Integrating experimental feedback improves generative models for biological sequences.

Area of Science:

• Computational biology
• Molecular biology
• Biomolecular engineering

Background:

• Generative probabilistic models show potential for designing artificial RNA and protein sequences.
• A major limitation is a high rate of false positives, where predicted functional sequences fail experimental validation.

Purpose of the Study:

• To address the false-positive challenge in generative biomolecular design.
• To explore the impact of reintegrating experimental feedback into model design.
• To improve the generation of functional biomolecular sequences.

Main Methods:

• Proposed a likelihood-based reintegration scheme.
• Conducted extensive computational experiments on RNA and protein datasets.
• Performed wet-lab experiments on Group I intron RNA self-splicing ribozymes.

Main Results:

• The feedback-driven approach significantly enhanced the model's capacity for generating functional sequences.
• Active designs increased from 6.7% to 63.7% (at 45 mutations) after integrating experimental data.
• The method demonstrated particular efficacy in designing self-splicing ribozymes.

Conclusions:

• Integrating recent experimental data directly tackles the false-positive challenge in biomolecular design.
• This feedback-driven approach offers a significant improvement for designing functional RNA and protein sequences.
• The proposed scheme enhances the reliability and success rate of generative biomolecular design.