Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Bioinformatics . 2026 Mar 4;27:84. doi: 10.1186/s12859-026-06405-3 Search in PMC Search in PubMed View in NLM Catalog Add to search Uncertainty-aware hybrid deep generative framework for robust and explainable drug discovery Saniya Gupta Saniya Gupta 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Chennai, India Find articles by Saniya Gupta 1 , A Sherly Alphonse A Sherly Alphonse 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Chennai, India Find articles by A Sherly Alphonse 1, ✉ , D Kavitha D Kavitha 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Chennai, India Find articles by D Kavitha 1 Author information Article notes Copyright and License information 1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Chennai, India ✉ Corresponding author. Received 2025 Nov 26; Accepted 2026 Feb 12; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13067696 PMID: 41782096 Abstract Conventional techniques in drug discovery are time-consuming and less accurate due to the vast chemical space and associated uncertainty. Artificial intelligence and machine learning have introduced new methods to improve the drug discovery process. But still, some challenges need to be overcome, like higher attrition rates and the need to optimize multiple molecular properties simultaneously. The proposed study overcomes these problems by introducing a novel hybrid computational framework that integrates Graph Convolutional Networks, Variational Autoencoders, and Uncertainty Aware Adaptive Multi-Objective Optimization-based Reinforcement Learning (UAAMOO-RL) for generating drug molecules with needed pharmacological properties. By combining uncertainty awareness and explainability, the proposed method generates new drug molecules, prioritizing reliability and reducing false positives, thereby addressing two challenges in drug discovery, such as robustness and interpretability. The proposed framework is trained and validated using the ChEMBL dataset, consisting of 1.5 million bioactive molecules with known properties. The results indicate that the proposed model achieves a QED pass rate of 88.8% at a threshold of = 0.60, which outperforms the state-of-the-art techniques while maintaining structural diversity, uniqueness, and validity. Results also show significant improvement on the Zinc250k (82.7%) and PDBbind (85.9%) datasets. The proposed approach enhances early-stage drug discovery by reducing exhaustive experimental screening. Keywords: Drug discovery, Explainability, Uncertainty, Reinforcement learning, Molecular graph Introduction Drug discovery is very complicated because of high costs, time expenditure, and significant attrition rates. Conventional methods employ heuristic-driven screening and have some limitations in exploring the vast chemical space. Artificial intelligence and machine learning introduced new data-driven approaches that help improve the drug discovery process [ 4 , 5 ]. Generative adversarial networks (GANs) are also used for molecular graph generation. De Cao and Kipf (2018) have proposed MolGAN, which combines reinforcement learning (RL) with GANs, for generating valid molecules [ 6 ]. Furthermore, You et al. [ 40 ] have used the Graph Convolutional Policy Network (GCPN) for optimizing the molecular structures after the computation of chemical property rewards. Evolutionary algorithms and reinforcement learning can also be applied to guide molecular generation. Popova et al. [ 23 ] and Wang et al. [ 32 ] demonstrated the adaptive multi-objective optimization techniques used to enhance the molecular design while balancing competing objectives such as efficacy, toxicity, and synthesizability. The proposed technique combines Graph Convolutional Networks (GCNs), Variational Autoencoders (VAEs), Uncertainty Aware Adaptive Multi-Objective Optimization (UAAMOO), and Reinforcement Learning (RL) to generate novel drug molecules. The molecular graphs are derived from SMILES representations, where atoms and bonds are encoded as nodes and edges, respectively. The GCN creates graph-level embeddings based on the structural and chemical properties of molecules. From these molecular graphs, VAE learns a continuous latent space that helps in the generation of new molecules [ 43 ]. Adaptive MOO techniques help optimize efficacy and novelty, improving the adaptive multi-objective nature of drug discovery. The RL is also applied to fine-tune the latent representations by giving rewards according to predefined metrics, such as drug likeness and synthetic accessibility. The results show the role of AI-driven molecular design in revolutionizing drug discovery, thereby reducing the time and costs taken with early-stage drug development. The proposed approach is a scalable and adaptable framework for computational chemistry. It also ensures that the generated molecules are chemically valid with suitable pharmacological properties [ 28 ]. In this work, the main contributions are as follows: GCN helps in the creation of graph-level embeddings. VAE learns the latent space from the molecular graphs and supports the generation of novel structures. The uncertainty calibration is also included in the Reinforcement Learning (RL) loop, thus making the Pareto front optimization penalize high-uncertainty molecules that are generated. The epistemic uncertainty from VAE and also the predictive uncertainty from MC dropout help in uncertainty modeling. The explainability enhances transparency in molecular properties. The proposed framework incorporates an uncertainty-based reward in UAAMOO-RL, which enhances reliability and eliminates hallucinated molecules. The proposed technique balances the Quantitative Estimation of Drug-likeness (QED), Synthetic Accessibility score (SA), and structural diversity. Failure-case analysis helps in the attainment of molecular generation with good reliability. The existing uncertainty-aware optimization methods incorporate uncertainty as a filtering technique or as a fixed regularization term. Belakaria et al. focus on incorporating uncertainty, estimated using surrogate models that guide the Pareto-optimal candidates [ 3 ]. Furthermore, Tuo and Wang use uncertainty as a fixed regularization mechanism in exploration [ 30 ]. But the proposed method is integrating uncertainty in the optimization loop itself, thereby balancing the exploration and molecular property objectives. The proposed UAAMOO-RL technique integrates uncertainty into the RL objective by applying a Lagrangian-inspired reward-based formula for influencing policy updates. Furthermore, the optimization of the parameters is based on maximizing Pareto-front diversity. Thus, the proposed system is an integrated pipeline that combines validity constraints, uncertainty computation, and explainability-based atom-level attributions that result in a robust and molecular design with better interpretability. Materials and methods The datasets used, and a detailed explanation of the proposed methodology