Peptide Property Prediction for Mass Spectrometry Using AI: An Introduction to State of the Art Models

Abstract

This review explores state of the art machine learning and deep learning models for peptide property prediction in mass spectrometry-based proteomics, including, but not limited to, models for predicting digestibility, retention time, charge state distribution, collisional cross section, fragmentation ion intensities, and detectability. The combination of these models enables not only the in silico generation of spectral libraries but also finds many additional use cases in the design of targeted assays or data-driven rescoring. This review serves as both an introduction for newcomers and an update for experienced researchers aiming to develop accessible and reproducible models for peptide property predictions. Key limitations of the current models, including difficulties in handling diverse post-translational modifications and instrument variability, highlight the need for large-scale, harmonized datasets, and standardized evaluation metrics for benchmarking.

Keywords: deep learning; machine learning; mass spectrometry; peptide property prediction; proteomics.

FIGURE 1

Citations over time for different digestibility models. As of February 18, 2025. The cutoff years are 2010 and 2024. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 2

Schematic architecture of DeepDigest [25].

FIGURE 3

Citations over time for different retention time models. As of February 18, 2025. The cutoff years are 2010 and 2024. Chronologer is a preprint. Prosit includes publications from 2019, 2022, and 2024 (preprint). Elude includes publications from 2010 and 2012. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 4

Schematic architecture of DeepLC [48].

FIGURE 5

Citations over time for different charge state and charge state distribution models. As of February 18, 2025. The cutoff years are 2010 and 2024. AlphaPeptDeep is not included as its predictor has not been part of a publication yet. CPM is an abbreviation for Charge Prediction Machine. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with: https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 6

Schematic architecture of CPred [63].

FIGURE 7

Citations over time for different ion mobility models. As of February 18, 2025. The cutoff years are 2010 and 2024. McKetney et al. is a preprint. DeepCCS is an abbreviation for DeepCollisionalCrossSection. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with: https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 8

Schematic architecture of ionmob [86].

FIGURE 9

Citations over time for different fragmentation ion intensity models. As of February 18, 2025. The cutoff years are 2010 and 2024. MS2PIP Original includes publications from 2013 to 2015. Prosit includes publications from 2019, 2021, 2022, 2024, and 2024 (preprint). pDeep includes publications from 2017 and 2019. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with: https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 10

Schematic architecture of Prosit [51].

FIGURE 11

Citations over time for different detectability (flyability) models. As of February 18, 2025. The cutoff years are 2010 and 2024. Pepper is the only flyability model. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with: https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 12

Schematic architecture of DeepDetect [102]. *Not clearly defined in the publication but assumed to be word embedding.

FIGURE 13

Schematic architecture of PepFormer [101].

FIGURE 14

Schematic architecture of Pepper [103].

FIGURE 15

Citations over time for different hydrophobicity models. As of February 18, 2025. The cutoff years are 2010 and 2024. ALOGPS includes publications from 2001, 2001, 2002, 2004, and 2004. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with: https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].

FIGURE 16

Citations over time for different 3D structure models. As of February 18, 2025. The cutoff years are 2010 and 2024. tFold includes publications from 2021, 2022 (preprint), and 2024 (preprint). Chai‐1 is a preprint. RoseTTAFold includes two publications from 2021. Omega‐Fold is a preprint. References to all models can be found in Table S1. Models using deep learning (DL) are blue. Models using only classical machine learning (ML) are gray. The exact values of the citation counts per tool and year can be found in Table S2. Created with: https://github.com/jesseangelis/Citation_vis/ and OpenAlex [32].