ToxinPred3

ToxinPred3 predicts peptide toxicity using computational models to distinguish toxic and non-toxic peptide sequences and to support safer therapeutic peptide design.

Key Features:

• Similarity/Alignment-Based Approach: Uses BLAST for sequence similarity and alignment-based toxicity inference, with coverage limitations addressed by integrating additional approaches.
• Motif-Based Approach: Identifies toxic-sequence-specific motifs using MERCI, providing high specificity for motif-driven toxicity signals.
• Deep Learning (ANN + LSTM): Implements an artificial neural network combined with LSTM layers, achieving a maximum AUROC of 0.93 and an MCC of 0.71.
• Extra Tree Machine Learning: Employs an extra tree model built on compositional peptide features, attaining an AUROC of 0.95 and an MCC of 0.78.
• Large Language Models: Applies large language model embeddings with ESM2-t33, reporting an AUC of 0.93.
• Hybrid/Ensemble Methods: Combines motif-based, alignment-based, and machine/deep learning models in ensemble frameworks, achieving up to AUROC 0.98 and MCC 0.81 on independent datasets.

Scientific Applications:

• Therapeutic Peptide Design: Predicts toxicity to support the design of non-toxic therapeutic peptides.
• Early Toxicity Screening: Enables early prediction of potential peptide toxicity during the development pipeline to reduce late-stage failures.
• Candidate Refinement and Prioritization: Aids refinement and prioritization of peptide candidates to lower the risk of failure in clinical development.

Methodology:

Computational methods explicitly include BLAST-based similarity/alignment, MERCI motif identification, alignment-free models comprising ANN+LSTM and an extra tree model using compositional features, large language model embeddings with ESM2-t33, hybrid/ensemble combinations of these approaches, and model training and testing using five-fold cross-validation on 80% of the dataset with evaluation on the remaining 20% as an independent test set.