The native-stop safety atlas
Readthrough should cross a premature stop—not a healthy protein's normal stop. This atlas measures that unwanted continuation for G418.
Measurement coverage
The model scores one MANE reference transcript per gene. Direct comparison also needs enough ribosome reads around its normal stop.
Prediction coverage versus prediction-and-measurement coverage, on the same denominator
BLUE CAN TEST THE MODEL · Grey is missing empirical evidence, not evidence of zero native-stop readthrough.
All later results use these 1,700 stops, in one cell line.
Prediction versus measurement
Each dot is one normal stop. Right means higher prediction; up means more measured continuation after G418.
1,700 matched genes · G418 versus control · HEK293T
UP AND RIGHT = MORE UNINTENDED CONTINUATION · Lower-left means lower on these two measures, not proven safe.
1 means identical ranking; 0 means none. The 95% interval is 0.24–0.33.
Where the model disagrees
Agreement means both values fall on the same side of descriptive cuts. Both-high still means greater burden.
Complete partition of the same 1,700 matched normal stops
AGREEMENT SUPPORTS RANKING · Disagreement shows where direct measurement changes the model-only story.
407 stops change category when measurement replaces prediction.
Highest-value next experiment
Only G418 has a measured atlas; five other drugs and every suppressor tRNA remain unknown.
Empirical coverage by therapy; dashed rows mean unknown, never zero
THE SAFETY LAYER IS INCOMPLETE · All suppressor-tRNA designs also lack a qualifying native-stop measurement.
Test one suppressor tRNA against its empty delivery construct: three biological replicates per arm, transcriptome-wide ribosome profiling, no invented pass/fail cutoff.
Direct scope · One suppressor tRNA scaffold in one cell line, measured across that cell line's expressed transcriptome · endpoint · Occupancy past each annotated stop, treated against control, transcriptome-wide