PharmaGNN: a model for odor prediction based on graph neural networks

Abstract

BACKGROUND

Artificial intelligence (AI)‐assisted odor prediction holds significant potential for applications in fragrance discovery, food quality control, and sensory evaluation. However, the accuracy of existing models is often limited by insufficient features and class imbalance across diverse odor labels. We propose the PharmaGNN model that integrates pharmacophore features to capture essential molecular structural characteristics for odor prediction. An asymmetric loss function is employed during model training to mitigate label imbalance.

RESULTS

Evaluations demonstrate that PharmaGNN achieves state of‐the‐art performance with the highest AUROC score among comparative models. Even when evaluated across 232 labels, the model maintains robust generalization performance. Ablation studies show that both the pharmacophore features and the molecular graph information are significant to the accuracy.

CONCLUSION

PharmaGNN serves as a practical tool to accelerate the design of novel flavor molecules and improve quality control and odor assessment in the fragrance and food industries. This model bridges the gap between cheminformatics and structural biology, offering a robust tool for the rational design of functional scents and the digitization of olfaction. © 2026 Society of Chemical Industry.

INTRODUCTION

Odor molecules are vital across diverse sectors, including perfumery, food, cosmetics, pharmaceuticals, and agriculture. Precise odor prediction is essential for the development of superior functional foods, sustainable perfumery, and palatable drug formulations.^{1, 2} A key challenge is the limited library of characterized odorants, ³ which leaves a vast portion of odor chemical space unexplored and insufficient for market demands. ⁴ Human sensory evaluation captures direct perception from mixtures, but suffers from low repeatability and safety limitations. ⁵ Instrumental methods provide accurate and objective data but are restricted by high operational costs and environmental sensitivities. ⁶ Machine learning has emerged as a robust and scalable solution for the high‐throughput prediction of odor characteristics from extensive molecular datasets. ⁷

Recently, machine learning has been widely used in prediction tasks. ⁸ For instance, the random forest model predicts odor label and scent retention grade by using molecular descriptors and fingerprints, and provides a reliable technical approach for digitalizing smell.^{9, 10, 11, 12} To improve accuracy in odor classification and sweetness prediction, the XGBoost and LightGBM models incorporate not only molecular descriptors and fingerprints but also olfactory receptor data.^{13, 14, 15, 16} Similarly, support vector machine models achieve high performance in specific tasks such as sweetener identification.^{17, 18} Diverse traditional machine learning algorithms were integrated in DREAM Winner ¹⁰ and FlavorMiner ¹⁹ to predict multiple odor tasks based on molecular descriptors and fingerprints. Nevertheless, these traditional approaches are generally limited to smaller datasets and rely on manually crafted feature representations. ²⁰

Deep learning has emerged as a powerful tool for predicting flavor and odor properties by handling complex chemical data.^{21, 22} Various architectures have been applied in specific prediction tasks. ²³ For instance, multilayer perceptron (MLP) and recurrent neural network integrate peptide sequences and structural features to discover more umami peptides.^{24, 25} Convolutional neural networks (CNNs) and deep neural networks (DNNs) leverage molecular descriptors and image‐based features to predict multiple odor labels, which aids in masking undesirable flavors in drugs.^{26, 27, 28} Graph neural networks (GNNs) learn atom‐ and bond‐level relationships on molecular graphs, and have been effectively applied to identify odor labels for uses including consumer‐tailored perfume design and sugar‐free food development.^{29, 30, 31, 32, 33} Variants like message‐passing neural networks (MPNNs) explicitly model the nonlinear propagation of structural information to map structure–odor relationships. ³⁴ Specifically, GNN‐QSOR ³⁵ captures a meaningful odor space representation by learning transferable odor embeddings from GNNs. D‐MPNN ³⁶ uses directed edge‐based message passing for molecular property prediction. In comparison, MPNN‐POM ³⁷ learns relationships between nodes to generate a principal odor map. These models can achieve expert‐level accuracy and have facilitated the digitization of odor space. Graph architectures that combine graph isomorphism networks and graph attention networks (GATs), such as ScentGraphX, MSGNN, capture both local subgraph patterns and global molecular features, improving multi‐label prediction in environmental gas detection.^{38, 39} Models based on three‐dimensional (3D) geometric features such as ComENet and SphereNet further capture spatial molecular conformation, and coordinate‐based methods like Mol‐PECO process full atomic coordinates to enrich structural representations.^{40, 41} Despite notable advancements in deep learning‐based odor prediction, model performance remains limited by persistent challenges, including data sparsity and label imbalance. ⁴² To address these issues, recent studies have introduced strategies at various levels, such as smote resampling, ⁴³ the focal loss function, ⁴⁴ and stratified sampling strategies. ⁴² Collectively, deep learning could improve the efficacy and reliability of odor prediction for screening novel odorants. ⁴⁵

