Three E16 partners share a passing profile at one damaged position.
See the finding →UNRESOLVEDWhat explains it?Standalone effects, interaction, and statistical dependence must be separated.
Inspect the interpretation →NEXT TESTOriginal. A. B. AB.Compare all four protein versions to measure the interaction.
Read the validation plan →Question → Evidence → Interpretation → Validation
The question, the approach,
and the current conclusion.
Objective. Identify second-site substitutions associated with repeated recovery of destabilized protein variants, while separating partner stabilization from interaction-specific effects and experimental-library structure.
Approach. Re-examine an existing screen derived from published protein folding-stability measurements. Recovery labels require agreement between physical and held-out isotonic correction methods for the same damaging-mutation / partner orientation. Candidate enrichment was screened conditionally on damaged substitutions and partner positions.
Observed results. The available selected-candidate summary contains 26 substitutions across 16 protein-domain identifiers. Four partners have negative median standalone effects under the project’s sign convention. The supplied 2K28 matrix shows E16F, E16W, and E16Y passing on the same 15 of 17 damaging R9 substitutions.
Interpretation. These observations motivate further study but do not establish general-purpose repair or a specific structural mechanism. Matched background base rates, per-combination interaction estimates, and dependence-aware significance have not yet been verified from the complete source data.
Counts above were recomputed directly from the downloadable 26-row summary. This is a selected subset, not the complete experimental library; its composition cannot estimate a global rescue prevalence.
New to protein stability?Open the plain-language guide and interactive examples
Can a second change
help a protein?
Proteins are chains of small building blocks called amino acids. Many proteins need to fold into a particular shape. We study whether a second change in that chain can improve the stability lost after a first change.
A protein is a folded chain.
Its sequence helps determine how stable its folded shape is. Stability is one property of a protein; it is not the same thing as biological function.
One change can reduce stability.
A mutation replaces a building block. Some replacements make the folded state less stable. In this project, that first change is the “damaging mutation.”
A second change may help.
A partner mutation may improve stability on its own, work differently alongside the first mutation, or do both. We need to distinguish these explanations.
We found patterns worth checking—not proven repairs.
Three second changes in one protein show the same passing pattern across 15 of 17 tested first changes. But those 17 first changes all occur at the same position. We still need to test whether other partners do just as well, why the pattern occurs, and whether it works elsewhere.
See the actual results and their limits →Two ways to improve stability.
Imagine the first change lowers stability by 2 units. Adjust the second change’s own contribution and the extra effect of combining the two. These units represent kcal/mol, but the numbers below are teaching examples—not measured candidate effects or predictions.
The lesson: improving on the first damaged variant does not necessarily restore the original stability. And an improvement alone does not tell us how much comes from an interaction. This additive example does not reproduce the study’s correction models.
A quick check: what does 15 out of 17 tell us?
Choose the statement best supported by the supplied results.
Five useful words, explained simply
- Mutation
- A change in a protein’s amino-acid sequence.
- Stability
- How favorable the folded state is relative to the unfolded state under specified conditions.
- Partner
- The second change tested alongside a damaging change.
- Interaction, or epistasis
- The amount by which the combined effect differs from adding the two individual effects.
- Candidate
- A result selected for further investigation, not an experimentally confirmed repair.
Start by defining what “help” means.
Recovery is improvement over the damaged variant. Interaction is the extra effect beyond adding the two mutations separately.
Recovery measures improvement over a damaged variant. Interaction asks whether the combined effect exceeds the sum of the two separate effects.
A double mutant can outperform a damaging single mutant because its partner is stabilizing, because the mutations interact, or both. The quantity of interest determines what counts as evidence.
Define the measurement scale
Let sA, sB, and sAB denote stability effects relative to the same wild-type reference, with positive values meaning stabilization in kcal/mol. A is the damaging substitution; B is the partner. These symbols define an additive reference scale, not an assumption about all correction outputs.
Define the target of inference
Recovered stability describes improvement relative to A. Epistasis describes departure from the sum of the two single-mutant effects. Positive epistasis does not necessarily restore wild-type stability, and recovery alone does not demonstrate positive epistasis.
