Glydentify: An explainable deep learning platform for glycosyltransferase donor substrate prediction

22 Glycosyltransferases (GTs) are a large family of enzymes that catalyze glycosidic linkages formation 23 between chemically diverse donor and acceptor molecules to regulate diverse cellular processes across all 24 domains of life. Despite their importance, the activated sugar donors (donor substrates) used by most GTs 25 remain unidentified, limiting our understanding of GT functions. To address this challenge, we developed 26 Glydentify, a deep learning framework that predicts donor usage across GT-A and GT-B fold 27 glycosyltransferases. Trained on large-scale UniProt annotations, Glydentify integrates protein sequence 28 embeddings learned from protein language models with chemical features derived from molecular 29 encoders trained on extensive chemical datasets. The resulting models achieve high predictive 30 performance, with precision–recall AUCs (PR-AUC) of 0.86 for GT-A and 0.91 for GT-B, surpassing 31 general enzyme–substrate predictors while requiring minimal manual curation. We employed Glydentify 32 to predict the donor specificity of uncharacterized plant GTs and experimentally tested the predictions 33 using in vitro biochemical assays. Furthermore, we demonstrate that the model utilizes a combination of 34 evolutionary, structural, and biochemical features to predict donor specificity through residue attention 35 score analysis. Together, these results establish Glydentify as a robust, explainable framework for 36 decoding donor-glycosyltransferase relationships and highlight its potential as a broadly applicable 37 framework for modeling enzyme classes that act on chemically diverse substrates. 38

39 Glycosylation is a ubiquitous and fundamental mechanism for regulating diverse biological processes 40 across all domains of life 1,2. The diverse biological roles of glycosylation are directly linked to the 41 remarkable chemical diversity of glycans3,4. Glycosylation is catalyzed by glycosyltransferases (GTs), a 42 large family of enzymes that transfer sugar moieties from activated nucleotide-sugar donors to a wide 43 range of acceptor molecules, including proteins, lipids, small molecules, and other carbohydrates. Over 44 evolutionary time, GTs have diversified to utilize a broad spectrum of chemically distinct nucleotide 45 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 3 sugars, shown in Error! Reference source not found.. Two major structural folds, GT-A and GT-B, 46 dominate the known GT landscape and rely on nucleotide sugars as donors5. Despite their central role in 47 glycan biosynthesis, donor usage remains unknown for the vast majority of putative GTs. The CAZy 48 database catalogs more than 700,000 GT entries from diverse organisms (to this date), grouped into 139 49 families. However, only a small fraction of these families have been experimentally characterized with 50 respect to their donor substrates. An incomplete understanding of GT sugar donor specificity presents a 51 major challenge in engineering GTs for biotechnological applications, such as antibody engineering, 52 vaccine development, and modulation of natural-product pharmacokinetics1. 53 Biochemical assays for donor usage are costly, low-throughput, and generally limited to a narrow panel of 54 candidate sugars, making comprehensive mapping impractical. Moreover, donor specificity cannot be 55 inferred reliably from sequence or structure alone. Sugar donor specificity is often governed by spatially 56 dispersed, co-evolving residues and epistatic interactions, which cannot be inferred through simple 57 sequence patterns or structural comparisons5. For instance, enzymes adopting entirely distinct structural 58 scaffolds, such as cellulose synthase from family GT2 (GT-A fold)6 and glycogen synthase, from family 59 GT3 or GT5 (GT-B fold)7, can both bind UDP-glucose; in contrast, in the ABO blood group transferase, 60 only four amino acid substitutions convert specificity from UDP–GalNAc to UDP–Gal 8. This challenge 61 is further compounded by the high physicochemical similarity of nucleotide-sugar donors, which share a 62 conserved nucleotide moiety (e.g., UDP) and differ only by subtle stereochemical changes at one carbon 63 in the hexose ring(Figure 1A). This structural similarity presents a significant challenge for predictive 64 modeling, as enzymes often exhibit strict stereochemical stringency, where even minor stereochemical 65 variations (e.g., the C4-epimerization between glucose and galactose) can completely abolish substrate 66 recognition. Such cryptic determinants highlight why homology-based annotations are inadequate for 67 predicting sugar donor specificity. Thus, there is a need for specialized tools to advance glycoenzyme 68 function prediction. 69 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 4 Previous computational approaches for predicting sugar donor specificity in GTs largely relied on 70 statistical models and handcrafted features, restricting their application to specific families or fold-types. 71 Yang et al.9 introduced GT-Predict to model Arabidopsis GT1 enzymes with decision trees that combined 72 physicochemical descriptors of both donors and acceptors. Taujale et al.10 predicted donor classes within 73 GT-A enzymes using residue-level physicochemical descriptors derived from curated sequence 74 alignments. Moreover, many supervised learning approaches have framed such problems as single-label 75 classification tasks, assuming one donor per enzyme. For instance, Hennen et al.11 trained GT-B 76 classifiers on curated data using sequence properties and AF2-based structural features, but assumed that 77 one GT could only utilize one donor sugar. This overlooks the biological reality that some GTs exhibit 78 dual or promiscuous donor specificity. While demonstrating feasibility, these models remain constrained 79 by the reliance on handcrafted input features, limited scope, and single-label formulations. 80 Conversely, the advent of deep learning and protein language models has revolutionized protein function 81 prediction tasks by learning biological representations directly from exiting data in an unsupervised way, 82 eliminating the need for manual feature curation. Initial efforts in the glycoenzyme space, such as 83 DeepGlycanSite12, GlyNet13, SweetNet14 , and CandyCrunch, leveraged Graph Neural Networks (GNNs) 84 to capture the subtle sequence–structure interactions governing glycan–protein interactions. Furthermore, 85 protein-language models like ESM-215 utilize self-supervised methods on millions of publicly available 86 sequences to extract context-aware representations for downstream tasks. Such pre-trained protein 87 language models have been successfully adapted by enzyme-substrate predictors such as ESP16 and 88 EZspecificity17. More recent protein-language models, such as ESM-18 and SaProt19 , further extend the 89 paradigm by incorporating explicit structural information during pretraining. 