2D-JCOG: Transforming 1D 1H NMR Spectra into J‑δ Correlation Maps via the Shared Splitting Theorem and Graph Neural Networks Authors: Carlos Cobas Published: 2026 ABSTRACT Scalar coupling constants and their associated coupling networks encode essential structural information in 1H NMR spectra, yet their automated extraction remains challenging. Traditional multiplet analysis relies on first-order approximations that fail under spectral congestion, impurities, or strong coupling conditions, while iterative quantum-mechanical (QM) methods require precise initial estimates and often suffer from convergence issues. We present 2D-JCOG (J-COrrelation via Graph neural networks), a deep learning framework that transforms 1D 1H NMR peak lists into 2D J-δ correlation maps revealing the coupling topology. The method is grounded in the Shared Splitting Theorem (SST), which establishes that coupled multiplets must share at least one common splitting regardless of coupling strength. This physical invariant provides the inductive bias for a heterogeneous Graph Neural Network architecture where splitting candidates are represented as nodes connected by SST-based similarity edges. The network employs a hybrid message-passing strategy: transformer-style attention layers with physics-aware edge weighting validate SST partnerships by learning to gate connections based on quantum mechanical consistency features, while mean-aggregation layers aggregate hierarchical context from multiplet regions. Training on QM-simulated spectra with exact ground truth, the model achieves 92–96% recall and 82–91% precision across varying spectral complexity levels. Validation on experimental spectra from the GISSMO database demonstrates 91.9% recall with a mean J-value error of 0.113 Hz. 2D-JCOG provides a robust alternative to classical multiplet analysis, extending applicability to moderately coupled systems, where first-order rules break down. While not intended to replace iterative QM methods for severely strongly coupled spin systems, it offers automated J-coupling extraction with direct visualization of coupling connectivity for routine spectral analysis. Introduction 1H NMR spectra encode molecular structural information in the form of spin systems, whose precise characterization via chemical shifts and scalar coupling constants is essential for establishing both the molecular constitution and configuration. , Conventionally, these parameters are determined via multiplet analysis, a procedure relying on heuristic protocols grounded in first-order theory, such as the “n + 1” rule, which is currently well-automated in specialized software using peak-picking or advanced deconvolution techniques. , However, this approach faces intrinsic limitations: it fundamentally requires first-order behavior and typically processes each multiplet in isolation, meaning the global topology of the coupling network is not directly leveraged during extraction but only inferred a posteriori. However, as signal congestion increases or systems exhibit strong coupling effects (where the ratio Δν/J decreases), the strict application of classical multiplicity rules becomes insufficient. Although computational methods can often accommodate simple intensity perturbations such as the roof effect, they become unreliable when higher-order interactions fundamentally alter the multiplet structure. For instance, when the breakdown of the “n + 1” rule manifests as a change in the number of observable transitions, standard peak analysis or deconvolution algorithms are unable to accurately resolve the underlying spin system. Furthermore, while sophisticated iterative approaches based on Quantum Mechanical (QM) spectral analysis can theoretically resolve these complex features by fitting the full spin Hamiltonian, they are characteristically computationally intensive and prone to convergence failures in the absence of precise initial estimates. On the experimental side, specialized 2D NMR correlations such as E.COSY, 2D J-resolved spectroscopy, or selective refocusing techniques (e.g., SERF) offer alternatives capable of simultaneously determining spin system topology and quantifying coupling constants. However, these methods are rarely employed in routine structural-analysis workflows. Their widespread adoption is hindered by substantial sensitivity penalties and prolonged acquisition times compared with standard 1D 1H NMR, often restricting their application to complex elucidation cases that require expert intervention. Recognizing these limitations, we present 2D-JCOG (2D J-COrrelation via Graph neural networks), a novel spectral analysis tool that maps 1D spectral data into a 2D correlation plot, effectively displaying coupling constants against chemical shifts. This transformation allows for the direct inspection of the coupling topology as well as the extraction of J values and chemical shifts. It is not intended to replace iterative Quantum Mechanical (QM) methods specifically tailored for strongly coupled systems (such as AA’BB’); rather, it serves as a complement and a superior alternative to the traditional multiplet analysis described above. While it does not explicitly solve for parameters in severe strong-coupling regimes, it proves robust in numerous scenarios where classical heuristic methods fail, with the distinct advantage of directly revealing the coupling connectivity. The conceptual basis of the present methodology builds upon the “J-Correlator” approach, which demonstrated that systematic detection of shared splittings across multiplets could enable automated coupling network reconstruction. This approach is grounded in the Shared Splitting Theorem (SST), a fundamental principle that will be formally detailed in the Theory section. Based on the premise that coupled spins must share at least one common splitting magnitude, the original algorithm generated two-dimensional correlation maps linking chemical shifts via their shared splittings. While theoretically enabling the direct visualization of the spin system topology, the practical application of this concept suffered from a high rate of false positive correlations caused by numerical coincidences lacking physical origin. Although subsequent attempts sought to suppress these artifacts through heuristic filtering, the combinatorial complexity of nontrivial spectra ultimately limited the tool’s effectiveness to simple molecular systems. Recent advancements in Deep Learning provide the ideal framework to overcome these limitations and evolve the original concept into a robust analytical solution. Recognizing that spin systems are naturally analogous to graph structures, Graph Neural Networks (GNNs) operating through iterative message passing are particularly well-suited for this task. The 2D-JCOG methodology therefore constructs a heterogeneous graph representation of the spectrum, formulated to learn splitting patterns that validate the Shared Splitting Theorem. By directly encoding physical constraints such as “near-J” edges and processing them via attention-based message passing that incorporates edge-level physical features, the system successfully distinguishes genuine scalar couplings from spectral artifacts, maintaining high precision even in chemically dense environments. Recently, Schmid et al. proposed MolDeTr, a detection transformer framework that addresses the full inverse problem of