Resolving Chemically Inequivalent ¹¹B NMR Sites via Interpretable Hybrid Machine Learning

Abstract

Organoboron compounds are widely used across pharmaceuticals and materials science, where ¹¹B NMR spectroscopy serves as a valuable tool for structural characterization. However, severe spectral line broadening induced by the quadrupolar nature of the boron nucleus often causes signal overlap, making it exceptionally difficult to experimentally resolve chemically inequivalent sites in complex multiboron architectures. While traditional density functional theory can resolve these ambiguities, it faces prohibitive computational bottlenecks, whereas data-driven alternatives remain constrained by the scarcity of high-quality data sets. Herein, we report a manually verified, solvent-annotated ¹¹B NMR data set constructed via a large language model (LLM)-assisted workflow. Interpretable machine learning identifies a strong correlation between the BCUT2D_MRLOW descriptor and the boron hybridization. Integrating these ML-derived features as prior knowledge, we developed a prior-guided Graph Transformer for accurate atom-level chemical shift prediction. Notably, the model provides a form of virtual spectral resolution, enabling the discrimination of chemically inequivalent boron sites that are difficult to resolve experimentally. We further deploy the framework as an open-access Web tool to support the rapid structural analysis of organoboron compounds.

Introduction

Organoboron chemistry has evolved from a specialized synthetic discipline into a versatile platform central to modern molecular science. Boryls and related organoboron species exhibit rich bonding patterns and reactivity manifolds, further expanding the chemical space and synthetic utility of boron compounds. − From the Nobel Prize-winning Suzuki–Miyaura coupling, to the development of boron-containing pharmaceuticals (e.g., Bortezomib and Tavaborole) and advanced materials like boron- and nitrogen-doped polycyclic aromatic hydrocarbons (BN-doped PAHs) and organic light-emitting diodes, the demand for precise structural engineering of boron compounds is at an all-time high (Figure a). In this context, ¹¹B NMR spectroscopy stands as a key tool for monitoring reactions and validating structures, offering a direct window into the electronic environment of the boron center. However, extracting precise structural details from ¹¹B NMR spectra remains far more challenging than routine ¹H or ¹³C NMR analysis. Whereas ¹H NMR benefits from well-established chemical shift ranges and coupling patterns that often allow empirical structural assignment, ¹¹B NMR is hindered by quadrupolar broadening and the lack of simple, widely applicable interpretive rules. As a result, synthetic chemists often rely on superficial comparisons, such as confirming the disappearance of starting-material signals or referencing a small number of familiar analogues. This approach readily fails for unfamiliar bonding motifs or unexpected byproducts, reducing spectral interpretation to subjective guesswork.

1.

(a) Examples of boron-containing drugs (tavaborole and bortezomib) and BN-embedded pi-conjugated emitters (DABNA , and B2N2-embedded corannulene). (b) Study workflow for δ­(¹¹B) NMR prediction. The pipeline encompasses LLM-assisted data set curation (left), an integrated hybrid modeling strategy combining molecular-level machine learning with a prior-guided graph neural network (center), and final deployment as an open-access interactive web server (right).

To circumvent these empirical limitations, rigorous quantum chemical calculations, specifically density functional theory (DFT), offer a powerful theoretical recourse. While optimized protocols (e.g., GIAO-ωB97XD) can achieve predictive accuracies of approximately 2 ppm for typical organoboron species, , they face a severe “precision-efficiency” dilemma that largely precludes their use as a routine tool for the average bench chemist. Accurate DFT predictions mandate computationally demanding setups, such as explicit solvent interactions or Boltzmann weighting for flexible conformers. As noted by Rzepa and co-workers, computing even moderately sized systems (∼125 atoms) can consume days on parallel clusters, rendering such methods entirely prohibitive for high-throughput screening or rapid, on-the-fly daily analysis. Against this backdrop, the rapid advancement of artificial intelligence provides a transformative alternative: it circumvents the unreliability of manual empirical matching and delivers DFT-quality predictions at a fraction of the computational cost. , Data-driven spectroscopic analysis has emerged as a practical alternative, holding the potential to digitize the “chemical intuition” needed for structural elucidation. , By providing instant, high-fidelity shift predictions, such intelligent frameworks empower nonspecialists to confidently assign complex multiboron structures, shifting the paradigm of organoboron research toward a more streamlined and accessible future.

