Fragmentation propensity & scoring modes¶
No trained intensity-prediction model (in the Prosit/MS2PIP sense) exists for intact-protein fragmentation. ProteoformTracker instead uses a literature-grounded, rule-based propensity score per backbone bond, calibrated against real matched b/y ions from published top-down datasets. Every b/y tick on the MS2 fragment ladder is colored by tier and shaded by this score.
What feeds into the score¶
Several effects combine into a per-bond propensity score:
- Residue-pair effects: cleavage N-terminal to proline and C-terminal to aspartate is strongly enhanced under collision-based fragmentation (HCD/CID).
- Positional (terminal-vs-internal) bias: terminal fragment ions statistically carry the majority of fragment-ion current for the first few backbone cleavages from either end, regardless of protein size — cleavages near the termini are weighted above deep-internal ones.
- Protein length: a real, calibrated multiplier — without it, a single very long proteoform's own length systematically suppresses every one of its bonds relative to shorter proteoforms, purely as an artifact of scale, not biology.
- Phospho-proximity: a calibrated multiplier for bonds near a phosphorylation site, fit directly from real proteoforms carrying PTMs in the validation data.
Two scoring modes¶
Pick between them in MS resolution parameters — they are not on the same numeric scale and must never be compared directly against each other's thresholds.
Calibrated (length-aware) — the default¶
The primary, auditable formula: a genuine 0-1 probability, with real fold-enrichment thresholds fit against matched b/y ions from two independent public top-down datasets, including the length multiplier. Tiers and their fold-enrichment over baseline:
| Tier | Score threshold | Fold enrichment |
|---|---|---|
| Elevated | > 0.10 | ~3.1× |
| High | ≥ 0.20 | ~4.7× |
The tradeoff: because length is baked in as a suppressive multiplier, a single very long proteoform can have every bond fall below even the "Elevated" threshold, even though some bonds are still meaningfully more likely to fragment than others within that same protein.
RF ranking (no length) — for long proteoforms¶
A random-forest model trained without a length term, so it stays useful for long proteoforms whose absolute calibrated scores are all suppressed. This is a relative ranking within one proteoform, not a calibrated absolute probability, and its numbers are not comparable in magnitude to Calibrated mode's. Its own tier thresholds:
| Tier | Score threshold | Fold enrichment |
|---|---|---|
| Elevated | > 0.08 | ~2.5× |
| High | ≥ 0.15 | ~3.5× |
Switching modes re-tiers every bond and re-colors the whole ladder; it also changes which bonds qualify for on-demand isotope-pattern inspection (RF mode's qualifying gate is score > 0.08 vs. Calibrated mode's score > 0.10).
Reading the ladder¶
Tick opacity encodes the raw propensity score (within whichever filter level is active); tick color encodes the unique/partial/common tier from comparing against the other checked proteoforms or confounders — see Tier colors for that separate axis. The fragment filter buttons let you show only bonds clearing a given tier, useful for finding which proteoform-distinguishing fragments actually matter.
What this model does not claim¶
This predicts cleavage likelihood, not observed ion intensity — there is no mature, generalizable intensity-prediction model for intact-protein fragmentation to build on. Treat tier labels as "worth checking whether this peak is informative," not as a guarantee any specific ion will actually appear in a real spectrum at a given signal level.