applying GCN, VAE, UAAMOO-RL, uncertainty estimation, and attribution score estimation are discussed in this section. Datasets ZINC-Drugs, the ChEMBL dataset, and PDBbind datasets were used in the experiments. The proposed work enhances the performance by avoiding the hallucinated molecules and by addressing the uncertainty while using UAAMOO-RL for optimized drug discovery. The hallucinated molecules are the chemically valid molecules, but have high uncertainty and fall outside the prediction regime. These molecules are different from chemically invalid structures, which are filtered using validity constraints and removed from further evaluation. This enables better explainability while generating new molecules. Methods The molecular data are taken as input, where the atoms are represented as nodes and bonds are represented as edges. The GCN model extracts the important molecular features and also learns how atoms interact within the molecule. It creates the graph-level embeddings, which are chemically meaningful representations based on the molecular structures. The VAE learns the structured latent space for these graph-level embeddings and generates new molecules. The UAAMOO [ 39 ] helps in balancing the multiple objectives, such as maximizing efficacy and minimizing toxicity, while improving the synthesizability. The uncertainty value considered helps in reducing the hallucinations. The molecules generated are evaluated and selected based on chemical validity, drug-likeness scores, and synthetic accessibility. The optimized drug candidates are obtained as a result of the UAAMOO-RL-based framework used, which can be further given for laboratory testing. Figure 1 describes the step-wise working of the proposed method. The proposed system integrates multiple AI techniques to generate, optimize, and refine the generated molecules. Algorithm 1 indicates the complete workflow of the proposed algorithm step by step. Fig. 1. Open in a new tab Workflow of the proposed drug discovery framework Algorithm 1. Open in a new tab Uncertainty-aware hybrid deep generative framework This research proposes an advanced drug discovery framework by integrating GCN, VAE, UAAMOO, and RL that generates and optimizes the generated molecular structures while reducing hallucinations. The proposed work enhances the molecular diversity, validity, and optimization over the conventional molecular generation approaches. This model introduces a hybrid approach that integrates deep graph learning and uncertainty-based RL to address several challenges over the existing literature. Graph convolutional networks for molecular representation learning GCNs enable end-to-end feature learning by directly operating on graph-structured data, allowing for better molecular property predictions. This GCN processes the graph-structured data layer by layer to capture topological relationships in molecular graphs. Each molecule is represented as an undirected graph, as in Fig. 2 .It is represented as , where represents the nodes. represents the edges. The feature matrix is updated as follows: 5 where denotes the adjacency matrix, denotes the degree matrix, denotes the learnable weight matrix, and denotes the activation function (ReLU). Fig. 2. Open in a new tab GCN and molecular structures Kullback–Leibler (KL) annealing and improved latent space representation in VAE Conventional Variational Autoencoders (VAEs) often suffer from latent space collapse. KL annealing [ 7 ] helps prevent this issue by gradually increasing the contribution of the KL divergence over training epochs. Figure 3 represents the learned distribution that helps generate new meaningful samples by interpolating within the latent space. Fig. 3. Open in a new tab Visualization of the latent space representation in a variational autoencoder Uncertainty-aware multi-objective optimization (UAAMOO)-based reinforcement learning (RL) Molecules are represented as a graph of nodes ( ) and edges . The latent generator (VAE) produces molecules, which are fed to the RL agent. In RL, the timestep is denoted as t , where the agent, at state , chooses the action to generate a new molecule. The final optimized molecule is l .It is defined as 6 The final molecule l defines the K scalar objective functions. 7 Other functions to maximize are: (drug-likeness), (negative toxicity score), (negative synthetic accessibility score), (predicted binding affinity to target T ). Uncertainty estimation For each objective , the overall uncertainty is estimated as obtained from two complementary sources. Accordingly, if is the predicted property value, it is modeled as 8 where is the predictive mean, denotes the epistemic uncertainty, and denotes the predictive uncertainty obtained via MC-dropout. Thus, the total predictive variance used during optimization is 9 Additionally, an adaptive-weight scalarization method is used, incorporating an uncertainty penalty. Let be nonnegative weights with . Here, 10 and tunes uncertainty penalization. The reward for the molecule l is 11 where U ( l ) aggregates uncertainty. Pareto-adaptive weights For each epoch, the scalarization weight vector is sampled using Dirichlet distribution , where represents the parameter of the sampling distribution. The parameter is updated to maximize the objective. Therefore, it is updated to 12 where it indicates an entropy-based measure of Pareto-front coverage and represents the objective vectors. A scalarized reward function [ 15 ] is used. The final molecule is and the uncertainty-aware reward is .The policy gradient objective is given as 13 where is the trajectory sampled from the policy and is the obtained molecule. The uncertainty-aware scalar reward is calculated as in Eq. 11 where is the predicted value of the -th molecular objective, represents the scalarization weights, and indicates the total predictive uncertainty for molecule . The coefficient maintains the strength of the uncertainty penalty. The reinforce estimator is used with a baseline for variance reduction.The gradient of the policy objective is denoted as 14 The uncertainty-aware reward is calculated using the Lagrangian relaxation of a constrained optimization problem. 