Recent work in computational olfaction has focused on understanding odorant–receptor interactions. ⁴⁶ Large language models such as OP Benchmark ⁴⁷ and Olfaction‐LLM ⁴⁸ effectively combine semantic knowledge from text with real sensor signals for odor‐related reasoning. Machine learning models, including NOSE, ⁴⁹ LORAX, ⁵⁰ AROMMA, ⁵¹ GODE, ⁵² DEGNN, ⁵³ and the three‐stage framework, ⁵⁴ learn effectively from molecular structures and receptor information, and achieve reliable performance across a range of odor‐related tasks. Besides, structure‐based methods incorporate receptor structures generated by AlphaFold2 to analyze odorant–receptor binding via molecular docking and molecular dynamics simulations.^{55, 56, 57, 58} These methods enable atomistic‐level interaction characterization, but they rely heavily on the availability of reliable receptor structures and involve high computational cost. These methods still have limitations in balancing high prediction accuracy, computational efficiency, and generalization ability.

We propose PharmaGNN, a model designed to enhance odor prediction by integrating a global molecular graph with local pharmacophore features. This design is motivated by the recognition that olfactory reception is governed by specific interactions between odor molecules and receptors, thereby establishing pharmacophore features as critically important.^{59, 60, 61, 62, 63} The architecture utilizes GAT and MLP components to capture multiscale chemical structure information pertinent to odor. An asymmetric loss (ASL) is implemented to recalibrate learning toward rare odor labels. The performance of PharmaGNN was rigorously benchmarked against established baseline models. Ablation studies were performed to quantify the contribution of pharmacophore features. The robustness of this model was assessed to validate its stability in extended odor classification tasks. Furthermore, t‐distributed stochastic neighbor embedding (t‐SNE) visualizations of odor labels were performed to organize odor labels in the latent feature space. This model offers valuable insights into odor molecule discovery and structure–odor relationships.

MATERIALS AND METHODS

Datasets and data splitting

Three datasets were employed for model development and evaluation. Dataset 1 was derived from the OpenPOM project, ³⁷ which integrates odorant data from the Good Scents ⁶⁴ and Leffingwell PMP 2001 ⁶⁵ databases. After standardization and removal of labels with fewer than 30 associated molecules, ⁶⁶ it yields a final set of 4983 molecules associated with 138 odor labels. Dataset 2 was extracted from FlavorDB ⁶⁷ and was filtered identically to obtain 13 088 molecules and 85 odor labels. To evaluate the generalization capability, Dataset 3 was constructed as an extended version of Dataset 2 with a broader set of odor labels. Compared to Dataset 2, Dataset 3 includes additional labels, each associated with more than 3 but fewer than 30 samples. The refined dataset comprises 13 203 molecules and 232 distinct odor labels, while retaining 99% of the original molecular entries. All datasets were split for model development using a reproducible random partition. A fixed seed was used to divide each dataset into training, validation, and test sets in a 7:1.5:1.5 ratio.

Feature extraction

A detailed summary of all feature dimensions is provided in supporting information, Tables S1–S3. The SMILES (simplified molecular input line entry system) strings of the molecules were used as input to convert them into molecular graphs and to optimize their 3D structures using MMFF94 ⁶⁸ in RDKit. Two primary feature types were generated. The global molecular graph features from the whole molecular graphs were encoded with 11‐dimensional node and 4‐dimensional edge features.