90 Despite these recent advancements in protein function prediction, representation learning approaches have 91 not yet been fully harnessed for GT donor substrate prediction. General frameworks, such as ESP16 , 92 broadly cover enzyme substrates, and the enzyme specificity predictor EZspecificit17 has been tested on 93 some glycosyltransferase acceptor substrates. However, these models lack the resolution to predict GT 94 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 5 donor specificity at scale, primarily due to the limited GT data available in the general enzyme databases 95 used for training. Additionally, these models lack interpretability, limiting their applications in 96 investigating the relationships connecting sequence, structure, and function in GTs. To bridge this gap, 97 we develop Glydentify, an explainable evotuned deep learning framework trained on GT-A and GT-B 98 fold families from diverse organisms. The model leverages pre-trained protein and donor embeddings in 99 an end-to-end framework, utilizing bi-directional cross-attention to support multi-label donor prediction 100 across different GT families and fold classes, while learning from existing data in an unsupervised 101 manner. Unlike earlier approaches, Glydentify does not rely on handcrafted descriptors or sequence 102 alignments; instead, it learns directly from the intrinsic structure and context of GTs and their donor 103 substrates. The framework estimates the probability of each donor sugar independently, enabling the 104 identification of enzymes with multiple donor specificities. Trained on large-scale UniProt annotations 105 and evaluated against CAZy entries, Glydentify achieves state-of-the-art predictive performance while 106 offering interpretability in identifying protein residues that contribute to donor prediction. We 107 demonstrate the application of Glydentify in predicting and experimentally evaluating donor substrates in 108 a subset of GT-A and GT-B plant transferases not used in training. By establishing a robust and scalable 109 baseline for donor prediction, Glydentify offers a practical tool for functional annotation of GTs and a 110 tool for guiding experimental discovery in glycobiology. 111 112

113 Glydentify: Attention-guided weighting of GT and donor fused representation 114 To address the challenges of predicting GT donor specificity at scale, we developed Glydentify, a 115 structure-aware deep learning framework that integrates protein and donor information. While explicitly 116 modeling enzyme-substrate complex structures can offer the most rigorous basis for defining specificity17, 117 the scarcity of experimentally verified complex structures makes it difficult to train a deep learning 118 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 6 model. Moreover, computational surrogates could introduce noise and eventually compromise a model's 119 performance. Generative models like AlphaFold 320 frequently exhibit stereochemical inaccuracies21,22, 120 whereas molecular docking optimizes for ground-state binding stability rather than the transition-state 121 geometry required for turnover. In contrast, Glydentify captures the high-dimensional chemical logic of 122 donor recognition without relying on these noisy structural priors. Two separate predictors were trained 123 for GT-A and GT-B enzymes. Each predictor has the same architecture as shown in Figure 1B. A 124 pretrained protein language model encodes each GTs information at the residue level. In parallel, a 125 pretrained molecule encoder represents donor sugars at atomic resolution. To integrate these modalities, 126 we employ a bi-directional cross-attention module that enables residues and donor atoms to exchange 127 information, assigning attention weights to prioritize informative features for donor recognition. A final 128 classifier produces probability scores for each GT-donor pair, which, when concatenated across donors, 129 yields a multi-label output vector for each GT. This design allows Glydentify to learn directly from 130 protein sequences and donor molecules, bypassing the need for handcrafted features or sequence 131 alignments. 132 We systematically evaluated different protein language models to understand how sequence-versus 133 structure-aware representations influence donor sugar prediction. We hypothesized that protein language 134 models trained under different objectives capture distinct biochemical features of glycosyltransferases, 135 leading to variations in how donor-recognition patterns are represented and ultimately predicted. 136 Specifically, ESM-2 was trained on large-scale natural protein sequences to capture evolutionary co-137 variation, ESM-C was contrastively trained to emphasize structural and physicochemical consistency, and 138 SaProt incorporated AlphaFold-derived embeddings to encode explicit three-dimensional geometric 139 information. To further enhance structural awareness across all representations, we generated 140 AlphaFold320 predictions for all training and test sequences. These were used to enrich the SaProt 141 sequence embeddings with structural context during the representation learning process. Donor molecules 142 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 7 were encoded from SMILES strings into 3D conformers using RDKit and UniMol V2, ensuring that the 143 chemical diversity of nucleotide sugars was consistently represented. 144 We trained Glydentify on a large set of UniProt entries annotated with donor usage (Figure 1C). To 145 reduce label noise, only entries with high annotation scores were retained, and each entry was categorized 146 as GT-A or GT-B based on HMM profiles. To prevent redundancy, we clustered sequences within each 147 donor type, and then selected representatives at 95–100% identity thresholds depending on donor sample 148 size. For evaluation, we constructed a benchmark test set from CAZy by extracting experimentally 149 verified GT–donor pairs, clustering at 90% identity, and removing any training homologs with > 90% 150 sequence identity to any test sequence. This split ensures that performance is assessed on clean labels and 151 different sequence similarities. 152 For each enzyme, the model outputs an independent probability for every donor in the vocabulary. 153 Training was performed under a closed-world assumption, where unannotated donors are treated as 154 negatives. Although this may introduce occasional false negatives, the effect is mitigated by the large 155 ratio of true negatives to positives, providing a consistent and reproducible setup across donor classes. 156 With this framework and dataset in place, we next evaluated Glydentify’s global performance on the 157 CAZy benchmark, comparing it to baseline sequence-only models and to a general enzyme–substrate 158 predictor. 159 Glydentify outperforms existing machine learning models for donor substrate 160 prediction 161 We assessed the performance of Glydentify by comparing its results with existing general-purpose 162 models. To further validate the effectiveness of the architecture, we also tested a standard classification 163 pipeline consisting of a pre-trained protein encoder backbone with an MLP head trained on our own data. 164 We found that the general enzyme substrate predictors ESP16 and EZspecificity17 perform only slightly 165 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 8 above random, reflecting the extreme sparsity of glycosyltransferase-donor data in its training data 166 (Figure 2A). This limitation motivated our curation of a focused GT donor dataset. A simple pre-trained 167 encoder with a multi-label classification head trained on our UniProt-derived GT sequences already 168 delivers strong gains (PR-AUC ≈ 0.70, ROC-AUC ≈ 0.80, MCC ≈ 0.63), indicating that sequence 169 information alone carries substantial signal. When integrating the rich-context embedding from the pre-170 trained protein encoder and small molecule encoder via cross-attention, the performance can be further 171 improved. Although the overall performances across different encoder choices are similar, Glydentify 172 with SaProt and UnimolV2 achieves the best performance across all metrics (PR-AUC: 0.82–0.90, ROC-173 AUC: 0.94–0.97, MCC: 0.71–0.81). Furthermore, these performance gains are consistent on both GT-A 174 and GT-B fold enzymes, demonstrating the models ability to resolve donor specificity under heavy class 175 imbalance. 