PTransIPs

PTransIPs predicts phosphorylation sites by integrating ProtTrans and EMBER2 pre-trained language model embeddings into a Transformer-CNN architecture trained with a TIM loss to improve identification of serine/threonine and tyrosine phosphorylation sites.

Key Features:

• Pre-trained language model embeddings: Uses ProtTrans for sequence embeddings and EMBER2 for structure embeddings to capture sequence and structural context.
• Transformer-CNN architecture: Combines a Transformer backbone with convolutional neural networks for enhanced feature extraction.
• TIM loss function: Employs a TIM loss during training to refine model predictions.
• Amino acid encoding for universality: Implements an amino-acid encoding that enables application to diverse peptide bioactivity tasks and addresses dataset size and overfitting challenges.

Scientific Applications:

• Phosphorylation site prediction: Predicts serine/threonine (S/T) and tyrosine (Y) phosphorylation sites with reported AUCs of 0.9232 (S/T) and 0.9660 (Y).
• Cellular signaling and disease mechanism studies: Facilitates identification of phosphorylation events relevant to cellular signaling pathways and disease mechanisms.
• Viral–host interaction research: Supports investigation of phosphorylation changes during viral infections to inform studies of viral-host interactions.
• Peptide bioactivity tasks: Serves as a universal framework applicable to broader peptide bioactivity prediction tasks via its amino-acid encoding.

Methodology:

Integrates ProtTrans and EMBER2 PLM embeddings and an amino-acid encoding, inputs them to a Transformer augmented with CNN layers, and trains the model using a TIM loss function.