extracting complete spin system parameters directly from 1D 1H NMR spectra. While MolDeTr targets the detailed parametrization of individual spin systems within defined spectral windows, the present work pursues a complementary objective: the construction of a global coupling correlation map across the entire spectrum, revealing the topology of scalar coupling networks. The two approaches are naturally synergistic, as the coupling maps generated by 2D-JCOG could serve as prior constraints for spin system fitting methods, while detailed solutions from approaches such as MolDeTr could refine local predictions within the correlation map. Theory The Shared Splitting Theorem as a Topological Invariant for Spin System Analysis The shared splitting theorem (SST) establishes a fundamental constraint on the structure of NMR spectra. Consider a two-spin fragment A, B embedded within a larger coupled spin system. As illustrated in Figure , four energy levels can be identified corresponding to the four possible spin state combinations of nuclei A and B, while all other nuclei maintain constant spin states. Energy level diagram for a two-spin subsystem. The four transitions (A0, A1 for nucleus A; B0, B1 for nucleus B) form a closed quadrilateral, implying A1 – A0 = B1 – B0. The closed quadrilateral geometry imposes a strict constraint: the splitting A1-A0 within multiplet A must exactly equal the splitting B1–B0 within multiplet B. This is the shared splitting theorem: If two multiplets are coupled, they MUST share a splitting regardless of the complexity of the spin system or the strength of the couplings . The power of the SST lies in its unconditional validity: two coupled multiplets will always share a splitting, regardless of coupling strength. However, an important distinction must be made between the existence of shared splittings and their correspondence to actual J-coupling constants. In first-order (weak coupling) spectra, shared splittings equal the J values exactly. Under moderate second-order conditions, they remain good approximations. In strongly coupled systems such as AA’BB’, the shared splittings may deviate significantly from the true coupling constants due to quantum state mixing. Nevertheless, the splittings remain sharedthe theorem holds exactly even when the spectral patterns become highly distorted. The SST guarantees a structural relationship in the spectrum that can always be exploited for pattern recognition, even when direct J extraction becomes approximate. This invariance transforms J-coupling extraction from local peak analysis into global pattern recognition: a candidate coupling is validated not by its local context alone, but by the existence of a matching splitting elsewhere in the spectrumits ″SST partner.″ The SST principle provides the foundational inductive bias for the entire 2D-JCOG methodology. From candidate generation to graph construction to model training, each component is designed to detect and exploit shared splittings across the spectra. Because this structural relationship holds universally, a model that learns to leverage it can generalize to any spin system regardless of molecular complexity. The specific implementation of these SST-aware components is detailed in the subsequent sections. Experimental Methods Two distinct data sources were employed in this work. For model training and initial evaluation, approximately 22,000 1H NMR spectra were quantum-mechanically simulated from predicted spin system parameters of organic molecules spanning diverse chemical space (see Training Data section below for details). For experimental validation, 1H NMR spectra from the GISSMO database were used, covering over 1,000 small molecules acquired predominantly at 400–600 MHz under documented sample conditions. These experimental spectra were processed automatically with Mnova: Global Spectral Deconvolution (GSD) was applied to generate peak lists, and peaks classified by the software as solvent signals, artifacts, chemical shift reference standards (e.g., TMS), and weak signals such as impurities or 13C satellites were automatically excluded prior to analysis. Full details on experimental data curation and evaluation criteria are provided in the Experimental Results section. The following subsections describe the complete computational pipeline from peak list to 2D δ-|J| correlation map. The 2D-JCOG methodology translates the Shared Splitting Theorem into a complete computational framework encompassing both the construction of training data from quantum-mechanical simulations and an inference pipeline that transforms experimental peak lists into a 2D δ-|J| correlation map. This section provides a conceptual overview of the complete system; more comprehensive details are provided in the Supporting Information. Figure illustrates both aspects of the methodology. The training pipeline (top) generates labeled examples from quantum-mechanically simulated spectra, where exact ground truth for all chemical shifts and coupling constants is known. The inference pipeline (bottom) processes experimental peak lists through six interconnected stages: (i) peak list generation, where spectral deconvolution extracts individual resonances from the raw spectrum; (ii) ROI segmentation, partitioning peaks into regions containing putative multiplets; (iii) candidate generation, enumerating all possible splittings within each region; (iv) feature extraction, encoding local spectral properties and SST support metrics; (v) graph construction, organizing candidates into a structure that explicitly represents SST relationships; and (vi) GNN scoring, where a trained neural network propagates evidence across the graph to assess each candidate and produce δ−|J| correlations with associated confidence scores. Overview of the 2D-JCOG pipeline. The shared processing pipeline (green) is identical for training and inference, differing only in the input source. During training (blue), GNN output probabilities P( J ) are compared against ground truth labels to optimize the network weights. During inference (orange), the trained model produces a 2D δ−|J| correlation map of validated J-couplings. Both pipelines share identical processing logic from the peak list onward, ensuring that the model encounters during training the same data representations it will process during inference. The following sections provide a conceptual description of each pipeline stage; detailed algorithmic specifications and implementation parameters are provided in the Supporting Information. Training Data: Quantum-Mechanical Spectral Simulations The foundation of 2D-JCOG is a training data set comprising approximately 22,000 quantum-mechanically synthesized 1H NMR spectra. These simulations solve the full spin Hamiltonian to generate synthetic spectra with exact ground truth for all chemical shifts and coupling constantsinformation that would be impossible to obtain at scale from experimental data. The input spin systems were derived from predicted NMR parameters of organic molecules spanning diverse chemical space, encompassing varying levels of spectral complexity and spin system topologies. All predictions and QM spectral synthesis were performed using Mnova software. An important design choice addresses the risk of overfitting to systematic biases in predicted parameters. Each base simulation was replicated with random perturbations