ST-PARM: Pareto-Complete Inference-Time Alignment for Multi-Objective Protein Design.

Area of Science:

• Protein engineering
• Computational biology
• Machine learning for science

Background:

• Protein engineering is inherently multi-objective, requiring generation of Pareto-optimal candidates.
• Existing methods like linear objective scalarization can under-explore non-convex Pareto regions and are sensitive to noise.

Purpose of the Study:

• To introduce Smooth Tchebycheff Preference-Aware Reward Model (ST-PARM), an alignment framework for controllable protein sequence generation.
• To improve Pareto coverage and controllability in multi-objective protein engineering.

Main Methods:

• Developed ST-PARM, an inference-time alignment framework using a lightweight, single-trained reward model.
• Incorporated uncertainty-aware, reward-calibrated pairwise preference loss.
• Utilized smooth Tchebycheff scalarization for improved trade-off coverage and latent-space pair-construction strategies.

Main Results:

• ST-PARM demonstrated broader Pareto coverage and stronger preference tracking than baseline methods on GFP and IL-6 nanobody design tasks.
• Achieved controllable sequence generation for GFP with a focus on fluorescence and stability.
• Showcased robustness, a three-objective extension, and generality in natural language alignment.

Conclusions:

• ST-PARM provides a practical foundation for controllable sequence generation under competing multi-objectives and noisy measurements.
• The framework enhances exploration of complex trade-off surfaces in protein engineering.
• Results indicate improved performance in achieving desired protein properties through advanced computational methods.