This study provides a large-scale empirical evaluation of multimodal feature fusion for molecular property classification by systematically integrating SMILES-based chemical language representations with fingerprint-based structural descriptors across 60 benchmark datasets.
Abstract
Accurate molecular property prediction is a cornerstone of modern chemical science, driving progress in drug discovery, materials design, and environmental research. Yet, most existing models remain unimodal, while multimodal approaches often rely on simple aggregation, leaving much of the complementary chemical information underexploited. In this work, we present a multimodal feature fusion framework that unites the strengths of deep chemical language processing (CLP) models and molecular fingerprints, integrating sequential and structural representations for more comprehensive molecular characterization. Unlike previous heuristic combinations, our framework systematically investigates the principles of effective cross-modal fusion. We benchmark ten CLP architectures and eight fingerprint types through exhaustive combinatorial search to identify the most synergistic configurations. This exploration shows that aggregating multiple models does not necessarily improve performance; instead, successful fusion requires data-aware design guided by feature integration and complementarity. The proposed strategy effectively couples sequential features learned from SMILES with structural information captured by molecular fingerprints, resulting in a coherent and chemically interpretable molecular representation. Evaluated across 60 datasets from MoleculeNet and TOXRIC, our fusion models deliver consistent and substantial gains over state-of-the-art baselines. Beyond outperforming existing architectures, this work provides conceptual insights and practical guidelines for multimodal fusion in molecular property prediction, highlighting the importance of efficient fusion strategies in building robust and generalizable molecular models.Scientific contributionThis study provides a large-scale empirical evaluation of multimodal feature fusion for molecular property classification by systematically integrating SMILES-based chemical language representations with fingerprint-based structural descriptors across 60 benchmark datasets. The framework introduces a data-aware combinatorial fusion strategy to identify task-specific complementary feature combinations, improving robustness and interpretability compared with unimodal models and baseline approaches. The results clarify how sequence-based and structure-based molecular representations complement each other, providing practical guidance for designing multimodal models in cheminformatics.
Computer-aided drug discovery has substantially accelerated modern pharmaceutical research, where accurate molecular property prediction plays a central role in identifying promising therapeutic candidates. Self-supervised learning (SSL), which exploits large-scale unlabeled molecular data to learn transferable representations, has recently emerged as a powerful paradigm well-aligned with the data characteristics of cheminformatics. Integrating chemical domain knowledge further enhances the ability of SSL models to capture structural, physicochemical, and functional properties of molecules. In this review, we provide a systematic overview of recent advances in SSL-based molecular property prediction. We summarize representative methodological developments and analyze how multimodal molecular representation learning─by integrating sequence, graph, three-dimensional structure, and textual information─can improve the quality and expressiveness of molecular representations. We further examine the synergistic relationship between multimodal modeling and SSL, highlighting how complementary modalities can enhance representation learning in low-label settings. To demonstrate the practical benefits of multimodal molecular properties, we compare their performance with conventional SSL models on two downstream benchmark tasks with distinct prediction objectives. Finally, we discuss key open challenges, including the scarcity of high-quality 3D molecular data, modality imbalance across data sets, and the limited interpretability of learned representations. We conclude by outlining promising research directions toward more robust, generalizable, and biologically meaningful frameworks for molecular property prediction.
Shuning Yang, Lei Deng· Journal of Chemical Informat...· 0 citations
A novel Dual-Attention Multimodal framework for Graphs and Sequence-based representations, so-called DAM-GS, which provides a promising solution for molecular property prediction with broad applications in drug discovery and computational molecular science.
Bay Van Nguyen, Vinh Truong, Ha Duong Thi Hong et al.· Journal of Chemical Informat...· 0 citations
Molecular property prediction is a critical task in accelerating drug discovery. While deep learning has shown promise, prevailing single-modal methods struggle to integrate multi-source (e.g., atomic graph and molecular fingerprints), heterogeneous chemical knowledge, thereby failing to holistically represent molecular structures and capture the high-order synergistic interactions governing their functions. To address these challenges, we present HyperMolFusion, a hypergraph-enhanced multi-modal fusion model for molecular property prediction. Compared with traditional graphs limited to pairwise atomic bonds, HyperMolFusion models chemical motifs as hyperedges to explicitly capture high-order structural correlations and encode complex molecular interactions. The framework comprises three core representation learning modules: AtomConv for local atomic interaction learning via attention-enhanced message passing, HyperConv for motif-level high-order correlation extraction via hypergraph convolution with GRU gating, and a mixed molecular fingerprint module that adaptively integrates MACCS, PubChem, and Pharmacophore fingerprints. A chemically guided attention (CGA) mechanism then dynamically fuses these multi-level features into hierarchical molecular representations, alleviating over-smoothing and preserving structural information effectively. Evaluated on eight MoleculeNet benchmarks (covering regression and classification tasks), HyperMolFusion achieves promising performance. For regression, it achieves an RMSE of 0.611 in lipophilicity, 0.653 in ESOL, and 0.951 in FreeSolv. For classification, it achieves a ROC-AUC of 0.935 in ClinTox, 0.907 in BBBP, and 0.689 in SIDER. This work provides a systematic and effective solution for molecular property prediction: by holistically integrating atomic, motif, and global fingerprint information via hypergraph modeling, HyperMolFusion offers a more reliable computational tool to enhance the efficiency and accuracy of drug development pipelines.
