Convolutional ProteinUnetLM competitive with long short-term memory-based protein secondary structure predictors

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Protein secondary structure (SS) prediction is crucial for understanding protein structure and function.
• Emerging protein language models (pLMs) offer accurate SS prediction without multiple sequence alignment (MSA).
• Existing pLM-based predictors like SPOT-1D-LM and NetSurfP-3.0 represent the state-of-the-art.

Purpose of the Study:

• To introduce ProteinUnetLM, a novel model for 8-class protein secondary structure (SS8) prediction.
• To evaluate ProteinUnetLM's performance against existing state-of-the-art methods, including LSTM-based predictors and AlphaFold2.
• To address the challenge of imbalanced SS8 datasets using advanced loss functions and evaluation metrics.

Main Methods:

• Developed ProteinUnetLM, a convolutional Attention U-Net architecture for SS prediction.
• Extended the loss function with the Matthews correlation coefficient to handle imbalanced SS8 data.
• Utilized the adjusted geometric mean (AGM) metric for robust performance assessment.
• Compared ProteinUnetLM against LSTM-based models (SPOT-1D-LM, NetSurfP-3.0) and AlphaFold2 on various datasets.

Main Results:

• ProteinUnetLM achieved prediction quality and inference times comparable to or better than leading LSTM-based models for SS8 prediction.
• The model demonstrated superior performance in adjusted geometric mean (AGM) and sequence overlap scores, particularly for rare structures (310-helix, beta-bridge, high curvature loop).
• ProteinUnetLM showed competitive results on challenging datasets, including those without homologs and free-modeling targets, and outperformed its MSA-based predecessor, ProteinUnet2.
• ProteinUnetLM provided better AGM than AlphaFold2 for a significant portion (1/3) of CASP14 proteins.

Conclusions:

• ProteinUnetLM represents a significant advancement in protein secondary structure prediction, offering high accuracy and efficiency.
• The model's ability to handle imbalanced data and rare structures makes it a valuable tool for protein scientists.
• ProteinUnetLM demonstrates competitive performance against leading structure prediction tools like AlphaFold2, highlighting its potential impact on the field.