# PROTEUS-REPAIR: research note

James Hong and Adam Shen | Student secondary analysis | Not peer reviewed
Contact: jameshongbc@gmail.com
Presentation updated: October 31, 2026. Source summary: October 30, 2026.

## Objective
Distinguish recurrent recovery by a partner substitution from standalone stabilization, position effects, and dependence among tested backgrounds.

## Experimental foundation
Tsuboyama K et al. Mega-scale experimental analysis of protein folding stability in biology and design. Nature 620, 434–444 (2023). DOI: 10.1038/s41586-023-06328-6.
The original investigators generated the experimental measurements. The student candidate analysis is separate.

## Descriptive findings
The available 26-row selected summary covers 16 domain identifiers. Four candidates have negative median standalone effects. These counts were recomputed from candidate-summary.csv.
The separately supplied 2K28 matrix reports E16F/W/Y passing on the same 15 of 17 R9 substitutions; R9A and R9N fail. All backgrounds are variants at one damaged position. This matrix has not been reconstructed from original experimental records.

## Quantities
On a consistent wild-type-relative scale, positive stability effects are stabilizing. Recovery R = s_AB - s_A; additive epistasis epsilon = s_AB - s_A - s_B; thus R = s_B + epsilon. Correction outputs require a documented consistent reference/null before this identity is used.
The summary reports standalone partner medians, not per-combination interaction estimates.

## Original conditional screen
Reported 9,908 substitutions tested with at least five backgrounds. Same-site strata fix protein, damaged substitution, and partner position. Poisson-binomial tails assume independent candidate indicators across backgrounds. Bonferroni corrects multiplicity if the individual p-values are valid; it cannot repair an invalid dependence model. Original adjusted p-values remain available but uncalibrated under alternative dependence assumptions.

## Unresolved analyses
Matched alternative-partner rates; distinct damaged-position coverage for most candidates; full event uncertainty rules; interaction-effect uncertainty/covariance; grouped held-out generalization; independent experimental validation.

## Proposed audit
Run %run -i gate9_audit.py in the original Colab containing orientation-level dataframe d. Code requires NumPy and pandas; syntax checked but not executed on the complete original dataframe. It verifies labels/duplicates and original conditional tests, then performs a block-column sensitivity test with its own exchangeability and between-block independence assumptions. It writes only new timestamped outputs.

## Proposed validation
Reconstruct all tested opportunities and label definitions. Estimate continuous effects on a consistent scale with shared-measurement covariance. Calibrate a structured null and assess threshold sensitivity. Hold out entire damaged positions or domains where feasible. With a laboratory partner, select a suitable stability assay for wild type, A, B, and AB, define controls and replication, and conduct independent measurements. No laboratory partnership, assay selection, or new experimental validation is claimed.