Pep2Prob: Probabilistic Peptide Models

Updated 8 July 2026

Pep2Prob is a family of peptide-centric probabilistic models that assign calibrated probabilities to peptide events in MS-based proteomics.
Its likelihood-based scoring pipeline uses theoretical and observed spectra with supervised maximum likelihood to achieve high calibration and correct identification rates.
Pep2Prob integrates machine learning and large-scale benchmarks to predict proteotypic peptide detectability and fragment-ion occurrence, enhancing protein quantitation.

Searching arXiv for Pep2Prob and related proteomics papers to ground the article in current records. Pep2Prob denotes a peptide-centric probabilistic orientation in MS-based proteomics in which probabilities are assigned to peptide-spectrum matches, peptide detectability, or peptide-specific fragment-ion occurrence, and are then used for identification, calibration, or downstream protein-level analysis. In the literature summarized here, the name is attached to three closely related but distinct formulations: a likelihood-based scoring pipeline for calibrated peptide-spectrum match probabilities, a machine-learning framework for predicting proteotypic peptide detection and absolute protein abundance, and a large-scale benchmark for predicting fragment-ion probability $P(f\mid p)$ for a precursor $p$ in tandem mass spectrometry (Li et al., 2013, He et al., 2013, Xu et al., 12 Aug 2025).

1. Conceptual scope

The unifying object across Pep2Prob formulations is a probability attached to a peptide-level event. In the likelihood-based peptide-identification setting, the target quantity is a posterior probability that a candidate peptide generated an observed spectrum. In the proteotypic-peptide framework, the target is $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ , where $y=1$ denotes “proteotypic” and $\mathbf{x}_p\in\mathbb{R}^{1088}$ is a physicochemical feature vector. In the fragment-ion benchmark, the target is the peptide-specific fragment-ion probability

$P(f\mid p)=\Pr[\text{“ion } f \text{ appears in } p\text{’s MS}^2 \text{ spectrum”}],$

with fragment ion $f=(t,c,n)$ , where $t\in\{a,b,y\}$ is the ion type, $c\in\mathbb{N}$ its charge, and $n$ the cleavage position along the peptide backbone (He et al., 2013, Xu et al., 12 Aug 2025).

These formulations occupy different levels of the proteomics pipeline. One level concerns peptide-spectrum matching; another concerns peptide observability across experiments; a third concerns fragmentation behavior conditional on a precursor. This suggests that Pep2Prob is best understood not as a single software artifact, but as a family of probabilistic peptide models whose outputs can support identification, quantitation, and protein inference.

2. Likelihood-based Pep2Prob for peptide-spectrum match probabilities

In the likelihood-based formulation associated with Li, Eng, and Stephens, one begins with a theoretical spectrum $p$ 0 for a candidate peptide and an observed spectrum $p$ 1. Each theoretical peak is $p$ 2 and each observed peak is $p$ 3, with distinct $p$ 4 and $p$ 5 after peak clustering and preprocessing. The central quantity is the likelihood

$p$ 6

where $p$ 7 is an unobserved emission configuration assigning each theoretical peak either to a matching observed peak or to “no emission” (Li et al., 2013).

The model factorizes as

$p$ 8

For an emission configuration with $p$ 9 emitting theoretical peaks,

$P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 0

Matched observed peaks are modeled by a truncated normal in m/z around the corresponding theoretical peak and a matched-peak intensity density $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 1, whereas noise peaks are distributed uniformly over the m/z range of width $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 2 with intensity density $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 3. The emission probability is parameterized on the logit scale as

$P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 4

with global slope $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 5 and spectrum-specific intercept $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 6 (Li et al., 2013).

Because exact evaluation is combinatorially large, scoring uses the most probable emission configuration:

$P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 7

For a candidate set $P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 8 and a uniform prior over candidates, posterior probabilities are approximated by normalized scores,

$P(\text{detect}\mid p)=P(y=1\mid \mathbf{x}_p)$ 9

Parameter estimation proceeds by supervised maximum likelihood on known spectrum-peptide pairs, alternating between configuration updates and parameter updates for $y=1$ 0 and the spectrum-specific intercepts $y=1$ 1 (Li et al., 2013).

The framework emphasizes calibration as much as ranking. Calibration curves compare nominal posterior bins with empirical correctness, while target-decoy analysis estimates

$y=1$ 2

On the ISB standard mixture, Pep2Prob achieved 96.6% correct top-rank identifications on the full test set, versus 87.6% for the similarity index; among the “confident” subset with $y=1$ 3, 98.7% were correct and 93.1% of spectra were assigned. On spectra whose top-10 SEQUEST candidates included at least one true peptide, the Pep2Prob correct rate was 87.6% overall, versus 80.2% for SEQUEST XCorr and 74.4% for the similarity index, and 98.1% on the $y=1$ 4 subset (Li et al., 2013).