SETH

SETH predicts per-residue continuous intrinsic disorder in protein sequences by using ProtT5 protein language model embeddings and a shallow convolutional neural network trained on CheZOD scores.

Key Features:

• ProtT5 embeddings: Uses embeddings generated by the ProtT5 protein language model as input representations for residues.
• MSA-free single-sequence operation: Operates solely on single sequences and does not require multiple sequence alignments (MSAs).
• Convolutional neural network: Processes ProtT5 embeddings with a relatively shallow convolutional neural network architecture.
• Continuous CheZOD prediction: Predicts residue-level continuous disorder values on the CheZOD scale rather than binary disorder assignments.
• IDR sensitivity: Captures nuanced variations in intrinsically disordered regions (IDRs) across residues.
• Proteome-scale performance: Capable of generating proteome-wide predictions in approximately one hour on a consumer-grade PC with an NVIDIA GeForce RTX 3060 GPU.
• AlphaFold2 quality indication: Identifies regions or proteins likely to correspond to low-quality 3D structure predictions from AlphaFold2.

Scientific Applications:

• Evolutionary analyses: Uses detailed disorder variation to provide insights into evolutionary differences between organisms and proteins.
• Structural prediction prioritization: Assists in quality filtering and prioritization of AlphaFold2 3D structure predictions by flagging disordered regions.
• Disease-related studies: Supports investigation of diseases associated with protein disorder, including Alzheimer's Disease, by mapping residue-level disorder.

Methodology:

Generates ProtT5 embeddings from single protein sequences, processes them with a relatively shallow convolutional neural network trained on CheZOD-labelled data, and outputs per-residue continuous CheZOD disorder scores in an MSA-free workflow; reported proteome-scale runtimes used an NVIDIA GeForce RTX 3060 GPU.