Bi-LSTM/Bi-GRU Amyloidogenicity Predictor

Whether a peptide will nucleate an amyloid fibril is determined largely by its sequence, and short six-residue stretches act as the hot spots that initiate aggregation. Predicting amyloidogenicity therefore matters both for understanding diseases of protein misfolding and for engineering biologics that must stay soluble. Most established predictors — AmyloGram, PASTA 2.0, APPNN, ReRF-Pred, ANuPP — derive their features from evolutionary motifs, amino acid physicochemical properties, or hand-designed compositional descriptors such as pseudo amino acid composition and tripeptide counts.

This model, from Zohra Yagoub and Hafida Bouziane at the Université des Sciences et de la Technologie d'Oran Mohamed Boudiaf, replaces that feature engineering with contextual representations from a protein language model. Per-sequence embeddings from a frozen ESM-2 checkpoint are fed to a small recurrent classifier — a bidirectional LSTM followed by a bidirectional GRU — trained once on WaltzDB 2.0 and then applied without retraining to two independent external benchmarks. The result is a compact demonstration that language model representations alone, with no additional descriptors, are competitive with purpose-built amyloid predictors on both peptide classification and whole-protein region annotation.

#Key Features

• Frozen language model features: 1,280-dimensional embeddings are obtained by averaging per-residue representations from the 650M-parameter, 33-layer ESM-2 model pretrained on UniRef, so no protein language model fine-tuning is required.
• Bidirectional recurrent head: A Bi-LSTM followed by a Bi-GRU, each with dropout, reads the embedding in both directions before a single sigmoid unit performs binary classification.
• Sliding-window protein scanning: Whole protein sequences are segmented into overlapping hexapeptides, scored, and the predictions redistributed to individual residues and averaged, which turns a peptide classifier into an aggregation-prone region annotator.
• Balanced operating point: The design targets sensitivity and specificity jointly rather than optimizing one at the other's expense, a failure mode visible in several comparator tools.

#Technical Details

Training used WaltzDB 2.0: 515 amyloidogenic and 901 non-amyloidogenic hexapeptides, split 80/20 into a training and cross-validation set and a held-out test set. The model achieved 84.5% accuracy under 10-fold cross-validation with sensitivity 74.5% and specificity 89.7%, and 83% accuracy on the test split with sensitivity 76%, specificity 87%, F1 0.77, and MCC 0.64 — the best MCC and F1 among nine classifiers the authors trained on the same embeddings, ahead of ExtraTrees, LightGBM, random forest, SVM, KNN, logistic regression, AdaBoost, and an MLP, which spanned 80% to 83% accuracy. On Pep-251, the 251-peptide set introduced with ReRF-Pred and evaluated here without retraining, a peptide is called amyloidogenic if any constituent hexapeptide is predicted positive; the model reaches accuracy 0.808, sensitivity 0.734, specificity 0.843, and MCC 0.566, against 0.801/0.552 accuracy and MCC for ReRF-Pred, 0.784/0.555 for AmyloGram, 0.769/0.542 for APPNN, and 0.733/0.440 for ANuPP. PASTA 2.0 posts a higher MCC (0.603) but at sensitivity 0.506. On AmyPro27, 27 full-length amyloid proteins with annotated aggregation-prone regions, the model scores segment overlap of 53.1 on aggregation-prone regions and 54.1 on non-aggregation-prone regions, an average of 53.6, and misclassifies only one protein as non-amyloid; ANuPP averages higher (56.7) but calls four proteins non-amyloid, and Waltz shows a 26.7 versus 61.0 split between the two region types.

#Applications

The predictor is aimed at researchers screening peptide libraries or protein sequences for aggregation liability: identifying the amyloidogenic hot spots in a disease-associated protein, flagging aggregation-prone stretches in a therapeutic protein or antibody before expression, or prioritizing candidates for the thioflavin-T and electron microscopy assays that ultimately confirm fibril formation. Because it consumes only sequence and needs no structure, alignment, or evolutionary profile, it can be run across a proteome cheaply.

#Impact

The work's significance is methodological rather than one of scale: it shows that a few-layer recurrent head on frozen protein language model embeddings matches specialized amyloid predictors built on decades of biophysical feature engineering, on both a peptide benchmark and a whole-protein region benchmark it never saw during training. The scope is narrow — a single institution, a 1,426-hexapeptide training set, and evaluation limited to the authors' own comparison against six published tools — and the balanced-prediction claim rests on aggregate metrics over 27 proteins in the region-annotation experiment. No code or model weights accompany the paper, so the reported numbers cannot be independently reproduced. The study appeared as a conference proceedings chapter at AISTC 2025 under a CC BY-NC 4.0 license and as an arXiv preprint.