176 Having established that models trained on a GT-specific training set substantially outperform boths 177 general enzyme-substrate predictors, we next examine how performance varies across biologically and 178 data-driven subgroups. This is critical since global averages can conceal systematic weakness on under-179 represented donors, rare taxa or remote homologs. Therefore, we stratified the CAZy test sets along three 180 different categories: sequence novelty (percent identity to the nearest training example, binned at 0–20%, 181 20–40%, 40–60%, 60-80%, 80-90%), donor sugar type, and taxonomic origin (kingdom). For each 182 subgroup and fold (GT-A/GT-B), we report PR-AUC as the primary metric. To aid interpretation, each 183 panel overlays the number of (positive) samples of that subgroup in the training set and testing set. 184 Subgroups with fewer than 10 test positives are excluded because PR-AUC becomes unstable at very 185 small values. Overall, our data showed that Glydentify outperforms the protein-encoder-only baselines 186 across all donor classes that meet the inclusion criterion. Across both folds, PR-AUC increases with 187 sequence identity, and Glydentify (red) outperforms the ESM-2–only baseline (blue) in every bin that 188 meets the inclusion criterion (Figure 2B). Gains are largest in the low–identity regime, which is precisely 189 where the distribution shift is strongest. The monotonic trend with identity and consistent gains in the 20–190 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 9 60% bins indicate that cross–attending SaProt embeddings with UniMol donor geometry mitigate 191 generalization gaps to remote homologs, while maintaining near–ceiling performance when close 192 sequence neighbors are present. For performance across different donor sugars, while the model achieves 193 high PR-AUC for well-represented donor types (Figure 2C), we observed a distinct performance drop-off 194 for rare donors (N<50) within the training set (Supplementary Table 2), similar to the limitations 195 observed in general enzyme-substrate models. To systematically define this data dependency, we 196 performed a few-shot fine-tuning analysis (Supplementary Text 1). Both the minor-donor evaluation and 197 the few-shot experiments converge on a similar finding: successful generalization requires a critical 198 threshold of annotated examples (approx. N>50), below which the structural logic of donor specificity 199 cannot be robustly inferred from sequence alone. As for the performance on different organisms, across 200 both folds, Glydentify consistently exceeds the ESM-2–only baseline (blue) in the major, well–201 represented clades (Figure 2D). Gains are largest where we have ample supervision or a clear donor–202 sequence signal. Performance on Metazoan and plant (Viridiplantae) GTs reaches near–ceiling PR–AUC, 203 with substantial improvements for GT–B fold enzymes (e.g., Metazoa +∼0.19 PR–AUC, Viridiplantae 204 +∼0.08). For proteins that adopt a GT-A fold, predictions on fungal GTs showed a pronounced 205 improvement (+∼0.25), while predictions on bacterial GTs (Pseudomonadota) improved modestly. 206 Mixed outcomes appear in very small bacterial/viral groups (e.g., Bacillota, Mycoplasmatota); their PR–207 AUCs have high variance due to the small number of positive training samples in those categories, so we 208 did not overinterpret those fluctuations. Overall, the kingdom-stratified analysis indicates that explicit 209 protein–donor interaction modeling transfers well across eukaryotic taxa and moderately across bacteria, 210 with variability in performance likely reflecting biological divergence among kingdoms rather than 211 differences in dataset size. 212 Evaluation on independent datasets demonstrates Glydentify’s 213 generalizability across diverse species and families 214 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 10 We evaluated Glydentify's performance on two independent datasets spanning distinct GT folds and GT 215 families to assess its generalizability beyond the training distribution (Table 1). For GT-A fold proteins, 216 we curated a literature-based set of experimentally validated GTs with known donor specificities. For 217 enzymes that adopt a GT-B fold, we focused on Spirodela polyrhiza and Spirodela intermedia enzymes 218 from the GT-47 family. The selection of plant GT47 enzymes for this study was due to their broad 219 functional diversity. More specifically, plant GT47 members are known to have the capacity to utilize a 220 diverse array of activated nucleotide sugar donor substrates, including UDP-β-L-Araf, UDP-β-L-Arap, 221 UDP-α-D-Xyl, UDP-α-D-Gal, and UDP-α-D-GalA. Furthermore, they are known or hypothesized to be 222 involved in the synthesis of all major classes of plant cell wall matrix polysaccharides; however, many of 223 their donor and acceptor specificities remain unknown, making functional characterization efforts 224 difficult23. To validate the donor prediction accuracy of the Glydentify model within this family, 225 candidate enzymes were heterologously expressed in HEK 293 cells and purified. Nucleotide sugar donor 226 specificity was subsequently determined by quantifying the release of UDP in the absence of an acceptor, 227 a byproduct of GT activity after a sugar from the activated donor is transferred to water24, providing an 228 experimental benchmark for the model's predictive performance. Enzymes that exhibited no hydrolytic 229 activity in the absence of acceptor were assayed with an appropriate acceptor substrate to identify donor 230 specificity. 231 Despite low sequence similarity to the current training set (Table 1), Glydentify accurately predicted 232 donor sugar identity across the newly characterized enzymes. Notably, the recently described CAZy 233 family GT-139 enzyme Cgm1 (Uniprot ID: J9VFZ3), a mannosyltransferase involved in the biosynthesis 234 of capsular glucuronoxylomannogalactan in Cryptococcus neoformans25, was correctly assigned GDP-235 Man as its preferred donor with 90% prediction score (Figure 3A). This result demonstrates Glydentify's 236 robust ability to generalize to even newly added CAZy families. 237 For those with GT-B folds, Glydentify correctly identified UDP-Xyl as the most probable donor for 238 Sp124-E, Sp415-C, and Sp124-C, with prediction scores ranging from 73.4% to 93.3%, in full agreement 239 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 11 with our biochemical assays (Figure 3C, 3D and 3E). Likewise, Sp146-A and Sp197-A, were predicted 240 to use UDP-Gal with high confidence (99%), again in agreement with biochemical evidence where those 241 were characterized to act on UDP-Gal26(Supplementary Figure 1). We also observed that B3GLCT, a 242 member of the GT31 family, was not correctly predicted despite its high sequence similarity to other 243 GT31 enzymes in the training set (Supplementary Figure 2). 