DeepTMHMM2

The lumen of the endoplasmic reticulum is topologically the outside of the cell once you trace a protein through the secretory pathway, so a topology predictor that emits "inside" and "outside" hands a cell biologist a label meaning the opposite of what it reads like. Worse, the standard topology alphabet has no symbol at all for segments that dip into the bilayer without crossing it: re-entrant regions, which fold back on the side they entered and often line the selectivity filter of an ion channel, and interfacial helices, which lie along the membrane surface.

DeepTMHMM2, from University of Copenhagen and the Technical University of Denmark, closes both gaps. It is the first topology predictor to place re-entrant regions and interfacial helices in its output alongside conventional α-helical and β-barrel spans, signal peptides, and transit peptides, and it predicts in parallel which of 17 biological membranes the protein resides in — from the archaebacterial membrane to the thylakoid to the lysosome. That second head is what makes the first interpretable: once the membrane type is known, the two topological sides can be named concretely, as "mitochondrial matrix" and "intermembrane space" rather than a generic inside and outside.

The model is a supervised head over a frozen protein language model, released as a fixed ensemble that runs without retraining. The authors applied it to all 574,627 Swiss-Prot proteins, and the result reframes how common these overlooked elements are.

#Key Features

• Non-crossing segments as first-class labels: Re-entrant regions and interfacial helices are predicted directly, at F1 scores above 0.7 for detection, against 0.38 and 0.31 for the prior methods that attempted them.
• 17-membrane localization: A separate multilabel head assigns proteins to specific membranes, and because a protein can occupy more than one, the head predicts a multi-hot vector rather than forcing a single choice.
• Grammar-constrained decoding: A conditional random field with an expanded state space encodes biophysical rules — a bilayer can only be crossed by a complete span of legal length, a re-entrant segment must return to the side it started on, helices and barrels cannot coexist — so every prediction is syntactically valid by construction.
• Learning from partial labels: Because the CRF marginalizes over an "undetermined" symbol, proteins with only partially resolved topologies still contribute training signal.
• Structure-bootstrapped training data: A Foldseek pipeline transfers curated topology annotations onto remote structural homologs in AlphaFold DB and the ESM Metagenomic Atlas, multiplying a small experimental set.

#Technical Details

Sequences are embedded by a frozen ESM-2 650M backbone, which is not fine-tuned; a convolution with kernel size 3 feeds an LSTM with hidden size 512, projected to CRF emissions. The membrane-type head is an attention-pooling classifier over the same embeddings. The experimental training set intersects OPM and PDBTM annotations through SIFTS, adds UniProt sorting signals, and was reduced to 20% maximum pairwise identity, leaving 8,094 sequences across five folds; 540 proteins were manually reannotated with interfacial helices and 98 with re-entrant regions. To this the authors added 9,660 α-helical and 5,316 β-barrel proteins with topology transferred by structural alignment, plus 13,277 UniProt membrane proteins with known location but unknown topology for the localization head. The shipped predictor ensembles the five cross-validation folds, averaging probabilities for membrane type and taking the highest-agreement topology. It reaches 85.6% and 76.1% correct topologies for α-helical and β-barrel proteins, comparable to TMbed and its own predecessor DeepTMHMM despite the larger label set, and processes the human proteome in roughly 65 minutes on an RTX 8000.

#Applications

Topology prediction remains the fastest route to a protein's orientation in a membrane, and DeepTMHMM2 fits both as a first-pass step ahead of structure prediction and as a complementary track alongside it. It tells an antibody engineer which loops of a multipass target are extracellular and therefore reachable, gives a channel biologist a candidate selectivity-filter re-entrant loop, and lets a proteomics group annotate an entire organism's membrane proteome overnight. It runs from a pip-installable command-line tool and Python API, or through a hosted web server.

#Impact

The Swiss-Prot sweep is the substantive finding: interfacial helices appear in 24.9% of predicted α-helical transmembrane proteins and re-entrant regions in 3.8%, with similar rates in the human proteome — features previously treated as curiosities turn out to be routine. Gene Ontology enrichment ties re-entrant regions to ion transport and gated channel activity, and interfacial helices to GPCRs and oxidoreduction-driven transporters, while pathogenic ClinVar variants are enriched in both relative to soluble regions. The authors state the limits plainly: recall is modest, at 61% and 64% of containing proteins receiving any prediction at all, and organelle membrane classes with few positive training examples — Golgi, vacuole, lysosome, nuclear inner — perform poorly. The work is a preprint awaiting peer review.