UniDL4BioPep

UniDL4BioPep performs binary classification of bioactive peptides using pretrained biological language models and convolutional neural networks to predict peptide bioactivities for peptide discovery and prioritization.

Key Features:

• Transfer Learning Framework: Implements a universal deep-learning architecture that enables transfer learning for bioactive peptide binary classification with a fixed architecture.
• Pretrained Biological Language Models (LMs): Uses pretrained deep learning–based LMs for protein/peptide sequence embedding to generate informative peptide embeddings.
• Convolutional Neural Network (CNN) Integration: Integrates LM-derived embeddings with a CNN to improve predictive performance on bioactivity tasks.
• Performance Metrics: Demonstrated improvements of 0.7–7% in accuracy, 1.23–26.7% in Matthews correlation coefficient (MCC), and 0.3–25.6% in area under the curve (AUC) over existing models in 15 out of 20 bioactivity dataset prediction tasks.
• Validation: Validated model representations and robustness using uniform manifold approximation and projection (UMAP) analysis.

Scientific Applications:

• Bioactive Peptide Discovery: Predicts peptide bioactivities to prioritize candidate peptides for experimental follow-up.
• In silico Screening: Facilitates high-throughput computational screening of peptide libraries to reduce the scale of wet-lab assays.
• Therapeutic Lead Identification: Assists in identifying peptide candidates with potential therapeutic applications based on predicted bioactivity.

Methodology:

Peptide sequences are embedded using pretrained biological LMs, these embeddings are input to a CNN within a transfer learning framework for binary classification, and UMAP is used for representation validation.