What would settle it

Three answers we did not get, and the experiments that would.

Two negatives and a dataset that could not answer its own question. Each open question carries a designed experiment, down to the replicate count.

5designed follow-up experiments
3nothing else beats on both axes
3with no measurement in existence today
01

Where to start

Most needed against most buildable.

Worth starting if nothing else is both more urgently missing and easier to run. Three of five clear that bar.

Three experiments sit on the frontier

Evidence gap against feasibility, for five designed programmes

rule
MOST NEEDED, STILL BUILDABLEpartly answeredindirect evidence onlyno measurement at allhardmoderatehow buildable it ishow little evidence exists today →SUPTRNA-REPLICATED: no measurement at all, moderate to run — on the frontierSUPTRNA-REPLICATEDUPSTREAM-RESIDUE: indirect evidence only, moderate to run — on the frontierUPSTREAM-RESIDUEG418-INDEPENDENT: partly answered, moderate to run — beaten by UPSTREAM-RESIDUE; SUPTRNA-SAFETYG418-INDEPENDENTSUPTRNA-SAFETY: no measurement at all, moderate to run — on the frontierSUPTRNA-SAFETYFUNCTION-RESCUE: no measurement at all, hard to run — beaten by SUPTRNA-REPLICATED; SUPTRNA-SAFETYFUNCTION-RESCUE
  • filled — on the frontier, start here
  • hollow — another programme beats it on both
Details & limits
Position is a design judgement, not a measurement. The hollow points are not bad experiments — their answers partly arrive for free if the filled ones run first.
02

The programmes

What each one measures, and what it only implies.

Each card keeps two numbers apart: what the experiment touches, and who it would reach only if the result carries.

Every programme measures one variant and implies tens of thousands

Direct scope against the population behind its generalisation, log scale

rule
1101001,00010,00070,000SUPTRNA-REPLICATEDUPSTREAM-RESIDUEG418-INDEPENDENTSUPTRNA-SAFETYFUNCTION-RESCUEMEASURED DIRECTLYREACHED ONLY IF IT GENERALISESMeasured directly: one variant, one gene, one cell line30,976 variants sit behind this generalisation, none of them measured30,976Measured directly: one variant, one gene, one cell line70,376 variants sit behind this generalisation, none of them measured70,376Measured directly: one variant, one gene, one cell line70,376 variants sit behind this generalisation, none of them measured70,376Measured directly: one variant, one gene, one cell line70,376 variants sit behind this generalisation, none of them measured70,376Measured directly: one variant, one gene, one cell line7,120 variants sit behind this generalisation, none of them measured7,120variants, log scale
  • filled — actually measured
  • hollow — reached only if it generalises
Details & limits
The dashed line is the inferential distance, and it runs to four orders of magnitude. Nothing on the right is evidence.

01 · SUPTRNA-REPLICATED

Does a suppressor tRNA actually get ribosomes past a premature stop in patient cells?

start here
Suppressor tRNA
Patient-stop reporter
Ribosome profiling
Readthrough signature

Measured directly

One nonsense variant (IDUA p.Trp402Ter) in one gene, one UAG stop, one tyrosine suppressor tRNA scaffold, one patient fibroblast line

Reached only if it generalises

30,976 variants · 2,930 genes · 4,340 conditions

How the replicate count was derived, and the rest of the design
What the lab does
Ribosome profiling of patient fibroblasts carrying a nonsense variant, transduced with a UAG-reading tyrosine suppressor tRNA, at four biological replicates per arm rather than the two the only public dataset used.
Compared against
Suppressor tRNA against the same delivery construct carrying EGFP instead, prepared in the same batch.
What counts as success
All three parts of the readthrough signature move together: occupancy past the stop rises, occupancy at the stop falls, and the reading frame past the stop shifts toward the gene's own coding frame with an interval excluding zero.
If it fails
Either the suppressor tRNA does not read through at a level ribosome profiling can see, or delivery failed — which the EGFP arm distinguishes. A null with four replicates would be informative; the existing null with two is not.
Why it matters
This is the modality the whole tRNA design layer recommends, and there is no measurement of it anywhere at a quality this project would accept. Every ranking of tRNA designs currently rests on prediction alone.
What generalisation it needs
That a UAG-reading suppressor tRNA behaves similarly at other UAG stops, in other genes and cell types, with other scaffolds and delivery systems. None of that is tested here
Why the evidence is missing
The one public dataset could not have answered this at any sequencing depth: 92-98% of its variance sits between libraries rather than in the counts, so the limiting factor is replication, not reads.
Assay
Ribosome profiling (Ribo-seq), transcriptome alignment, P-site calibration
Model system
Patient-derived fibroblasts carrying a UAG nonsense variant
Endpoint
Downstream in-frame occupancy, termination occupancy, and downstream frame composition, each against the library's own coding sequence
Decision rule
The frozen three-part rule of ADR-0010, unpaired, with the frame condition required to clear zero
Replicates
4 per arm, derived from the observed between-library variance of GSE179274 against a G418-sized reference effect (ADR-0019)
Replicate endpoint
frame_gap: downstream frame composition minus the library's own coding frame composition
Effect size assumed
0.183, the frame_gap difference G418 produces in GSE144140 over its selected length set
Variance source
GSE179274's own between-library variance, decomposed against Poisson counting error (ADR-0019)
Design
unpaired, two arms
Alpha and power
5% two-sided, 80% power, design degrees of freedom rather than Welch
Method
smallest integer n per arm whose minimum detectable effect falls below the reference effect; rounded up
Records
ADR-0010, ADR-0019; results/exploratory/detectability/gse179274_fibroblast

