Manufacturing-aware generative models enable petascale synthesis of designed DNA.

Area of Science:

• Biotechnology
• Synthetic Biology
• Machine Learning

Background:

• Generative modeling is powerful for designing DNA, RNA, and protein sequences.
• Large-scale physical synthesis of these designed sequences is currently too expensive.

Purpose of the Study:

• To introduce an efficient method for synthesizing DNA designs generated by machine learning models.
• To overcome the cost limitations of current DNA synthesis methods.

Main Methods:

• Integrated machine learning with wet lab procedures.
• Physically implemented generative sampling algorithms using DNA oligo synthesis and controlled chemical reactions.
• Synthesized approximately 10^16 DNA designs from generative models.

Main Results:

• Achieved DNA synthesis with realism and diversity comparable to state-of-the-art protein language models.
• Verified designed DNA sequences through sequencing.
• Expressed designed antibody single-chain fragment variable (scFv) regions in human cell lines.
• Identified potential therapeutic chimeric antigen receptors through high-throughput screening against multiplexed human leukocyte antigen (HLA)-presented intracellular proteins.
• Confirmed method generalizability by synthesizing designs for Taq polymerase and the HLA-presented peptidome.

Conclusions:

• The developed method enables efficient, large-scale DNA synthesis from generative models.
• This approach facilitates the discovery of novel functional proteins, including therapeutic antibodies.
• The method is broadly applicable to various biological sequence design challenges.