However, the exploration of this data-driven paradigm remains uneven across different nuclei. While machine learning has made substantial progress in predicting ¹H and ¹³C NMR parameters, − extending these successes to heteronuclei remains a distinct frontier. For instance, Cheng et al. recently established a unified benchmark and the NMRNet framework, setting a rigorous standard for deep learning-based prediction across diverse chemical systems. Significant strides have also been made in specific heteronuclear domains: Xue et al. developed a robust graph convolution network (GCN) for ¹⁹F NMR prediction using specialized fluorine fingerprints, while Geitner and co-workers introduced Ilm-NMR-P31, a comprehensive open-access database and graph neural network (GNN)-based predictive tool that unlocked the chemical space of ³¹P NMR. In contrast, the domain of ¹¹B NMR remains disproportionately underexplored. Although Berezin and co-workers demonstrated the feasibility of predicting ¹¹B shifts for BODIPY dyes with high accuracy using Random Forest regression, their model is restricted to a narrow chemical space and specific scaffold descriptors. Crucially, such niche models struggle to resolve the structural complexity of multiboron environments, failing to deliver the atom-level precision required to distinguish chemically inequivalent sites. The scarcity of comprehensive, structurally diverse data sets capturing these diverse architectures has created a “precision gap” that prevents the generalization of AI tools to broader organoboron systems.

Herein, we report a comprehensive framework that addresses the scarcity of solvent-annotated ¹¹B NMR data by unifying literature curation with an interpretable hybrid AI strategy − (Figure b). We constructed a rigorously verified data set encompassing diverse solvent environments, derived from recent publications and a reference compilation via a large language model (LLM)-assisted workflow. To couple interpretability with predictive performance, we developed an integrated hybrid modeling strategy consisting of two synergistic stages. (1) Extraction of interpretable prior knowledge: we evaluated classical machine learning algorithms across diverse molecular representations, ranging from classical fingerprints and physicochemical descriptors to modern pretrained molecular embeddings. By applying SHAP analysis to the best-performing model, we aligned statistical features with physical-organic principles, specifically identifying BCUT2D_MRLOWan electronic-topological descriptoras a quantitative proxy for boron hybridization (sp² vs sp³). (2) Prior-guided atom-level deep learning: to decode complex multiboron architectures, we systematically benchmarked various GNN architectures. By fusing the top-ranked global interpretable features extracted in the first stage with a self-attention-based message-passing mechanism, the best-performing Graph Transformer model achieved superior atom-level prediction, effectively predicting the distinct chemical shifts of individual boron atoms within complex molecular graphs. To further promote accessibility and practical utility, we have deployed this framework as an open-access interactive web server, thereby bridging advanced deep learning methods with routine structural elucidation in synthetic chemistry.

Results and Discussion

Data Set Construction

High-fidelity data underpins robust machine learning. To construct a curated data set with explicit solvent annotations, we integrated experimental values from a reference compilation and the Supporting Information of recent publications. As illustrated in Figure a, this effort yielded 2190 validated entries from the compilation and an additional 3300 entries from the literature. The curation process employed a semiautomated pipeline utilizing Prompt Engineering with LLMs, followed by a rigorous human validation protocol. We ensured that all chemical shifts were referenced to the external standard BF₃·Et₂O (δ = 0.0 ppm) and restricted the data set to measurements recorded at 298 K to minimize temperature-dependent shift variations. After removing duplicates and verifying structure-spectrum correspondence, the final high-quality data set comprises 5490 entries.

2.

Construction and descriptive statistics of the solvent-annotated ¹¹B NMR data set. (a) Workflow for converting unstructured sources (journal SI and books) into structured records via keyword-guided localization, LLM-based extraction, and rigorous human validation. (b) Global distribution of ¹¹B NMR chemical shifts across all 5490 records. (c) Distribution of molecular size represented by atom counts per compound. (d) Data distribution across the 10 distinct solvent environments. (e) Statistical analysis of boron centers: (i) counts of sp² vs sp³ hybridized boron atoms. Note that the total atom count (N = 6, 686) exceeds the number of data set entries due to the inclusion of multiboron systems. (ii) Structural composition showing the proportion of single-boron vs multiboron compounds (17.5% multiboron species).