No measurement exists · moderate to run · bears on safety · would settle IDUA-W402X-SUPTRNA

02 · UPSTREAM-RESIDUE

Does the amino acid immediately before a premature stop change how well a drug reads through it, differently for each of the three stop codons?

start here
Codon–residue pairs
Synthetic reporters
Matched library
Context effect

Measured directly

Sixty amino-acid-by-stop-codon combinations in a synthetic reporter, with several synonymous codons each, in one cell line

Reached only if it generalises

70,376 variants · 3,603 genes · 5,974 conditions

How the replicate count was derived, and the rest of the design
What the lab does
A reporter library in which the codon before the stop is varied across all twenty amino acids, crossed with all three stop codons, with several different codons per amino acid so that residue identity and codon identity can be told apart.
Compared against
Each amino-acid-by-stop-codon combination against the library mean, under each drug and untreated.
What counts as success
Readthrough differs by amino acid in a way that depends on the stop codon, and synonymous codons for the same amino acid behave alike.
If it fails
If synonymous codons differ from each other as much as amino acids do, the signal is codon-level after all and Wave 2's interpretation was wrong. Either outcome resolves the question the observational data cannot.
Why it matters
Wave 2 found a signal that tracks the amino acid before the stop rather than how fast its codon is read. If real, it is a term the published model does not carry, and it would improve every prediction rather than one therapy's.
What generalisation it needs
That an effect measured in a reporter construct transfers to endogenous genes, and that the six drugs tested behave like others of their class
Why the evidence is missing
The existing observation is post-hoc, has sixty parameters against the five it beats, and has had none of the shuffle controls. It cannot be strengthened by further analysis of the same dataset.
Assay
Massively parallel reporter assay with readthrough read out by fluorescence ratio
Model system
HEK293T, matching the library the baseline was fitted on
Endpoint
Readthrough fraction per construct, per drug
Decision rule
Declared before the assay runs; the interaction is supported only if it holds under a split that holds out whole amino acids
Replicates
not estimated
Replicate endpoint
not estimated
Effect size assumed
not estimated
Variance source
not estimated
Design
not estimated
Alpha and power
not estimated
Method
not estimated
Records
ADR-0006, ADR-0020, ADR-0021; docs/notebook.md

Indirect evidence only · moderate to run · would settle KINETICS-BEYOND-SEQUENCE

03 · G418-INDEPENDENT

Does the readthrough signature we measured for G418 hold up in a different laboratory?

beaten by another
G418 + stall control
Independent cell line
Ribosome profiling
Signature repeats

Measured directly

One compound and one stalling comparator, in one cell line, under one library preparation chemistry differing from the published one

Reached only if it generalises

70,376 variants · 3,603 genes · 5,974 conditions

How the replicate count was derived, and the rest of the design
What the lab does
Ribosome profiling of G418-treated and vehicle-treated cells, three replicates per arm, in a laboratory using different linker chemistry from the published dataset.
Compared against
G418 against vehicle, with a stalling agent as a second arm so the assay must discriminate rather than merely respond.
What counts as success
The same three-part signature reproduces, and the stalling agent fails it in the direction its mechanism predicts.
If it fails
The signature is an artefact of one protocol, and every result in this package that depends on the assay being sensitive has to be reread as unvalidated.
Why it matters
Every negative result this project reports about other therapies rests on the assay being able to see a working one. That rests on a single laboratory and a single library preparation chemistry.
What generalisation it needs
That an assay validated for G418 is sensitive to other readthrough modalities, including suppressor tRNAs, which it has never been shown on
Why the evidence is missing
The current evidence is confirmatory within one laboratory and protocol family. A shared-protocol artefact would survive it, and no public dataset profiles a readthrough compound from a second laboratory at three replicates.
Assay
Ribosome profiling (Ribo-seq)
Model system
Any human cell line expressing the reporter panel; the laboratory and chemistry are what must differ
Endpoint
The same three quantities, computed identically
Decision rule
The frozen rule of ADR-0010, applied unchanged
Replicates
3 per arm, matching the design the rule was defined for
Replicate endpoint
the three-part readthrough signature of ADR-0010
Effect size assumed
the effect sizes GSE144140 already measured for G418
Variance source
GSE144140's three-replicate design, matched rather than recomputed
Design
unpaired, three arms including the stalling comparator
Alpha and power
as ADR-0010 froze them; the rule is applied unchanged
Method
matched to the design the rule was defined for, not derived afresh
Records
ADR-0009, ADR-0010; results/readthrough/gse144140

