Deep learning for NAD/NADP cofactor prediction and engineering using transformer attention analysis in enzymes

Area of Science:

• Biochemistry
• Bioengineering
• Computational Biology

Background:

• NAD(P)-dependent oxidoreductases are vital enzymes with broad natural distribution.
• Manipulating their cofactor preferences is crucial for bioengineering applications.
• Current methods for identifying cofactor preferences and designing mutants are complex and limited in scale.

Purpose of the Study:

• To develop a novel deep learning model for predicting the nicotinamide adenine dinucleotide (phosphate) [NAD(P)] cofactor preferences of oxidoreductases.
• To enable large-scale analysis and facilitate the design of enzymes with altered cofactor specificities.
• To provide an interpretable model for identifying key residues involved in cofactor binding.

Main Methods:

• Development of DISCODE, a transformer-based deep learning model.
• Training the model on a dataset of 7,132 NAD(P)-dependent enzyme sequences.
• Leveraging whole-length enzyme sequences for prediction without structural or taxonomic constraints.
• Analyzing transformer attention layers to identify key residues influencing cofactor specificity.

Main Results:

• DISCODE achieved high prediction accuracy (97.4%) and F1 score (97.3%) for NAD(P) cofactor preferences.
• Attention layer analysis successfully identified residues critical for NAD(P) interaction and specificity.
• Identified key residues demonstrated high consistency with known cofactor switching mutants.
• The model's interpretability facilitates understanding of cofactor specificity determinants.

Conclusions:

• DISCODE offers an accurate and efficient method for predicting NAD(P) cofactor preferences in oxidoreductases.
• The model's interpretability aids in identifying key residues for enzyme engineering.
• DISCODE, integrated with attention analysis, provides a fully automated pipeline for redesigning enzyme cofactor specificity.