of ±5% applied independently to all chemical shifts and coupling constants, creating an augmented data set. To ensure that training data faithfully represent experimental conditions, peak lists were generated using global spectral deconvolution (GSD) rather than directly employing the quantum-mechanical transition frequencies (see Peak List Generation section for details). This approach introduces realistic peak characteristics, including the natural peak-picking uncertainties encountered in experimental workflows. The use of GSD during training is essential: it ensures the model learns from the same peak representations it will encounter during inference on experimental data, rather than from idealized transition lists that never occur in practice. The data set was stratified into three complexity tiersTractable, Mixed, and Challengebased on the presence of congested spectral regions. This stratification enables systematic evaluation of model performance across different difficulty levels, from well-resolved multiplets to heavily overlapped spin systems with strong coupling effects (see section S1.1.2 for detailed stratification criteria). Peak List Generation Both training and inference pipelines begin with peak list generationthe transformation of a continuous spectrum into a discrete set of resonances characterized by chemical shifts, intensities, and line widths. The quality of J-coupling assignment is fundamentally bounded by the quality of the underlying peak list; no downstream algorithm can recover couplings whose constituent peaks were missed or incorrectly characterized during this initial step. 2D-JCOG is designed to operate on peak lists from any source, but an optimal performance requires a deconvolution algorithm capable of resolving partially overlapped resonances. For this work, we employed global spectral deconvolution (GSD), which models the spectrum as a sum of generalized Lorentzian (or mixed Lorentzian–Gaussian) line shapes and optimizes peak parameters globally to minimize residual signal. The critical requirements for the input peak list are (i) accurate chemical shift determination, as systematic position errors propagate directly into J-value errors; (ii) reliable detection of overlapped peaks, since missed components lead to incomplete multiplet characterization; and (iii) reasonable intensity estimates, which inform the geometric weighting of candidate splittings. While GSD satisfies these requirements for most experimental scenarios, alternative algorithms with comparable resolution capabilities could be substituted without modification to the downstream pipeline. From Peaks to Splitting Candidates Given a peak list, the inference pipeline proceeds through ROI segmentation: following intensity normalization and noise filtering, peaks are clustered into Regions of Interest (ROIs), which are contiguous frequency intervals separated by gaps exceeding the maximum expected J-coupling (∼18 Hz). Each ROI typically corresponds to one or more overlapping multiplets. The candidate generation stage then performs exhaustive splitting enumeration within each ROI by computing all pairwise frequency differences between peaks. For an ROI containing n peaks, this generates n(n – 1)/2 candidate splittings, each representing a hypothesis that the frequency difference corresponds to a true J-coupling. A geometric weight based on constituent peak intensities is assigned to each candidate, reflecting the expected correlation between coupling strength and peak prominence. An important architectural decision concerns the treatment of SST evidence during candidate generation. An obvious strategy would use SST as a hard filter, discarding candidates lacking reciprocal support at the putative partner’s position. While effective at reducing false positives, this aggressive filtering inevitably eliminates true couplings in cases where the partner multiplet is obscured by overlap or where peak-picking failures prevent detection of the reciprocal splitting. 2D-JCOG adopts a ″soft SST″ strategy: all candidates satisfying basic physical constraints are retained regardless of SST support, with the reciprocal partner evidence encoded as learned features rather than hard filters. A systematic evaluation on thousands of simulated spectra demonstrated that SST filtering caused non-negligible recall loss while providing only marginal precision improvements. By delegating the precision/recall trade-off to the neural network, 2D-JCOG enables context-dependent decisions that rigid filtering cannot achieve. To manage the high redundancy inherent in exhaustive generation, a greedy nonmaximum suppression (NMS) algorithm consolidates overlapping candidates, retaining only the highest-weighted instance within tolerance windows. The surviving candidates form the nodes of the graph passed to the neural network (see section 1.2 of the Supporting Information for complete algorithmic details). Feature Engineering The feature extraction stage characterizes each candidate splitting with a comprehensive feature vector encoding local spectral properties, regional contexts, and global consistency measures. The features are specifically designed to capture both the physical plausibility of a proposed coupling and its consistency with the shared splitting theorem. Local candidate features describe intrinsic properties including the splitting magnitude, geometric mean intensity of constituent peaks, and intensity balance metrics quantifying the symmetry expected in first-order multiplets. ROI context features encode information about the parent multiplet region: its width, peak count, peak density, and the candidate’s position relative to the ROI center. SST support features implement the soft encoding strategy central to 2D-JCOG’s design. Rather than filtering candidates based on SST compliance, multiple features quantify the degree of reciprocal support: the count of other ROIs containing similar J values, quality metrics for the best-matching partner, and bilateral consistency scores. This detailed representation allows the network to learn nuanced relationships between SST evidence strength and coupling validity. Several features specifically target common sources of false positives. Harmonic detection features identify candidates whose J value approximates twice another candidate in the same ROIa signature of overtone artifacts. Combinatorial closure features flag candidates arising from arithmetic relationships between other splittings, which frequently occur as coincidental frequency differences rather than true couplings. The complete feature set comprises 60 engineered descriptors, enumerated with full definitions in Supporting Information Table S1. Graph Representation The graph construction stage organizes the feature-enriched candidates into a heterogeneous graph that explicitly encodes the SST relationships at the structural level. The graph comprises two node types that serve distinct semantic roles. Candidate nodes (termed ″circle nodes″) represent individual splitting hypotheses, each characterized by the full feature vector. ROI nodes represent spectral regions, providing aggregation points for multiplet-level information computed by pooling statistics over the constituent candidates. Two edge types encode fundamentally different physical relationships. Membership edges establish the hierarchical containment