DivPro: diverse protein sequence design with direct structure recovery guidance.

Area of Science:

• Computational biology
• Protein engineering
• Bioinformatics

Background:

• Structure-based protein design aims to create novel proteins with desired structures and functions.
• Current deep learning methods often rely on sequence recovery metrics, overlooking sequence-structure ambiguity and limiting diversity.
• This limitation is particularly evident in designing remote homologous proteins.

Purpose of the Study:

• To develop a model, DivPro, that generates diverse protein sequences capable of folding into specified structures.
• To address the limitations of existing methods in producing varied sequences while preserving structural integrity.

Main Methods:

• DivPro learns a probabilistic sequence space, enabling sampling of diverse sequences for a given structure.
• It incorporates in silico protein structure prediction results for training guidance, ensuring sequence-structure reliability.
• The model was evaluated on three sequence design benchmarks using structure prediction tools like AlphaFold2.

Main Results:

• DivPro significantly enhances sequence diversity compared to existing methods.
• The model maintains high structure recovery, ensuring designed sequences fold into the target structure.
• Experimental results validate DivPro's effectiveness across multiple benchmarks.

Conclusions:

• DivPro offers an advancement in protein design by balancing sequence diversity and structural accuracy.
• The probabilistic approach overcomes limitations of fixed sequence representations in deep learning models.
• This method holds promise for designing proteins with novel functions and improved properties.