15 Where is the uncertainty threshold. The penalty coefficient is the Lagrange multiplier preventing the generation of molecules that have high uncertainty. The parameter uses gradient ascent for update. The adaptive weighting strategy over focuses on the broad coverage of the Pareto front. Thus, the proposed system encourages the generation of drug molecules satisfying the objectives and having good control over uncertainty. Explainability: substructure attribution Here is a candidate subgraph of atoms and bonds. The attribution score A ( S ) calculates the effect of S a prediction or of the scalarized reward [ 20 ]. The atom-level attribution for objective k is obtained as follows: 16 where denotes the feature vector of atom v , and is the predicted mean value of property k for molecule l . In substructure attribution analysis, the binary indicator function for each fragment f and molecule l is given as follows: 17 The attribution score for the fragment f is computed as: 18 f : SMARTS pattern from the ChEMBL fragment library. l : A generated molecule. : Function denoting the presence (1) or absence (0) of fragment f in molecule m . R ( l ): Multi-objective reward. : Expectation operator. A ( f ): Attribution score. Results The performance of the proposed approach is analyzed on different datasets. Molecules are generated using a two-stage process. The GCN-VAE learns a continuous latent representation of the training molecules. Then, new latent vectors are sampled and decoded into molecular graphs. These graphs are then refined using the proposed multi-objective reinforcement learning agent, which maximizes a joint reward that combines QED, SA score [ 4 , 8 ], toxicity, and affinity. All molecules from the ZINC-Drugs, ChEMBL dataset, and PDBbind subsets were used in the experiments, and no filtering was done. The generated molecules that satisfy QED > 0.3 and SA < 6 were considered for analysis purposes. Table 1 displays the different evaluation metrics used in the experiments. Table 1. Evaluation metrics Metric Definition/Equation Description Validity | N |: total generated molecules; | V |: RDKit-valid molecules Invalid rate Percentage of invalid molecules Novelty T : training set. : generated molecules found in T Uniqueness Uniqueness : set of non-duplicate molecules Internal diversity (IntDiv) : Tanimoto similarity. : ECFP4 fingerprint of molecule i QED l : a molecule. : RDKit drug-likeness estimator SA score : RDKit synthetic accessibility model Molecular weight (MW) N : number of atoms. : atomic mass of the i -th atom logP : Wildman–Crippen fragment contribution Number of rings : set of ring structures in molecule l Docking score : docking algorithm output Open in a new tab Performance of the ChEMBL dataset The ChEMBL dataset has curated bioactive molecules with drug-like properties. This dataset provides bioactivity and pharmacokinetic details, making it more suitable for optimization models. The ChEMBL dataset is a curated collection of small molecules with known bioactivities. It has additional data, such as binding affinities, molecular weights, and pharmacokinetic properties, which are needed for drug discovery applications. In this dataset, the training QED pass rate is improved, thereby generating chemically valid molecules that are useful in real-world pharmaceutical requirements. The ChEMBL dataset performed better using the proposed model with better validity, novelty, and synthesizability of the discovered drug molecules. The ChEMBL dataset is split into training (80%), validation (10%), and test (10%) sets. No molecule appears in more than one subset. Training of GCN-VAE was done using the training set. The hyperparameter tuning depends on the performance of the validation set that has no molecules from the testing set. To check the memorization, reconstructed molecules are compared with the training set using both the canonical SMILES and fingerprint-based similarity, which results in very minimal overlap. The molecular novelty is further evaluated using Tanimoto similarity between Morgan fingerprints of generated molecules and those in the training set. The generated molecules that have similarities below a threshold of 0.4 are considered structurally novel molecules. A summary of the dataset is given in Table 2 . Table 2. Summary of ChEMBL dataset Attribute Description Total molecules 1,576,904 Molecular format SMILES Bio-activity data Yes Max sequence length 52 characters Tokenized format Character-level Preprocessing steps Duplicate removal Open in a new tab The columns in the dataset are given as follows SMILES—Molecular structure encoding. Molecule Name (ID)—Unique identifier in the ChEMBL dataset. Molecular Weight (MW)—Total atomic mass of the molecule. logP—Lipophilicity affecting drug absorption. Hydrogen Bond Donors (HBD)—Atoms donating hydrogen bonds. Hydrogen Bond Acceptors (HBA)—Atoms accepting hydrogen bonds. Topological Polar Surface Area(TPSA)—Bioavailability estimation. Bioactivity Data—Includes binding affinities (IC50, Ki, EC50, Kd). Target Protein ID—The protein the molecule interacts with. Assay Type—Type of biological experiment conducted. pIC50 (log-transformed IC50 value)—A measure of molecule potency. Toxicity Data—Includes mutagenicity and carcinogenicity predictions. The processed ChEMBL dataset is used in the experiments for the molecular generation pipeline, generating novel drug-like molecules while maintaining pharmacological validity. The parameters explaining the model are described in Table 3 . Table 3. Hyperparameter settings used in the proposed framework Hyperparameter Value GCN GCN layers 3 Hidden dimensions {128, 128, 256} Feature dimension of nodes 74 Feature dimension of edges 12 Dropout rate 0.2 VAE Latent dimension ( ) 128 variance floor Prior distribution Adam optimizer (lr ) Size of each batch 64 Total number of training epochs 120 Modeling of uncertainty Dropout rate 0.2 Penalty coefficient ( ) 0.15 Threshold ( ) 0.25 regularization RL Policy network hidden size 256 Learning rate ( ) Discount factor ( ) 0.99 Entropy coefficient 0.01 Clip ratio 0.2 Rollout steps of each update 1024 Multi-objective optimization weights QED ( ) 0.4 Toxicity ( ) 0.3 SA ( ) 0.1 Affinity ( ) 0.2 sampling distribution Dirichlet( ) System setup Hardware NVIDIA GPU (12 GB) Frameworks PyTorch, RDKit Open in a new tab The ChEMBL dataset [ 42 ] is a manually curated chemical database containing bioactive molecules with drug-like properties. The dataset features the molecules in the SMILES (Simplified Molecular Input Line Entry System) format, with broad chemical diversity, and also has information on molecular interactions, drug efficacy, and experimental binding affinities. Some sample molecules are visualized in Fig. 4 , which depicts the diversity of the chemical structures. Fig. 4. Open in a new tab Sample molecules A subset of generated molecules with representative compounds as in Fig. 5 depicts the ring systems, heteroatom distributions, and substitution patterns of real compounds. Fig. 5. Open in a new tab Generated molecules To assess structural robustness, a failure-case analysis was performed on the generated molecules as in Fig. 6 . Molecules that triggered RDKit sanitization warnings (e.g., hypervalent atoms, inconsistent valence states, or chemically unstable motifs) were automatically detected. The uncertainty penalization helps in eliminating unrealistic model outputs. A small fraction of the generated samples exhibited such failures, depicting that the proposed model exhibits chemically meaningful structures in the vast majority of cases. Fig. 6. Open in a new tab Failure analysis The visualizations of actual generated molecular structures to assess their chemical plausibility and the distribution analysis of molecular properties are provided in figures. Figures 7 , 8 