1
Experimental glycoprofiles
30 glycoengineered CHO lines (17 glycosyltransferase knockout/knock-in combinations) produced 10 recombinant proteins. HILIC-UPLC yielded 309 N-glycan profiles and 68 structures.
Why a lectin profile is not a full glycoprofile
Lectins recognize epitopes (local glycan features), such as core fucose, α2,6-sialylated LacNAc, biantennary structures, or bisecting GlcNAc. Figure 1 of the paper simulated lectin signals from 2,912 random glycoprofiles and found that lectin profiles with Pearson R > 0.95 could still map to at least seven distinct N-glycan compositions. The same lectin pattern can therefore correspond to many intact glycan mixtures.
LeGenD (Lectin to Glycoprofile ENhanced with Data-driven methods) does not simply add more lectins. It learns the constraints of glycoprofiles that cells actually produce. The CHO enzyme repertoire, monosaccharide supply, and protein steric limits shrink the possible structure space, so a neural network can map an 8-dimensional lectin vector onto relative abundances of 68 N-glycans.
Lectin array
Many lectins are immobilized on a chip; fluorescent glycoproteins or lysates bind to give a high-throughput RFU map. The paper notes that this format can also use LeGenD, but it needs its own training set.
ELLA / ELISA
The main assay in the paper is an enzyme-linked lectin assay: coat protein in 96-well plates, add biotin-lectin → ABC-HRP → TMB, and read at 450 nm. Signals are background-subtracted and normalized to sum to 1.
FACS
Fluorescent lectins stain the cell-surface glycocalyx; MFI is the binding intensity. The relative lectin vector is the same idea as ELLA, but cell surfaces also carry O-glycans and glycolipids, so interpretation is more cautious.
Original model: 8 lectin inputs → 4 fully connected hidden layers (20 nodes each) → softmax over 68 N-glycan abundances. Best validation RMSE was about 3.60×10−2. SHAP interpretation found LCA and MAL-I most important for IgG. This teaching page cannot ship the unpublished ANN weights, so it uses the paper's Equations 1–2 (glycan-feature matrix × binding-rule matrix) for an interpretable reconstruction, with the same 0.02 denoising threshold.
How the database was built from the paper
The core data gap is that almost no public set has both full UPLC glycoprofiles and experimental lectin profiles on the same samples. The authors therefore used experimental glycoprofiles as ground truth and simulated lectin signals as training inputs, then calibrated with real ELLA of IgG and Fetuin B.
2
Feature decomposition
Each glycan is counted into LinearCode features: core Fuc, terminal Manα1,2, α2,3/α2,6 Sia, bisecting GlcNAc, β1,6 branching, terminal type-2 LacNAc, and others.
3
Binding rules
Specificities of eight lectins for about ten N-glycan features come from Bojar et al. 2022 (machine-learning annotation of 57 lectins). P-values were converted to z-scores to build the binding-rule matrix (Table S3).
4
Simulated lectin profiles
LP = GP × (feature matrix × binding rules). Linear regression against experimental ELLA of IgG / Fetuin B calibrated the simulation. IgG RMSE = 2.63×10−3; Fetuin B = 3.85×10−3.
5
Training and expansion
The ANN learns simulated lectin → true glycoprofile. Thirty Fetuin-like profiles were also drawn from a Dirichlet distribution (~10% of the training set) to reduce false positives on unseen structures.
Validation proteins (also used as demos here)
- Human serum IgG: Mostly core-fucosylated biantennary glycans (G0F / G1F / G2F), with some bisecting and α2,6-sialylated forms. High LCA, low MAL-I and PHA-L. The same IgG lectin ELISA can also estimate Glycan Age.
- Bovine Fetuin B: Highly sialylated tri-/tetra-antennary complex N-glycans with almost no core Fuc. High MAL-I, SNA, and PHA-L; low LCA. The model cannot tell trisialylated glycans from polysialic acid (neither SNA nor MAL-I binds α2,8).
Source: Li H, Peralta AG, Schoffelen S, et al. LeGenD: determining N-glycoprofiles using an explainable AI-leveraged model with lectin profiling (bioRxiv 2024; PMC10996628). Binding rules from Bojar et al., ACS Chem. Biol. 2022。
Select an assayed lectin to list matching glycan structures
This table follows Table S1: eight lectins jointly cover about ten common N-glycan features. Click a lectin card or use the dropdown.
Enter lectin array / ELISA / FACS data to reconstruct a glycan profile
Relative intensities are enough. The tool subtracts background if provided and normalizes the lectin vector to sum 1, as in the paper's ELLA processing. Empty lectins are ignored.
Or paste CSV / TSV
Column names may be lectin, signal / intensity / MFI / OD / RFU, plus optional background.
Worked examples (from the paper and common N-glycomes)
Demo lectin signals are simulated from literature-style reference glycoprofiles with the same forward model as the paper, so reconstructions can be compared with the reference mix. This also recreates LeGenD's logic of training on simulated lectin profiles.