The statistical characteristics of this data set demonstrate both broad chemical coverage and structural diversity. The distribution of ¹¹B NMR chemical shifts (Figure b) spans a wide range from −46.2 to 90.0 ppm, effectively covering the spectral window of typical organoboron compounds. In terms of molecular size, the distribution of atom counts per compound (Figure c) confirms the inclusion of diverse architectures ranging from small fragments to complex systems. Crucially, the data set is explicitly solvent-annotated, featuring data across 10 distinct solvent environments (Figure d). While CDCl₃ remains the predominant medium, significant subsets in C₆D₆, DMSO-d ₆, and acetone-d ₆ provide critical variance for learning specific solvent–solute interactions. Regarding electronic environments, an atom-level analysis (Figure e­(i)) reveals a composition of 4577 sp²-boron and 2109 sp³-boron atoms. Furthermore, to address the challenge of spectral complexity in multicenter systems, often overlooked in prior benchmarks, the data set includes a substantial proportion of multiboron compounds (Figure e­(ii)), which constitute 17.5% of the total collection. Notably, 273 of these compounds feature chemically inequivalent boron sites, and this specific structural heterogeneity provides a robust test bed for atom-level shift discrimination.

Extraction of Interpretable Prior Knowledge via Classical ML

To decode the electronic and structural factors governing ¹¹B NMR chemical shifts, we conducted a systematic evaluation of the representational efficacy. Our objective was to identify an optimal classical model configuration that couples predictive accuracy with chemical interpretability, enabling the robust extraction of the global prior knowledge. We assessed 11 distinct molecular representations spanning explicit physicochemical descriptors, topological fingerprints, and latent embeddings across 7 regression algorithms (e.g., XGBoost, random forest, and lightGBM).

This comparative analysis indicates that for this data set and regression setting, hybrid representations consistently outperformed single-modality baselines. Among the evaluated configurations (Table ), the combination of topological fingerprints with physicochemical descriptors (Morgan-RDKit) paired with XGBoost demonstrated the highest performance in this benchmark, achieving a test set R ² of 0.921, root-mean-square error (RMSE) of 6.18 ppm, and mean absolute error (MAE) of 2.76 ppm (averaged over three independent random splits). This performance compares favorably to models relying solely on explicit physicochemical features (R ² = 0.897) or pretrained embeddings (e.g., UniMol, R ² = 0.370). These results suggest that accurate ¹¹B NMR prediction benefits from integrating physicochemical (electron density-related) descriptors (captured by RDKit) with local substructural environments (captured by Morgan fingerprints). Consequently, the Morgan-RDKit + XGBoost model was selected as the representative model for subsequent interpretability analysis.

1. Comparative Performance of the Top-Performing Machine Learning Models for Each Molecular Representation in Predicting ¹¹B NMR Chemical Shifts (Ranked by Test-Set R ²)­ .

descriptor set	best algorithm	test R 2	test RMSE (ppm)	test MAE (ppm)
Morgan-RDKit	XGBoost	0.921	6.18	2.76
MACCS	XGBoost	0.911	6.60	2.92
UniMol-RDKit	HGBR	0.899	7.02	3.39
RDKit	XGBoost	0.897	7.06	3.16
MACCS-RDKit	XGBoost	0.893	6.91	3.64
Morgan	XGBoost	0.882	7.61	3.63
MolT5-RDKit	HGBR	0.877	7.72	3.89
MolT5-Morgan	LightGBM	0.875	7.78	4.06
MolT5-MACCS	XGBoost	0.871	7.91	3.83
MolT5	CatBoost	0.760	10.84	6.46
UniMol	LightGBM	0.370	16.55	11.73

^a Note: The benchmarking was performed on the subset of compounds featuring chemically equivalent boron sites (n = 5217), derived from the full data set (N = 5490) by excluding multiboron systems with inequivalent sites. For the multiboron compounds retained in this subset, since all boron atoms are chemically equivalent and share an identical chemical shift, this single value was used directly as the target for molecular-level prediction.

^bModels utilizing only topological fingerprints or pretrained embeddings were evaluated on the full benchmark set (n = 5217).

^cModels incorporating RDKit physicochemical descriptors were evaluated on a compatible subset (n = 4680). This subset was generated by filtering the benchmark set (n = 5217) to remove compounds containing specific heavy heteroatoms (e.g., Se and Sn) that caused calculation errors (NaNs) for specific descriptors. All performance metrics represent the averages of three independent random splits.