244 Glydentify leverages structural, evolutionary, and biochemical cues for donor 245 recognition 246 To assess whether Glydentify learns generalizable principles of donor recognition, we conducted a series 247 of interpretability analysis inspired by prior works27 demonstrating that attention-based models can 248 recover latent structural determinants of enzyme function. In most enzymes, substrate specificity arises 249 from a complex interplay of interactions: first-shell contacts that form direct hydrogen bonds, ionic 250 interactions, or steric complementarity with the ligand; second-shell residues that do not contact the 251 substrate directly28 and even long-range or allosteric effects that influence active-site dynamics through 252 cooperative networks across the protein scaffold29. These multi-layered interactions collectively shape 253 substrate binding and catalysis, and often cannot be inferred from primary sequence alone. We therefore 254 examined whether Glydentify, trained solely on sequence embeddings, nevertheless recapitulates these 255 canonical principles of enzyme–substrate recognition in its internal representation of donor-sugar 256 specificity. 257 We analyzed the attention scores derived from the cross-attention module to identify which residues drive 258 donor-specific recognition. From each selected sample, we extracted the per-residue attention weights 259 (α! (#)) from the output of the cross-attention layer, where 𝑖 the indices correspond to residues in the 260 aligned sequence. To investigate what signals Glydentify relies on to infer donor sugar specificity, we 261 mapped the residue-level attention scores onto predicted three-dimensional structures and analyzed their 262 spatial distribution. 263 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 12 We first performed a global analysis across true-positive GT–donor pairs by plotting the spatial 264 distribution of high-attention residues (normalized attention score > 0.8) relative to the donor molecule. 265 Interestingly, for the GT-A fold, attention was concentrated in residues located primarily within ~10–25 266 Å (peak at 16.6 Å) of the donor (Figure 4A). Given that distances are measured from the backbone Cα, 267 this indicates that Glydentify prioritizes the immediate active site environment and the surrounding 268 structural scaffold. While for the GT-B fold, attention scores exhibited a bimodal distribution (peaks at 269 13.6 Å and 22.7 Å) (Figure 4B). The proximal peak (~13.6 Å) highlights residues surrounding the active 270 sites, while the distal peak (~22.7 Å) likely captures the flexible hinges that allow the enzyme to close 271 around and trap the donor. Detailed per-donor attention concentration distributions showing these 272 consistent fold-specific trends can be found in Supplementary Figure 3. Additionally, we found 273 instances where Glydentify uses the biochemical cue to predict the correct glycosyltransferase-donor 274 sugar pair. For example, in glucuronosyltransferases (CAZy family GT 43), residues with high-attention 275 score were found proximal to the negatively charged, C5-carboxylate of UDP-GlcA, highlighting basic 276 side chains positioned within 4-6 Å from carboxylate (Figure 4C). This ionic interaction between 277 residues and donor sugar, captured by attention score pairs, suggests that Glydentify understands the 278 biochemical cues of donor sugar recognition by GTs. Together, these structural and quantitative analyses 279 indicate that the model emphasizes residues in spatial proximity to the donor, capturing biochemical cues 280 relevant to sugar specificity. 281 The GT-47 family exemplified that donor sugar specificity can be shaped by more complex principles 282 rather than by first-shell interactions alone (Figure 4D). Instead of focusing on residues in direct contact 283 with the donor sugar, Glydentify consistently highlighted a set of spatially aligned positions outside the 284 donor sugar binding pocket, suggesting that Glydentify may detect co-evolutionary signals rather than 285 simple contact-based features (Supplementary Figure 4 and supplementary Table 3). A striking 286 example is a conserved cysteine located in a highly variable region of the central beta-sheet strands, 287 present in all GT-47 homologs, forming disulfide bonds (Supplementary Figure 5 and 6) with a 288 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 13 spatially proximal cysteine. In Galactosyltransferase Sp146-A and Sp 197-A, a spatially aligned 289 phenylalanine residue is consistently found in the vicinity of this conserved cysteine, suggesting a co-290 evolved, second-shell pair interaction for recognizing UDP-Gal. Conversely, the first spatially aligned 291 position is occupied by hydroxyl amino acids, serine and threonine, accompanied by branched 292 hydrophobic residues (Ile and Val) at the second spatially aligned position, again adjacent to the 293 conserved cysteine (Supplmentary Figure 6). Together, these observations suggest that donor specificity 294 in the GT-47 family may be a result of co-evolved second-shell residue networks. 295 Overall, our analysis of residues with high attention scores revealed that Glydentify does not rely on a 296 single simplistic feature; instead, the model integrates multiple layers of information, including direct 297 biochemical interactions, structural context, and patterns conserved across evolution to predict donor 298 specificity. 299

300 301 Our study introduces Glydentify, a deep learning framework for predicting glycosyltransferase (GT) 302 donor specificity. By training directly on noisy UniProt annotations and evaluating against manually 303 curated CAZy entries, we show that reliable donor-sequence relationships can be learned from large, 304 imperfect datasets. Previous ML-based approaches have been restricted to specific folds or families, such 305 as the GT-A10 or GT-1 family9,11, limiting their scalability. General enzyme–substrate predictors such as 306 ESP, EZsepecificity16,17 achieve strong results across diverse reaction classes but underperform for GTs. 307 Our models achieve PR-AUCs of 0.8638 for GT-A and 0.9124 for GT-B, surpassing general enzyme–308 substrate predictors and performing on par with, or better than, curated family-specific approaches. To 309 our knowledge, Glyidentify is the first end-to-end model designed to learn donor usage across the entire 310 sequence space of the GT without requiring extensive manual feature engineering. A major contributor to 311 Glydentify’s performance is the use of SaProt embeddings, which encode 3D structural context absent in 312 sequence-only models. Integrating SaProt representations with UniMol sugar geometries consistently 313 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 14 improved accuracy for well-supported donors, demonstrating that the choice of protein language model 314 significantly influences GT donor classification and that structural priors are crucial for distinguishing 315 sugar donors. Furthermore, the model even captured the donor sugar promiscuity of glycosyltransferases 316 (Supplementary Figure 7 and 8). For instance, the processive diacylglycerol β-glycosyltransferase from 317 Mycoplasma genitalium was predicted to utilize both UDP-Glc (93.5%) and UDP-Gal (99.2%) with high 318 confidence (Supplementary Figure 7), consistent with kinetic measurements showing Km = 50 μM for 319 UDP-Glc and Km = 243 μM for UDP-Gal30. 