A pharmacophore is an abstract group of molecular features and spatial constraints essential for biological activity. ⁶⁹ The binding between odorants and olfactory receptors arises from multiple intermolecular interactions, including hydrophobic interactions, hydrogen bonding, and electrostatic interactions.^{70, 71} To capture these interactions, pharmacophore features were extracted from the 3D molecular structures using RDKit's feature factory, with each region represented as a 10‐dimensional feature vector comprising physicochemical and geometric descriptors. For physicochemical descriptors, electrostatic contributions are encoded by the averaged Gasteiger charge (avg_Gasteiger) ⁷² ; hydrophobic interactions are captured by total molar refractivity (totalMR), and averaged logarithm of the partition coefficient (avgLogP), and hydrogen bonding is reflected by local topological polar surface area (local_TPSA). The geometric descriptors characterize the spatial position and shape of each region through the center coordinates avg_Pos_(x,y,z) and three distances to the surface (farthest, nearest, average). Finally, the global (graph) and local (pharmacophore) feature sets were integrated to form a complementary molecular representation.

Architecture of PharmaGNN

Figure 1 depicts details of the proposed PharmaGNN framework. The architecture includes feature extraction from a global molecular graph and local pharmacophores, as well as a multi‐label binary classification prediction module. Each of these components is described in detail in the following subsections.

Figure 1.

Architecture of PharmaGNN.

Feature extraction from a global molecular graph

In the molecular graph, atoms are represented by a vector v , while bonds are represented by a vector e . These raw features are first processed through a 128‐dimensional linear layer, followed by normalization and ReLU layers. The GATv2Conv ⁷³ is then employed as an encoder, with five 128‐dimensional layers and a hierarchical multi‐head design (with 8, 8, 4, 4, and 2 heads, respectively) to capture molecular structural information. Residual connections are incorporated between adjacent GAT layers to facilitate gradient flow and stabilize training. A global mean pooling operation aggregates all node features into a 128‐dimensional graph embedding Z _g to represent the molecule's global topological structure.

Feature extraction from pharmacophores

In this module, key features of the pharmacophore are identified from the 3D molecular coordinates generated by RDKit. These features are aggregated through multichannel pooling (max, mean, and sum), and the resulting vectors are concatenated to capture comprehensive statistical information. The concatenated feature vector is processed through a two‐layer MLP with dimensions of 256 and 128. Each layer incorporates residual connections, layer normalization, and dropout for stable training. Ultimately, the module generates a 128‐dimensional pharmacophore embedding vector Z _p .

Feature fusion and prediction module

The global graph embedding Z _g and the pharmacophore embedding Z _p were concatenated to a 256‐dimensional fused feature vector Z . It serves as input to a six‐layer MLP predictor. The first three hidden layers have 512, 256, and 128 dimensions, respectively, and each uses a ReLU activation. The final linear layer reduces the dimension to the number of target odor labels, after which a sigmoid activation is applied per label to produce independent probability scores for the multi‐label prediction.

Asymmetric loss

To mitigate the class imbalance prevalent in multi‐label odor datasets, we employ the ASL, ⁷⁴ as shown in Eqn (1). The ASL adapts the focal loss framework by introducing asymmetric focusing parameters, ‘γ+’ and ‘γ−’, which independently modulate the contribution of positive and negative samples to the gradient. p and p _m denote the network's output probability and the shifted probability, respectively. ⁷⁴

$$\mathrm{ASL}=\left\{\begin{array}{c}{L}_{+}={\left(1-p\right)}^{{\gamma }_{+}}\log \left(p\right)\\ L_{-}={\left(p_m\right)}^{{\gamma }_{-}}\log \left(1-p_m\right)\end{array}\right.$$ (1)

Model training was performed using the Adam optimizer, ⁷⁵ with a learning rate of 0.0005, a weight decay of 0.0005, and a batch size of 64. The ASL function was employed with hyperparameters γ+ = 1 and γ− = 4 (supporting information, Table S5). The model was trained for 200 epochs with early stopping (patience = 25).

Evaluation metrics

Model performance was evaluated using the AUROC (the area under the receiver operating characteristic curve). The calculation of AUROC requires first defining true positives TP, true negatives TN, false positives FP, and false negatives FN. AUROC measures the model's discriminative ability across different classification thresholds. It is computed from the true positive rate (TPR) and false positive rate (FPR), as defined in Eqns (2) and (3), which represent the ROC curve (FPR versus TPR). The AUROC is the area enclosed under this curve.

$$\mathrm{TPR}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$ (2)

$$\mathrm{FPR}=\frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}}$$ (3)