Building upon the quantitative correlation observed in the test set (Figure a, representative test split R ² = 0.930), we sought to uncover the chemical basis driving these predictions. We employed SHAP (SHapley additive exPlanations) analysis to deconstruct the feature contributions. As visualized in the SHAP summary plot (Figure c), the analysis reveals a hierarchical distribution of the feature importance.

3.

Performance and interpretability of the ML model. (a) Parity plot of predicted vs experimental ¹¹B NMR chemical shifts for the Morgan-RDKit + XGBoost model on the test set (n = 468, colored by solvent). (b) Correlation between the top-ranked descriptor, BCUT2D_MRLOW, and experimental shifts. High values correspond to electron-deficient sp² centers (blue), while low values indicate saturated sp³ centers (orange). The discrimination is statistically significant (p = 2.26 × 10^–42). (c) SHAP summary plot ranking the top 20 features by global importance. Each dot represents a sample; color indicates feature value (red = high, blue = low). (d) Representative substructures for key Morgan fingerprints identified in the SHAP analysis.

Global Feature Contribution

On the basis of the normalized sum of mean absolute SHAP values, solute-derived features account for 96.7% of the total importance, whereas solvent descriptors contribute 3.3%. This pattern is consistent with physical intuition: the intrinsic electronic structure of the organoboron species dominates while solvent effects modulate the environment.

Hierarchical Feature Analysis

A detailed examination of the top-ranked features (Figure c) suggests that the model prioritizes chemical information through three distinct layers:

• (1)Hybridization (tier 1): the feature with the highest impact, BCUT2D_MRLOW (Rank 1, Mean |SHAP| = 4.38), correlates with the boron hybridization state. This descriptor represents the lowest eigenvalue of the Burden matrix weighted by molar refractivity (a proxy for polarizability). As shown in Figure b, high values correlate with electron-deficient sp²-boron centers (low-field), while low values correspond to saturated sp³-boron centers (high-field). Statistical analysis confirms that this discrimination is highly significant (p = 2.26 × 10^–42), consistent with the link between electronic softness/polarizability and shielding contributions.
• (2)Structural motifs (tier 2): complementing the global baseline, specific Morgan fingerprints act as structural indicators (Figure d). These assignments are obtained by mapping the activated Morgan bits back to the representative substructures. Morgan_bit_690 (rank 2) is associated with cyclic boronate esters, contributing to the deshielding effect (+25 ppm); Morgan_bit_271 (rank 3) captures features of acyclic sp² boron and heterocyclic carbon (+20 ppm); and Morgan_bit_390 (rank 6) corresponds to borohydride-like sp³ environments, associated with significant shielding effects (−41 ppm).
• (3)Electronic fine-tuning (tier 3): finally, physicochemical descriptors account for subtle variations. MaxPartialCharge (rank 5) captures local electron density fluctuations, while SlogP_VSA7 (rank 4) appears to function as a proxy for steric effects, modulating the observed chemical shift.

While solute features dominate the prediction, the remaining contribution from solvent descriptors provides a minor but non-negligible effect that helps to account for solvent-induced shift variations. Detailed analysis suggests that the model captures solvent effects compatible with Lewis acid–base interactions. The top-ranked solvent features, Solvent_MinPartialCharge (SHAP = 0.143) and Solvent_BCUT2D_CHGLO (SHAP = 0.075), are proxies related to charge distribution and thus to donor–acceptor tendencies. A more negative partial charge implies a stronger tendency for the solvent to donate electron density to the Lewis acidic boron center, thereby increasing the shielding. Additionally, the topological index Solvent_Chi0v (SHAP = 0.100) reflects the steric bulk and connectivity of the solvent molecules. Although the global weight of these features is low, their inclusion enables the model to account for solvent-induced shift variations, which are essential for distinguishing chemically similar environments.

Prior-Guided Graph Learning for Chemically Inequivalent Boron Sites

To overcome the limitations of molecular-level descriptors in resolving chemically inequivalent boron sites within multiboron systems, we constructed a node-level regression framework based on GNNs. We systematically evaluated the nine distinct GNN architectures. To ensure robust evaluation, models were trained using a 5-fold cross-validation strategy within the training split, and the final ensemble predictions were evaluated on a held-out 20% test set. Within this data set, the 273 compounds with inequivalent boron sites were split into 215 for training and 58 for testing.