Yawen Lin, Sheng Lian, Shaoxin Bian et al.· IEEE journal of biomedical a...· 0 citations
Accurate prediction of molecular properties is fundamental to environmental chemistry, yet remains challenging when experimental data are limited. Multimodal fusion provides a promising strategy for integrating complementary molecular representations; however, the relative contributions of molecular representation, fusion strategy, and learning algorithm to the predictive accuracy and uncertainty remain poorly understood. Five molecular modalities (RDKit descriptors, Mol2Vec embeddings, graph neural network embeddings, SMILES representations, and MS2 fragmentation spectra) were evaluated by using early and late fusion strategies with four learning algorithms (LightGBM, RF, AttentiveFP, and DMPNN). Across 14 physicochemical properties, multimodal models exhibited modest numerical improvements over the best unimodal models, although these differences were generally not statistically significant. In contrast, uncertainty quantification revealed clearer distinctions among the modeling strategies. Multimodal integration significantly improved the alignment between prediction error and estimated uncertainty. The fusion strategy had a modest influence on epistemic uncertainty, with significant early versus late differences observed only for selected modality combinations, whereas meta-learner selection had the greatest effect on uncertainty calibration. Ablation, grouped SHAP, and RDKit descriptor reduction analyses showed that RDKit descriptors remained consistently informative despite substantial descriptor reduction, while Mol2Vec, SMILES, GNN embeddings, and MS2 contributed in a property-dependent and partially redundant manner. Computational cost increased substantially with multimodal complexity, whereas predictive accuracy exhibited diminishing returns, indicating that intermediate multimodal configurations often provided the most favorable balance among computational efficiency, predictive performance, and uncertainty reliability. Overall, the results demonstrate that successful multimodal learning depends on the coordinated selection of complementary molecular representations, fusion strategy, and learning algorithm rather than simply increasing the number of integrated modalities. Multimodal integration may provide particular value by improving the reliability of uncertainty estimation.
J. Wasswa, G. W. Kajjumba, Bharath Ramsundar· Journal of Chemical Informat...· 0 citations
Motivation Accurate ADMET prediction is essential for prioritizing compounds before costly experimental validation, yet ADMET tasks are highly heterogeneous. Properties such as solubility, permeability, protein binding, clearance, transporter activity and toxicity are governed by different molecular signals, ranging from local functional groups and physicochemical descriptors to bonded topology and three-dimensional geometry. Consequently, a single molecular representation or backbone is unlikely to be optimal across all ADMET tasks. Results We present Trimole-Hybrid, a task-wise multimodal framework that addresses ADMET heterogeneity by selecting or combining predictors built from complementary molecular representations. Trimole-Hybrid constructs a candidate pool of SMILES-, graph-, geometry-sensitive EPT/3D- and chemical descriptor-based predictors. For each task, Trimole-Hybrid selects the best-performing predictor to obtain the final prediction. On 22 Therapeutics Data Commons ADMET benchmarks, Trimole-Hybrid exceeded the public TDC top-1 methods on 10 tasks and ranked within the top 10 for 21 tasks. Ablation studies confirmed the contribution of both complementary multimodal molecular representations and task-specific ensemble strategies. In two small-molecule case studies, Trimole-Hybrid shows sensitivity to changes in essential functional motifs, suggesting its ability to capture ADMET-relevant molecular substructures. Availability and implementation Source code, supplementary tables, and audit files are available at the project repository: https://github.com/dchen0212/trimole_hybrid. Contact liyu@cse.cuhk.edu.hk and qzyu22@cse.cuhk.edu.hk Supplementary information Supplementary data are available with the submitted manuscript.
Accurate molecular property prediction is fundamental to drug discovery and is critically governed by molecular representations. While most existing approaches primarily focus on small molecules, extending reliable prediction to structurally complex macrocyclic compounds remains challenging due to their conformational flexibility and nonlocal interactions. To bridge this gap, we developed automated molecular property prediction (AutoMPP), an automated machine learning‐based pipeline that automates model selection and systematically evaluates fingerprint combinations across 75 molecular property prediction tasks. The results demonstrate that multifingerprint fusion significantly improves predictive robustness, with a four‐fingerprint combination achieving superior performance across diverse molecular tasks. Using this optimized representation strategy, AutoMPP outperforms leading models Uni‐Mol and fingerprints and graph neural networks (FP‐GNN), securing top performance on 68% (51/75) of tasks. Notably, AutoMPP generalizes effectively to macrocycles, such as cyclic peptide, attaining a Pearson correlation coefficient of 0.794, outperforming Uni‐Mol (0.692) and FP‐GNN (0.681). Furthermore, by integrating SHapley Additive exPlanations, the framework offers chemist‐intelligible insights into the structural determinants. Together, these results establish AutoMPP as a robust and adaptable framework for molecular property prediction, capable of identifying task‐specific optimal fingerprint combinations and learning architectures for both small molecules and complex macrocycles.
Qiyang Qian, Shihang Wang, Ruifeng Li et al.· ChemMedChem· 0 citations