Generating experimentally unrelated target molecule-binding highly functionalized nucleic-acid polymers using machine...

Area of Science:

• Biochemistry
• Molecular Biology
• Computational Chemistry

Background:

• In vitro selection is crucial for discovering functional polymers from vast sequence spaces.
• Exploring novel sequence space beyond experimental variants is challenging.
• Limitations in selection and sequencing restrict the discovery of diverse functional biopolymers.

Purpose of the Study:

• To develop an integrated approach combining in vitro selection and machine learning.
• To discover novel highly side-chain-functionalized nucleic acid polymers (HFNAPs) with high affinity for daunomycin.
• To generate diverse HFNAP sequences unrelated to experimentally derived variants using machine learning.

Main Methods:

• In vitro selection was used to identify HFNAPs with daunomycin binding activity.
• A conditional variational autoencoder (CVAE) machine learning model was trained on selection data.
• The CVAE model generated novel HFNAP sequences with predicted high daunomycin affinities.

Main Results:

• HFNAPs with potent daunomycin affinities (K_D = 5-65 nM) were discovered.
• The CVAE model generated diverse HFNAP sequences with improved daunomycin affinities (K_D = 9-26 nM).
• Generated sequences were unrelated to those obtained through experimental selection, expanding sequence space exploration.

Conclusions:

• Coupling in vitro selection with machine learning enables efficient discovery of functional biopolymers.
• This integrated approach allows for the direct generation of active variants beyond experimental limitations.
• Demonstrates a powerful new strategy for biopolymer engineering and functional discovery.