ProtRAP-LM

Membrane proteins make up roughly a fifth to a third of most proteomes and are the targets of more than half of known drugs, yet they remain hard to identify computationally. Most existing tools work by predicting transmembrane spans, which systematically misses membrane-anchored proteins, peripheral proteins, and proteins that touch the lipid bilayer without crossing it. ProtRAP-LM takes a different route: it predicts, for every residue, the probability of direct contact with the hydrophobic acyl chains of lipids — a quantity its authors call membrane contact probability (MCP) — together with relative solvent-accessible surface area.

ProtRAP-LM was developed by the Chen Song group at Peking University and posted as a preprint in January 2025, with the peer-reviewed version appearing in Genomics, Proteomics & Bioinformatics. It is the third model in a line that began with an MCP predictor and continued with ProtRAP, both of which required a multiple sequence alignment and a position-specific scoring matrix as input. Building the alignment was the bottleneck: it dominated runtime and capped analysis at a few hundred proteins. ProtRAP-LM removes it entirely by feeding embeddings from ESM-2 into a lightweight supervised head, cutting per-sequence cost from 32.75 seconds to 0.10 seconds.

That 300-fold speedup is what makes the model interesting beyond its accuracy: the authors used it to annotate the proteomes of 48 organisms spanning five kingdoms, a scale out of reach for the alignment-based predecessors.

#Key Features

• Membrane contact rather than transmembrane topology: Predicting which residues touch lipid, instead of where a membrane-spanning helix starts and ends, lets the model flag membrane-anchored and tail-anchored proteins that span-based tools cannot represent.
• Five accessibility outputs in one run: A single forward pass yields relative accessible surface area and MCP directly, from which relative lipid accessibility, relative solvent accessibility, and relative buried surface area are derived.
• No alignment at inference: Dropping the MSA requirement is what enables proteome-scale screening on a single workstation GPU.
• Ten-model ensemble with uncertainty: Predictions are averaged across ten models trained on the folds of a 10-fold cross-validation split, and the per-residue standard deviation is reported alongside the mean.
• Generalizes to unseen protein classes: Intermembrane lipid transfer proteins such as Atg2A, TamB, and YebT were absent from the training set, yet the model correctly localizes their lipid-contacting residues.

#Technical Details

Input sequences are encoded by the 650M-parameter ESM-2 model into 1,280-dimensional per-residue vectors. These pass through a linear projection, a bank of one-dimensional convolutions with kernel sizes 3, 5, 7, and 9 whose outputs are concatenated and added back to the input, then a three-layer transformer encoder with four attention heads, and finally a sigmoid output layer. Training used 1,362 membrane protein chains from MemProtMD (filtered below 40% identity and 1,000 residues) plus 7,440 soluble chains below 25% identity, with RASA labels computed by FreeSASA at a 1.4 Å probe radius and soluble-protein MCP set to zero. Optimization used AdamW-family AMSGrad at a 1e-4 learning rate, batch size 32, and mean squared error loss over 10 warm-up and 20 normal epochs.

On the 184-protein MemProtMD_2022 held-out set, residue-level Pearson correlations reach 0.915 for MCP, 0.782 for RASA, 0.863 for relative lipid accessibility, and 0.803 for relative solvent accessibility — improvements of 0.053, 0.068, 0.023, and 0.056 over the MSA-based predecessors, and 0.089 over NetSurfP-3.0 on RASA. Across the 48 proteomes, ProtRAP-LM predicted 26.5% of human proteins as membrane-associated versus 24.2% for the span-based tool TMbed, while still recovering 99.4% of TMbed's own calls and 96.9% of annotated human single-pass transmembrane proteins.

#Applications

Structural biologists and drug discovery teams use ProtRAP-LM to triage which proteins in a newly sequenced proteome are worth pursuing as membrane targets, and to locate the specific residues likely to sit in the bilayer before committing to expression and purification. Because it scores lipid and solvent accessibility separately, it is also useful for choosing construct boundaries, designing detergent or nanodisc conditions, and interpreting predicted structures where the membrane position is unknown. A public web server accepts single sequences for groups without local GPU access, and the released ensemble runs on CPU for small jobs.

#Impact

The model's proteome-wide screen produced concrete leads: 78 human proteins flagged as likely membrane-associated that are absent from UniProt membrane annotations, unidentified by TMbed and SignalP, and lacking membrane-related Gene Ontology terms. Two of the highest-scoring, PLIN1 and PLIN5, have independent experimental evidence of membrane localization, which supports the list as a genuine discovery set rather than a noise floor. The code, the ten-model ensemble, and the per-organism membrane protein lists are released under GPL-3.0. The main constraint is scope: ProtRAP-LM is a per-residue accessibility regressor, not a topology predictor, so it reports where a protein contacts lipid without saying how many times it crosses the membrane or in which orientation.