between candidates and their parent ROIs, instantiated bidirectionally to enable information flow in both directions: candidates inform multiplet-level reasoning, while regional context influences individual candidate assessment. Similarity edges (labeled “SST edges” in Supporting Information Figure S1) directly implement the Shared Splitting Theorem at the graph topology level. Two candidate nodes are connected if their J values differ by less than an optimized tolerance threshold regardless of their spatial positions in the spectrum. This connectivity pattern embodies the SST principle: candidates representing the same physical coupling should exhibit mutual support through graph connections, while spurious candidates will lack such a reinforcement. Similarity edges carry physics-aware features encoding the quality of each potential SST partnership. These include normalized J-value agreement, intensity balance similarity, integral ratio compatibility, and roof effect consistencya binary indicator verifying whether intensity asymmetries in both multiplets point toward each other as dictated by second-order effects. This edge-level encoding reflects the inherently relational nature of the SST: the theorem makes statements about pairs of multiplets, including those within the same ROI when it encompasses multiple overlapping spin systems (see sections 1.4–1.5 of the Supporting Information for complete graph construction details). Neural Network Architecture: Physics-Aware Message Passing The GNN scoring stage employs a graph neural network architecture specifically designed to exploit the dual nature of the coupling assignment problem: rigorous physics-based validation of SST consistency combined with hierarchical aggregation of multiplet-level context. For similarity edges encoding potential SST partnerships, the network employs graph attention layers based on the transformer architecture, which incorporate edge features directly into the attention computation. Unlike standard graph convolutions that treat all edges uniformly, this mechanism learns to modulate attention weights based on the physical compatibility features carried by each edge. The attention coefficient between two candidates incorporates their spectral features (J value, position, and symmetry) alongside the quantum-mechanical validation metrics on the connecting edge. This formulation acts as a “physics gate”: the network learns to suppress attention when edge features indicate violations of spectral consistency, such as mismatched roof effects or incompatible integral ratios. For hierarchical membership edges between candidates and ROIs, the network employs GraphSAGE-style convolutions with mean aggregation. Since these edges represent structural containment rather than physical interaction, a simpler aggregation strategy suffices while maintaining the computational efficiency. This allows ROI nodes to accumulate summary statistics about their constituent candidates, providing contextual information on the local multiplet environment. The hybrid architecture applies expensive physics-aware reasoning selectivelyonly on SST edges where quantum-mechanical validation is criticalwhile maintaining a lightweight backbone for hierarchical context aggregation. Five message-passing layers enable information to propagate across the spectrum, allowing SST evidence from distant multiplets to influence local candidate assessment. Residual connections with layer normalization ensure stable gradient flow across the network depth. Following the final message-passing layer, each candidate receives additional context through dual pooling over its parent ROI, computing both mean and max aggregations to capture different aspects of the multiplet environment. The enriched candidate representation is passed through a classification head yielding a probability score (see section 1.6 of the Supporting Information for complete architectural specifications). Training Protocol Training employs a hybrid objective function addressing two challenges inherent to the coupling assignment problem. The extreme class imbalancetrue couplings constitute a small minority of generated candidatesis handled through focal loss, which down-weights easy negatives to focus learning on difficult examples. The competitive nature of the problem, where multiple false candidates often vie with the true coupling within the same ROI, motivates an additional pairwise ranking loss. This component explicitly enforces correct ordering: for every positive–negative candidate pair within the same multiplet region, the model incurs a penalty if the true coupling is not ranked higher than spurious alternatives. This ensures robust selection, even when absolute probability scores are uncertain. The training set of ∼22,000 quantum-mechanical simulations was split into training (80%) and validation (20%) partitions. Training utilized automatic mixed precision on GPU hardware, with hyperparameters optimized through Bayesian search over learning rate, dropout, focal loss parameters, and ranking loss weight. Early stopping based on validation average precision prevented overfitting, with the best model typically emerging within 200 epochs (Section 1.6.6 of the Supporting Information details the complete training protocol). Ensemble Strategy Beyond the single-model configuration, 2D-JCOG implements an ensemble approach by combining five independently trained models. Each model in the ensemble is initialized with a different random seed and trained with varied hyperparameter configurations, selected on the basis of sensitivity analysis (ANOVA) to identify parameters with the strongest influence on model variance. This diversity in training conditions produces models that, while individually performant, capture complementary aspects of the data. The ensemble strategy serves two purposes: First, aggregating predictions across multiple models improves the overall accuracy by reducing variance and mitigating individual model biases. Second, the distribution of predictions across ensemble members provides a measure of uncertainty: candidates receiving consistent scores across all models can be assigned with high confidence, while those showing substantial disagreement warrant careful inspection. Detailed performance comparisons between single-model and ensemble configurations are presented in the Results section. Output The final stage of the pipeline produces a probability score for each candidate, representing confidence that the splitting corresponds to a true J-coupling. A threshold (default 0.5) converts these scores to binary predictions. The output consists of a list of validated couplings, each characterized by the coupling constant |J| in Hz, chemical shifts of the coupled nuclei in ppm, and their associated probability scores. This list can be used to reconstruct 2D δ−|J| correlation maps for visual inspectionexamples of such maps are presented in the Results sectionor can be directly employed in downstream applications such as structure elucidation, spectral verification, or molecule-spectrum assignments. Results and Discussion This section presents a comprehensive evaluation of the proposed J coupling prediction system. We begin by introducing the evaluation framework, including the rationale for our metric design and the mathematical definitions of the performance measures employed. Subsequently, we