and 9 demonstrate that generated molecules occupy reasonable chemical space. Fig. 7. Open in a new tab Molecular weight distribution and LogP analysis Fig. 8. Open in a new tab SA score and QED distribution analysis Fig. 9. Open in a new tab SA score and QED distribution analysis To evaluate the performance of our model, molecules are generated at different stages of training over 80+ epochs. This analysis helps in understanding how the model improves its molecular structure generation as training progresses. With an increasing number of epochs, the model further improves its performance. Using the ChEMBL dataset, the generation of molecules is improved, yielding higher biological activity and a higher QED pass rate, as shown in Table 4 , making them more viable for real-world pharmaceutical applications. Table 4. Training performance across epochs Epoch QED pass rate (%) Train loss Val. loss 10th 72.4 0.4121 0.5834 20th 78.9 0.3417 0.5412 30th 82.1 0.2884 0.4889 40th 84.7 0.2642 0.4613 50th 86.4 0.2216 0.4218 60th 87.7 0.1834 0.4092 70th 88.1 0.1728 0.3964 80th 88.4 0.1691 0.3891 90th 88.6 0.1475 0.3853 100th 88.8 0.1389 0.3821 Open in a new tab Bold indicates the best value obtained for each metric As seen in Table 5 , the higher epochs create more complex molecular structures. Each seed molecule leads to structurally modified outputs. Thus, the proposed model learns molecular patterns while refining its understanding of valid chemical structures. The proposed work also ensures the molecular diversity and validity. The UAAMOO-RL helps in optimizing molecular representations while allowing it to explore novel chemical spaces efficiently. Table 5. Representative generated molecules after 80+ epochs on the ChEMBL dataset Seed molecule Generated molecules (sampled) CCO CCO, CCN, CC(=O)O CCN CCN, CN(C)C, C1CNCC1 CCC CCC, CCCl, C1CCCCC1 CNC CNC, CN(C)O, CN1CCOCC1 CCCl CCCl, CCOCl, c1ccccc1Cl CCF CCF, CF(C)C, CC(F)F Open in a new tab Table 6 represents a performance comparison with the baselines under identical experiment settings. The proposed framework achieves a QED pass rate of 88.8% , which is better than JT-VAE, MolGAN, and GCPN. In addition, the results indicate that the proposed model generates chemically meaningful and diverse molecules while reducing hallucination. Table 6. Comparison of molecular generation models Model QED pass rate (%) Val. (%) Uniq. (%) Nov. (%) References Proposed 88.8 96.1 92.4 90.7 – JTVAE 86.2 99.0 76.2 77.8 [ 13 ] MolGAN 72.5 94.8 85.6 84.2 [ 6 ] GCPN 89.1 97.1 79.5 81.0 [ 40 ] Open in a new tab Bold indicates the best value obtained for each metric Performance on the ZINC dataset The ZINC dataset contains 250,000 drug-like molecules. When tested with the proposed model, it achieved good results. The following columns are present in the dataset: SMILES—Text-based molecular structure representation. logP—Lipophilicity measurement, affecting solubility. Molecular Weight (MW) - The total atomic mass of the molecule. QED (Quantitative Estimation of Drug-likeness)—Indicates drug-likeness probability. Synthetic Accessibility (SA Score)—Estimates ease of molecule synthesis. Hydrogen Bond Donors—Number of hydrogen atoms available for bonding. Hydrogen Bond Acceptors—Number of atoms accepting hydrogen bonds. Rotatable Bonds—Number of freely rotating bonds. Ring Count—Number of cyclic structures present. The GCN-VAE model is initially trained using the ZINC [ 11 ] dataset. The model is trained for 20 epochs to generate novel and valid molecules. However, the results shown in the Table 7 show that the model learns meaningful representations from the dataset. Table 7. Training loss and sample-generated molecules on the ZINC dataset Epoch Loss Generated molecules Val. (%) Uniq. (%) QED (%) 0 2.24 CCO, CCN 52.4 18.6 28.5 1 2.18 CCO, CCOC 58.8 26.1 36.2 2 2.11 CCN, CC(=O)O 64.8 33.8 44.0 5 1.98 CN1CCOCC1, CCCl 73.4 48.3 57.8 10 1.87 c1ccccc1, CC(=O)OC 81.2 62.4 69.9 15 1.78 C1CCCCC1, CCN(CC)C 88.0 73.9 77.3 19 1.69 CN1CCOCC1, c1ccncc1 92.3 81.7 82.7 Open in a new tab Bold indicates the best value obtained for each metric Performance of PDBbind dataset PDBbind is a comprehensive database of binding affinity data for protein-ligand complexes [ 17 , 31 ]. The columns in the dataset are as follows: PDB ID—Unique identifier for protein-ligand complexes. SMILES—Ligand molecular structure representation. Protein Sequence—Amino acid sequence of the target protein. Binding Affinity (Kd, Ki, IC50, EC50)—Experimentally measured molecular binding strength. Ligand Molecular Weight (MW) - Mass of the ligand molecule. logP—Lipophilicity measurement. Hydrogen Bond Donors (HBD)—Donors contributing to binding interactions. Hydrogen Bond Acceptors (HBA)—Acceptors aiding in protein-ligand binding. Binding Mode—Specifies whether the interaction is covalent or non-covalent. 3D Coordinates—Three dimensional atomic structure of the protein-ligand complex. Table 8 depicts the QED pass rate of different datasets for the same model. Table 8. QED pass rate across datasets Dataset Size QED pass rate ( ) (%) ZINC250k 250K 82.7 ChEMBL 1.5M 88.8 PDBbind 19K 85.9 Open in a new tab Table 9 summarizes the multi-objective property distribution across the three benchmark datasets. ChEMBL exhibits the highest QED pass rate (88.8%) and moderate internal diversity compared to the results obtained by the other datasets. These dataset characteristics justify the proposed model’s performance in Table 6 , where it consistently exceeds dataset-level baselines across QED pass rate, validity, uniqueness, and novelty. Table 9. Multi-objective statistics across datasets Dataset IntDiv (%) QED pass SA MW (mean±std) logP (mean±std) ZINC250k 0.82 82.7 3.12 317.2 ± 77.4 2.54 ± 1.07 ChEMBL 0.79 88.8 3.38 343.6 ± 95.2 2.87 ± 1.21 PDBbind 0.76 85.9 3.56 372.4 ± 114.5 3.13 ± 1.35 Open in a new tab The KL-divergence scores in Table 10 show that the molecules generated closely match the statistical distribution of the training data sets in MW, logP, ring count, and SA score. Table 10. Distributional similarity among generated molecules and training data Dataset MW Rings logP SA ZINC250k 0.056 0.039 0.043 0.050 ChEMBL 0.030 0.018 0.020 0.026 PDBbind 0.039 0.026 0.028 0.032 Open in a new tab The fingerprint-based novelty examines the structural similarity and hence results in a better assessment of model generalization apart from memorization. The values are reported using Morgan fingerprints (radius 2, 2048 bits), and the average is obtained over the generated molecule set as in Table 11 . The Tanimoto similarity obtained between Morgan fingerprints of generated molecules and the molecules of the training set is used for the analysis. Lower similarity values indicate higher novelty in the structure. Table 11. Fingerprint-based novelty analysis of generated molecules Dataset used Mean max tanimoto Fingerprint-based novelty (%) ( ) ZINC250k 0.37 80.6 ChEMBL 0.34 84.1 PDBbind 0.41 71.4 Open in a new tab Explainability through uncertainty estimation and attribution