The checks behind every number

Before you believe a prediction, believe the instrument.

Six checks ask whether our ribosome measurements work—including the control that rejected our own hypothesis.

24,776,791ribosome positions mapped
9 of 9libraries calibrated without a failure
486transcripts deep enough to measure
01

Periodicity

A ribosome moves three letters at a time.

Real ribosome footprints repeat every three nucleotides; random RNA fragments do not.

The three-nucleotide beat is visible in the raw density

Footprint density by distance from the codon, averaged within each treatment arm

measured
normalised ribosome footprintsaround the start codon-20020400%2%4%6%8%DMSO (control): normalised ribosome footprints around the start codonG418: normalised ribosome footprints around the start codonSRI-37240: normalised ribosome footprints around the start codondistance from start codon (nucleotides)around the stop codon-40-20020DMSO (control): normalised ribosome footprints around the stop codonG418: normalised ribosome footprints around the stop codonSRI-37240: normalised ribosome footprints around the stop codondistance from stop codon (nucleotides)EXPECTED QUALITY SIGNAL: PEAKS REPEAT EVERY 3 NT
  • DMSO (control)
  • G418
  • SRI-37240

PASS · All three treatment arms show the repeating three-nucleotide pattern expected from genuine ribosome footprints.

Details & limits
Arms are averaged separately, never pooled — pooling would erase the stop-codon difference the next section depends on.
02

Selection

We did not choose the read lengths. The data did.

Keep only read lengths that preserve the expected three-nucleotide reading frame.

Only three footprint lengths carry the reading frame

Share of reads landing in the coding frame, per length, across 9 libraries

measured
reads in the correct coding frame0%20%40%60%80%20 nt — 35.9% in frame, 1.0% of reads21 nt — 82.1% in frame, 4.2% of reads · kept22 nt — 70.7% in frame, 4.6% of reads · kept23 nt — 46.8% in frame, 3.1% of reads24 nt — 33.7% in frame, 1.8% of reads25 nt — 36.9% in frame, 1.8% of reads26 nt — 40.2% in frame, 2.9% of reads27 nt — 48.0% in frame, 6.4% of reads28 nt — 41.5% in frame, 12.5% of reads29 nt — 41.6% in frame, 18.6% of reads30 nt — 42.0% in frame, 19.0% of reads31 nt — 49.6% in frame, 13.7% of reads · kept32 nt — 43.4% in frame, 6.7% of reads33 nt — 50.3% in frame, 2.3% of reads34 nt — 43.1% in frame, 0.8% of reads35 nt — 42.3% in frame, 0.3% of reads36 nt — 37.0% in frame, 0.1% of reads37 nt — 32.7% in frame, 0.1% of reads38 nt — 34.1% in frame, 0.0% of reads39 nt — 31.7% in frame, 0.0% of reads40 nt — 32.1% in frame, 0.0% of readswhat a random fragment would givekeptkeptkept202122232425262728293031323334353637383940footprint length (nucleotides)

PASS · Lengths 21, 22, 31 rise above the random-fragment baseline and are retained.

Details & limits
Lengths 21, 22, 31 clear the bar; the rest sit near the one-in-three a random fragment would give. Selecting on the frame rather than on the length distribution is what makes the codon-level results below meaningful.
03

Discrimination

One drug reads through. The other piles up at the stop.

A trustworthy assay must separate a readthrough drug from a stalling one.

