community Nextflow module for genome annotation pipelines.

#Impact

DeepTMHMM made the case that a general protein language model plus a small, well-curated labelled set beats decades of specialised alignment-based topology machinery, and it consolidated three tasks that had been separate tools into one. Its stated limits matter for interpretation: it is set up primarily for prokaryotic beta barrels, since eukaryotic beta-barrel structural data is scarce, and its state space constrains barrels to an even number of strands, so the 19-stranded human VDAC is predicted with 18. The work remains a preprint. The software is free for academic users and licensed for a fee to commercial users, with a version commercialised by DTU; local copies are supplied on request rather than as an open download, though the training set is published.