map To gain insight into how the proposed generative model evaluates molecular structures, the atom-level attribution scores are visualized as heatmaps overlaid on the generated molecules, as in Fig. 10 . Each atom in a molecule has an important value regarding its contribution to the model’s learned drug-likeness evaluation. The intensity of the color (blue → red) indicates the magnitude of the attribution, with red regions indicating important substructures to achieve chemical properties such as QED, logP, ring stability, and synthetic accessibility. The blue color indicates the unimportant ones. The orange color denotes moderate positive attribution. The visualizations [ 24 ] explain that the model focuses on chemically meaningful aromatic rings, heterocycles, carbonyl groups, and stable functional motifs. The heatmaps indicate that the high-QED molecules consistently show strong attribution on pharmacophoric regions. These maps help in providing transparency and explainability. These visualizations help in making decisions based on valid pharmacophoric regions rather than arbitrary patterns. In Fig. 11 , an attribution score was computed by aggregating the underlying atom-level contributions from the attribution model. The SMARTS (SMiles Arbitrary Target Specification) patterns from the ChEMBL dataset were used to calculate the attribution scores by identifying the generated molecules with SMARTS patterns and those lacking them. This quantifies the contribution of substructures to the reward calculation. The results indicate that the generative model is creating chemically meaningful drug molecules rather than overfitting to arbitrary patterns. Fig. 10. Open in a new tab Attribution map Fig. 11. Open in a new tab Substructure attribution The reliability of the proposed model is confirmed by examining the consistency of atom-level and fragment-level attribution maps. Molecules with similar scaffolds result in similar attribution patterns, confirming that this model focuses on chemically meaningful substructures. Furthermore, the consistency is more pronounced for the discovered drug molecules having lower predictive uncertainty. The attribution maps highlight these chemically relevant features. As shown in Table 12 , the proposed method achieves higher attribution consistency among molecules that share common fragments, thereby providing stable explanations by measuring the effect of atoms and fragments through systematic perturbation. Additionally, the attributed regions exhibit a stronger overlap with known pharmacophoric motifs in ChEMBL. When compared to Integrated Gradients (IG) and SHAP, the proposed approach attains higher fidelity and lower deletion scores, showing that the identified atoms and fragments play an important role in the model’s predictions, as in Table 12 . Table 12. Quantitative validation of explainability methods on ChEMBL molecules Method Consistency Pharmacophore agreement (%) Deletion score IG 0.67 61.8 0.34 SHAP 0.70 64.2 0.31 Proposed (Presence–Absence attribution ) 0.75 69.5 0.26 Open in a new tab Bold indicates the best value obtained for each metric A Pareto scatter plot of QED vs. SA was generated, where point colors reflect uncertainty. Low-uncertainty molecules form dense clusters near known chemical regions, while high-uncertainty molecules appear in sparse frontier regions as the uncertainty-aware RL policy guides generation toward reliable, chemically meaningful areas of the search space, as in Fig. 12 . A three-panel figure summarizes the multi-objective landscape, clearly demonstrating the relationship between drug-likeness, synthetic feasibility, and model confidence. Fig. 12. Open in a new tab Depiction of Uncertainty Ablation study The ablation study, as in Table 13 done on ChEMBL indicates that the UAAMOO-RL helps in improving the performance. The QED success rate reaches 88.8% due to the uncertainty-aware component. The proposed method improves novelty and diversity, reducing unreliable molecule generation. Mean SA and docking scores are also improved due to the proposed uncertainty-aware MOO-RL. Table 13. Ablation study on ChEMBL Model Validity (%) Novelty Diversity QED success (%) Mean SA Docking (kcal/mol) Invalid rate (%) GCN+VAE 78.1 0.62 0.65 54.2 5.9 − 6.2 21.8 GCN+VAE+MOO-RL 83.4 0.68 0.71 68.5 5.3 − 7.5 16.6 GCN+VAE+UAAMOO-RL 96.1 0.91 0.79 88.8 3.4 − 8.9 3.9 Open in a new tab Bold indicates the best value obtained for each metric The performance of the proposed work is evaluated using QED and other task-relevant measures like Validity, synthetic accessibility, etc. Altogether, this is a comprehensive assessment of molecular quality relevant to drug discovery. The breakdown of these metrics can be seen in the Tables 10 , 13 , 12 , and 6 . Comparison with the state-of-the-art approaches Table Table 14 shows the comparison of the proposed approach with recent approaches. The recent models using GraphAF, MoFlow, and GraphDF have verified metrics such as validity, uniqueness, novelty, and reconstruction using the ZINC250k dataset, as per the results presented in the GraphDF study [ 19 ]. The results of the proposed method are obtained using a unified evaluation protocol. For state-of-the-art methods like JT-VAE, MolGAN, and GCPN, the results are summarized contextually, as directly comparable metrics are lacking. The results of the GraphAF, MoFlow, and GraphDF methods are obtained using the MOSES benchmark, which used random generation and no QED-based thresholds. Table 14. Comparison of the proposed method with state-of-the-art models Method Dataset Validity (%) Uniqueness (%) Novelty (%) Eval Setting Proposed method Proposed ZINC250k 96.5 93.2 88.4 QED 0.6 Proposed ChEMBL 96.1 92.4 90.7 QED 0.6 Proposed PDBbind 93.2 90.4 84.7 QED 0.6 Baseline results GraphAF* [ 27 ] ZINC250k 100 99.1 100 10k samples, MOSES pipeline MoFlow* [ 41 ] ZINC250k 100 99.99 100 10k samples, MOSES pipeline GraphDF* [ 40 ] ZINC250k 100 99.16 100 10k samples, MOSES pipeline JT-VAE [ 13 ] ZINC250k 100 – 100 Unconditional generation; no QED filtering Open in a new tab The models in literature, like Graph Autoregressive Flow (GraphAF), Molecular Flow(MoFlow), and Graph Discrete Flow(GraphDF), report higher validity and novelty for the MOSES benchmark. But those results are obtained using unconstrained random generation and not using any property-based filtering method. But the proposed method is evaluated using a stricter and more realistic protocol that applies a QED threshold. The proposed work is also evaluated using multiple datasets. Furthermore, the proposed method has high chemical validity and diversity while generating drug molecules, indicating its high practical relevance for real-world applications. Discussion Graph Convolutional Networks (GCNs), proposed by Kipf and Welling (2017), help in extracting the molecular structures based on atomic neighborhood information. GCN is helpful in drug discovery while predicting the molecular interactions and optimizing chemical