PLPTP

Curated peptide toxicity datasets are lopsided: in the training data shipped with PLPTP, only about one peptide in three-and-a-half is toxic. A classifier trained on that split with ordinary cross-entropy can post a high accuracy while quietly failing at the only thing the task is for, because the cheapest way to be right most of the time is to call everything safe. Sensitivity on the minority class is where these predictors live or die, and a headline accuracy hides it.

PLPTP treats that imbalance as a modeling decision rather than a preprocessing one. It keeps the natural class ratio and swaps the loss function instead, using focal loss to shrink the gradient contribution of examples the model already gets right so that training attention stays on the hard, rare toxic sequences. The representation on top of which that loss operates comes from ESM-2: peptides are embedded by the pretrained 150M-parameter protein language model, a bidirectional LSTM re-reads that embedding sequence in both directions, and a small fully connected head makes the binary call. A separate motif analysis over the toxic and non-toxic sets identifies the recurring sequence patterns associated with toxicity, which the paper uses to argue that the classifier is keying on interpretable chemistry rather than dataset artifacts.

The model comes from Zilong Zhang's group at Hainan University, with collaborators at UESTC, Wuhan University of Science and Technology, Wuhan Textile University, Macao Polytechnic University and Xiamen University, and was published in Journal of Molecular Biology in 2025. It belongs to the same wave as NTxPred2 and AntiCP3, which likewise replaced hand-built composition descriptors with protein language model embeddings for a narrow peptide-property task.

#Key Features

• Fine-tuned language model backbone: The ESM-2 encoder is trained end to end with the classifier rather than frozen, so the 640-dimensional per-residue representations adapt to toxicity rather than staying generic.
• Bidirectional recurrence over embeddings: A two-layer BiLSTM passes over the embedded sequence in both directions, giving every residue both N- and C-terminal context before the classification head sees it.
• Focal loss instead of resampling: Down-weighting easy examples preserves the real class distribution, avoiding the information loss of undersampling and the overfitting risk of duplicating minority sequences.
• Motif-level interpretation: The paper reports motifs enriched in the toxic and non-toxic sets and uses them to argue that the classifier keys on recognizable sequence patterns; the analysis is a result in the write-up rather than code or output in the release.
• Sequence in, answer out: Prediction needs no structure, no alignment and no descriptor calculation — an amino acid string is the only input the model takes.

#Technical Details

The backbone is the 150M-parameter, 30-layer ESM-2 checkpoint (esm2_t30_150M_UR50D), producing 640-dimensional residue embeddings for sequences tokenized to a fixed length of 50 — comfortably above the length of the therapeutic peptides in scope. Those embeddings feed a two-layer bidirectional LSTM with 320 hidden units per direction and 0.3 dropout, then a three-layer fully connected head narrowing 640 to 128 to 64 to two logits. Training uses focal loss with α = 0.6 and γ = 2, Adam at a learning rate of 1e-4 with 5e-4 weight decay, batches of 128, and 40 epochs under five-fold cross-validation, with the best-scoring epoch of each fold saved as that fold's checkpoint. The released benchmark ships as two dataset variants, each pairing a training set of roughly 6,400 peptides — about 1,800 of them toxic — with an independent test set of 1,126. Performance is reported in the balanced metrics this literature uses: sensitivity, specificity, balanced accuracy, MCC and AUC, rather than raw accuracy.

#Applications

The intended user has a list of candidate peptides and a limited synthesis budget. Peptide therapeutics fail late and expensively on toxicity, and both designed sequences from generative models and natural sequences mined from venom or antimicrobial libraries need triage before they reach a bench. PLPTP is meant for that filtering step, but acting on that intent currently takes work: the prediction server the project page links to does not respond, and the repository ships training code only — train.py hardcodes a local ESM-2 path, does not parse as published, and has no inference script beside it. The motif analysis is where the value beyond ranking would sit, since a designer told which pattern drove a toxic prediction has something concrete to edit, but that analysis lives in the paper rather than in anything a user runs.

#Impact

Peptide toxicity prediction is a crowded field with a long tool lineage, and PLPTP's contribution is narrow but specific — it isolates class imbalance as the thing worth fixing, where contemporaries have reached instead for structural features or ensembles of language models. Its wider significance is as another data point that a mid-sized protein language model plus a light recurrent head is now a strong default for short-peptide property prediction, displacing the composition-descriptor pipelines that dominated the task for a decade. The practicalities temper that: the code is public but carries no license, and the trained checkpoint is distributed through a consumer cloud-drive link rather than a model hub.