present quantitative results comparing the single-model and deep-ensemble configurations across different spectrum complexity levels. The section concludes with illustrative examples from both simulated and experimental spectra to demonstrate the practical utility of the approach. Evaluation Framework The Class Imbalance Issue The J-coupling prediction task presents a severely imbalanced classification problem. For a typical 1H NMR spectrum, the number of candidate coupling pairs generated by the combinatorial pairing of splittings across multiplet regions vastly exceeds the number of true couplings. In our evaluation data set comprising 2,052 spectra, the global ratio of negative to positive candidates is 11:1, with median per-spectrum ratios ranging from 6:1 for tractable spectra to 12:1 for the most complex cases. This imbalance renders standard accuracy metrics uninformative, as a naive classifier predicting all candidates as negative would achieve misleadingly high accuracy scores. Furthermore, the objective of this work is not the precise numerical determination of coupling constant values but rather the construction of a two-dimensional coupling map that reveals the topology of spin–spin interactions within the molecule. This visualization-oriented goal influences our evaluation strategy: what matters most is whether the system correctly identifies the presence and approximate location of couplings in the (chemical shift, J value) space, rather than achieving subhertz accuracy in J value estimation. Tolerance Region-Based Evaluation The coupling correlation mapping task presents an additional challenge beyond class imbalance: multiple candidates may yield practically degenerate solutions that are extremely difficult for the GNN to distinguish. This degeneracy arises from several sources, including signal overlap affecting peak position determination, limited digital resolution, and second-order effects that introduce systematic variations in the apparent splitting values. Consequently, candidates lying very close to the ground-truth coordinates can exhibit nearly identical feature representations, making fine-grained discrimination unreliable. During model development, we observed that employing very tight tolerancesunderstood here as the Euclidean distance between ground truth and putative correlationshindered training convergence. When the model is penalized for predictions that are essentially correct but fall slightly outside a narrow acceptance window, the learning signal becomes noisy and optimization struggles to find stable solutions. After extensive analysis, we found that moderately relaxed tolerances during training allow the network to achieve optimal performance in terms of both recall and precision. The final tolerance values were optimized using Optuna to maximize both metrics simultaneously. From a practical standpoint, these tolerances also provide value for end users interpreting 2D correlation maps: a cluster of predictions in the vicinity of a true coupling serves as a robust indicator of a genuine J-correlation, even if individual predictions do not coincide exactly with the ground truth coordinates. To provide a comprehensive assessment of model performance, we report quantitative results at multiple tolerance levels throughout this section, enabling evaluation under both standard and more stringent criteria. Formally, we define a Tolerance Region T(J true) around each GT coupling J true = (δ A ,J) as the set of points (δ, j) satisfying: T(Jtrue)={(δ,j):|δ−δA|≤Δppmand|j−J|≤ΔJ} where Δppm and Δ J are the tolerance thresholds for chemical shift and coupling constant, respectively. For standard evaluation, we employ Δppm = 0.075 ppm and Δ J = 0.32 Hz. For comparative purposes, we also report metrics using tighter thresholds (Δppm = 0.05 ppm, Δ J = 0.25 Hz) to demonstrate performance under more demanding conditions. Note that for each true coupling (δ A , δ B , J), two tolerance regions are defined: one centered at (δ A , J) and another at (δ B , J), corresponding to the two multiplets involved in the coupling. A prediction is considered valid if it falls within either region Performance Metrics On the basis of the tolerance region framework, we define the following evaluation metrics. Tolerance-Based Recall (Coverage) This metric quantifies the fraction of true couplings that are detected by at least one positive prediction within their tolerance region: Recall=|{Jtrue:∃ p∈P+,p∈T(Jtrue)}||JGT| where P+ denotes the set of positive predictions and J GT is the set of ground-truth couplings. High recall is critical for structure elucidation, as missed couplings may lead to incorrect connectivity assignments. Tolerance-Based Precision This metric measures the fraction of positive predictions that fall within the tolerance region of at least one true coupling: Precision=|{p∈P+:∃ Jtrue∈JGT, p∈T(Jtrue)}||P+| Under this definition, multiple prediction clusters near a single true coupling all contribute positively to precision, reflecting the practical scenario where redundant markers in a valid region do not impair interpretation. Spurious Prediction Rate While precision captures predictions outside the standard tolerance region, we introduce an additional metric to specifically quantify predictions that are unambiguously erroneous. The spurious rate measures the fraction of positive predictions falling outside an extended tolerance region: Spurious Rate=|{p∈P+:∀ Jtrue∈JGT, p∉Text(Jtrue)}||P+| where Text is an extended tolerance region with Δppm = 0.15 ppm and Δ J = 0.80 Hz. Predictions within the extended region but outside the standard region represent borderline cases that may still provide useful information. In contrast, spurious predictions are definitively misleading markers that could confuse interpretation. Evaluation Data Set Model performance was evaluated on the held-out test set, stratified by spectral complexity as described in the Training Data section. The evaluation subset comprises 2,056 spectra distributed across three complexity tiers: tractable (N = 1121), mixed (N = 493), and challenge (N = 442). This stratification enables an assessment of model robustness from well-resolved patterns to severely overlapped regions. Quantitative Results Single Model Performance Table presents the performance metrics for the single GNN model evaluated at probability threshold τ = 0.5. The model achieves high recall across all complexity categories, with values exceeding 91% even for tractable spectra where the combinatorial challenge is greatest relative to the number of true couplings. 