G418 meets all three conditions; SRI-37240 meets none

Mean difference against control with a 95% interval, per condition

measured
Ribosomes continuing past the stopDESIRED → more ribosomes passno changeG418G418 · Ribosomes continuing past the stop: +0.00962 [+0.00505, +0.01418] — condition held+0.0096SRI-37240SRI-37240 · Ribosomes continuing past the stop: −0.00024 [−0.00243, +0.00195] — condition did not hold−0.0002NOT METRibosomes sitting at the stopfewer ribosomes stall ← DESIREDno changeG418G418 · Ribosomes sitting at the stop: −0.01160 [−0.02619, +0.00299] — condition held−0.0116SRI-37240SRI-37240 · Ribosomes sitting at the stop: +0.00093 [−0.01452, +0.01637] — condition did not hold+0.0009NOT METReading frame kept past the stopDESIRED → more stay in frameno changeG418G418 · Reading frame kept past the stop: +0.18059 [+0.11964, +0.24154] — condition held+0.1806SRI-37240SRI-37240 · Reading frame kept past the stop: −0.06284 [−0.40343, +0.27775] — condition did not hold−0.0628NOT MET
  • filled — the condition was met
  • hollow — the condition was not met

PASS · G418 moves in the desired direction on all three conditions; SRI-37240 does not. The G418 termination interval spans zero, but that condition is registered as a direction check rather than an interval-exclusion check.

Details & limits
Readthrough is a redistribution, so it must show in more than one place at once: density beyond the stop rises, density at the stop falls, and what continues stays in the coding frame. 3 of 3 hold for G418 and 0 of 3 for SRI-37240.

Every library, drawn separately

Ribosomes continuing past the stop codon, one dot per sequencing library

measured
0.0020.0040.0060.0080.0100.0120.014G418SRI-37240SEPARATES · READTHROUGH SIGNALOVERLAPS · NO READTHROUGH SIGNALgse144140_dmso_rep1_riboseq — 0.00278 (control)gse144140_dmso_rep2_riboseq — 0.00270 (control)gse144140_dmso_rep3_riboseq — 0.00238 (control)gse144140_g418_rep1_riboseq — 0.01208 (treated)gse144140_g418_rep2_riboseq — 0.01413 (treated)gse144140_g418_rep3_riboseq — 0.01049 (treated)gse144140_dmso_rep1_riboseq — 0.00324 (control)gse144140_dmso_rep2_riboseq — 0.00321 (control)gse144140_dmso_rep3_riboseq — 0.00278 (control)gse144140_sri37240_rep1_riboseq — 0.00340 (treated)gse144140_sri37240_rep2_riboseq — 0.00186 (treated)gse144140_sri37240_rep3_riboseq — 0.00323 (treated)ribosomes past the stop (fraction of transcript signal)
  • control (DMSO)
  • treated

PASS · Treated G418 libraries separate from control; SRI-37240 overlaps control.

Details & limits
G418's treated libraries separate completely from their controls; SRI-37240's interleave. The same evidence as above, with nothing summarised away.
04

Recovery

The pipeline finds biology nobody told it about.

An untuned measurement should recover known amino-acid differences in ribosome occupancy.

Occupancy varies by amino acid, in the order the literature reports

Mean relative occupancy per amino acid, both ribosome sites, 61 sense codons

measured
amino acid (one-letter code)0.81.01.21.41.6YEHWDLIKQPSRNGAMFTCVaverage codonY · P site 1.00 (2 codons)Y · A site 1.62 (2 codons)E · P site 0.87 (2 codons)E · A site 1.51 (2 codons)H · P site 0.94 (2 codons)H · A site 1.35 (2 codons)W · P site 0.93 (1 codons)W · A site 1.33 (1 codons)D · P site 1.13 (2 codons)D · A site 1.33 (2 codons)L · P site 0.90 (6 codons)L · A site 1.15 (6 codons)I · P site 0.88 (3 codons)I · A site 1.15 (3 codons)K · P site 0.72 (2 codons)K · A site 1.15 (2 codons)Q · P site 0.99 (2 codons)Q · A site 1.05 (2 codons)P · P site 1.44 (4 codons)P · A site 0.98 (4 codons)S · P site 0.86 (6 codons)S · A site 0.96 (6 codons)R · P site 1.05 (6 codons)R · A site 0.89 (6 codons)N · P site 1.04 (2 codons)N · A site 0.87 (2 codons)G · P site 1.25 (4 codons)G · A site 0.84 (4 codons)A · P site 0.94 (4 codons)A · A site 0.82 (4 codons)M · P site 0.94 (1 codons)M · A site 0.81 (1 codons)F · P site 0.97 (2 codons)F · A site 0.80 (2 codons)T · P site 0.99 (4 codons)T · A site 0.79 (4 codons)C · P site 1.00 (2 codons)C · A site 0.77 (2 codons)V · P site 1.12 (4 codons)V · A site 0.74 (4 codons)relative occupancy (farther right = ribosome waits longer)
  • A site — where the next amino acid is read
  • P site — where the chain is joined

