Graph neural networks can predict ketosynthase substrate specificity.

Area of Science:

• Biochemistry
• Computational Biology
• Synthetic Biology

Background:

• Modular polyketide synthases (PKS) produce diverse natural products, including medicines.
• Engineering PKS for novel compounds is challenging due to ketosynthase (KS) proofreading, which rejects non-native substrates.
• Understanding KS proofreading is crucial for successful PKS pathway manipulation.

Purpose of the Study:

• To explore ketosynthase (KS) proofreading using Graph Neural Networks (GNNs).
• To predict when KS proofreading occurs and identify strategies to mitigate it.
• To assess the potential of GNNs in guiding PKS engineering for novel polyketide synthesis.

Main Methods:

• Trained Graph Neural Networks (GNNs) on AlphaFold structures of ketosynthases (KS).
• Used GNNs to predict substrate features, including β-carbon reduction states and α-methylation.
• Conducted proof-of-concept experiments to test GNN-guided mutation strategies for mitigating KS proofreading.

Main Results:

• GNN models accurately distinguished six out of ten β-carbon reduction state pairwise combinations (81-92% AUC).
• GNNs partially predicted the α-methylation state of polyketide substrates from KS structure (79% AUC).
• GNNs successfully predicted beneficial mutations to KS enzymes receiving non-native substrates, demonstrating their utility in engineering.

Conclusions:

• Graph Neural Networks are effective tools for analyzing KS proofreading mechanisms.
• GNNs can guide protein engineering efforts to overcome limitations in modular polyketide synthases.
• This approach holds promise for the rational design of PKS pathways to produce novel therapeutics.