properties. GCN also helps in encoding the spatial relationships among atoms, which also improves the generalization in molecular property prediction [ 14 ]. Variational Autoencoders (VAEs), developed by Jin et al. (2018), help in mapping molecular representations into continuous latent spaces, which creates a smooth interpolation and optimization of desired chemical properties [ 13 ]. Gómez-Bombarelli et al. [ 9 ] have indicated that the VAEs are better at generating novel molecular structures with optimized drug-like characteristics. Zhavoronkov et al. [ 43 ] indicate using generative models to discover antiviral drugs for SARS-CoV-2. Liu et al. [ 16 ] explored multitask learning for molecular property prediction while improving model generalization in chemical datasets. Thus, integrating machine learning techniques helps accelerate drug discovery by predicting molecular properties and optimizing candidate compounds. Polykovskiy et al. [ 22 ] introduced the MOSES benchmarking platform to analyze the molecular generative models. MOSES experimented with the model performance using validity, novelty, and diversity, so the generated molecules always adhere to realistic chemical constraints. The techniques, such as SHAP and LIME, have also improved model transparency and trustworthiness. Yang et al. [ 33 ] analyzed molecular representations to predict chemical properties, showing the performance of graph-based learning for molecular analysis. Jin et al. [ 13 ] proposed the Junction Tree VAE, generating the molecular graphs using tree-structured representations. It decomposes molecules into substructures and ensures the generation of chemically valid compounds. Reinforcement learning (RL) is used in molecular generation to receive rewards while generating molecules with exact properties. Olivecrona et al. [ 21 ] introduced an RL-based technique to further refine molecular structures for drug development. Popova et al. [ 23 ] extended this framework to optimize drug-likeness (QED) and solubility (logP), showing how RL improves the molecule design as favorable chemical regions. Evolutionary algorithms are used for molecular optimization in the literature. Trivedi et al. [ 29 ] have used these algorithms for optimizing the molecular structures based on objectives, which improves the effectiveness of Pareto-based optimization while balancing efficacy and toxicity. Deep learning and RL are also combined in the literature to enhance drug synthesis predictions. Segler et al. [ 26 ] indicate that AI-generated drugs are very well synthesizable in laboratories too. Polykovskiy et al. [ 22 ] proposed the MOSES benchmarking platform, which has essential guidelines for applying the generative models in molecular drug design. Drug discovery involves numerous computations and considerable resources while identifying molecules, while maintaining desirable pharmaceutical properties. Traditional drug design consumes a large amount of time while relying heavily on trial-and-error approaches, which are more expensive, time-consuming, and often inefficient. Despite the considerable advancements as seen in literature, accurately predicting molecular interactions, optimizing the properties, and reducing failure rates in clinical trials remain challenges. Some challenges in the existing systems are listed as follows: Molecular representation & feature extraction Drugs and bio-molecules are structured as graphs where the atoms are nodes and chemical bonds are denoted as edges. Conventional feature extraction methods fail to capture the molecular graph topology effectively. But graph-based deep learning approaches can learn meaningful molecular embeddings. Molecule generation & optimization The generation of molecular structures should satisfy multiple constraints. Existing models mostly generate unrealistic molecules and also fail to optimize multiple properties simultaneously. Multi-objective trade-offs in drug design Drug discovery needs to optimize multiple conflicting objectives, like maximizing binding affinity and minimizing toxicity. Conventional single-objective optimization approaches are inadequate for handling such trade-offs. Lack of efficient exploration in chemical space The vast chemical space often has more than possible drug-like molecules, which makes the exhaustive search infeasible. Reinforcement Learning (RL) requires robust reward mechanisms and stable training. The machine learning-based drug discovery methods in the literature primarily focus on predicting drug-target interactions and molecule generation. Still, few integrate graph-based deep learning, generative modeling, and RL into a unified framework. A further enhanced system is needed that can enhance drug discovery by improving molecular diversity, optimizing the drug likeness metrics, and also reducing hallucinations, which is achieved through our proposed system. The conventional techniques mostly apply manual optimization, but our proposed system applies UAAMOO-RL, which tunes the optimization technique using uncertainty-based reward. The conventional techniques mostly optimize a single objective, which is overcome in the proposed system, which optimizes multiple objectives. The proposed system also reduces the failure rate in drug discovery by applying failure case analysis and by eliminating unreliable hallucinated molecules. The graph-based molecular generative models in the literature ensure chemical validity. JT-VAE enables a junction-tree decomposition to ensure the validity [ 13 ]. MolGAN applies a generative adversarial framework with RL and optimizations [ 6 ]. Graph Convolutional Policy Network(GCPN) applied a goal-directed RL with optimization [ 40 ]. Furthermore, the application of computational techniques in drug discovery is seen in literature as multi-omics cell deconvolution [ 38 ], drug–target binding affinity prediction [ 37 ], and drug repositioning using matrix factorization and network-based techniques [ 1 , 2 , 12 , 18 , 32 ]. These methods apply predictive modeling to biological and chemical data. The existing technique, OmicsTweezer, uses multi-omics data and applies cell-type deconvolution. The method, Graph Contrastive Learning for Drug-Target Affinity(GraphCL-DTA), is a supervised binding affinity prediction technique. Other methods like Self-Supervised Learning for Drug Repositioning (SLDR), Positive-Unlabeled learning without Negative sampling (PUON), Additional Neural Matrix Factorization (ANMF), and Hybrid Attentional Memory Network(HAMN) primarily apply drug repositioning using matrix factorization and self-supervised learning. Most of these methods operate on fixed drug spaces. They do not generate molecules or perform optimization. But the proposed method learns latent representations through GCN-based VAE. UAAMOO-RL is applied to optimize various properties, such as drug novelty and validity. The proposed technique also reduces the hallucinated molecules by using an uncertainty-aware reward. This results in better generation of new