|
Abstract INTRODUCTION:
Clinical application of mass spectrometry (MS) and proteomics is primarily aimed at detecting and measuring targeted proteolytic peptides associated with the biomarker proteins. Intact protein and top-down proteomics approaches can capture global modifications including sequence variation and post-translational modifications (PTM). With high-resolution MS, intact protein methods can be used to profile relative abundance of various proteoforms associated with proteins of interest. Existing analysis workflows are based on exact mass matches to pre-defined protein sequences and modifications, thus have limited capability in biomarkers with multiple sequence variants or variable modifications. Here we developed a flexible bioinformatic workflow to annotate and profile glycosylation features of alpha-fetoprotein (AFP) based on intact protein measurement. The workflow can establish a proteoform database to annotate glycoforms, calculate their relative abundance, utilize machine learning to select relevant features, and detect differentiating features relevant to patient characteristics.
METHODS:
AFP was first enriched from human serum or plasma samples via immune-enrichment workflow. Chromatographic separation was performed using PLRP 300Å column (Agilent), followed by intact protein analysis on an Exploris 480 Orbitrap (Thermo). Multiple charge envelopes corresponding to chromatographic peaks of AFP were deconvoluted using the ReSpec algorithm in BioPharma FinderTM software (Thermo). The deconvolution was performed with target m/z range between 500-3200 Da. Intact masses, sum peak intensities, retention time, and deconvolution scores were exported.
An initialized glycan residues database was audited based on a small training patient dataset (n=30), and non-viable glycan residues were removed based on the feasibility of the composition. To generate initial AFP proteoforms database, we consider 60 combination of mammalian glycan modification and a variable modification of oxidation (for methionine). The glycan modifications were further reduced to 43 based on structural viability. Then deconvoluted peaks were compared with AFP proteoform database by binary search. There were two rounds for the search with a mass tolerance of 7 Da: the first was based on wild type AFP; if we could not find any match, we continued to do second-round search on AFP with oxidized methionine. Furtherly, an individualized retention time (RT) window was applied to each sample to rule out deconvoluted peaks outside of the set tolerance. For each mass assigned as an AFP proteoform, we calculated minimum mass, maximum mass, mean of mass, range of mass, accuracy of mass mean, and frequency of proteoforms among all examined patient samples.
Consequently, the combinations of features were tested by K-Mean elbow method, which could determine the optimal number of clusters(k). As a result, dimensionality reduction algorithms such as PCA, t-SNE, and UMAP were used for the feature extraction. Additionally, LightGBM (Light Gradient-Boosting Machine) model was explained with SHAP (SHapley Additive exPlanations) values for the feature selection. Finaly, differentiating glycan features were associated with different disease phenotypes.
RESULTS:
Among 242 authentic patient samples, there were 2789 deconvoluted intact mass peaks. After database search, 1242 peaks were annotated, which correspond to 59 AFP proteoforms (40 corresponding to wild type AFP, and 19 to AFP with oxidized methionine). The average deconvoluted mass accuracy observed was -6.28 Da, with a standard deviation of 2.53 Da. We use AFP_5401 to represent the AFP glycoprotein with glycan composition of 5 “H” (Mannose or Galactose), 4 “N” (N-Acetylglucosamine), 0 “F” (Fucose), 1 “A” (N-Acetylneuraminic Acid, sialic acid); where “5401” represents the total numbers of “HNFA”, respectively. Four most identified glycoforms were AFP_5401, AFP_5402, AFP_5411, AFP_5412, with frequency of 55.8%, 72.7%, 68.6%, and 66.5%.
RT window was used to further rule out deconvoluted peaks outside of the set tolerance in each sample. For each annotated AFP proteoforms, we defined confidence score as 100*(1-abs(tm-em)/w), where tm denotes theoretical mass, em denotes experimental mass, and w denotes mass tolerance 7 Da. For each sample, we defined the RT window based on the proteoforms with confidence score above 60, use mean of RTs from these selected proteoforms as a baseline, and a tolerance of 0.3 was set to filter out other proteoforms. After filtering, there were 116 false positive proteoforms filtered across these 1242 peaks. As a result, the total number of remaining annotated proteoforms was 1126. The glycosylation features were extracted, to include the total intensity of peaks corresponding to the identified glycoforms, fractional abundance of AFP_5401, AFP_5402, AFP_5411, AFP_5412, AFP-L3, S1, S2 in each of the samples. Where AFP-L3 represents the glycoproteins contain a single fucose (F), S1 represents the glycoproteins contain a single sialic acid (A), S2 represents the glycoproteins contain 2 A’s.
Among all study samples, seven consistent AFP glycoforms were identified, while their relative abundances varied.
CONCLUSION:
The developed algorithm involved establishing a proteoforms database based on training patient dataset and filtering based on RT windows. Feature extraction and feature selection were used to improve model performance by handling redundant or irrelevant features. The results demonstrate that the developed algorithm allows to facilitate data processing to identify protein glycosylation profiles and extract pertinent features for disease association. To the best of our knowledge, our algorithm is the first flexible bioinformatics pipeline to accommodate fit-for purpose proteoforms database and RT filter to annotate PTM of targeted proteins in intact protein MS. |
= Discovery stage. (53.14%, 2025)
= Translation stage. (22.33%, 2025)
= Clinically available. (24.53%, 2025)
