AlphaCD: a machine learning model capable of highly accurate characterization for 21,335 cytidine deaminases

Area of Science:

• Biochemistry and Molecular Biology
• Genomics and Proteomics
• Bioinformatics and Computational Biology

Background:

• Sequence databases contain numerous proteins with limited functional data, hindering the identification of those with specific activities.
• Apolipoprotein B mRNA-editing enzyme, catalytic polypeptide (APOBEC)-like family cytidine deaminases (CDs) are crucial for various biological processes, but their functional diversity is not fully characterized.
• Efficiently predicting protein function is essential for advancing biotechnology and understanding biological systems.

Purpose of the Study:

• To experimentally characterize the functional properties of a large set of APOBEC-like cytidine deaminases.
• To develop a machine learning model (AlphaCD) for accurate prediction of protein function based on sequence, structure, and experimental data.
• To demonstrate the utility of AlphaCD in predicting and optimizing protein functions for biotechnological applications.

Main Methods:

• Experimentally characterized catalytic efficiency, target site preference, and off-target activity of 1100 APOBEC-like CDs fused with nCas9 in HEK293T cells.
• Constructed a machine learning model, AlphaCD, integrating experimental data with sequence, 3D structure, and other protein features.
• Validated AlphaCD's predictive accuracy using independent datasets and applied it to predict functions for over 21,000 CDs in UniProt.

Main Results:

• Generated the largest dataset of experimentally validated functions for the APOBEC-like CD family to date.
• AlphaCD achieved high prediction accuracy for catalytic efficiency (0.92), off-target activity (0.84), target windows (0.73), and catalytic motifs (0.78).
• Successful application of AlphaCD in predicting functions for a large protein set and guiding protein engineering for high-fidelity base editing.

Conclusions:

• AlphaCD provides a powerful tool for high-accuracy, high-throughput functional characterization of APOBEC-like CDs.
• The study demonstrates a successful strategy for accelerated protein characterization and engineering using machine learning.
• This approach can be extended to accelerate the functional characterization of other protein families, advancing various fields of biological research and biotechnology.