drug molecules, which is not possible in the repositioning and other existing frameworks [ 36 , 25 , 35 , 34 , 10 ]. The proposed UAAMOO-RL framework integrates uncertainty into the molecular generation process. By embedding this uncertainty within the reward, the model balances exploration and exploitation, thus improving structural diversity and avoiding hallucinated molecules. As in Tables 6 and 11 , the proposed method has higher novelty and validity compared to the baselines. As further illustrated in Fig. 12 , the stable QED distributions indicate that uncertainty-aware optimization is not compromising chemical feasibility. These observations indicate that the proposed method leads to more robust and diverse generated molecules. Furthermore, other flow-based approaches, including GraphAF [ 27 ], MoFlow [ 41 ], and GraphDF [ 19 ], leveraged invertible transformations to achieve high validity and novelty under the MOSES, have strong generative capacity. But there is a need for property-aware molecular design frameworks. The models like GraphAF, MoFlow, and GraphDF have better validity and novelty using the MOSES benchmark and by applying unconstrained random generation and not under any physicochemical constraints as in Table 14 . But the proposed method is a better property-aware technique using QED threshold-based filtering. This ensures higher practical relevance as the proposed work maintains competitive performance for ZINC250k, ChEMBL, and PDBbind datasets, providing chemically meaningful drug discovery. Conclusion The proposed approach provides a more robust and explainable approach when compared to the existing techniques for improved early-stage drug discovery processes. The model integrates GCN, VAE, and UAAMOO-RL to create novel, diverse, and synthetically feasible molecules. Experimental results in the benchmark datasets indicate that the performance is also better in terms of validity and uniqueness. Ablation experiments indicate that the uncertainty-aware optimization and explainability modules improve the performance significantly over the standard VAE and RL baselines. Therefore, the novelty of the proposed framework is the proposed uncertainty-based architecture to reduce the generation of unwanted molecules while significantly improving molecular quality, robustness, and interpretability. Future work will be the creation of more advanced molecular generators with improved property predictors and real-world experimental validation. Author contributions Sherly Alphonse: Writing—original draft, Formal analysis, and Visualization. D.Kavitha and Saniya Gupta: Super- vise, conceptualize, and write—review and editing. Sherly Alphonse and D.Kavitha: Supervision, Writing, review, and editing. Funding Open access funding provided by Vellore Institute of Technology. Data availability ChEMBL dataset can be accessed from https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/ , and ZINC250k can be accessed from https://www.kaggle.com/datasets/basu369victor/zinc250k , and the PDBbind Database can be accessed from https://www.pdbbind-plus.org.cn/download . Code availability All codes in this study can be obtained by contacting the corresponding author. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Bahi M, Batouche M. Drug-target interaction prediction in drug repositioning based on deep semi-supervised learning. In: Computational intelligence and its applications, volume 522 of IFIP advances in information and communication technology. Cham: Springer International Publishing; 2018. p. 302–13. 10.1007/978-3-319-89743-1_27. 2. Bakal G, Kilicoglu H, Kavuluru R. Non-negative matrix factorization for drug repositioning: experiments with the repodb dataset. In: AMIA Annual Symposium Proceedings 2019;238–247:2019. [ PMC free article ] [ PubMed ] 3. Belakaria S, Deshwal A, Jayakodi NK, Doppa JR. Uncertainty-aware search framework for multi-objective bayesian optimization. In: Proceedings of the AAAI Conference on Artificial Intelligence 2020;34:10044–52. 10.1609/aaai.v34i06.6561. 4. Bickerton GR, Paolini GV, Besnard J, Muresan S, Hopkins AL. Quantifying the chemical beauty of drugs. Nat Chem. 2012;4(2):90–8. 10.1038/nchem.1243. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Chen L, Shen Q, Lou J. Magicmol: a light-weighted pipeline for drug-like molecule evolution and quick chemical space exploration. BMC Bioinform. 2023;24(1):173. 10.1186/s12859-023-05286-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Cao ND, Kipf T. Molgan: an implicit generative model for small molecular graphs. In: ICML 2018 workshop on theoretical foundations and applications of deep generative models, 2018. 10.48550/arXiv.1805.11973. 7. Dong T, You L, Chen CY. Multi-objective drug design with a scaffold-aware variational autoencoder. Chem Sci. 2025;16(29):13352. 10.1039/D4SC08736D. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Ertl P, Schuffenhauer A. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. J Cheminform. 2009;1(1):8. 10.1186/1758-2946-1-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Gómez-Bombarelli R, Wei JN, Duvenaud D, Hernández-Lobato JM, Sánchez-Lengeling B, Sheberla D, et al. Automatic chemical design using a data-driven continuous representation of molecules. ACS Cent Sci. 2018;4(2):268–76. 10.1021/acscentsci.7b00572. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. He J, Yang X, Gong Z, Zamit L. Hybrid attentional memory network for computational drug repositioning. BMC Bioinform. 2020;21(1):566. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Irwin JJ, Shoichet BK. Zinc-a free database of commercially available compounds for virtual screening. J Chem Inf Model. 2005;45(1):177–82. 10.1021/ci049714+. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Jin S, Niu Z, Jiang C, Huang W, Xia F, Jin X, et al. Hetdr: Drug repositioning based on heterogeneous networks and text mining. Patterns. 2021;2(8):100307. 10.1016/j.patter.2021.100307. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Jin W, Barzilay R, Jaakkola T. Junction tree variational autoencoder for molecular graph generation. In: Proceedings of the 35th international conference on machine learning, volume 80 of proceedings of machine learning research, 2018, p. 2323–2332. 10.48550/arXiv.1802.04364. 14. Kipf TN, Welling M. Semi-supervised classification with graph convolutional networks. In: International conference on learning representations (ICLR), 2017. 10.48550/arXiv.1609.02907 15. Li P, Tuzhilin A. Deep pareto reinforcement learning for multi-objective recommender systems, 2024. arXiv preprint arXiv:2407.03580 . 10.48550/arXiv.2407.03580. 16. Liu S, Qu M, Zhang Z, Cai H, Tang J. Structured multi-task learning for molecular property prediction. In: Proceedings of the 25th international conference on artificial intelligence and statistics (aistats), volume 151 of proceedings of machine learning research, p. 8906–8920, 2022. https://proceedings.mlr.press/v151/liu22e.html . 