Protein language-model embeddings for fast, accurate, and alignment-free protein structure prediction

Area of Science:

• Computational biology
• Structural bioinformatics
• Artificial intelligence in protein science

Background:

• Advanced protein structure prediction often relies on multiple sequence alignments (MSAs) for evolutionary information, which are not always available.
• Current artificial intelligence (AI) methods using single sequences lack the accuracy needed for practical applications.

Purpose of the Study:

• To develop a novel, accurate, and computationally efficient method for protein structure prediction using only single sequences.
• To leverage pre-trained protein language models (pLMs) and convolutional neural networks (CNNs) for predicting inter-residue distances.

Main Methods:

• Utilized embeddings from the pre-trained protein language model (pLM) ProtT5, focusing on its attention heads.
• Inputted these embeddings into a shallow convolutional neural network (CNN) to predict 2D protein structure (inter-residue distances).
• Developed a new method named EMBER2, which does not require MSAs or co-evolutionary data.

Main Results:

• EMBER2 achieved performance comparable to methods that rely heavily on co-evolutionary data.
• The method demonstrated competitive accuracy, approaching AlphaFold2's performance at a significantly lower computational cost.
• EMBER2's protein-specific predictions may better capture unique structural features compared to family-averaged predictions.

Conclusions:

• EMBER2 provides a viable, cost-effective alternative for protein structure prediction when MSAs are unavailable.
• The study validates the potential of using pLM embeddings with CNNs for accurate, single-sequence-based structure prediction.
• Protein engineering and deep mutational scanning experiments support the capability of EMBER2 to capture specific protein structural characteristics.