Protein Optimization Feedback Shift

Summary

protein-optimization-feedback-shift is Vipin’s AI4S project on uncertainty-aware closed-loop protein sequence optimization under feedback distribution shift. Its thesis is that calibration and decision quality can separate once the optimization loop changes its own candidate distribution.

Exact local project-name entry: uncertaintyprotein-ai4s.

Key Facts

  • EXTRACTED: The local root is D:/Research/UncertaintyProtein-AI4S.
  • EXTRACTED: The executable codebase is D:/Research/UncertaintyProtein-AI4S/protein_bo_conformal.
  • EXTRACTED: The project asks when calibration stops tracking decision quality, when weighted conformal acts as recovery under shift, and when proposal/batch policy explains observed UQ effects.
  • EXTRACTED: The main panel has 12 tasks: 3 FLIP, 3 FLIP2, and 6 ProteinGym tasks.
  • EXTRACTED: The appendix robustness panel has 10 tasks.
  • EXTRACTED: Strong supported conclusions include that greedy is a strong baseline, raw UCB is not universally better, improved coverage does not automatically imply improved decisions, and weighted conformal is condition-dependent.
  • EXTRACTED: The project is in final paper-preparation mode, with the remaining major task described as essay/-level LaTeX manuscript assembly and final server-side asset refresh.

Relationships

  • Related to analog-agent through closed-loop optimization and simulator/oracle-backed decision loops.
  • Related to uncertainty through calibration, decision quality, and uncertainty semantics.
  • Adds the AI4S branch of Vipin’s research map.

Open Questions

  • Which figures/tables need final server refresh before writing?
  • How should weighted conformal be framed so it is not overstated as a universal optimization improvement?
  • Which local data/provenance details should remain out of the public wiki?