Performance Comparison and Ablation Analysis

We report the comparative performance of all nine GNN architectures on the held-out test set in Table . Notably, the integration of global interpretable priors yielded architecture-dependent effects. For the top-performing attention-based models (graph transformer, GATv2) and the foundational GCN, the prior knowledge improved predictive performance. Among the nine architectures, the prior-guided graph transformer achieved the highest R ² (0.946) and the lowest RMSE (5.217 ppm), while GATv2 yielded a marginally lower MAE (1.971 ppm versus 1.979 ppm). Overall, the graph transformer provided the strongest balance across the three evaluation metrics. Among the 3D equivariant models, PaiNN achieved both the highest R ² (0.942) and the lowest MAE (2.093 ppm), although it did not outperform the leading Graph Transformer and GATv2 models, possibly reflecting geometric uncertainty associated with the use of a single RDKit-derived conformer. Conversely, for certain architectures such as GINE and NNConv, the direct concatenation of global features led to slight performance degradation, likely due to feature redundancy. This pattern suggests that attention mechanisms (Graph Transformer, GATv2) show greater compatibility in dynamically weighing and integrating macroscopic physical-organic descriptors with local graph topologies. By filtering out structural noise and modeling longer-range context, the prior-guided Graph Transformer translates these interpretable priors into a more accurate and chemically informed predictive model, representing a substantial leap over the best classical ML baseline (R ² = 0.921, MAE = 2.76 ppm). To assess the effect of training-set expansion, we expanded the training set from 4392 to 17,897 records by adding 13,505 nonoverlapping ¹¹B records from NMR exp, which yielded only modest, architecture-dependent performance changes on the unchanged held-out test set (Table S21).

2. Performance Comparison of Nine Different GNN Architectures Evaluated on the 20% Held-Out Test Set .

model	R 2	RMSE	MAE	MAE
graph transformer	0.946	5.217	1.979	2.015
GATv2	0.945	5.268	1.971	2.042
PNA	0.943	5.344	2.083	2.090
PaiNN	0.942	5.430	2.093	2.102
NNConv	0.940	5.532	2.190	2.188
GINE	0.939	5.543	2.074	2.042
EGNN	0.938	5.607	2.374	2.373
GCN	0.936	5.713	2.231	2.291
MACE	0.934	5.783	2.343	2.371

^a Note: Evaluated on the 20% held-out test set; MAE and RMSE are in ppm.

^bModels were augmented with the top 20 SHAP features.

^cBaseline models without prior ML features were used.

^d3D equivariant neural network architectures: EGNN, PaiNN, and MACE.

Furthermore, to assess generalization beyond random splitting, we performed Tanimoto–Butina cluster-based hold-out evaluations using Morgan fingerprints (Table S16 and Figure S2). Across six similarity thresholds (0.3–0.8), the model retained strong predictive performance on cluster-held-out test molecules (MAE = 0.66–1.68 ppm), supporting its robustness under structure-based splitting and its potential utility for exploring less familiar chemical space.

Prior-Guided Graph Transformer Architecture

The strong performance of the Graph Transformer in this benchmark stems from its ability to capture both neighbor-dependent local environments and long-range topological context (Figure a). The model processes the input graph G = (V, E) through a three-stage pipeline: (1) Attention-based message passing: node and edge attributes are projected into a high-dimensional latent space (d = 256) and processed by stacked transformer-style attention convolution layers (TransformerConv) (see Section S5 in the Supporting Information for network details). Unlike isotropic convolution, the multihead self-attention mechanism computes learnable attention scores for each neighbor, enabling the model to dynamically prioritize critical substructures based on feature relevance rather than mere topological proximity. (2) Global information aggregation: to mitigate “over-smoothing” and limited receptive fields, we instantiated a Virtual Node connected to all atoms. This global node is updated at each layer via pooled graph representations (AddPool) and broadcast back, establishing a shortcut for information propagation to facilitate the learning of long-range dependencies. (3) Prior-guided multi-modal fusion: to construct the final predictive representation, the updated node embeddings are concatenated with two crucial contextual components: a learnable solvent embedding (d _s = 32) to capture environmental effects, and the top-ranked global interpretable features (derived from the classical ML stage) to inject macroscopic physical-organic priors. This fused representation is then fed into a Multilayer Perceptron prediction head. Empirical evaluation on the held-out test set confirms the efficacy of this architecture (R ² = 0.946, Figure b), demonstrating the model’s ability to accurately map complex molecular structures to specific ¹¹B chemical shifts.

