DeepPFP: a multi-task-aware architecture for protein function prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Predicting protein function from sequence is challenging due to complex sequence-function relationships.
• Deep learning models struggle with transfer learning across different protein types and tasks.
• Protein function is influenced by structural features, not just sequence information, necessitating models that capture shared characteristics.

Purpose of the Study:

• To develop a generalized model for multi-task protein function prediction.
• To address the limitations of domain-specific models in transfer learning.
• To improve the ability of models to capture shared features across diverse sequence-function mapping tasks.

Main Methods:

• Utilized Model-Agnostic Meta-Learning integrated with the Evolutionary Scale Modeling protein language model.
• Trained the architecture on five out-of-domain deep mutational scanning (DMS) datasets.
• Evaluated performance across four key dimensions, focusing on generalization and few-shot learning capabilities.

Main Results:

• The proposed architecture demonstrated satisfactory generalization performance and an effective few-shot learning strategy.
• Achieved an approximate 0.31% increase in Pearson's correlation coefficient (PCC) compared to baseline results.
• Successfully predicted SARS-CoV-2 binding affinity scores using transfer learning, with a notable 0.11 PCC improvement on a subset of the Ube4b dataset.

Conclusions:

• The developed conceptual architecture shows significant promise for multi-task protein function prediction.
• The model's ability to generalize and perform few-shot learning offers a robust solution for diverse biological tasks.
• This approach advances the field of predicting protein function by leveraging meta-learning and advanced language models.