320 Despite the overall strong performance, we observed two notable exceptions in the GT-A model: UDP-321 GlcNAc and UDP-GalNAc showed negligible or slightly reduced performance gains. This is likely due to 322 a combination of their subtle chemical differences and the limitations of UniProt-derived labels. 323 GlcNAc and GalNAc differ only at the C4 stereocenter, suggesting that it is difficult for UniMol to 324 separate the epimeric difference. In addition, donor annotations for these sugars in UniProt are often 325 inferred rather than experimentally confirmed, which may further contribute to the lower accuracy 326 observed for these two donors. We also observed that B3GLCT, a member of the GT31 family, was not 327 correctly predicted despite its high sequence similarity to other GT31 enzymes in the training set 328 (Supplementary Figure 2), suggesting that high global sequence similarity within GT31 does not always 329 translate into clear donor-specific signals, and closely related sequences with divergent donor sugar 330 specificity may cause difficulty for the model to resolve during training. 331 UniProt labels contain substantial noise from homology-based inference, and CAZy covers only a small 332 fraction of GT–donor space. Rare donor classes continue to present challenges, with many represented by 333 fewer than ten validated sequences. Our analysis of these minor donors, such as UDP-GalNAc, further 334 confirms this challenge: even with a specialized architecture pre-trained on other diverse sugars, including 335 UDP-GlcNAc, performance remains unattainable with limited data, indicating the strict stereochemical 336 determinants of specificity cannot be robustly migrated from chemically similar donors without sufficient 337 training samples (Supplementary Figure 10). With the current depth and quality of available 338 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 15 annotations, this likely represents the practical performance ceiling achievable from existing resources. 339 Future advances will depend on the generation of higher-quality, large-scale data—particularly from 340 emerging high-throughput assays capable of directly profiling GT activity. Such datasets could provide 341 the clean, experimentally grounded labels needed for models to learn more accurate sequence–function 342 mappings and iteratively refine themselves through feedback from correct predictions. Combined with 343 improved structural modeling, such as explicit integration of active-site representations, these 344 developments promise to substantially enhance both the accuracy and biological interpretability of GT 345 donor prediction. More broadly, the framework exemplifies how deep learning can close sequence-to-346 function gaps in enzymology, particularly in cases where data are abundant but noisy, and curated 347 annotations remain scarce. 348

349 Data Collection and Preprocessing 350 We assembled a large-scale training set of GT sequences with donor annotations from UniProt by 351 querying major donor-type enzyme names (e.g. “glucosyltransferase,” “galactosyltransferase,” etc.). To 352 ensure annotation quality, we retained only entries with an annotation score greater than 3, a threshold 353 that reflects substantial manual curation and reduces the prevalence of purely computationally inferred 354 labels. This initial query yielded 126,451 sequences annotated with at least one donor sugar. 355 We assigned each sequence to the GT-A or GT-B structural fold using hidden Markov model (HMM) 356 profiles from Pfam (PF00535 for GT-A, PF01062 for GT-B) and the HMMER suite (E-value < 1e-5)31,32. 357 From the UniProt dataset, we identified 11,492 GT-A and 15,235 GT-B sequences. To train our model 358 efficiently and focus on the catalytic core, each candidate training sequence was cropped to the detected 359 HMM domain, plus 15 flanking residues at both the N- and C-termini of the conserved catalytic domain. 360 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 16 Because donor classes were highly imbalanced, we reduced sequence redundancy within each donor type 361 using mmseqs33 clustering at identity thresholds tailored to class prevalence: 95% for donor types with > 362 500 sequences and 100% for <100 sequences. After selecting one representative per cluster, the final 363 training set comprised 4,938 GT-A and 7,611 GT-B sequences. 364 For evaluation, we built a high-confidence test set from CAZy by scraping all experimentally 365 characterized GT entries34. Since CAZy does not include explicit donor sugar annotations, we utilize a 366 Large Language Model (LLM) to infer the donor sugar used from the protein name or the related 367 publication abstract. GPT 5.1 was used with temperature set to 1.0. The system prompt of the LLM 368 included the input specifications and some examples of inputs and expected output pairs. In case an 369

is provided, the LLM was instructed to extract a subset of lines as evidence showing where the 370 donor sugars are mentioned. The full prompt and code are provided in the supplementary data. We 371 clustered sequences at 90% identity and then removed any remaining training sequences with an identity 372 greater than 90% to a test sequence to prevent homology leakage. The resulting test set contains 174 GT-373 A and 194 GT-B sequences. All test sequences were retained at full length to evaluate the model’s 374 performance on complete proteins without additional truncation. 375 To further enrich our representations with three-dimensional context, we incorporated predicted structures 376 for all GT sequences. Because the vast majority of these proteins lack experimentally resolved 377 coordinates in the Protein Data Bank, we generated models using AlphaFold20. Predictions were 378 performed with default parameters, and the resulting coordinate files served as the structural inputs for 379 downstream representation learning and model evaluation. 380 To obtain 3D molecular inputs for the donor encoder, we extracted the exact donor sugar molecule from 381 the UniProt “catalytic activity” or “reaction” fields when available; if no explicit annotation existed, we 382 assigned the most common sugar for that enzyme class (e.g. “glucosyltransferase” → UDP-Glucose). 383 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 17 Each SMILES string was converted to an energy-minimized 3D conformer using RDKit, then encoded by 384 UniMolV2 to produce a fixed donor embedding. 385 Because we only observe positive GT-donor annotations in UniProt and CAZy, we treat every 386 unannotated GT-donor pair as a negative example during training. Concretely, if a sequence is labeled 387 with 𝑘 donors out of the 𝐶 total in our vocabulary (𝐶 ≈ 10 ), we generate (𝐶 − 𝑘) negatives by pairing 388 that sequence with each of the other donors. 389 We acknowledge that some true GT–donor pairs may be missing from UniProt/CAZy annotations, 390 resulting in them being erroneously labeled as negatives. However, these false negatives are likely rare 391 for our focused set of major donors. UniProt and CAZy both aim for high coverage of well-studied GTs 392 and any remaining noise is mitigated by the large number of true negatives. In practice, this simple 393 closed-world assumption yielded stable training and reliable generalization to the high-confidence CAZy 394 test set. 395 Encoding GTs and Donor Sugars 396 Given a protein of length 𝐿, we tokenize the input as: 397 𝑥seq = 0[𝐶𝐿𝑆], 𝑟(, 𝑟), … , 𝑟*, [𝐸𝑂𝑆]: . 