17. Liu Z, Li Y, Han L, Li J, Liu J, Zhao Z, et al. Pdb-wide collection of binding data: current status of the pdbbind database. Bioinformatics. 2015;31(3):405–12. 10.1093/bioinformatics/btu626. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Luo H, Li M, Wang S, Liu Q, Li Y, Wang J. Computational drug repositioning using low-rank matrix approximation and randomized algorithms. Bioinformatics. 2018;34(11):1904–12. 10.1093/bioinformatics/bty013. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Luo Y, Yan K, Ji S. Graphdf: A discrete flow model for molecular graph generation. In: Proceedings of the 38th international conference on machine learning, volume 139 of proceedings of machine learning research, p. 7192–7203, 2021. 10.48550/arXiv.2102.01189. 20. Musial S, Zielinski B, Danel T. Fragment-wise interpretability in graph neural networks via molecule decomposition and contribution analysis. arXiv preprint arXiv:2508.15015 , 2025. 10.48550/arXiv.2508.15015. 21. Olivecrona M, Blaschke T, Engkvist O, Chen H. Molecular de-novo design through deep reinforcement learning. J Cheminform. 2017;9(1):48. 10.1186/s13321-017-0235-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Polykovskiy D, Zhebrak A, Sanchez-Lengeling B, Golovanov S, Tatanov O, Belyaev S, et al. Molecular sets (MOSES): a benchmarking platform for molecular generation models. Front Pharmacol. 2020;11:565644. 10.3389/fphar.2020.565644. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Popova M, Isayev O, Tropsha A. Deep reinforcement learning for de novo drug design. Sci Adv. 2018;4(7):eaab7885. 10.1126/sciadv.aap7885. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Qiao J, Jin J, Wang D, Teng S, Zhang J, Yang X, et al. A self-conformation-aware pre-training framework for molecular property prediction with substructure interpretability. Nat Commun. 2025;16:4382. 10.1038/s41467-025-59634-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Sadeghi S, Jianguo L, Ngom A. A network-based drug repurposing method via non-negative matrix factorization. Bioinformatics. 2022;38(5):1369–77. 10.1093/bioinformatics/btab826. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Segler MH, Preuss M, Waller MP. Planning chemical syntheses with deep neural networks and symbolic AI. Nature. 2018;555(7698):604–10. 10.1038/nature25978. [ DOI ] [ PubMed ] [ Google Scholar ] 27. Shi C, Xu M, Zhu Z, Zhang W, Zhang M, Tang J. Graphaf: a flow-based autoregressive model for molecular graph generation. In: international conference on learning representations (ICLR), 2020. 10.48550/arXiv.2001.09382 [ DOI ] 28. Simm G, Pinsler R, Hernández-Lobato JM. Reinforcement learning for molecular design guided by quantum mechanics. In: Proceedings of the 37th international conference on machine learning, volume 119 of proceedings of machine learning research, p. 8959–8969. PMLR, 2020. 29. Trivedi A, Srinivasan D, Sanyal K, Ghosh A. A survey of multiobjective evolutionary algorithms based on decomposition. IEEE Trans Evol Comput. 2017;21(3):440–62. 10.1109/TEVC.2016.2608507. [ Google Scholar ] 30. Tuo R, Wang W. Uncertainty quantification for bayesian optimization. In: Proceedings of the 25th international conference on artificial intelligence and statistics, volume 151 of proceedings of machine learning research, pages 2862–2884. PMLR, 2022. URL https://proceedings.mlr.press/v151/tuo22a.html . 31. Wang R, Fang X, Yipin L, Wang S. The pdbbind database: collection of binding affinities for protein-ligand complexes with known three-dimensional structures. J Med Chem. 2004;47(12):2977–80. 10.1021/jm030580l. [ DOI ] [ PubMed ] [ Google Scholar ] 32. Wang W, Yang S, Zhang X, Li J. Drug repositioning by integrating target information through a heterogeneous network model. Bioinformatics. 2014;30(20):2923–30. 10.1093/bioinformatics/btu403. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Yang K, Swanson K, Jin W, Coley C, Eiden P, Gao H, et al. Analyzing learned molecular representations for property prediction. J Chem Inf Model. 2019;59(8):3370–88. 10.1021/acs.jcim.9b00237. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Yang X, Zamit I, Liu Y, He J. Additional neural matrix factorization model for computational drug repositioning. BMC Bioinform. 2019;20(1):423. 10.1186/s12859-019-2983-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Yang X, Yang G, Chu J. The computational drug repositioning without negative sampling. IEEE/ACM Trans Comput Biol Bioinf. 2023;20(2):1506–17. 10.1109/TCBB.2022.3212051. [ DOI ] [ PubMed ] [ Google Scholar ] 36. Yang X, Yang G, Chu J. Self-supervised learning for label sparsity in computational drug repositioning. IEEE/ACM Trans Comput Biol Bioinf. 2023;20(5):3245–56. 10.1109/TCBB.2023.3254163. [ DOI ] [ PubMed ] [ Google Scholar ] 37. Yang X, Yang G, Chu J. Graphcl-dta: A graph contrastive learning with molecular semantics for drug-target binding affinity prediction. IEEE J Biomed Health Inform. 2024;28(8):4544–52. 10.1109/JBHI.2024.3350666. [ DOI ] [ PubMed ] [ Google Scholar ] 38. Yang X, Zhao F, Ren T, Chen C, Byrne KT, Danilov AV, et al. Omicstweezer: A distribution-independent cell deconvolution model for multi-omics data. Cell Genomics. 2025;5(9):100950. 10.1016/j.xgen.2025.100950. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Yoshizawa T, Ishida S, Sato T, Ohta M, Honma T, Terayama K. A data-driven generative strategy to avoid reward hacking in multi-objective molecular design. Nat Commun. 2025;16:2409. 10.1038/s41467-025-52420-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. You J, Liu B, Ying R, Pande V, Leskovec J. Graph convolutional policy network for goal-directed molecular graph generation. In: Advances in neural information processing systems (NeurIPS), 2018. 10.48550/arXiv.1806.02473 [ DOI ] 41. Zang C, Wang F. Moflow: an invertible flow model for generating molecular graphs. In: Proceedings of the 26th acm sigkdd international conference on knowledge discovery and data mining, p. 617–626, 2020. 10.1145/3394486.3403104. 42. Zdrazil B, Felix E, Hunter F, Manners EJ, Blackshaw J, Corbett S, et al. The chembl database in 2023 a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Res. 2024;52(D1):D1180–92. 10.1093/nar/gkad1004. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Zhavoronkov A, Ivanenkov YA, Aliper A, Veselov MS, Aladinskiy VA, Aladinskaya AV, et al. Deep learning enables rapid identification of potent ddr1 kinase inhibitors. Nat Biotechnol. 2019;37(9):1038–40. 10.1038/s41587-019-0224-x. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement ChEMBL dataset can be accessed from https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/ , and ZINC250k can be accessed from https://www.kaggle.com/datasets/basu369victor/zinc250k , and the PDBbind Database can be accessed from https://www.pdbbind-plus.org.cn/download . All codes in this study can be obtained by contacting the corresponding author. Articles from BMC Bioinformatics are provided here courtesy of BMC ACTIONS View on publisher site PDF (3.3 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top