Predicting Transmembrane Topology with TMHMM

Transmembrane helices · Hidden Markov model · Topology prediction · QTY code

Published 2026 · Last updated 2026-09-28 · Hailu Zhou

What question does this answer? After QTY substitution, are the transmembrane helices still there? The QTY code replaces hydrophobic residues (L, I, V, F) with polar ones (T, Q, Y) so that a membrane protein can fold in water without detergent. Before any cell is lysed or any resin is packed, we must ask whether that substitution destroyed the very helices that define the fold. TMHMM gives a first, zero-cost answer.

What TMHMM actually computes

TMHMM is a hidden Markov model that labels each residue of a protein sequence as inside, transmembrane (helical), or outside, jointly for the whole chain [1]. A hidden Markov model is a probabilistic machine with hidden states (the true topology) that emit observable symbols (the amino acids). Because a membrane protein is a single connected chain that must traverse the bilayer an integer number of times, TMHMM does not score helices independently; it finds the most probable topology as a whole. This global constraint is why it outperforms simple hydrophobicity windows: it forbids a helix sitting alone in the middle of an aqueous loop.

The model was trained on proteins of known three-dimensional structure, and its key advance was allowing variable helix length and modelling the loop regions between helices. For a G-protein-coupled receptor such as CCR5 or CXCR4, the expected output is seven transmembrane segments, alternating with three intracellular and three extracellular loops plus the N-terminal extracellular tail and the C-terminal intracellular tail.

Why a computational pre-check comes first

Every wet-lab step later in the pipeline — yeast two-hybrid screening, affinity capture, size-exclusion chromatography, thermophoresis — consumes plasmid, competent cells, media, and instrument time. A topology prediction consumes a text file and seconds of CPU. If QTY substitution has erased a helix, no downstream assay can rescue the construct, and the earlier we learn that, the less is wasted. This is the rational-design argument for ordered validation: cheap and deterministic checks gate expensive and stochastic ones.

The pre-check is not merely a gate, it is also a design instrument. By running TMHMM on the wild-type sequence and on each candidate variant side by side, we confirm that the helices removed from the membrane-facing set are the ones we intended to solubilise, and that no unexpected transmembrane segment appeared.

How to read the output

TMHMM reports several complementary views. The summary line gives the predicted number of transmembrane helices and the expected number of amino acids in each compartment. The per-residue table lists, for each position, the probability of being inside, in the membrane, or outside, and the predicted label. The probability plot draws these three curves along the sequence.

Output elementWhat to look forReason
Predicted helix countRemains 7 for a chemokine receptorThe seven-helix bundle is the structural signature of the fold
Helix boundariesShift by less than a few residuesSmall register shifts are acceptable; large shifts move loop anchors
Inside/outside label patternAlternating TM/loop/TM as in wild typeThe connectivity, not just the count, defines topology
Probability curvesTM probability near 1 across helix coresAmbiguous plateaus near 0.5 flag weak helices
N- and C-terminal tailsCorrect side assignmentConfirms the global orientation of the model

What "topology preserved" means after QTY substitution

QTY substitution is not random: it targets only residues with aliphatic or aromatic hydrophobic side chains and replaces them with threonine, glutamine, or tyrosine. Those replacements make the sequence far less hydrophobic, which is the whole point — a helix that cannot partition into a membrane can now be handled in buffer. The consequence for prediction is that the absolute hydrophobicity of each helix drops, so a naive sliding-window method might fail to call a helix at all. A hidden Markov model, having learnt the statistical grammar of membrane proteins, is more tolerant: as long as the pattern of alternating hydrophobic and charged residues survives, the helix is still recognised.

We therefore define topology preserved as: the same number of transmembrane helices, the same inside/outside assignment of the intervening loops, and helix boundaries within a few residues of the wild type. When these three conditions hold, the variant is a legitimate candidate for the wet-lab screen. The prediction tells us that the water-solubility change did not come at the cost of the fold architecture.

Troubleshooting and pitfalls

Where this fits in the pipeline

Topology prediction is the first of four questions the pipeline asks, and it rules out the first class of artefact: a construct that was mis-designed before it was ever made. It costs nothing, it is deterministic, and it protects every expensive step that follows. Only after the helices are confirmed present does it make sense to ask whether the variant binds a ligand inside a living cell — the question answered by yeast two-hybrid screening.

References

  1. Krogh A, Larsson B, von Heijne G, Sonnhammer ELL. Predicting transmembrane protein topology with a hidden Markov model: application to complete genomes. J Mol Biol. 2001;305(3):567–580. doi:10.1006/jmbi.2000.4315
  2. Waterhouse A, Bertoni M, Bienert S, et al. SWISS-MODEL: homology modelling of protein structures and complexes. Nucleic Acids Res. 2018;46(W1):W296–W303. doi:10.1093/nar/gky427
  3. Pándy-Szekeres G, Munk C, Tsonkov TM, et al. GPCRdb in 2018: adding GPCR structure models and ligands. Nucleic Acids Res. 2018;46(D1):D440–D446. doi:10.1093/nar/gkx1109
  4. Zhang S, Tao F, Qing R, et al. QTY code enables design of detergent-free chemokine receptors that retain ligand-binding activities. Proc Natl Acad Sci USA. 2018;115(37):E8652–E8659. doi:10.1073/pnas.1811031115
  5. Krieger E, Vriend G. YASARA View—molecular graphics for all devices—from smartphones to workstations. Bioinformatics. 2014;30(20):2981–2982. doi:10.1093/bioinformatics/btu426
  6. Maier JA, Martinez C, Kasavajhala K, et al. ff14SB: improving the accuracy of protein side chain and backbone parameters from ff99SB. J Chem Theory Comput. 2015;11(8):3696–3713. doi:10.1021/acs.jctc.5b00255
  7. Zhang S, Qing R, et al. QTY code for designing water-soluble membrane proteins. Proc Natl Acad Sci USA. 2018;115(37):E8652–E8659.