Mutation notation and units
E16F means glutamate (E) at sequence position 16 is replaced by phenylalanine (F). A “damaging background” here is a specific single substitution, such as R9D—not an independent organism, replicate experiment, or necessarily a distinct position. Labels such as 2K28 and 3DKM identify source protein domains.
E16F: glutamate at position 16 is replaced by phenylalanine.
ΔΔG conventions differ between publications. This page consistently uses the project convention: positive is stabilizing, negative is destabilizing.
Trace the route from measurements to candidates.
Both correction methods must pass on the same damaging-mutation orientation. Agreement is useful, but is not independent replication.
The experimental dataset, correction models, and student screening analysis are separate steps. Each contributes assumptions that matter for interpretation.
The experimental foundation is Tsuboyama et al. (2023), who used cDNA display proteolysis to estimate protein-domain folding stability. The candidate rankings, correction comparisons, and interpretations on this page are the student project’s secondary analysis, not claims attributed to the original investigators. [1]
Specify the damaging mutant
Map each combination to its damaging and partner substitutions. Preserve direction: the two correction methods must agree for that same orientation. Exclude ambiguous mappings rather than infer identity from approximate matches alone.
Apply correction criteria
Include damaging single-mutant effects below −1.0 kcal/mol. The reported recovery screen uses a minimum corrected gain of 1.0 kcal/mol under each method. Additional uncertainty conditions remain to be reconstructed.
Control tested opportunities
Count distinct damaging substitutions tested with each partner, including failures. Compare candidate outcomes within the same protein and damaged substitution, then additionally within the same partner position.
Analysis units, duplicates, and shared measurements
The original summary aggregates each protein × partner substitution × damaging substitution to one observation. Its earlier code used maximum labels when collapsing duplicates. The downloadable audit instead stops on conflicting duplicate labels or partner effects so they can be examined.
Different substitutions at the same damaging position may share measurements, correction fits, and structural context. Repeated observation of a partner against these substitutions does not establish independent evidence across positions or proteins. The audit reports distinct damaged positions as well as substitution counts.
The approximate orientation mapping reproduces the original screen’s matching tolerances. A completed audit should verify mutation identity directly from source keys and inspect ambiguous or equal-effect mappings.
Why correction agreement is useful—and limited
Agreement between two correction procedures is a sensitivity check for those procedures. Because they share the underlying experiments, it is not independent experimental replication. Held-out fitting must also be checked at the correct grouping level to prevent related observations from crossing training and evaluation boundaries.
Read the pattern before judging its meaning.
E16F, E16W, and E16Y share the same 15 passing R9 substitutions; the result covers one damaged position, not 17 different positions.
The following observations describe selected candidates. They do not yet establish a mechanism or predict performance on new damaged positions.
The most interpretable example is a shared passing profile for three E16 aromatic substitutions in 2K28. Its meaning depends on the backgrounds tested and the alternatives available at that site.
Matched R9 outcomes in 2K28
Each row is a damaging substitution at one position, R9. A pass means the combination meets the project’s physical and isotonic event labels on the same orientation.
| Damaging background | E16F | E16W | E16Y |
|---|
Standalone stabilization is insufficient as a universal explanation
Four candidates have negative median standalone effects: E16F, R27Y, H37F, and H37Y in their respective domains. Their reported recurrence is not explained solely by a positive median partner effect. This observation does not quantify interaction magnitudes or resolve correction artifacts.
A global rate is not the relevant null
A candidate’s raw passing fraction should be compared with partners tested on the same damaging substitutions, particularly substitutions at the same partner site. The selected 26-row summary lacks the unsuccessful alternatives needed to reconstruct that comparison.
Move from the headline to individual observations.
Search the original selected summary, compare candidates, and inspect the reported passing fractions. These are screening outcomes, not validated repair probabilities.
All 26 original candidates, with their denominators and standalone effects. Search, filter, or select a substitution to inspect what is—and is not—available for it.
| Domain | Partner | Both / tested | Passing fraction | Standalone kcal/mol | Original adjusted p |
|---|
Compare two candidates.