4.

Architecture and performance of the prior-guided Graph Transformer with solvent input. (a) Schematic of the graph transformer architecture (see Section S5 in the Supporting Information for detailed model architecture), illustrating node/edge encoding, virtual node updates, and the integration of the extracted global ML priors alongside a 32-dimensional learnable solvent embedding. (b) Parity plot of the prior-guided Graph Transformer predictions versus experimental chemical shifts on the 20% held-out test set. (c) Qualitative t-SNE visualization of the 32-dimensional solvent embeddings, annotated by solvent polarity (color) and proticity (shape).

Effect of Including Solvent Input

The architecture utilizes end-to-end optimized, learnable solvent embeddings. Qualitative t-SNE visualization (Figure c) shows partial grouping of some solvent classes, suggesting that the learned embeddings retain information related to solvent identity and broad solvent properties. To rigorously quantify their impact, we conducted an expanded ablation study across four random seeds (Table S15). Adding solvent input yielded a reproducible MAE improvement (2.045 ± 0.284 vs 2.092 ± 0.293 ppm; mean paired decrease = 0.046 ± 0.013 ppm), alongside marginal gains in RMSE and R ². This indicates that while molecular structure dictates the primary shift, solvent embeddings provide beneficial secondary fine-tuning. We further evaluated this via solvent-swapping predictions across the held-out test set (n = 1098; Table S18). The results exhibit a systematic dependence on solvent polarity: protic/high-polarity solvents generally induce shielding, whereas aromatic/nonpolar solvents induce deshielding. The maximum average predicted variation of approximately 3.55 ppm suggests model sensitivity to subtle environmental contributions. Finally, to evaluate the impact of the highly imbalanced solvent distribution, we report solvent-stratified test-set performance in Table S17. The model achieves robust accuracy for well-represented solvents, notably CDCl₃ (MAE = 1.37 ppm) and DMSO-d ₆ (MAE = 0.94 ppm). However, estimates for sparsely sampled solvents (e.g., THF-d ₈, D₂O, and toluene-d ₈) carry higher uncertainty due to limited test targets, making these specific metrics less statistically robust.

Case Studies: Overcoming Experimental Bottlenecks with Virtual Resolution

To further assess the practical utility of our framework in real-world synthetic scenarios, particularly in situations where traditional manual interpretation is hindered by intrinsic physical limitations, we applied the model to a series of literature-reported (1–5) and in-house synthesized organoboron compounds (6–10) (Figure a,b), which collectively encompass 18 distinct boron sites. In structurally complex multiboron systems, chemical shift assignments are frequently obscured by severe line broadening induced by quadrupolar relaxation. Such experimental ambiguity often makes it exceptionally difficult for chemists to distinguish subtle electronic differences between spatially proximate boron centers. For example, in the triboron species 5, our prior-guided Graph Transformer successfully discriminated the unique tetrahedral sp³ center (B1) from the two chemically equivalent trigonal sp² centers (B2/B3) with high precision. Across this entire 18-site subset, the prior-guided Graph Transformer achieved an MAE of 0.19 ppm, compared with 1.12 ppm for the unscaled DFT predictions (Table S20). Crucially, the DFT calculations required an average wall-clock time of 34.36 h per molecule (14.31 days in total for the 10-compound set), whereas the graph model’s inference required only seconds. This highlights the practical value of our integrated framework for rapid, atom-resolved ¹¹B NMR assignment.

5.

Representative examples and web deployment of the boron NMR prediction model. (a) Representative literature test-set compounds (1–5). (b) Wet-lab compounds synthesized and measured in this work (6–10). Boron centers are labeled as B1–B3. For each boron center, δ­(¹¹B) values are listed in the order Lit./ML/DFT for compounds 1–5 and Exp./ML/DFT for compounds 6–10, corresponding to literature or experimental values, machine-learning predictions, and DFT-calculated values, respectively. All values are reported in ppm. (c) Web deployment interface of the boron NMR prediction platform, showing molecule input, solvent selection, model inference, and prediction output. The model-inference schematic is simplified for web deployment; prior features, graph-derived molecular embeddings, and solvent embeddings are concatenated in parallel at the feature-fusion step.