Partly answered already · moderate to run · bears on safety · would settle G418-READTHROUGH

04 · SUPTRNA-SAFETY

Does a suppressor tRNA also read through the body's normal stop codons, and how much?

start here
Suppressor tRNA
Normal stop codons
Ribosome profiling
Burden map

Measured directly

One suppressor tRNA scaffold in one cell line, measured across that cell line's expressed transcriptome

Reached only if it generalises

70,376 variants · 3,603 genes · 5,974 conditions

How the replicate count was derived, and the rest of the design
What the lab does
Ribosome profiling of cells expressing a suppressor tRNA, measuring occupancy past every normal stop codon in the transcriptome rather than at the target variant.
Compared against
Suppressor tRNA against the same delivery construct without it, three replicates per arm.
What counts as success
There is no success direction here. The output is a measured burden across normal stops, reported whatever it is.
If it fails
A high burden would constrain which designs are usable and would change the coverage frontier's recommendations, which currently optimise reach without a measured safety cost for this modality.
Why it matters
A therapy that gets past the patient's premature stop but also runs past thousands of normal stops is producing extended proteins everywhere. The safety atlas measures this for G418 and for nothing else.
What generalisation it needs
That native-stop burden measured for one scaffold in one tissue predicts burden for other scaffolds, other inserted amino acids and the tissues a therapy would actually reach
Why the evidence is missing
The native-stop atlas exists only for G418. The safety layer for suppressor tRNAs is prediction with no measurement behind it.
Assay
Ribosome profiling, native-stop occupancy across the transcriptome
Model system
A human cell line tolerant of sustained suppressor tRNA expression
Endpoint
Occupancy past each annotated stop, treated against control, transcriptome-wide
Decision rule
Reported as a burden distribution rather than a pass or fail; no threshold is declared because none is defensible yet
Replicates
3 per arm, matching the atlas the G418 measurement used
Replicate endpoint
occupancy past each annotated native stop, treated against control
Effect size assumed
not estimated; no threshold is declared because none is defensible yet
Variance source
matched to the G418 native-stop atlas rather than derived
Design
unpaired, two arms
Alpha and power
not applicable; the output is a burden distribution rather than a test
Method
matched to the existing atlas design
Records
ADR-0010; results/atlas/gse144140_native_stops.tsv

No measurement exists · moderate to run · bears on safety · opens a new question

05 · FUNCTION-RESCUE

When a drug does get the ribosome past the stop, does the protein it makes actually work?

beaten by another
Dose response
Patient-relevant cells
Protein activity
Function restored

Measured directly

A small number of genes chosen for assay availability, each with its own functional readout

Reached only if it generalises

7,120 variants · 1,430 genes · 2,169 conditions

How the replicate count was derived, and the rest of the design
What the lab does
Enzyme or functional assays on the protein produced under readthrough, for variants where the message is predicted to survive decay and readthrough is predicted to be high.
Compared against
Readthrough product against wild-type protein and against the untreated truncated product.
What counts as success
Measurable restoration of function, reported as a fraction of wild-type rather than as a yes or no.
If it fails
Readthrough without function would mean the scored table's top rows are not the best candidates, and the function layer would have to gate them rather than annotate them.
Why it matters
Every prediction in this package stops at readthrough. Whether the resulting protein folds and functions is a separate question, and a variant can score well on readthrough and produce nothing useful.
What generalisation it needs
That function restored in one gene's assay says anything about the others, which it largely does not — this programme informs its own genes and generalises worst of the five
Why the evidence is missing
The function layer is residue tolerance only. Nothing in this package measures whether a rescued protein works, and the layers deliberately share no ground truth.
Assay
Gene-specific functional or enzymatic assay
Model system
Patient-derived cells or a heterologous expression system, gene by gene
Endpoint
Function as a fraction of wild-type
Decision rule
Per gene, against a threshold set by that disease's known genotype-phenotype relationship
Replicates
not estimated
Replicate endpoint
not estimated
Effect size assumed
not estimated
Variance source
not estimated
Design
not estimated
Alpha and power
not estimated
Method
not estimated
Records
ADR-0018; results/amenability_landscape.tsv

No measurement exists · high to run · opens a new question