1Single Model Performance by Spectrum Complexity (Threshold τ = 0.5)Scope N Recall (%)Precision (%)Spurious (%)Tractable112191.6889.134.23Mixed49396.3784.664.78Challenge44295.9780.085.30 Single Model Performance by Spectrum Complexity (Threshold τ = 0.5) The apparent increase in recall for more complex spectra (mixed and challenge) reflects the higher candidate density in these cases: with more predictions generated, the probability of covering each ground-truth region increases. However, this comes at the cost of reduced precision, which drops from 89.1% for tractable spectra to 80.1% for challenging cases. The spurious rate remains controlled below 6% across all categories, indicating that severely erroneous predictions are rare. The results presented in Table employ the standard tolerance thresholds defined in the Tolerance Region-Based Evaluation section. To assess model robustness under more stringent evaluation criteria, Supporting Information Table S2 presents performance metrics across four tolerance levels, ranging from permissive to strict. As expected, both recall and precision decrease with tighter tolerances, reflecting the inherent uncertainty in peak positions discussed in the Tolerance Region-Based Evaluation section. Notably, the spurious rate remains constant across all tolerance levels, confirming that predictions falling outside the extended tolerance region are genuinely erroneous, rather than borderline cases affected by threshold selection. The effect of the probability threshold τ on model performance is examined in Supplementary Table S4. Increasing τ from 0.5 to 0.7 yields modest improvements in precision (up to 3 percentage points) at the cost of slightly reduced recall (1–2 percentage points), demonstrating that the model produces well-calibrated probability scores. Users can adjust this threshold according to their specific requirements: lower thresholds favor comprehensive coverage, while higher thresholds prioritize prediction confidence. Summary of single-model performance on the held-out test set. (A) Performance metrics across spectrum complexity categories at the default operating point (τ = 0.5, FAT tolerance). The model achieves recall exceeding 91% and precision above 80% across all categories, with spurious rates below 6%. (B) Effect of tolerance stringency on recall (solid lines) and precision (dashed lines). Both metrics decrease as tolerances tighten, with precision showing steeper degradation, particularly for complex spectra. Deep Ensemble Performance To improve prediction robustness and provide uncertainty estimates, we implemented a deep ensemble combining five independently trained models with diverse hyperparameter configurations selected through an ANOVA-based sensitivity analysis. The ensemble output is computed as the mean probability across models, with the standard deviation serving as an uncertainty measure for each prediction. Supplementary Table S11 presents the ensemble performance across spectrum complexity levels. Compared to the single-model results (Table ), the ensemble achieves consistent improvements: precision increases by 1.5–2.0 percentage points across all categories, while the spurious rate decreases by approximately 1 percentage point. Recall remains essentially unchanged, indicating that the ensemble primarily enhances prediction quality rather than coverage. A key finding is the remarkable agreement among ensemble members. Analysis of prediction uncertainty reveals that the median standard deviation across the five models is effectively zero for all complexity categories, indicating that ensemble members produce identical classifications for the vast majority of candidates. This high degree of coherence suggests that the learned representations are robust and not artifacts of particular training configurations. The cases where models disagree (90th percentile σ ≈ 0.21–0.26) correspond to genuinely ambiguous candidates near the decision boundary, providing users with actionable uncertainty information. Detailed uncertainty statistics are presented in Supplementary Table S5. The ensemble also enables flexible adjustment of the operating point through the probability threshold τ. Increasing τ from 0.5 to 0.7 improves precision by approximately 4 percentage points at the cost of 2–3 points in recall, allowing users to prioritize either comprehensive coverage or prediction confidence depending on their application requirements. A complete threshold sensitivity analysis is provided in Supplementary Table S4 Illustrative Simulated Examples To demonstrate the practical application of 2D-JCOG, we present several illustrative examples using simulated spectra. These examples showcase the model’s performance across different molecular complexity levels and coupling patterns. Melosine The first example corresponds to the simulated 1H NMR spectrum of melosine-d (Figure ), an indole alkaloid featuring a complex polycyclic scaffold that gives rise to 22 distinct 1H–1H scalar coupling interactions. The coupling constants in this molecule span a wide range from 0.90 Hz (long-range allylic couplings) to 16.36 Hz (geminal couplings in methylene groups), providing a representative test case for relevant J-values. 2D-JCOG analysis of the simulated 1H NMR spectrum of Melosine-d. Upper panel: reconstructed 1D spectrum from GSD peaks. Lower panel: δ versus |J| correlation map showing ground-truth couplings (red circles), positive predictions (green circles), and rejected candidates (gray squares). The molecular structure is shown as inset. Figure presents the analysis results in a two-panel visualization. The upper panel displays the reconstructed 1H NMR spectrum generated from GSD peaks calculated on the Mnova-predicted spectrum. The lower panel shows the two-dimensional δ versus |J| correlation map, where each potential coupling interaction is represented according to its chemical shift position (x-axis) and absolute coupling constant magnitude (y-axis). In this representation, red circles indicate ground-truth couplings from the resulting predicted spin system, green filled circles denote correlations classified as positive J-couplings by 2D-JCOG (probability ≥ 0.5), and light gray squares represent candidate correlations that were not classified as true couplings. The visual inspection of Figure reveals excellent agreement between predicted couplings (green circles) and ground truth (red circles), with virtually all green markers enclosed within their corresponding red circles. The rejected candidates (gray squares) form a characteristic pattern throughout the δ−J space, representing the numerous peak-pair combinations that arise from spectral complexity but do not correspond to actual scalar couplings. The clear separation between accepted and rejected predictions demonstrates the discriminative power of 2D-JCOG. Quantitatively, the model correctly identified all 22 ground-truth couplings with mean prediction errors of 0.043 Hz for coupling constants and 0.0009 ppm for chemical shift positions, demonstrating its ability to accurately localize coupling interactions in the δ−J space. Since any practical matching between predictions and ground truth requires defining tolerance margins to account for numerical differences, we adopted the evaluation framework described in the Tolerance Region-Based Evaluation section, with tolerances designed to reflect the accuracy level required for a useful analytical tool rather than pursuing arbitrarily high resolution. As can be observed in Figure , some positive predictions (green circles) do not coincide exactly with their corresponding ground-truth positions (red circles), but fall within the defined tolerance region. These predictions, while not pixel-perfect matches, represent correct identifications of the underlying coupling interactions and are therefore counted as true positives. Under this evaluation scheme, 2D-JCOG achieves 100% recall (all 22 couplings detected) and 100% precision (no false positive predictions outside tolerance margins). An additional simulated example evaluating 2D-JCOG on a para-substituted aromatic compound exhibiting an AA′XX′ spin system is presented in Supplementary