Multitask Contrastive Learning with Attention Mechanisms for Neuropeptide Prediction Using ESM Representations.

Area of Science:

• Biochemistry
• Computational Biology
• Neuroscience

Background:

• Neuropeptides are vital signaling molecules regulating physiological and cognitive functions.
• Accurate neuropeptide identification from sequence is difficult due to high diversity and weak motifs.
• Current computational methods lack robustness and generalization for neuropeptide prediction.

Purpose of the Study:

• To develop a robust and generalizable computational framework for neuropeptide identification.
• To improve the accuracy and reliability of predicting neuropeptides from primary sequences.
• To advance neuropeptide discovery for therapeutic and drug development.

Main Methods:

• Integrated ESM-derived protein representations with a BiLSTM encoder and multihead self-attention.
• Employed attention-based pooling, knowledge distillation, and contrastive representation learning.
• Utilized a unified multitask learning framework (NeuroPred-MTCL) for enhanced generalization.

Main Results:

• NeuroPred-MTCL achieved 93.6% accuracy and 0.977 AUROC on an independent test set.
• Demonstrated a balanced precision (92.9%) and recall (94.4%), resulting in a 0.936 F1-score.
• Effectively captured discriminative sequence characteristics, enhancing neuropeptide identification reliability.

Conclusions:

• NeuroPred-MTCL offers a robust and generalizable approach for computational neuropeptide identification.
• The method significantly advances the ability to discover and analyze neuropeptides.
• This work facilitates the elucidation of neural communication and peptide-based therapeutic development.