EXPERIMENTS AND RESULTS

Comparisons with baseline

The performance of PharmaGNN was evaluated and compared with established baseline models on two datasets. To validate stability, experiments were repeated under three different random seeds (34, 52, 123). The results from Table 1 showed highly consistent AUROC values across seeds for both Dataset 1 (0.8819, 0.8813, 0.8811) and Dataset 2 (0.9690, 0.9708, 0.9702), indicating its minimal sensitivity to data splitting. Notably, PharmaGNN consistently outperformed all baselines. On Dataset 1, the mean AUROC (0.88) exceeded those of DREAM Winner, ScentGraphX, GNN‐QSOR, and MPNN‐POM, all of which had AUROC > 0.8. A more substantial lead was observed on Dataset 2, where PharmaGNN (mean AUROC = 0.97) significantly surpassed FlavorMiner and Mol‐PECO. In summary, PharmaGNN demonstrates both robust and state‐of‐the‐art predictive performance.

Table 1.

Performance evaluation of PharmaGNN with random seeds and baseline models on two datasets

Dataset 1		Dataset 2
Method	AUROC	Method	AUROC
DREAM Winner 10	0.8304	FlavorMiner 19	0.8800
GNN‐QSOR 35	0.8618	Mol‐PECO 34	0.8130
DNN‐CNN 28	0.7851	PharmaGNN (34)	0.9690
MPNN‐POM 37	0.8296	PharmaGNN (52)	0.9708
MSGNN 31	0.6460	PharmaGNN(123)	0.9702
ComENet 41	0.7789
SphereNet 40	0.7796
DMPNN 36	0.6523
ScentGraphX 38	0.8626
PharmaGNN (34)	0.8819
PharmaGNN (52)	0.8813
PharmaGNN(123)	0.8811

Note: Bold values represent best performance.

Compared to these baseline models, the enhanced predictive performance of PharmaGNN can be attributed to advances in both model architecture and feature representation. Unlike traditional machine learning models (DREAM Winner, FlavorMiner) that use static fingerprints, PharmaGNN employs a GAT‐MLP framework to dynamically learn discriminative features from raw molecular graphs and pharmacophore descriptors. Compared to GNNs that use only molecular graphs (GNN‐QSOR, MPNN‐POM, MSGNN, DMPNN, and ScentGraphX), our explicit incorporation of pharmacophore information guides the model to prioritize odor‐relevant functional groups. While 3D geometric models (ComENet, SphereNet) capture spatial conformation, our approach complements this by preserving essential atom and bond information through graph structure, which is crucial for molecular characterization. Furthermore, while methods like Mol‐PECO process full atomic coordinates, our pharmacophore‐based approach selectively emphasizes odor‐relevant functional groups, improving both relevance and computational efficiency.

Ablation experiments

Ablation experiments were conducted to evaluate the effectiveness of multilevel molecular feature extraction by removing features from the pharmacophore and molecular graph. As evident from Table 2, when only the molecular graph was employed, the performance AUROC is 0.8695 on Dataset 1 and 0.9660 on Dataset 2. The AUROC value has surpassed all baseline models in Table 2. Interestingly, when only the pharmacophore module was used, the model also achieved a high AUROC of 0.8420 on Dataset 1 and 0.9544 on Dataset 2. Thus, the pharmacophore module can also outperform most baseline models. The observed performance gains show that pharmacophore features provide complementary spatial and functional information beyond topological graph representations. Therefore, integrating the molecular graph and the pharmacophore module can achieve high precision in odor prediction. Furthermore, the binary cross entropy (BCE) loss and ASL are compared on two datasets under random seed 34 in supporting information, Table S4. The results show that ASL yields higher AUROC values (0.8819 and 0.9690) than BCE loss (0.8749 and 0.9670) on both datasets. These results indicate that the ASL functions could improve performance in model training.

Table 2.

Comparison of AUROC in ablation studies

Method		AUROC
Molecular graph	Pharmacophore	Dataset 1	Dataset 2
√		0.8695	0.9660
	√	0.8420	0.9544
√	√	0.8811	0.9702

Note: Bold values represent best performance.

