PLTNUM

Protein lifetimes inside a cell span orders of magnitude: cell-cycle regulators are cleared in minutes, histones persist for days. Much of that spread is written into the sequence as degrons — motifs, often fewer than ten residues, that recruit the ubiquitin-proteasome machinery. The sequence-based predictors available for this question mostly answer an adjacent one. DEGRONOPEDIA scores a stability index derived from terminal-peptide experiments; TemStaPro and DeepSTABp predict thermal melting points. Melting temperature in a tube and turnover in a living cell are different properties, and applied to measured half-life data all three land close to chance.

PLTNUM — Protein Lifetime Neural Model — takes the direct route: fine-tune a pretrained protein language model on proteome-wide half-lives measured by mass spectrometry. It was built by Tatsuya Sagawa and colleagues at Kyoto University, the National Institutes of Biomedical Innovation, Health and Nutrition and RIKEN IMS, and posted as a preprint in September 2024.

The backbone choice is what makes structure available at no inference cost. PLTNUM fine-tunes SaProt AF 650M, whose vocabulary pairs each residue with a Foldseek structural state read off an AlphaFold 2 model, so 3D context arrives as tokens rather than through a folding step at prediction time. The paper's ablation makes the case that this matters: a two-layer LSTM and an un-pretrained ESM-2 650M perform about equally poorly, pretrained ESM-2 improves on both, and the structure-aware backbone improves again.

#Key Features

• Half-life as the training target: The model is fit against SILAC-measured turnover rates rather than melting temperature or a terminal-peptide proxy, which is why it separates short- from long-lived proteins where thermal-stability predictors do not.
• Structure without folding at inference: Structural information enters through Foldseek tokens precomputed from AlphaFold models, so prediction needs only a tokenized sequence.
• Unbiased degron nomination: SHAP values over the input tokens rank the amino acids and positions driving each prediction. Because training used half-lives rather than a curated degron list, candidate motifs are not restricted to previously reported ones.
• Structure tokens are interpretable too: The Foldseek states that most shorten predicted half-life concentrate in intrinsically disordered surface regions, matching independent reports.
• Checkpoints for two cell lines and two backbones: SaProt- and ESM-2-based weights trained on the mouse NIH3T3 and human HeLa datasets are released alongside the training and SHAP code.

#Technical Details

The training set is the NIH3T3 mouse embryo fibroblast half-life dataset of Schwanhäusser et al., restricted to the 4,162 proteins with an entry in the AlphaFold Protein Structure Database. Labels are binary — short- or long-lived relative to the dataset median — and the head is a dropout plus fully connected layer over the backbone's 1,280-dimensional <cls> embedding, trained with binary cross-entropy under 10-fold cross-validation for 10 epochs at a learning rate of 2e-5. Regularization is aggressive: dropout of 0.8, multi-sample dropout, and augmentation that masks 5% of tokens with 20% probability. Accuracy rises with the N-terminal window supplied and plateaus near 500 residues, so inputs are fixed at 510, giving roughly 71% accuracy and an ROC AUC of 0.737 on held-out folds. Applied without refitting to a HeLa half-life dataset from a different species, cell line and laboratory, one cross-validation checkpoint reached AUC 0.729, against 0.472 for DEGRONOPEDIA, 0.543 for TemStaPro and 0.565 for DeepSTABp. SHAP analysis identifies cysteine as the largest contributor to shortened half-life and shows contribution magnitude rising toward the N-terminus, consistent with the N-end rule. Eight of the ten top-ranked candidate degron motifs were carried by proteins with shorter measured half-lives than composition-matched controls holding the same residues in a different order.

#Applications

The model is most useful as a triage step in degradation-focused work: ranking which proteins in a proteome are likely to turn over quickly, screening designed or mutant sequences before committing to a pulse-SILAC experiment, and generating degron candidates for targeted-degradation programs that would otherwise start from a small set of literature motifs. Clustering the learned embeddings recovers groupings that align with localization and function — mitochondrial, nuclear, cell-cycle and PTM-related sets — so the representation carries context beyond raw composition. A hosted demo returns a prediction for a pasted sequence.

#Impact

PLTNUM is a narrow artifact rather than a general protein model: the output is one binary call thresholded at a dataset median, not a numeric half-life, and the authors are explicit that accuracy stalls near 71% because half-life depends on biological context — organelle, cell type, partner availability — that sequence cannot supply. They recommend it as a filter rather than an oracle, and the gap between prediction and measurement is itself a signal for which proteins are context-governed. Its more transferable contribution is methodological: a structure-aware protein language model, fine-tuned on a few thousand labeled proteins and read through SHAP, can surface sequence-level determinants of degradation without being told what a degron looks like.