MHTAPred-SS: A Highly Targeted Autoencoder-Driven Deep Multi-Task Learning Framework for Accurate Protein Secondary...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Accurate protein secondary structure prediction (PSSP) is vital for biopharmaceutics and disease diagnosis.
• Current PSSP methods struggle with feature space utilization and performance degradation when homologous proteins are scarce.
• Limitations exist in feature extraction and learning strategies of existing stacked network architectures.

Purpose of the Study:

• To develop a novel PSSP framework, MHTAPred-SS, that overcomes limitations of current methods.
• To improve feature representation and prediction accuracy through a multi-feature fusion approach.
• To enable robust PSSP independent of homologous protein availability.

Main Methods:

• Proposed MHTAPred-SS framework integrates six distinct features, including embeddings from a pre-trained protein language model.
• Introduced a highly targeted autoencoder (HTA) for homologous protein-independent sequence encoding.
• Developed a protein secondary structure prediction model utilizing multi-task learning (PSSP-MTL) guided by biological knowledge.

Main Results:

• MHTAPred-SS achieved state-of-the-art performance on six independent test sets, including the TEST2016 dataset.
• Achieved high Q3 (88.14%), SOV3 (84.89%), Q8 (78.74%), and SOV8 (77.15%) metrics on TEST2016.
• Demonstrated significant advantages in predicting single-category and boundary secondary structures, finely capturing segment distributions.

Conclusions:

• MHTAPred-SS offers a significant advancement in protein secondary structure prediction accuracy and robustness.
• The framework's ability to perform well with limited homologous sequences broadens its applicability.
• Improved secondary structure prediction contributes to enhanced downstream tasks in structural biology and drug discovery.