A combinatorial mutational map of active non-native protein kinases by deep learning guided sequence design

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Mapping protein sequence-function relationships is challenging due to limitations in exploring diverse mutations.
• Existing methods often restrict exploration to small mutational steps or evolutionarily conserved sequences.

Purpose of the Study:

• To overcome limitations in exploring protein sequence-function landscapes.
• To generate novel, functional protein sequences with highly combinatorial mutations using deep learning.

Main Methods:

• Applied deep-learning guided redesign to a natural protein tyrosine kinase.
• Measured activities and concentrations of 537 redesigned sequences using cell-free assays.
• Developed a regression model to identify sequence determinants of function.

Main Results:

• Generated functional sequences with an average of 37 mutations, retaining activity in 85% of variants.
• Explored 436 unique mutations across 76 positions in the kinase domain.
• Identified key sequence determinants and predicted function of unseen sequences.

Conclusions:

• Deep-learning guided redesign enables functional exploration of highly combinatorial sequence-function landscapes.
• The integrated approach allows exploration at unprecedented mutational scales.
• This methodology expands the possibilities for protein engineering and design.