Figure S2. Despite the second-order character of the aromatic signals, the model successfully identified all ground-truth couplings, including the characteristic ortho coupling (8.52 Hz), meta coupling (2.0 Hz), and the long-range couplings to the methylene protons (0.85 Hz), with mean prediction errors of 0.015 Hz for coupling constants and 0.0004 ppm for chemical shift positions. Experimental Results In this section, we present the performance evaluation of the model by using data derived from experimental NMR spectra. Since reliable ground truth is essential for quantitative assessment, we employed the GISSMO library, , which is publicly available at https://gissmo.bmrb.io/library. This database provides spin system parameters that were carefully derived from the experimental 1H NMR spectra. Prior to evaluation, we performed systematic data curation to ensure a meaningful comparison. Spectra containing only singlet signals were removed. Couplings between magnetically equivalent protons and multiplets corresponding to strongly coupled systems such as AA′BB′ were not considered in the analysis because they fall outside the scope of applicability of the proposed method. Finally, small long-range couplings (|J| < 0.5 Hz) were not included in the evaluations Results Overview Table summarizes the overall performance metrics. A prediction was considered a true positive if it matched a ground-truth coupling within tight tolerances (ΔJ ≤ 0.25 Hz, Δδ ≤ 0.02 ppm). For precision calculations, relaxed tolerances (ΔJ ≤ 0.5 Hz, Δδ ≤ 0.1 ppm) were employed to avoid penalizing predictions that correctly identify a coupling but with minor positional deviations. 2Performance Summary on Experimental Data from the GISSMO DatabaseMetricValueSpectra processed1060Ground-truth couplings evaluated7738Recall (sensitivity)91.9%Precision86.4%F1-Score0.890Mean Absolute Error (J-value)0.113 HzPearson correlation (R 2)0.998 Performance Summary on Experimental Data from the GISSMO Database The model achieved a recall of 91.9%, correctly identifying 7110 of the 7738 ground-truth couplings. Precision was 86.4%, yielding a combined F1-score of 0.890. The mean absolute error for J-values was 0.113 Hzwell below typical experimental uncertainties in J-coupling measurements (0.1–0.5 Hz depending on spectral resolution). The strong correlation between predicted and ground-truth values (R2 = 0.998) confirms that the model not only detects couplings but also accurately quantifies their magnitudes. Detailed error statistics and percentile distributions are provided in Supplementary Tables S6 and S7. Performance by Coupling Magnitude Analysis of performance stratified by J-coupling magnitude (Supplementary Table S8) reveals that the highest recall (98.0%) was achieved for couplings in the 6–8 Hz range, corresponding to typical vicinal proton–proton couplings. Performance remained consistently high (>91%) for couplings between 1–10 Hz, which represent the most commonly encountered values in organic molecules. The mean absolute error was remarkably consistent across all magnitude ranges, typically around 0.11–0.12 Hz. Lower recall was observed for small couplings (0.5–1 Hz, 20.3% recall) and for couplings in the 10–15 Hz range (75.9% recall). Error Distribution The distribution of J-value errors (Supporting Information Table S9) shows that the majority of predictions (72.1%) exhibited errors below 0.15 Hz, and 92.2% were within 0.25 Hz of the true value. The median error of 0.092 Hz indicates that half of all predictions were accurate to within approximately 0.1 Hz. Only 1.6% of predictions showed errors exceeding 0.4 Hz, demonstrating consistent accuracy across the data set. The experimental validation demonstrates that the proposed model achieves a high accuracy in extracting J-coupling constants from real NMR data. The overall recall of 91.9% indicates the successful identification of the vast majority of coupling interactions, while the prediction error of 0.113 Hz is well below typical experimental uncertainties. The false positive rate of 13.6% represents predictions that do not correspond to any ground-truth coupling within relaxed tolerances. These likely arise from (i) peak picking limitations in severely overlapped regions; (ii) borderline strong coupling cases exhibiting sufficient second-order character to confuse the classifier; and (iii) inherent classification errors in complex multiplet patterns. Selected Cases The example shown in Figure A examines biphenyl, a molecule comprising two equivalent phenyl rings that each constitute an AA′BB′C spin system. In this system, the three magnetically nonequivalent protons give rise to complex multiplet patterns due to strong coupling effects and therefore traditional multiplet interpretation rules cannot be applied. 2D-JCOG analysis of strongly coupled aromatic spin systems from the GISSMO database. (A) Biphenyl, featuring a strongly coupled AA′BB′C spin system. Upper panel: reconstructed 1D spectrum showing the complex aromatic multiplet pattern. Lower panel: δ versus |J| correlation map. (B) N1-(2-thienylmethyl)­benzenesulfonamide, containing an AA′XX′Y benzene system and a strongly coupled ABC thiophene system. Upper panel: reconstructed 1D spectrum. Lower panel: δ versus |J| correlation map. Molecular structures are shown as insets. The spin system contains five distinct observable coupling interactions: two 3 J ortho couplings (J(A–B) = J(A′–B′) and J(B–C) = J(B′–C), both 7.5 Hz) and three 4 J meta couplings (J(A–C) = J(A′–C), J(A–A′), and J(B–B′), all 2.0 Hz). The 5 J para couplings (0.1 Hz) fall below the detection threshold and are excluded from the analysis. 2D-JCOG successfully identified all 5 ground-truth couplings with mean prediction errors of 0.052 Hz for coupling constants and 0.0012 ppm for chemical shift positions. The model achieved 100% recall and 100% precision with no false positive predictions within the tolerance criteria. The fourth example presents N1-(2-thienylmethyl)­benzenesulfonamide (Figure B), a molecule containing two distinct aromatic ring systems that provide complementary test cases for 2D-JCOG’s handling of complex multiplet patterns. The monosubstituted benzene ring constitutes an AA′XX′Y spin system, where the ortho protons (A, A′ at δ = 7.83 ppm) and meta protons (X, X′ at δ = 7.59 ppm) are sufficiently separated (Δν/J ≈ 17) to exhibit near first-order behavior between them, while the para proton (Y at δ = 7.68 ppm) lies at intermediate chemical shift, creating a more complex coupling pattern. This differs from the AA’BB’C systems examined previously, where all protons were closely spaced in chemical shift. The thiophene ring presents an ABC spin system where protons A and B (δ = 6.899 and 6.874 ppm, respectively) are separated by only 0.025 ppm, resulting in strong coupling with Δν/J ≈ 4.4 at 600 MHz. This generates the complex overlapping multiplet observed in the 6.87–6.90 ppm region. Proton C (δ = 7.29 ppm) couples to A with J = 5.19 Hz, contributing additional fine structure. Figure B shows the 2D-JCOG analysis results. The model successfully identified all 5 unique ground-truth couplings (after excluding 2 magnetically equivalent interactions) with mean prediction errors of 0.059 Hz for coupling constants and 0.0012 ppm for chemical shift positions, achieving 100% recall and 100% precision. Notably, despite the strongly overlapping