398 The SaProt encoder produces contextual embeddings 399 𝐻seq = =ℎseqCLS, ℎseq ((), … , ℎseq (*), ℎseqEOS? ∈ ℝ(*0))×2seq 400 where ℎ denotes encoder outputs. We use the [𝐶𝐿𝑆] embedding as a global sequence summary and the 401 residue embeddings ℎ345 ((:*) for token–leve interaction. Similarly, for each donor 𝑑 ∈ {1, … , 𝐶}, we obtain 402 an RDKit conformer from its SMILES and feed it to UniMol V2 to obtain atom–level embeddings. 403 UniMol uses special tokens as well: 404 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 18 𝑥mol,2 = E[𝐶𝐿𝑆], 𝑎2 ((), … , 𝑎2 (;$), [𝐸𝑂𝑆]G, 405 𝐻2 mol = =ℎmol,2 CLS , ℎmol,2 (() , … , ℎmol,2 (;$) , ℎmol,2 EOS ? ∈ ℝ(;$0𝟚)×2mol . 406 Both Saprot and UniMol as well as the generated embeddings are frozen to reduce overfitting and training 407 cost. To adapt embeddings, map raw encoder embeddings ℎ to adapter outputs 𝑧 in their native 408 dimensions via: 409 𝑍seq = 𝑓seq(𝐻seq), 𝑍2 mol = 𝑓molK𝐻2 molL, 410 where 𝑓seq𝑎𝑛𝑑𝑓mol are MLP (Multi-Layer Perceptrons) layers consist of a linear projection followed by a 411 GELU activation and Layer Normalization. We write: 412 𝑧mol,2 CLS ∈ ℝ2mol , 𝑧mol,2 ((:;$) ∈ ℝ;$×2mol 413 Bi-directional cross-attention module and classification head 414 We compute donor–specific interactions in both directions. In 𝑚𝑜𝑙 → 𝑠𝑒𝑞, the donor [𝐶𝐿𝑆] queries 415 residue tokens; in 𝑠𝑒𝑞 → 𝑚𝑜𝑙, the protein [𝐶𝐿𝑆] queries atom tokens. Importantly, keys/values exclude 416 [𝐶𝐿𝑆] [𝐸𝑂𝑆] and use only 1: 𝐿 residues or 1: 𝐴2 atoms: 417 𝑧2 molWWWWW = 𝑀𝐻𝐴K𝑄 = 𝑧=>?,2 @*A , 𝐾 = 𝑍(:* 345 , = 𝑍(:* 345 L, 418 𝑧2 345WWWWWW = 𝑀𝐻𝐴K𝑄 = 𝑧345@*A, 𝐾 = 𝑍2,(:;$ =>? , 𝑉 = 𝑍2, (:;$ mol L. 419 We apply residual connections and layer normalization: 420 𝑟2 =>? = 𝐿𝑁 ]𝑧=>?,2 @*A + 𝐷𝑟𝑜𝑝𝑜𝑢𝑡c𝑧2 =>?WWWWWWde, 421 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 19 𝑟2 345 = 𝐿𝑁 ]𝑧345@*A + 𝐷𝑟𝑜𝑝𝑜𝑢𝑡 c𝑧2 345WWWWWWde. 422 We project each side to a shared width 𝑑∗ and fuse: 423 𝑔2 =>? = 𝑊=>?  𝑟2 =>? , 𝑔2 345 = 𝑊345  𝑟2 345 , 424 𝑢2 = 𝐿𝑁([𝑔2 345 ‖𝑔2 =>? ]) ∈ ℝ)2∗ . 425 A two–layer MLP produces a donor–specific logit 𝑠2 ∈ 𝑅 and probability 𝑦2k = σ(𝑠2): 426 𝑠2 = 𝑤) D tanh(𝑊(𝑢2 + 𝑏() + 𝑏), 𝑦2k = σ(𝑠2). 427 Stacking all donors yields a multi–label prediction vector 𝑦s ∈ [0 ,1]@. 428 Loss function and optimization 429 Because negative labels vastly outnumber positives, we adopt ASL (Asymmetric Loss)35 to down–weight 430 negatives while preserving the positive signal. Let 𝑥 ∈ 𝑅@ be the logits, 𝑦 ∈ {0 ,1}@ the targets, and 𝑝 =431 σ(𝑥). Denote class index by 𝑐 and choose focusing parameters γ0 ≥ 0 (positives) and γE > 0 (negatives). 432 Define an asymmetric clipping margin 𝑚 ∈ [0 ,1) applied only to the negative term: 433 𝑞Fx = min{ 1, 1 − 𝑝F + 𝑚}. 434 The per–example ASL is: 435 ℒ𝒜𝒮ℒ = −[𝑦 (1 − 𝑝)J)  log 𝑝F + (1 − 𝑦) 𝑝J*  log 𝑞]. 436 The positive term receives standard or mildly focused weighting via (1 − 𝑝F )J), while the negative term 437 is aggressively damped by 𝑝F J* and the clipping 𝑞Fx, which reduces gradients from easy negatives (small 438 𝑝F). 439 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 20 440 Training Details of Glydentify 441 We formulated the prediction task as a binary classification problem, where the model takes a 442 glycosyltransferase (GT) sequence and a donor substrate as input pairs. The model was trained using ASL 443 with γ0 = 0 and γE = 2. We optimized the network using AdamW with default parameters, a learning 444 rate of 5×10 −5 , and a weight decay of 1×10 −4. The training batch size was set to 16 and ran for a 445 maximum of 30 epochs, utilizing early stopping with a patience of 5 epochs to prevent overfitting. 446 Protein Expression and Purification 447 The expression constructs for all plant GTs used in this study (details in Table 1) were generated as 448 truncated catalytic domains as NH2-terminal fusion proteins in the pGEM2 expression vector, as 449 described in prior studie36. Briefly, the fusion protein coding region was comprised of a 25-amino acid 450 signal sequence, an His8 tag, AviTag, the “superfolder” GFP, the 7-amino acid recognition sequence of 451 the tobacco etch virus (TEV) protease followed by the truncated coding region of each individual plant 452 GTs under study. Coding sequences for Spirodela polyrhiza (v2) and Spirodela intermedia were obtained 453 from Phytozome 14 and the National Center for Biotechnology Information, respectively. The truncated 454 catalytic domains of Sp124-E (Spipo24G0012400; amino acids 37-801), Sp415-C (Spipo8G0041500, 455 amino acids 38-564), Sp124-C (Spipo2G0124300, amino acids 25-403), Sp291-D (Spipo1G0029100, 456 amino acids 45-531), and Sl728-C (SICAA7408728.1; amino acids 44-555) without the transmembrane 457 domains (predicted using DeepTMHMM - 1.035) were cloned into pTwist Gateway ENTR via gene 458 synthesis (Twist Biosciences). The genes were transferred into the mammalian expression vector pGEn2-459 DEST36,37 using GatewayTM recombination. 460 The expression constructs were then used for transient transfection of suspension culture HEK293-F cells 461 (FreeStyleTM 293-Fcells, Thermo Fisher Scientific, Waltham MA) maintained at 0.5–3.0x106 cells/ml in 462 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 21 a humidified CO2 platform shaker incubator at 37 °C with 50 % relative humidity and 125 RPM. 463 Transient transfection was performed using HEK293-F cells (2.5-3.0x106 cells/ml) in the expression 464 medium comprised of a 9:1 ratio of FreestyleTM293 expression medium (Thermo Fisher Scientific, 465 Waltham MA) and EX-Cell expression medium including Glutmax (Sigma-Aldrich). Transfection was 466 initiated by the addition of plasmid DNA and polyethyleneimine as transfection reagent (linear 25-kDa 467 polyethyleneimine, Polysciences, Inc., Warrington, PA). Twenty-four hours post-transfection, cell 468 cultures were diluted with an equal volume of fresh media supplemented with valproic acid (2.2 mM final 469 concentration), and protein production was continued for an additional 5 days at 37 °C, 125 RPM and 5% 470 CO2. The cell cultures were harvested, clarified by sequential centrifugation at 1200 rpm for 10 min and 471 3500 RPM for 15 min at 4 °C, and passed through a 0.8 µM filter (Millipore, Billerica, MA). The crude 472 protein was then subjected for further purification using immobilized metal affinity chromotography 473 (IMAC) using an AKTA 25L system (Cytiva) at 4 °C with HisTrap FF prepacked 5 ml or 1ml columns 474 (Cytiva)36,37. Purified proteins were also loaded onto an SDS-PAGE gel to confirm the expected 475 molecular weight. All proteins were buffer exchanged into 50 mM HEPES Sodium Salt (pH 8) with 100 476 mM sodium chloride and transferred to -80 °C for long-term storage. 477 Glycosyltransferase Activity Assay 478 UDP-sugar donors used in reactions were CIAP-treated (Calf Intestinal Alkaline Phosphatase) according 479 to Sheikh and Wells, 200638 The UDP-Glo™ Glycosyltransferase Assay (Promega) was carried out in a 480 total volume of 10 μL, consisting of 250 mM Tris+MES+MOPS pH 7 or pH 5.5, 0.5 mM UDP-sugar 481 donor, with and without 0.1 mg/mL of acceptor (when indicated), and 1 μM GT enzyme. Reactions were 482 set at room temperature for either 2 hours or overnight as indicated in the legend. For UDP quantitation, 5 483 μL of the reaction was subsequently mixed with 5 μL of UDP-Glo™ Detection Reagent directly in white 484 polystyrene, low-volume, 384-well assay plates (Corning) and incubated for 60 minutes. Luminescence 485 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 22 was recorded with a multifunctional microplate reader BioTek Synergy LX Multi-Mode Reader (Agilent). 486 Data was quantified relative to a standard curve of UDP. 487 Data Availability 488 The raw protein sequences were retrieved from the UniProt database (https://www.uniprot.org). The raw 489 information used to generated testing set can be found in CAZy database 490 (https://www.cazy.org/GlycosylTransferase-family). 491 Code Availability 492 The source code for Glydentify, including model architecture, training scripts, and inference pipelines, as 493 well as the processed training and testing data are is available for peer review at 494 https://github.com/RuiliF/Glydentify. A permanent DOI will be minted via Zenodo upon acceptance. 495

