henhaobio
Sign in

bioRxivBioinformaticsPreprint

DeepTMHMM predicts alpha and beta transmembrane proteins using deep neural networks

Jeppe HallgreniD, Konstantinos D. TsirigosiD, Mads D. Pedersen, José Juan Almagro ArmenterosiD, Paolo MarcatiliiD, Henrik NielseniD, Anders KroghiD, Ole WintheriD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

Transmembrane proteins span the lipid bilayer and are divided into two major structural classes, namely alpha helical and beta barrels. We introduce DeepTMHMM, a deep learning protein language model-based algorithm that can detect and predict the topology of both alpha helical and beta barrels proteins with unprecedented accuracy. DeepTMHMM (https://dtu.biolib.com/DeepTMHMM) scales to proteomes and covers all domains of life, which makes it ideal for metagenomics analyses.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (10 Apr 2022) · CC BY-NC-ND 4.0

Combining an ESM-1b encoder with a conditional random field decoder, DeepTMHMM classifies alpha-helical and beta-barrel transmembrane proteins and their topology, outperforming most compared tools, including on signal-peptide cleavage sites.

Transmembrane proteins make up roughly 30% of a proteome and are major drug targets but are scarce in the PDB, so topology must be predicted from sequence, and signal peptides are often mistaken for membrane segments. DeepTMHMM extends the widely used TMHMM, which handled only alpha-helical topology, with a protein language-model encoder and a conditional random field decoder that jointly predict signal peptides, alpha-helical and beta-barrel topology.

Key results

  1. ShownFor sorting proteins into five types (alpha TM with or without a signal peptide, beta barrel, globular with or without a signal peptide), DeepTMHMM beat all compared methods overall.

    Five-fold cross-validation on a 3,574-protein homology-reduced set; competitors were scored on the full set, which overlaps their training data. TMHMM was slightly better only for alpha TM proteins, which the authors put down to its conservative two-type output. Values are in supplementary tables not included in this text. · Main

  2. ShownDeepTMHMM was the best tested method for detecting beta-barrel proteins and ranked first for beta-barrel topology.

    BOCTOPUS2 and PRED-TMBB2 followed closely on topology, although their training sets substantially overlap the benchmark; the beta-barrel class holds 81 prokaryotic proteins. Numeric scores are not given in this text. · Main

  3. ShownAlpha-helical topology accuracy is level with the best existing tool, TOPCONS2, rather than clearly ahead: slightly behind it for alpha TM proteins and marginally ahead when a signal peptide is present.

    Alpha TM set: slightly below TOPCONS2 and slightly above CCTOP and DMCTOP (whose training sets include more than 30% of the test proteins), with other methods substantially worse. SP+alpha TM set: best overall, marginally ahead of TOPCONS2. · Main

  4. ShownSignal-peptide cleavage sites are located far more accurately than by other topology predictors, slightly better even than SignalP 6.0, without over-predicting signal peptides.

    Reported from comparisons in a display item and Supplementary Tables 5a-d that are not reproduced here; no cleavage-site metrics appear in this text. · Main

  5. SuggestedAlthough set up mainly for prokaryotic beta barrels, the model gives plausible topologies for eukaryotic beta barrels.

    For human VDAC (PDB 2JK4, 19 beta strands) it predicts 18 strands, because its state model forces an even count, and labels the first strand as signal peptide; other eukaryotic testing was small-scale. · Main

Methods, data and reproducibility

DeepTMHMM encodes sequences with the pre-trained ESM-1b protein language model, a bidirectional LSTM and a dense layer, then decodes per-residue labels (signal peptide, inside, alpha membrane, beta membrane, periplasm, outside) with a conditional random field constrained to follow transmembrane grammar. It was evaluated by five-fold cross-validation against existing topology predictors.

  • Homology-reduced set (CD-HIT, 30% identity within type): 2,000 globular, 1,000 SP+globular, 387 alpha TM, 106 SP+alpha TM and 81 prokaryotic beta-barrel proteins · n = 3,574 sequences
  • Human voltage-dependent anion channel (VDAC) structure, a eukaryotic beta-barrel test case · human · PDB 2JK4
Methods & tools
DeepTMHMM, ESM-1b, TMHMM, TOPCONS2, CCTOP, DMCTOP, BOCTOPUS2, PRED-TMBB2, SignalP 6.0, CD-HIT
Experimental conditions
The web server accepts up to 10,000 sequences per submission, and the authors recommend running locally for large-scale analyses. No hardware or runtime figures are given.
Reproducibility gaps
The paper says DeepTMHMM and its training dataset can be downloaded, but the links are not in this text; the tool is free only for academic users and licensed for a fee to commercial users.
Limitations (6)
  • Eukaryotic beta-barrel structures are scarce, so the model targets prokaryotic beta barrels and its eukaryotic testing was small-scale. · stated by the authors
  • The state-space model forces an even number of beta strands, so odd-stranded barrels such as VDAC (19 strands, 18 predicted) cannot be predicted exactly. · stated by the authors
  • TMHMM stays slightly better at classifying alpha TM proteins, and TOPCONS2 slightly better at alpha TM topology. · stated by the authors
  • Competitors were evaluated on the full set, which overlaps their training data, while DeepTMHMM was evaluated by cross-validation, so the comparison mixes evaluation regimes. · stated by the authors
  • The beta-barrel and SP+alpha TM classes are small (81 and 106 proteins), and the main text reports rankings without numeric scores, so per-class gains are hard to judge. · noted in this summary
  • The benchmark was designed and run by the developers, who share revenue from commercial licences of the tool. · noted in this summary

Read this if you annotate membrane proteins or signal peptides at proteome or metagenome scale and want one predictor covering alpha-helical and beta-barrel topology.

Citations
1,528

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTeufel '26Bernhofer '22Benavides '23Wang '19Gutierrez '22Li '21
80% semantic matchBioinformatics
Felix Teufel, Jeppe Hallgren, Henrik Nielsen+3 more·bioRxiv·Aug 27, 2026