Deep Learning Proteins using a Triplet-BERT network.

Area of Science:

• Proteomics
• Computational Biology
• Machine Learning

Background:

• Vast proteomic data fuels deep learning in biology.
• Challenges remain in modeling protein properties with limited labeled data.
• Interpretable deep learning is crucial for understanding protein function.

Purpose of the Study:

• To leverage a pre-trained BERT model for protein property prediction using minimal data.
• To fine-tune the BERT model using a triplet network structure.
• To evaluate the model's performance on diverse downstream tasks.

Main Methods:

• Utilized a BERT model pre-trained on extensive proteomic data.
• Employed a triplet network to fine-tune the BERT model for specific regression tasks.
• Evaluated performance on plasma membrane localization, thermostability, peak absorption wavelength, and enantioselectivity prediction.

Main Results:

• Achieved significant performance improvements over baseline BERT and previous state-of-the-art models.
• Demonstrated the efficacy of triplet network fine-tuning on limited datasets.
• Visualized model attention to identify critical protein regions and modifications.

Conclusions:

• The fine-tuned BERT model with a triplet network effectively predicts protein properties from limited data.
• This approach advances interpretable deep learning in proteomics.
• The method offers insights into protein function and modification impacts.