PPLN

Two peptide ions can share an m/z and still be told apart in a trapped ion mobility cell, because they drift at different speeds when their gas-phase shapes differ. That separation dimension only helps identification if the collision cross section (CCS) of a candidate sequence can be predicted and compared against the measured value. The deep-learning predictors that do this well are sequence models trained from scratch on hundreds of thousands of ions, and the available corpora are dominated by doubly charged tryptic peptides under 30 residues — so longer, more highly charged peptides are exactly the population those models handle worst.

PPLN — the pretrained protein language model-based network — attacks that from the training side rather than the architecture side. Instead of learning peptide representations from CCS labels, it borrows them: a peptide passes once through a frozen, general-purpose protein language model, the per-residue features are collapsed into a fixed-size vector, and only a small feed-forward regressor is fitted on top, with the ion's charge and mass as extra inputs. The bet is that a model trained on a large protein sequence database already encodes enough about local structure to explain what determines an ion's cross section.

The work comes from Kyoto University, the Nagoya Institute of Technology and the National Institutes of Biomedical Innovation, Health and Nutrition, posted to bioRxiv in September 2024 and published in Communications Chemistry in May 2025. The feature extractor used is ESM-1b, though the architecture treats the backbone as interchangeable.

#Key Features

• Frozen backbone, trained head: the language model is used purely as a feature extractor with its output layer removed. Of roughly 661M parameters, 650M are frozen and about 11M are fitted.
• Bidirectional positional encoding: rather than averaging residue features, PPLN builds two encodings — one indexed from the N-terminal, one from the C-terminal — and concatenates them, on the reasoning that terminal residues dominate an ion's shape.
• Charge and mass as explicit inputs: withholding charge is the single most damaging ablation, raising RMSE from 15.1 to 31.8 Ų.
• A dataset built for hard peptides: 91,677 unique peptide ions measured in-house and deliberately biased toward long, multiply charged species.

#Technical Details

ESM-1b emits 1280-dimensional features per residue. PPLN multiplies these element-wise by a sinusoidal position encoding, averages the N-terminal and C-terminal halves separately, and concatenates the two summaries; charge and mass are appended and fed to a 10-layer fully connected network of width 1000 with a scalar output, trained with Adam (learning rate 0.0003), MSE loss, batch size 200 and 400 epochs.

The corpus was produced by digesting HeLa lysate with seven proteases (trypsin, LysargiNase, Lys-C, Lys-N, Glu-C, Asp-N, chymotrypsin), enriching phosphopeptides, then dephosphorylating them to yield unmodified peptides with more missed cleavages; CCS values came from timsTOF Pro PASEF runs. The 91,677 ions span 289–1162 Ų and m/z 381–1798, with 11% singly, 57% doubly, 25% triply and 7% quadruply charged, split 73,342 for training and 18,335 for testing. PPLN reached Pearson r = 0.9930 and RMSE 15.12 Ų there, against 0.9925 / 15.64 Ų for a bidirectional LSTM trained from scratch on the same data and 0.9857 / 23.73 Ų for length-specific multiple linear regression; over ten random splits the authors report PPLN as competitive with the LSTM rather than uniformly better. The clear win is cost: head training took 1/78 to 1/18 of the LSTM's time, or 1/4 to 1/3 once the 0.058 s-per-ion feature extraction is included, and test-set inference took 0.5 s against 44.0 s.

#Applications

The demonstration application is rescoring in peptide identification. A PPLN trained on the HeLa dataset was applied unchanged to tryptic peptides from E. coli K12 BW25113, and the difference between predicted and measured CCS was added as a feature to a Mascot/Percolator search. At 1% FDR that raised peptide-spectrum matches from 84,580 to 99,463 and stripped sequences from 8,963 to 9,797. Because the fitted component is small, the same recipe suits a proteomics group wanting a CCS predictor calibrated to its own instrument and digestion protocol without a GPU cluster, and the authors argue it should transfer to retention time, fragmentation pattern and detectability prediction.

#Impact

PPLN is a compact demonstration that a general protein language model carries usable structural signal about short peptides, not just folded domains, and that a small head can convert it into a competitive physical-property predictor at a fraction of the training cost. The limits are stated plainly by the authors: the model cannot handle chemically or post-translationally modified residues, because the underlying language model has no tokens for them, and feature extraction becomes the bottleneck at scale. Larger backbones such as ESM-2 and xTrimoPGLM are proposed as substitutes that would trade preprocessing time for accuracy. The code is released under the MIT license with the trained checkpoint included, and the mass spectrometry data are deposited in ProteomeXchange.