Choose substitutions to compare their observed coverage, passing counts, and standalone effects. Differences in tested backgrounds prevent a direct claim that one partner is intrinsically better.
Observed passing fractions
Chart follows the table’s filters and sort order. Color denotes the sign of the median standalone partner effect. Click a label to inspect the candidate.
A small p-value depends on the right comparison.
Partner-site control is present. Dependence among related backgrounds still needs to be addressed before treating the original p-values as calibrated evidence.
The original tests include partner-site controls. The remaining question is whether their assumptions adequately represent dependence among related backgrounds.
The original screen tests 9,908 partner substitutions meeting a coverage threshold of at least five tested backgrounds, according to the supplied screen output. It assigns each combination a null pass probability from a comparison stratum and uses a Poisson-binomial upper tail for the candidate’s count. Bonferroni adjustment is applied across the tested substitutions.
What the strata control
The background comparison holds protein and damaging substitution fixed. The same-site comparison additionally fixes the partner position. This controls the observed opportunity structure at that resolution; it does not control all biochemical features or establish causal mutation specificity.
Where calibration may fail
Correlated variants at one damaged site, shared single-mutant estimates, and common correction fits can violate trial independence. Bonferroni does not require independent candidate tests when their individual p-values are valid, but it cannot repair invalid individual p-values.
Candidate inclusion in its own stratum is not automatically leakage
Including the tested candidate in a stratum’s observed pass fraction is compatible with conditioning on the stratum’s total successes under a suitable within-stratum exchangeability null. It should not be labeled an error solely for that reason. The substantive questions are whether the exchangeability null is appropriate and whether assignments across backgrounds can be treated independently.
The computed tail is exact for its specified Poisson-binomial model. “Exact” does not imply that this model matches the data-generating process.
Proposed dependence-aware sensitivity analysis
The downloadable audit exchanges complete partner-substitution outcome columns within each protein × partner position × damaged position block. It restricts comparators to substitutions observed on all of the candidate’s backgrounds in that block, then computes a count distribution over the available profiles.
This preserves recurrence across variants within one damaged-site block. It tests a different null and still assumes exchangeability among the retained profiles and independence between blocks. Restriction to complete comparators changes the comparison set and must be reported.
For a candidate spanning only one block with n fully observed alternative profiles, the smallest attainable non-randomized upper-tail probability is at least 1/n when the candidate profile is included. A high fraction across many variants at that one site therefore need not imply equally strong evidence under this block-level null.
Status: audit code is provided; its results have not been computed on the complete source dataframe.
Separate what is observed from what remains open.
These are candidates for further testing. A reusable repair mechanism has not been established.
This ledger distinguishes recomputed summary facts, reported outputs, inspected code, and claims that still require analysis or experiments.
| Statement | Evidence status | Scope |
|---|---|---|
| 26 candidates across 16 domain identifiers | Recomputed | Counts from the public selected summary. |
| E16F/W/Y share 15 passing R9 substitutions | Reported matrix | Supplied project output; not reconstructed from original double-mutant records. |
| Four candidates have negative standalone medians | Recomputed | Summary values; per-combination effect consistency remains to be checked. |
| Same-site control is present | Code inspected | Original strata include protein, damaged substitution, and partner position. |
| Enrichment over matched alternative partners | Not established here | Full opportunity table and calibrated dependence model required. |
| Interaction-specific rescue and structural mechanism | Unresolved | Interaction estimates, uncertainty propagation, and targeted validation needed. |
| Generalization to new positions or domains | Untested | Grouped held-out evaluation and independent measurements needed. |
Additional threats to interpretation
- Selection effects: candidates were selected using the same outcomes later summarized. Their high observed rates are not unbiased estimates of future performance.
- Threshold sensitivity: stability uncertainty and observations near a gain cutoff can change binary classifications. Continuous effects and full uncertainty distributions should accompany labels.
- Limited opportunity structure: tested backgrounds are determined by the experimental library; untested combinations are missing, not failures.
- Shared corrections: agreement between methods does not provide two independent experiments.
