TL;DR
One target per card: its library, its measured cohort, its predictions — and an explicit flag when there is no measured cohort to calibrate against.
Use it / Skip it
Programs answers the question that decides whether a score means anything: does this target have wet-lab data behind it? A target with a measured cohort has predictions calibrated on its own clones. A target without one is flagged out of domain, and the flag travels with every number produced for it.
Use when
Before trusting any score for a target, and when deciding which target is ready for a scoring-driven campaign.
Don't use for
Do not read a large library as readiness. Library size and measured cohort are different columns for a reason.
Inputs
- Target
- Selected from the program cards; UniProt accession shown on the card.
Outputs
- Library
- How many sequences exist for this target.
- Measured clones
- How many have wet-lab measurements behind them.
- Predictions
- How many predicted values are on record.
- Domain flag
- Whether predictions here come from a model calibrated on a different cohort.
Walkthrough
1. Read measured clones first, not library size
Read measured clones first, not library size.
2. If the card says no measured cohort, treat every score for that target as triage order
If the card says no measured cohort, treat every score for that target as triage order.
3. Pick a target with a cohort when the decision depends on the number itself
Pick a target with a cohort when the decision depends on the number itself.
Under the hood
- The domain flag is derived from whether a calibration exists for this target, not asserted by hand.
- The same flag is attached to individual measurements, so it survives export.
Worked example
Pitfalls
- A target can have a million sequences and zero measurements. The first number does not rescue the second.
- Out of domain does not mean wrong. It means undefended, and it must be shown rather than quietly averaged in.