More importantly, the model can break structural deadlocks where experimental resolution fundamentally fails. For the newly synthesized bis-boronates 9 and 10, standard ¹¹B NMR measurements yielded only a single, severely broadened peak, rendering it impossible to experimentally confirm whether the two boron sites were chemically equivalent in solution. Applying our framework to these species provided a powerful form of “virtual resolution”. The model confidently predicted two close yet distinct chemical shifts (33.18 and 33.41 ppm for 9; 33.53 and 33.43 ppm for 10), capturing an intrinsic site inequivalence driven by subtle long-range topological asymmetry that the physical spectrometer could not resolve. The accompanying DFT calculations also predicted distinguishable site-specific shifts for these compounds, providing an independent quantum-chemical reference for the ML-based virtual resolution. To facilitate broad community access, we have deployed this prior-guided, site-specific prediction model as an open-access web server (Figure c; available at https://modelscope.cn/studios/AI4Chem/boron-nmr-prediction), providing a practical tool to assist in routine structural elucidation.

Conclusions

In summary, we have established a robust data-driven paradigm for ¹¹B NMR spectroscopy that offers a balanced approach, combining computational efficiency with chemical accuracy. By constructing a rigorously verified, solvent-annotated data set of 5490 entries through an LLM-assisted workflow, we addressed the data scarcity that has limited the application of artificial intelligence in heteronuclear NMR. Our integrated hybrid modeling strategy successfully harmonized predictive performance with interpretability: molecular-level machine learning identified the electronic-topological descriptor BCUT2D_MRLOW as a quantitative proxy for boron hybridization, aligning learned feature importance with physical organic intuition. Furthermore, the implementation of a prior-guided Graph Transformer architecture established an effective atom-level modeling framework, achieving strong predictive performance (test R ² = 0.946, RMSE = 5.217 ppm, MAE = 1.979 ppm) by capturing neighbor-dependent local environments and long-range topological dependencies within complex multiboron systems. Notably, the framework also incorporates solvent information to account for solvent-dependent variations in ¹¹B NMR chemical shifts. While the intrinsic molecular scaffold remains the primary determinant of the chemical shift, the inclusion of solvent descriptors enables a secondary adjustment that partially captures the environment-dependent effects.

Beyond statistical benchmarking, the practical utility of this framework was validated through the successful characterization of both literature-reported and in-house synthesized compounds. The model provides a form of “virtual resolution”, enabling the precise assignment of chemically inequivalent boron sites, even in cases where experimental spectra are obscured by quadrupolar relaxation-induced line broadening. This capability offers a rapid complement to computationally demanding DFT calculations for structural elucidation. By deploying this framework as an open-access web server, we provide the synthetic community with a widely accessible tool for real-time spectral analysis. Furthermore, our LLM-assisted workflow offers a generalizable blueprint for unlocking literature “dark data”, providing a scalable solution to data scarcity across other challenging scientific disciplines. Ultimately, this work not only streamlines the structure assignment and routine analysis of organoboron compounds but also lays the foundation for the broader digitization of main-group chemistry, facilitating future discoveries in pharmaceuticals and advanced materials.

The solvent-annotated NMR data set constructed and analyzed during the current study is openly available from https://github.com/SII-ChemNext/boron-nmr-predictor/tree/main/data. The source code and pretrained models of our predictive framework are available at https://github.com/SII-ChemNext/boron-nmr-predictor. Additionally, the site-specific NMR prediction model can be used online as a user-friendly web server via https://modelscope.cn/studios/AI4Chem/boron-nmr-prediction.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jacsau.6c00935.

• Experimental details; data set construction and the LLM-assisted extraction and validation workflow; model architecture and training procedures; supplementary benchmark and ablation results; additional figures and tables; and prediction results for the in-house synthesized compounds (PDF)

#.

P.L. and B.G. contributed equally to this work. Y.L. and G.Y. conceived the concept, directed the project, and provided critical feedback and revisions on the manuscript. P.L. and B.G. established the LLM-assisted workflow to mine solvent-annotated NMR chemical shift data from the literature to construct the complete data set, designed the integrated hybrid AI framework, and built the interactive web server. W.L. participated in the discussion of the experiments and provided assistance with the experimental procedures. S.W. and Z.W. conducted the wet-lab experiments to verify the predictive performance of the model.

Safety Statement: No unexpected or unusually hazardous experimental procedures were encountered.

The authors declare no competing financial interest.