GENERALIZATION

The generalization capability of PharmaGNN was evaluated on the expanded Dataset 3, which contains 232 odor labels. Figure 2 shows that the AUROC experienced only a slight decrease from 0.9725 to 0.9609 as the label count expanded from 20 to 232. Notably, the model maintained high predictive accuracy even for labels with very few samples. The observed performance on scarce‐sample labels is likely attributable to the model's learning mechanism. By learning quantitative structure–odor relationships from labels with abundant samples, the model could identify key and potentially generalizable structural features. This effective representation learning may contribute to improved performance when predicting labels with scarce samples.

Figure 2.

AUROC for various numbers of labels.

t‐SNE visualization of various odor labels

t‐SNE is a dimensionality reduction technique that maps high‐dimensional data into two dimensions while preserving the local structure, with molecules of similar feature representations projected close together in the plot. To decode the structural organization of the odor perception space learned by our model, the 256‐dimensional latent space of concatenated features Z was visualized using t‐SNE for two datasets, as shown in Fig. 3. The plot represents each compound as a point, colored according to one of five representative odor labels. These labels were selected with a sufficient number of samples to ensure both clear visualization and statistical reliability. Figure 3(a) reveals distinct spatial clustering patterns for five odor labels from Dataset 1, with well‐separated clusters suggesting the model's ability to capture the broad odor classifications. The observed cluster separation can be attributed to the concatenated feature Z 's ability to preserve the inter‐category variance of odors. The integration of pharmacophore information refines the feature representation, enabling more precise classification. Interestingly, both ‘alcoholic’ and ‘winey’ clusters are located in the lower area of the figure, which may imply a similarity in their underlying olfactory or gustatory recognition pathways. Figure 3(b) shows the t‐SNE results for the five odor labels in Dataset 2. By incorporating both global and local structural information, the method produces compact yet distinct clusters with minimal overlap, capturing subtle differences. This results in a more concentrated distribution than in Fig. 3(a), which can be attributed to the structural divergence among these five specific odors. In summary, the t‐SNE plots show that PharmaGNN generates effective molecular structural feature embeddings, which capture the intrinsic relationships among odors.

Figure 3.

t‐SNE visualization of compound distribution in learned odor space. (a) Compounds from Dataset 1 labeled as roasted (547), winey (358), honey (198), amber (161), and alcoholic (163). (b) Compounds from Dataset 2 labeled as sulfurous (245), balsam (177), nutty (163), minty (121), and apple (173).

CONCLUSION

We proposed PharmaGNN, a model for odor prediction that integrates features of pharmacophore and molecular graph into a hybrid representation. It systematically combines global molecular information with local key 3D physicochemical descriptors that collectively capture the overall structure and functional group‐level characteristics of the molecules. To address class imbalance in odor datasets, we implemented an ASL function during training. Benchmarking on two datasets demonstrated superior performance compared to existing methods. PharmaGNN shows exceptional scalability and generalization on a larger dataset (AUROC of 0.9609 across 232 labels). Visualizations display well‐separated odor clusters, indicating learned discriminative representations. PharmaGNN achieves superior predictive performance due to its novel architecture that integrates molecular graphs with pharmacophore features. These results offer direct applications in flavor and fragrance development. Moreover, the pharmacophore‐integration strategy provides a transferable framework for tasks like drug‐binding affinity prediction. Future research will extend to integrating knowledge‐graph embeddings ⁷⁶ in chemistry and biology to explicitly model pharmacophore–receptor interactions, thereby gaining a more mechanistic understanding of the structure–odor relationship.

CONFLICT OF INTEREST

The authors declare no conflicts of interest.

AUTHOR CONTRIBUTIONS

Yunshu Liu: Writing – original draft, Writing – review & editing, Methodology. Qiumeng Song: Validation, Writing – review & editing. Muhammad Shoaib: Data curation, Visualization. Xiaofei Nan: Methodology, Writing – review & editing. Ji Ma: Resources. Kai Cui: Funding acquisition, Writing – review & editing. Yuan Shang.: Funding acquisition, Supervision. Jinshuai Song: Writing – review & editing, Supervision, Funding acquisition.

Supporting information

Table S1. Node features in the representation of molecules.

Table S2. Edge features in the representation of molecules.

Table S3. Pharmacophore features.

Table S4. Evaluating the performance of the PharmaGNN model with Binary Cross Entropy Loss and Asymmetric Loss on two datasets under random seed 34.

Table S5. AUROC performance of the PharmaGNN model with Asymmetric Loss under different γ ⁻/γ ⁺ combinations on two datasets (random seed 34).

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.