Controllable protein design via autoregressive direct coupling analysis conditioned on principal components

Area of Science:

• Computational biology
• Protein engineering
• Statistical modeling

Background:

• Direct Coupling Analysis (DCA) is a statistical method for protein sequence modeling.
• Existing generative models may lack biological context or fine-grained control.
• Protein design requires methods that balance generative accuracy with biological relevance.

Purpose of the Study:

• To introduce FeatureDCA, a novel statistical framework for protein sequence modeling and generation.
• To extend DCA by incorporating biologically meaningful conditioning for improved protein design.
• To demonstrate FeatureDCA's ability to guide sequence generation toward specific functional or structural properties.

Main Methods:

• FeatureDCA extends Direct Coupling Analysis (DCA) with conditioning on biological information (e.g., phylogeny, temperature, principal components).
• An autoregressive implementation of FeatureDCA was developed for sequence generation.
• Generated sequences were validated using structural prediction tools (AlphaFold, ESMFold) and compared against experimental data (deep mutational scanning).

Main Results:

• FeatureDCA matches or surpasses established models in generative accuracy for higher-order sequence statistics across multiple protein families.
• Generated sequences maintain substantial diversity and adopt biologically plausible folds consistent with wild-type targets.
• In a case study of Response Regulators, FeatureDCA accurately reproduced class-specific architectures when conditioned on subtype-specific principal components.
• FeatureDCA predictions showed accuracy comparable to unconditioned models for deep mutational scanning data, indicating capture of local functional constraints.

Conclusions:

• FeatureDCA offers a flexible and transparent approach for targeted protein sequence generation.
• The framework effectively bridges statistical fidelity, structural realism, and interpretability in protein design.
• FeatureDCA demonstrates potential for fine-grained structural control and accurate modeling of functional constraints in protein engineering.