- Mechanistic ambiguity: glycine, aromaticity, packing, and position-specific effects are hypotheses that cannot be resolved from recurrence counts alone.
- Endpoint mismatch: folding-stability recovery does not establish restoration of biological function, cellular fitness, or therapeutic utility.
Turn the candidate pattern into a falsifiable test.
Independent stability measurements of the original, A, B, and AB variants, together with a verified interaction estimate and uncertainty.
The next steps connect computational checks to independent measurements of the original protein, both single mutants, and their combination.
The following is a proposed validation sequence, not a completed experiment or an existing laboratory collaboration.
- Reconstruct the full opportunity table. Verify identities, orientation, missingness, repeated records, correction labels, and all uncertainty conditions. Retain successful and unsuccessful tested combinations.
- Estimate continuous interaction effects. Use matched wild-type, single-mutant, and double-mutant measurements on a documented scale. Propagate covariance from shared measurements and correction fits.
- Calibrate a structured null. Compare within matched sites, preserve related mutation profiles, and account for candidate selection and multiplicity. Report sensitivity to the baseline model and gain threshold.
- Evaluate prospective generalization. Hold out entire damaged positions or domains where coverage permits; fit corrections and select candidates using only training groups. Report a lack of coverage when this is not possible.
- Plan independent measurements with a laboratory partner. For a small candidate panel, measure wild type, A, B, and AB with appropriate controls. Equilibrium chemical denaturation monitored by fluorescence or circular dichroism is a possible approach if the domain and experimental assumptions are suitable; the assay is not yet selected.
Replication, sample size, assay suitability, and quality controls must be defined with the laboratory before experiments. No new experimental validation has been performed by this project.
CONTINUE THE ARGUMENTInspect the research materials →Make the evidence easy to inspect.
Download the source summary, audit script, and provenance record. Execution limits are documented alongside the materials.
The files below are served directly with this website and require no sign-in. The candidate summary is preserved as retrieved. Derived descriptive counts and website figures do not modify the source data.
Run the audit in the original Colab environment
The script requires the existing orientation-level dataframe d and the project’s source correction labels. Upload the Python file to that Colab and execute:
%run -i gate9_audit.py
Dependencies: NumPy and pandas. The script writes only to a newly created, timestamped audit folder. It validates label types, stops on conflicting duplicates, reproduces the original conditional tests, reports damaged-position coverage, and performs a block-column sensitivity test.
Execution status: syntax checked, not executed on the complete original source dataframe. The downloadable summary is insufficient to reconstruct the full inference. The evidence package is a partial reproducibility record, not a complete executable reproduction of the study.
Provenance and data dictionary
Source summary: gate9c_candidate_robustness.csv from the project’s Google Drive, retrieved during the October 30, 2026 review. The public copy is named candidate-summary.csv. The R9 outcome matrix was supplied separately in the project chat.
WT_name: source domain identifier; partner_mut: partner substitution; tested: distinct damaging substitutions tested; both: combinations passing both methods; physical and isotonic: method-specific passing counts; partner_effect: median standalone stability effect; adjusted_p: original same-site Bonferroni-adjusted screening p-value.
The source experimental publication and the student analysis are distinct provenance layers. Raw measurements, correction fitting, and all event-label code are not included in this package. Read the original evidence notes.
Experimental foundation.
[1] Tsuboyama, K., Dauparas, J., Chen, J., Laine, E., Mohseni Behbahani, Y., Weinstein, J. J., Mangan, N. M., Ovchinnikov, S. & Rocklin, G. J. Mega-scale experimental analysis of protein folding stability in biology and design. Nature 620, 434–444 (2023). doi:10.1038/s41586-023-06328-6 · Open-access full text · BibTeX ↓
Credit for the experimental measurements belongs to the original investigators. This website reports a student-led secondary analysis and does not imply endorsement, affiliation, or collaboration with those investigators.
Research team.
James Hong
Project team
Adam Shen
Project team
Project correspondence: jameshongbc@gmail.com
Open to discussion with researchers about statistical validation, protein biophysics, and feasible independent stability assays. No institutional affiliation or laboratory partnership is claimed.