PASS · The pipeline recovers the known amino-acid occupancy pattern without being trained on it.

Details & limits
Nothing was tuned to produce this ordering, so it checks the instrument rather than reporting a finding.
05

Ceiling

The model is nearly as good as the experiment allows.

Model accuracy is judged against the reproducibility ceiling of its own experiment.

Every drug sits close to its own reproducibility ceiling

Held-out accuracy against the agreement between repeats of the same experiment

predicted
DESIRED: BAR ENDS CLOSE TO BLACK CEILING TICK0.00.20.40.60.81.0ClitocineDAPG418SRISJ6986CC90009Clitocine: held-out r² 0.888 against a ceiling of 0.941, over 10 rounds0.89 of 0.94DAP: held-out r² 0.869 against a ceiling of 0.949, over 10 rounds0.87 of 0.95G418: held-out r² 0.763 against a ceiling of 0.953, over 10 rounds0.76 of 0.95SRI: held-out r² 0.760 against a ceiling of 0.970, over 10 rounds0.76 of 0.97SJ6986: held-out r² 0.706 against a ceiling of 0.911, over 10 rounds0.71 of 0.91CC90009: held-out r² 0.557 against a ceiling of 0.771, over 10 rounds0.56 of 0.77prediction accuracy, r² (higher is better; black tick = assay ceiling)
  • held-out accuracy
  • ceiling set by the experiment's own repeats

PASS · Each model approaches the maximum accuracy permitted by that drug's experimental repeats.

Details & limits
The distance that matters is to the black tick, not to 1.0.
06

Our own hypothesis

We tested an idea of ours, and it lost.

Three shuffles ask whether codon speed adds signal beyond local sequence. The decisive control says no.

The shuffle that keeps the amino acid reproduces the gain

Best observed improvement against 199 permutations, per shuffle family

measured
0.0000.0010.0020.0030.0040.0050.006Shuffled inside each geneShuffled everywhereShuffled between synonymous codonspreserves gene compositionchance range: mean 0.00033 ± 0.00040SJ6986: observed gain 0.00356, familywise p = 0.005GAIN EXCEEDS SHUFFLEp = 0.005preserves the distribution onlychance range: mean 0.00072 ± 0.00086SJ6986: observed gain 0.00356, familywise p = 0.02GAIN EXCEEDS SHUFFLEp = 0.02preserves the amino acidchance range: mean 0.00299 ± 0.00067SJ6986: observed gain 0.00356, familywise p = 0.21SHUFFLE EXPLAINS GAINp = 0.21grey band = what chance produces · dashed = the best chance ever reachedimprovement in r² over sequence alone (farther right = more gain)
  • what chance produces
  • hollow — chance explains it

NOT SUPPORTED · The same-amino-acid shuffle reproduces the gain, so codon speed does not explain it.

Details & limits
Shuffling inside genes and shuffling everywhere both fail to reproduce the gain. But shuffling only between codons that carry the same amino acid reproduces most of it — so the signal is which amino acid sits before the stop, not how quickly its codon is read. 199 of 999 permutations are complete, so p-values resolve only to 0.005.
0.21familywise p-value

The decisive control is not significant. The failed hypothesis remains a result.