Designing realistic regulatory DNA with autoregressive language models.
Avantika Lal1, David Garfield2, Tommaso Biancalani3
1Biology Research|AI Development, gRED Computational Sciences, Genentech, South San Francisco, California 94080, USA; lal.avantika@gene.com eraslan.gokcen@gene.com.
Researchers developed regLM, a framework using language models to design synthetic gene regulatory elements like promoters and enhancers. This tool creates DNA sequences with specific activities for biomanufacturing and therapies.
Area of Science:
- Genomics and Molecular Biology
- Synthetic Biology
- Bioinformatics
Background:
- Cis-regulatory elements (CREs), including promoters and enhancers, control gene expression through specific DNA sequences.
- CRE activity depends on the arrangement, composition, and spacing of transcription factor (TF) binding motifs.
- There is a growing need for synthetic CREs with tailored properties for applications in biomanufacturing and advanced therapies like cell and gene therapy.
Purpose of the Study:
- To introduce regLM, a novel computational framework for designing synthetic CREs with predictable and desired functional properties.
- To enable the creation of CREs with specific activity levels (high, low) or cell type-specific expression patterns.
- To leverage autoregressive language models and supervised sequence-to-function models for CRE design.
Main Methods:
- Developed the regLM framework integrating autoregressive language models with supervised sequence-to-function models.
- Utilized regLM to design synthetic yeast promoters and human enhancers with cell type-specific activity.
- Employed computational models to predict the functionality and biological relevance of the designed synthetic CREs.
Main Results:
- Successfully designed synthetic yeast promoters and human enhancers with targeted activity profiles using the regLM framework.
- Demonstrated that the computationally designed CREs are predicted to exhibit the desired functionality.
- Showed that the synthetic CREs possess biological features comparable to experimentally validated natural CREs.
Conclusions:
- regLM provides an effective computational approach for designing synthetic cis-regulatory elements with specific functional characteristics.
- The framework facilitates the creation of realistic regulatory DNA sequences for diverse biotechnological and therapeutic applications.
- regLM offers valuable insights into the complex cis-regulatory code governing gene expression.
Related Concept Videos
Regulation of Expression Occurs at Multiple Steps
Cooperative Binding of Transcription Regulators
Regulation of Expression at Multiple Steps
Master Transcription Regulators
Cis-regulatory Sequences
Co-activators and Co-repressors