496 1. Essentials of Glycobiology. (Cold Spring Harbor laboratory press, Cold Spring Harbor, New York, 497 2017). 498 2. Moremen, K. W., Tiemeyer, M. & Nairn, A. V. Vertebrate protein glycosylation: diversity, synthesis 499 and function. Nat. Rev. Mol. Cell Biol. 13, 448–462 (2012). 500 3. Chang, A., Singh, S., Phillips, G. N. & Thorson, J. S. Glycosyltransferase structural biology and its 501 role in the design of catalysts for glycosylation. Curr. Opin. Biotechnol. 22, 800–808 (2011). 502 4. Boltje, T. J., Buskas, T. & Boons, G.-J. Opportunities and challenges in synthetic oligosaccharide and 503 glycoconjugate research. Nat. Chem. 1, 611–622 (2009). 504 5. Lairson, L. L., Henrissat, B., Davies, G. J. & Withers, S. G. Glycosyltransferases: Structures, 505 Functions, and Mechanisms. Annu. Rev. Biochem. 77, 521–555 (2008). 506 6. Morgan, J. L. W., Strumillo, J. & Zimmer, J. Crystallographic snapshot of cellulose synthesis and 507 membrane translocation. Nature 493, 181–186 (2013). 508 7. Buschiazzo, A. et al. Crystal structure of glycogen synthase: homologous enzymes catalyze glycogen 509 synthesis and degradation. EMBO J. 23, 3196–3205 (2004). 510 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 23 8. Yamamoto, F., Clausen, H., White, T., Marken, J. & Hakomori, S. Molecular genetic basis of the 511 histo-blood group ABO system. Nature 345, 229–233 (1990). 512 9. Yang, M. et al. Functional and informatics analysis enables glycosyltransferase activity prediction. 513 Nat. Chem. Biol. 14, 1109–1117 (2018). 514 10. Taujale, R. et al. Deep evolutionary analysis reveals the design principles of fold A 515 glycosyltransferases. eLife 9, e54532 (2020). 516 11. Hennen, S. G., Bomble, Y. J., Urbanowicz, B. R. & Bharadwaj, V. S. Decoding substrate 517 specificity determining factors in glycosyltransferase-B enzymes – insights from machine learning 518 models. Digit. Discov. 4, 2214–2228 (2025). 519 12. He, X. et al. Highly accurate carbohydrate-binding site prediction with DeepGlycanSite. Nat. 520 Commun. 15, 5163 (2024). 521 13. Carpenter, E. J., Seth, S., Yue, N., Greiner, R. & Derda, R. GlyNet: a multi-task neural network 522 for predicting protein–glycan interactions. Chem. Sci. 13, 6669–6686 (2022). 523 14. Nasir, W. et al. SweetNET : A Bioinformatics Workflow for Glycopeptide MS/MS Spectral 524 Analysis. J. Proteome Res. 15, 2826–2840 (2016). 525 15. Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language 526 model. Science 379, 1123–1130 (2023). 527 16. Kroll, A., Ranjan, S., Engqvist, M. K. M. & Lercher, M. J. A general model to predict small 528 molecule substrates of enzymes based on machine and deep learning. Nat. Commun. 14, 2787 (2023). 529 17. Cui, H. et al. Enzyme specificity prediction using cross-attention graph neural networks. Nature 530 647, 639–647 (2025). 531 18. ESM Team. ESM Cambrian: Revealing the mysteries of proteins with unsupervised learning. 532 ESM Cambrian: Revealing the mysteries of proteins with unsupervised learning. 533 https://evolutionaryscale.ai/blog/esm-cambrian. (2024). 534 19. Su, J. et al. SaProt: Protein Language Modeling with Structure-aware Vocabulary. Preprint at 535 https://doi.org/10.1101/2023.10.01.560349 (2023). 536 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 24 20. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. 537 Nature 630, 493–500 (2024). 538 21. Huang, C., Kannan, N. & Moremen, K. W. Modeling glycans with AlphaFold 3: capabilities, 539 caveats, and limitations. Glycobiology 35, cwaf048 (2025). 540 22. Xu, S. et al. Benchmarking all-atom biomolecular structure prediction with FoldBench. Nat. 541 Commun. https://doi.org/10.1038/s41467-025-67127-3 (2025) doi:10.1038/s41467-025-67127-3. 542 23. Zhang, L. et al. Glycosyltransferase family 47 (GT47) proteins in plants and animals. Essays 543 Biochem. 67, 639–652 (2023). 544 24. Engel, L., Alves, J., Hennek, J., Goueli, S. A. & Zegzouti, H. Utility of Bioluminescent 545 Homogeneous Nucleotide Detection Assays in Measuring Activities of Nucleotide-Sugar Dependent 546 Glycosyltransferases and Studying Their Inhibitors. Molecules 26, 6230 (2021). 547 25. Kadooka, C. et al. Cgm1 is a β-galactoside α-(1 → 4)-mannosyltransferase involved in the 548 biosynthesis of capsular glucuronoxylomannogalactan in Cryptococcus neoformans. J. Biol. Chem. 549 301, 110632 (2025). 550 26. Corulli, C. J. et al. Biochemical characterization of xyloglucan galactosyltransferases MUR3 and 551 XLT2 from Spirodela polyrhiza. Plant J. (2025). 552 27. Gado, J. E. et al. Machine learning prediction of enzyme optimum pH. Nat. Mach. Intell. 7, 716–553 729 (2025). 554 28. Manigrasso, J., De Vivo, M. & Palermo, G. Controlled Trafficking of Multiple and Diverse 555 Cations Prompts Nucleic Acid Hydrolysis. ACS Catal. 11, 8786–8797 (2021). 556 29. Brissos, V. et al. Distal Mutations Shape Substrate-Binding Sites during Evolution of a Metallo-557 Oxidase into a Laccase. ACS Catal. 12, 5022–5035 (2022). 558 30. Andrés, E., Martínez, N. & Planas, A. Expression and Characterization of a Mycoplasma 559 genitalium Glycosyltransferase in Membrane Glycolipid Biosynthesis. J. Biol. Chem. 286, 35367–560 35379 (2011). 561 31. Eddy, S. R. Profile hidden Markov models. Bioinformatics 14, 755–763 (1998). 562 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 25 32. Paysan-Lafosse, T. et al. The Pfam protein families database: embracing AI/ML. Nucleic Acids 563 Res. 53, D523–D534 (2025). 564 33. Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the 565 analysis of massive data sets. Nat. Biotechnol. 35, 1026–1028 (2017). 566 34. Drula, E. et al. The carbohydrate-active enzyme database: functions and literature. Nucleic Acids 567 Res. 50, D571–D577 (2022). 568 35. Ridnik, T. et al. Asymmetric Loss For Multi-Label Classification. in 2021 IEEE/CVF 569 International Conference on Computer Vision (ICCV) 82–91 (IEEE, Montreal, QC, Canada, 2021). 570 doi:10.1109/ICCV48922.2021.00015. 571 36. Moremen, K. W. et al. Expression system for structural and functional studies of human 572 glycosylation enzymes. Nat. Chem. Biol. 14, 156–162 (2018). 573 37. Prabhakar, P. K. et al. Heterologous expression of plant glycosyltransferases for biochemistry and 574 structural biology. in Methods in Cell Biology vol. 160 145–165 (Elsevier, 2020). 575 38. Sheikh, M. O. et al. Rapid screening of sugar-nucleotide donor specificities of putative 576 glycosyltransferases. Glycobiology 27, 206–212 (2017). 577 578 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711695doi: bioRxiv preprint 26 579 Figure 1. Donor diversity and overview of Glydentify model architecture and training. A. Chemical 580 diversity of the sugar donor. Chemical structure of sugar donor used in the current prediction was shown. 581 B. A full schematic of our data curation pipelines. C. The architecture of the Glydentify model classifies 582 GT-Donor pairs as positive or negative interactions. 