N‐glycan fingerprint predicts alpha‐fetoprotein negative hepatocellular carcinoma: A large‐scale multicenter study

Abstract

Alpha-fetoprotein (AFP)-negative hepatocellular carcinoma (ANHCC) patients account for more than 30% of the whole entity of HCC patients and are easily misdiagnosed. This three-phase study was designed to find and validate new ANHCC N-glycan markers which identified from The Cancer Genome Atlas (TCGA) database and noninvasive detection. Differentially expressed genes (DEGs) of N-glycan biosynthesis and degradation related genes were screened from TCGA database. Serum N-glycan structure abundances were analyzed using N-glycan fingerprint (NGFP) technology. Totally 1340 participants including ANHCC, chronic liver diseases and healthy controls were enrolled after propensity score matching (PSM). The Lasso algorithm was used to select the most significant N-glycan structures abundances.Three machine learning models [random forest (RF), support vector machine (SVM) and logistic regression (LR)] were used to construct the diagnostic algorithms. All 13N-glycan structure abundances analyzed by NGFP demonstrated significant and was enrolled by Lasso. Among the three machine learning models, LR algorithm demonstrated the best diagnostic performance for identifying ANHCC in training cohort

Serum N-glycan profiling as a diagnostic biomarker for the identification of hepatitis B virus-associated hepatocellular carcinoma

“…Moreover, Log (peak 9/peak 7) combined with AFP had the best accuracy in the diagnosis of HCC. Due to the difficulty of ultrasound imaging in diagnosing HCC in patients with LC, serum glycan markers can be used as a valuable supplement to AFP in the diagnosis of HCC (30). The main limitation of this study is the small sample size.…”

Development of Machine Learning-Based Personalized Predictive Models for Early Detection of Hepatocellular Carcinoma in HBV-Related Cirrhosis Patients with Low Levels of Serum Alpha-Fetoprotein

“…Our model showed a satisfactory calibration curve and excellent agreement between the predicted and observed in the external test set and is comparable with other HCC risk prediction models. Prior attempts to increase the accuracy of HCC prognostic and diagnostic prediction have mostly relied on tissue-based, genomics, or imaging-assisted quanti cation of research biomarkers (29)(30)(31)(32)(33) However, there were several limitations to this study which should also be noted. The rst limitation was that we used a retrospective, single-center design, and thus selection bias was unavoidable.…”

Glycoinformatics in the Artificial Intelligence Era

“…119,122 Trained models in these tasks not only can be used to infer the properties of newly discovered glycans but also can be used to retrieve motifs that are important to endow a glycan with a property, such as human-like glycan motifs used in molecular mimicry by pathogenic Escherichia coli strains. Other notable efforts in this area, using machine learning, are the investigation of the association of clinical characteristics with glycans from cancer patients 138 or the prediction of αfetoprotein (AFP)-negative hepatocellular carcinoma using glycan fingerprints, 139 glycobiology is the lack of labeled data of glycan sequences with known information about their properties or functions that could be used to train a model. Often this information may exist in sufficient quantity across the literature yet is scattered and would require exhaustive manual curation that may be prohibitively expensive.…”

“…Current examples of predicting glycan properties from sequences include prediction of glycan class, taxonomy, immunogenic potential, and association with bacterial pathogenicity. , Trained models in these tasks not only can be used to infer the properties of newly discovered glycans but also can be used to retrieve motifs that are important to endow a glycan with a property, such as human-like glycan motifs used in molecular mimicry by pathogenic Escherichia coli strains. Other notable efforts in this area, using machine learning, are the investigation of the association of clinical characteristics with glycans from cancer patients or the prediction of α-fetoprotein (AFP)-negative hepatocellular carcinoma using glycan fingerprints, both of which could offer a promising target for deep learning in the future. The key limiting factor in all applications of this type of supervised learning in glycobiology is the lack of labeled data of glycan sequences with known information about their properties or functions that could be used to train a model.…”