DisorderUnetLM: Validating ProteinUnet for efficient protein intrinsic disorder prediction

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Predicting intrinsic disorder regions in proteins is crucial for understanding protein function, dynamics, and interactions.
• Novel protein-protein interactions are key targets for drug design and enzyme engineering.
• Emerging protein language models (pLMs) offer accurate predictions without computationally intensive multiple sequence alignments (MSAs).

Purpose of the Study:

• To introduce DisorderUnetLM, a novel predictor for intrinsic disorder regions in proteins.
• To evaluate DisorderUnetLM's performance against existing state-of-the-art predictors, including those using MSAs and pLMs.
• To provide a publicly available and reproducible tool for disorder prediction.

Main Methods:

• DisorderUnetLM utilizes an Attention U-Net convolutional network architecture.
• The model incorporates features derived from the ProtTrans protein language model (pLM).
• Performance was assessed on the CAID-2 benchmark, comparing against other predictors.

Main Results:

• DisorderUnetLM achieved state-of-the-art accuracy, outperforming predictors that rely on MSAs and other pLMs.
• On the CAID-2 benchmark, it ranked 1st for the NOX subset (ROC-AUC: 0.844) and 2nd for the AP metric (0.596).
• For the CAID-2 PDB subset, it ranked in the top 10, with a ROC-AUC of 0.924 and AP of 0.862.

Conclusions:

• DisorderUnetLM represents a significant advancement in predicting intrinsic disorder regions.
• The model demonstrates superior performance compared to existing methods, particularly on challenging benchmark datasets.
• The public availability of the code and model facilitates further research and application in protein science.