583 BhChOhN…hCLShSEPhC Asymmetric LosshMhEhVhq…hCLShSEPCross AttentionProjection HeadPrediction Score Donor GTYesBinary labeltokenizetokenizePre-trainedProtein EncoderPre-trainedMolecule Encoder catCross AttentionProjection HeadQKVQKVDropoutAttention Score per GT ResidueAttention Score per Donor Atom Residue RepresentationAtom Representation Classification Head A Transferase nameAnnotation score > 3 UniProtDonor Annotation CAZyCharacterized GTsFold Classification & Extract Domains Infer donor from namesSelect common donors Sequence Identity Cluster & Filter Extract Donor Molecule Energy Minimization Training GTSequences Molecule 3D Structure3D Structure Prediction Testing GTSequences Training SetHomologyFiltering Mapping Testing Set C 27 584 Figure 2. Evaluation of Glydentify performance. A. Overall Performance comparing with ESP, 585 EZspecificity, predicting with only pre-trained protein language models and our designs with different 586 pre-trained protein language models. All pre-trained encoders are frozen and only the shallow MLP probe 587 or fusion head is trained. B. Performance as a function of sequence similarity between each test enzyme 588 and its closest training sequence (Methods), binned as <20%, 20–40%, 40–60%, 60–80%, and 80–90%. 589 Bars report PR-AUC; the dashed line (right y-axis) shows the number of GTs per bin. C. Donor-wise 590 performance for GT-A (left) and GT-B (right). Bars show PR-AUC for the models indicated in the 591 legend; the dashed line (right y-axis) gives the number of positive sequences available for each donor in 592 the evaluation set. D. Performance by taxonomic group for GT-A (left) and GT-B (right). Bars report PR-593 AUC; the dashed line (right y-axis) indicates the number of GTs in each kingdom. Donor/taxon/identity 594 categories with fewer than 10 test positives are omitted because PR–AUC is unreliable at very limited 595 sample sizes. 596 0.20270.1650.8180.8620.7300.9110.9040.909 0.16640.1250.3790.6830.2810.8100.8160.864 00.20.40.60.81ESPEZSpecificityESM2 + MLPESMC + MLPSaProt + MLPGlydentify (ESM2)Glydentify (ESMC)Glydentify (SaProt) PR_AUC Score On CAZy Testing SetGT-AGT-B 0.770.280.370.500.740.890.300.740.891.00 0.000.200.400.600.801.001.20 <20%(n=84)20-40%(n=33)40-60%(n=26)60-80%(n=45)80-90%(n=56) PR_AUC Score GT-A Fold ESM2+MLPGlydentify(SaProt)0.370.740.670.880.580.880.911.00 0.000.200.400.600.801.001.20 20-40%(n=24)40-60%(n=97)60-80%(n=78)80-90%(n=76) GT-B Fold ESM2+MLPGlydentify(SaProt) 0.6370.2530.6600.267 0.9120.8780.9060.770 0.00.20.40.60.81.0 Metazoa(n=75)Plants(n=22)Fungi(n=24)Bacteria(n=39) PR_AUC Score GT-A Fold ESM2 + MLPGlydentify (SaProt) 0.8100.8030.8100.259 0.9040.9600.9940.708 0.00.20.40.60.81.0 Metazoa(n=82)Plant(n=128)Fungi(n=13)Bacteria(n=42) PR_AUC Score GT-B Fold ESM2 + MLPGlydentify (SaProt) 0.580.670.540.550.860.570.390.910.890.750.780.990.920.91 GDP-ManUDP-GalUDP-GlcUDP-GlcAUDP-GalNAcUDP-GlcNAcUDP-Xyl0.000.200.400.600.801.001.20 0.001.002.003.004.005.006.007.008.00 PR_AUC Score GT-A Fold ESM2+MLPGlydentify (SaProt) 0.890.750.910.800.430.990.930.950.830.93 GDP-FucGDP-ManUDP-GlcUDP-GlcAUDP-GlcNAc0.000.200.400.600.801.001.20 0.001.002.003.004.005.006.00 GT-B Fold ESM2+MLPGlydentify (SaProt) A B C D 28 597 Figure 3. Structural models and experimental validation of donor prediction in uncharacterized 598 plant GTs. A. Model of Cgm1 (GT 139) docked with the predicted donor GDP-mannose and manganese, 599 generated by AlphaFold3 (left), and active-site structure showing the DXD motif interacting with the 600 diphosphate group (right) B. Model of Sp415-C (GT 47) docked with the predicted donor UDP-xylose and 601 its acceptor homogalacturonan (DP 6), generated by AlphaFold3 (left), and active-site structure showing 602 the two arginines interacting with the diphosphate (right). C, D, E. Glycosyltransferase activity in the 603 absence (C) or presence of homogalacturonan oilgosaccharides (DP 5-15) as acceptors in an overnight 604 reaction Sp415-C (D) and Sp124-C (E). Data is represented as means ± standard deviation of three 605 biological replicates. Individual data points are shown. Letters indicate statistical significance accomplished 606 by a one-way ANOVA followed by a Tukey’s multiple comparison test (P≤0.001). 607 608 609 610 UDP-GalUDP-GlcUDP-GalAUDP-GlcAUDP-ArafUDP-Xyl0 5 10 15 20pmol of UDP A BB CB C B CC UDP-GalUDP-GlcUDP-GalAUDP-GlcAUDP-ArafUDP-Xyl0.0 0.1 0.2 0.3 0.4pmol of UDP A BBBBB UDP-GalUDP-GlcUDP-GalAUDP-GlcAUDP-ArafUDP-Xyl0.0 0.5 1.0 1.5 2.0 2.5pmol of UDP A BBBBB Sp124-CSp415-CSp124-EHydrolysisAcceptor+ Acceptor+ B. Sp415-C (GT47) A. Cgm1 (GT139) C D E !"#$%&' !"#$%&( !"#$%&% "#'$()( *+,$-./ *+,$-&0 29 611 Figure 4: Explainability analysis of donor prediction using attention. A, B. Structural colocalization 612 of model attention with the donor binding site. Kernel density estimates show the distribution of 613 Euclidean distances between high-attention residues (normalized score > 0.8) and the donor substrate for 614 (A) GT-A and (B) GT-B fold enzymes. Analysis was performed on true positive samples using 615 AlphaFold 3-generated structures. Distances were measured from the residue α-carbon (Cα)to the ligand 616 anomeric carbon (C1). C. AlphaFold3–predicted structure of B3GAT3 (Uniprot ID: Q9WU47) modeled 617 with UDP-glucuronic acid and Mn²⁺ reveals that an arginine, the residue with the highest cross-attention 618 score, positioned adjacent to the C5-carboxylate of UDP-GlcA. The proximity of this basic residue to the 619 acidic C5-carboxylate suggests a strong ionic interaction is the main determinant of GlcA specificity. D. 620 AlphaFold3-predicted ternary complex of a representative GT-B GT47 enzyme (Sp415-C) with UDP-Xyl 621 and acceptor galacturonan mapped with Glydentify attention scores. High-attention residues across GT47 622 enzymes were found at spatially aligned positions, suggesting that these conserved sites co-evolved to 623 shape donor-sugar specificity. 624 625 626 627 628 629 630 631 632 633 634 30 FOLD NAME OR UNIPROT ID ORGANISM VALIDATED DONOR SUGAR MAX SEQ. IDENT. (%) MODEL'S PREDICTION (%) GT-A J9VFZ3 Cryptococcus neoformans GDP-Man 26.5 GDP-Man (90.9); UDP-Glc (54.4) Q72K30 Thermus thermophilus HB27 GDP-Man 49.4 GDP-Man (79.0) P53059 Saccharomyces cerevisiae S288c GDP-Man 48.1 GDP-Man (99.6) GT-B Sp146-A Spirodela polyrhiza UDP-Gal 64 UDP-Gal (98.4) Sp197-A Spirodela polyrhiza UDP-Gal 45 UDP-Gal (99.4) Sp291-D Spirodela polyrhiza UDP-Xyl 63.2 UDP-GlcA (77.0); UDP-Xyl (51.3) Sp415-C Spirodela polyrhiza UDP-Xyl 50 UDP-Xyl (93.3) Sp124-C Spirodela polyrhiza UDP-Xyl 44.3 UDP-Xyl (84.3) Sp124-E Spirodela polyrhiza UDP-Xyl 37.6 UDP-Xyl (73.4); UDP-GlcA (52.1) Si728-C Spirodela intermedia UDP-Xyl 48.8 UDP-Xyl (93.3); UDP-GlcA (61.9) 635 Table 1. Independent validation of Glydentify predictions on experimentally verified and new GTs. 636 The table lists representative GT-A and GT-B enzymes from independent datasets, along with their 637 known donor sugars, sequence similarity to the training set, and predicted donor-specificity scores. 638 639 640 Acknowledgements (optional) 641 This work was supported by funding from the NSF BioFoundry (2400220/U.S. National Science 642 Foundation BioFoundry: Glycoscience Research, Education and Training). Funding for BU from U.S. 643 DOE, Office of Science, BER program, GSP grants DE-SC0023223 and DE-SC0026057 is 644 acknowledged. 645 Ethics declarations 646 Competing interests 647 31 Submission of a competing interests statement is required for all content of the journal. 648