nature of the thiophene ABC systema scenario for which 2D-JCOG was not specifically designedthe model correctly extracted the coupling information. However, examination of the expanded thiophene region (Figure B, inset) reveals a characteristic limitation. In the center of the strongly coupled ABC multiplet around 6.89 ppm, several predictions appear at positions that do not correspond to ground-truth couplings. While these fall within tolerance margins of nearby true couplings and are thus counted as true positives, they represent redundant detections arising from the complex peak pattern, where individual lines are closely spaced with similar intensities. This behavior is observed consistently in strongly coupled systems: the classifier successfully rejects most spurious candidates but occasionally generates multiple detections for the same underlying coupling interactions. As a final example demonstrating 2D-JCOG’s applicability to real experimental data, we analyzed the 1H NMR spectrum of (S)-2-amino-4-methylpentan-1-ol (Leucinol) acquired at 400 MHz in DMSO-d 6. Leucinol presents an instructive spin system with two pairs of diastereotopic methylene protons (H-2′/H-2″ and H-6′/H-6″) arising from the chiral center. These diastereotopic pairs exhibit both large geminal couplings (2 J = 10.25–13.36 Hz) and distinct vicinal couplings to their neighboring protons. The ground truth comprises 10 unique J-couplings ranging from 4.61 to 13.36 Hz. Using the standard tolerances defined in the Tolerance Region-Based Evaluation section, 2D-JCOG achieved 100% recall (24/24 positive predictions matching ground truth) and 100% precision at threshold 0.5, correctly rejecting 179 spurious candidates. Closer inspection of the δ 1.0–1.1 ppm regionwhere the diastereotopic H-2′ and H-2″ protons resonate as overlapping multipletsreveals limitations that merit discussion. In this congested region, the multiplets of the diastereotopic H-2′ and H-2″ protons are closely spaced but not truly overlapped. However, because this gap falls below the algorithm’s fixed ROI clustering threshold (18 Hz), both multiplets are grouped into a single ROI. This merging is the root cause of false positives observed in this region. When two distinct multiplets are treated as a single ROI, the algorithm generates “cross-multiplet” candidates connecting peaks from different spin systems. These spurious candidates are particularly difficult to reject because they compete directly with true positives within the same ROI, sharing similar local statistical features. The solution is straightforward: if the ROI is subdivided into its two constituent multipletseither manually by the user or automatically through a more sophisticated clustering algorithm that detects outlier gapsthe cross-multiplet candidates are never generated in the first place. Furthermore, the geminal coupling (J ≈ 13.4 Hz) between H-2′ and H-2″ becomes highly detectable, as SST support from both separate multiplets strongly reinforces these predictions. We verified this experimentally by manually splitting the ROI at the 13 Hz gap. The result was a complete elimination of the false positives while maintaining 100% recall of all true couplings, including the geminal J. This demonstrates that the observed limitations stem not from any fundamental algorithmic deficiency, but from a conservative ROI clustering threshold that can be refinedeither through user intervention or adaptive segmentation algorithmswhen higher precision is required in specific spectral regions Conclusions and Future Work We have presented 2D-JCOG, a deep learning framework that transforms conventional 1D 1H NMR peak lists into 2D J-δ correlation maps, enabling direct visualization and extraction of scalar coupling networks. The method addresses a long-standing challenge in NMR spectral analysis: automated J-coupling determination in scenarios where traditional multiplet analysis fails due to spectral congestion or second-order effects. The central innovation of 2D-JCOG lies in the integration of the shared splitting theorem (SST) as the foundational physical constraint guiding a heterogeneous graph neural network architecture. Rather than treating the SST as a hard filter, which inevitably discards true couplings obscured by overlap, the framework encodes the SST relationships as graph connectivity. This design enables the network to weigh SST evidence in context, distinguishing genuine couplings from coincidental splitting matches on the basis of learned spectral features. The hybrid message-passing strategy, combining transformer-style attention for physics-aware SST edge validation with mean aggregation layers for hierarchical multiplet context, allows the model to reason simultaneously about local spectral features and global coupling consistency. Evaluation on both simulated and experimental data demonstrates robust performance within the tolerance criteria established for this work: recall exceeding 92% and precision above 80% across varying complexity levels. Experimental validation of the GISSMO database confirms a mean J-value error of 0.113 Hz. These results establish 2D-JCOG as a practical tool for routine spectral analysis, particularly effective for first-order and moderately coupled spin systems, where classical multiplet analysis often struggles. Several limitations warrant acknowledgment. First, while the model successfully identifies coupling regions, predictions occasionally cluster around true values rather than converging precisely, reducing quantitative accuracy and introducing ambiguity in crowded correlation maps. Second, performance degrades for severely strongly coupled systems (e.g., AA′BB′ and higher-order spin systems) where the correspondence between observed splittings and true coupling constants breaks down. Although the SST remains mathematically valid in these regimes, extracting accurate J-values from such systems requires further methodological development. Additionally, the current method is inherently limited to homonuclear couplings, as the SST requires both coupled multiplets to be present in the observed spectrum. Heteronuclear couplings (e.g., 1H–19F, 1H–31P) produce splittings in the 1H spectrum that lack a reciprocal SST partner, and are therefore not validated by the current framework. Future work will focus on two primary directions. First, improving prediction precision by refining candidate consolidation strategies and incorporating additional constraints that penalize redundant detections near identical coupling coordinates. This may involve postprocessing algorithms or modified loss functions that explicitly discourage clustering behavior. Second, extending applicability to strongly coupled systems remains an open challenge. The SST’s topological guarantees hold regardless of coupling strength, suggesting that pattern recognition approaches may still prove viable; however, the optimal strategy for leveraging this invariance in regimes where splittings diverge from true J-values requires further investigation. In its current form, 2D-JCOG serves as a useful complement to existing NMR analysis workflows, providing automated coupling extraction with intuitive 2D visualization for the majority of routine analytical scenarios. As the methodology matures, we anticipate a broader applicability across increasingly challenging spectral conditions. Supplementary Material